# QRS (Quantitative Research Society): all pages Every page of https://qrsntu.org as Markdown, in the order of https://qrsntu.org/llms.txt. Pages longer than about 8 KB are shortened at a section boundary and name their full Markdown copy. Not included: translations, the Math Notes course pages with the download link of each file (see https://qrsntu.org/math_notes/index.md), and the NeurIPS 2026 guide (see https://qrsntu.org/service/neurips26/llms.txt). To find one fact, search https://qrsntu.org/data/site-index.json instead. # Quantitative Research Society (QRS) > QRS connects students and researchers across institutions to advance research in quantitative finance, machine learning, and related technical fields. Web page: https://qrsntu.org/ (content last changed 2026-11-12) Site index for agents: https://qrsntu.org/llms.txt QRS connects students and researchers across institutions to advance research in quantitative finance, machine learning, and related technical fields. [Projects](https://qrsntu.org/projects/) [Programmes](https://qrsntu.org/programmes/) [Publications](https://qrsntu.org/publications/) ## Latest News Research, releases and events from QRS. 11 Nov 2026 · Research ### [FRTBTrace: new project page](https://qrsntu.org/projects/frtbtrace/) FRTBTrace, our work on exact trade-level attribution of FRTB-IMA capital, now has its own project page: why one attribution vector cannot answer every question about capital, how one reverse pass through the full formula gives a ledger on (date, trade) positions, and the evidence from a pre-registered synthetic benchmark. Two FRTBTrace workshop papers were accepted at ICAIF 2026. 11 Nov 2026 · Research ### [MathResearchOS: new project page and usage guide](https://qrsntu.org/projects/mathresearchos/) A command-line system that turns a mathematical objective into a recoverable, verifier-gated run across Codex and Claude models. The new project page has a quick start and a usage guide covering install, model choice, human gates, campaigns and troubleshooting. 11 Nov 2026 · Release ### [NeurIPS 2026 Attendee Guide for Sydney, Atlanta and Paris](https://qrsntu.org/news/neurips-2026-attendee-guide/) A free, unofficial guide for everyone attending NeurIPS 2026 at any of its three sites: the full schedule, a guide to each city, a personal plan, hotels compared by price and commute, deadlines and templates for presenters, and the data for AI assistants. [All news](https://qrsntu.org/news/) ## Current Work A selection of active QRS research, projects, and competitions. [View all projects](https://qrsntu.org/projects/) ## Supported by [About our sponsors](https://qrsntu.org/about/#sponsors) --- # About > About QRS@NTU, which develops and conducts student-led research, and its place in the wider QRS network. Web page: https://qrsntu.org/about/ (content last changed 2026-10-15) Site index for agents: https://qrsntu.org/llms.txt ## Our Vision QRS@NTU develops and conducts student-led research in quantitative finance, machine learning, and related technical areas. ## QRS and QRS@NTU **QRS** is the Society's cross-institutional network. Research is conducted under the broader QRS identity to enable collaboration across universities and institutions. **QRS@NTU** is its originating campus community at Nanyang Technological University, where student-led research is developed and conducted alongside NTU-based events and local activities. ## Core Focus Areas ### Quantitative Research Original research and projects spanning quantitative finance, machine learning, and related technical areas. See [Projects](https://qrsntu.org/projects/) and [Publications](https://qrsntu.org/publications/). ### Researcher Development Entry-stage research training through the [QRS Research Academy](https://qrsntu.org/programmes/#research-academy), with a path to QRS Researcher roles. ### Academic & Industry Activity Talks, workshops, and competitions, listed on the [Events](https://qrsntu.org/events/) page. ### Open Academic Resources Publish notes, problem sets, and past-year papers for NTU modules on [Math Notes](https://qrsntu.org/math_notes/). ### Workshop & Research Community Support Web, publicity, outreach, and organisational support for academic workshops and research communities. See [Service](https://qrsntu.org/service/). ## Sponsors & Partners The largest interview preparation platform for trading and quant finance worldwide. Purpose-built for every stage of the trading and quant interview, with mental maths drills, probability and reasoning tests, market-making games, cognitive games and coding problems. Many trading firms also use Tradermath to screen their own applicants. [tradermath.org](https://www.tradermath.org) --- # Contact > Contact the Quantitative Research Society (QRS) for enquiries and collaboration. Web page: https://qrsntu.org/contact/ (content last changed 2026-10-15) Site index for agents: https://qrsntu.org/llms.txt Get in touch with QRS. ## Get in Touch For general inquiries, partnership opportunities, or questions about our activities, please reach out to us. We look forward to hearing from you. [contact@qrsntu.org](mailto:contact@qrsntu.org) We aim to respond to all inquiries within 2-3 business days. --- # Using This Site with AI Assistants > How to use qrsntu.org with AI assistants such as Claude, ChatGPT, Gemini and Codex: a ready prompt, llms.txt, Markdown copies of every page, the site search index and the data files. Web page: https://qrsntu.org/agents/ (content last changed 2026-11-12) Site index for agents: https://qrsntu.org/llms.txt How Claude, ChatGPT, Gemini, Codex and other assistants can read qrsntu.org. ## Ask your assistant Copy this prompt into any assistant that can read web pages, then add your question. It tells the assistant to read this site instead of answering from memory, and to give a link for each fact. ``` Answer my questions about the Quantitative Research Society (QRS) from its website, https://qrsntu.org, not from memory. Read https://qrsntu.org/llms.txt first: it lists every page. To find the right pages, search https://qrsntu.org/data/site-index.json (every page and section as plain text, with its URL). For the full text of a page, add index.md to its address (https://qrsntu.org/about/ becomes https://qrsntu.org/about/index.md). Give the page URL for each fact. If the site does not answer a question, say so instead of guessing. My question: [write it here] ``` The search box at the top of every page (or the / key) offers the same prompt with your search words filled in. ## What assistants can read Everything below is a plain file on the site. No account, API key or JavaScript is needed. - [llms.txt](https://qrsntu.org/llms.txt): an index of every page, by section, with a one-line description of each. - **index.md** beside every page: the page as Markdown, without the layout. Lists that pages draw with JavaScript, such as people, projects and publications, are written out in full. - [data/site-index.json](https://qrsntu.org/data/site-index.json): every page and section as plain text with its URL, plus the Math Notes courses, publications and events. The site's search box reads this file. - [llms-full.txt](https://qrsntu.org/llms-full.txt): all pages in one file. Pages longer than about 8 KB are shortened there and name their full Markdown copy. - **summary.md** on each research project, for example [GeoAFT](https://qrsntu.org/projects/geoaft/summary.md): a short account of the design, the results and the limitations, written from the evidence behind the project pages. - The data behind the pages as JSON: [people](https://qrsntu.org/data/people.json), [projects](https://qrsntu.org/data/projects.json), [publications](https://qrsntu.org/data/publications.json), [events](https://qrsntu.org/data/events.json), [news](https://qrsntu.org/data/news.json) and [Math Notes courses](https://qrsntu.org/courses.json), with a download link for every past paper. ## By assistant - **Claude, ChatGPT and Gemini apps:** turn on web search or browsing if your plan has it, then paste the prompt above. Assistants without web access cannot read the site and will answer from memory. - **Claude Code, Codex and other coding agents:** ask the agent to fetch https://qrsntu.org/llms.txt, or to download data/site-index.json and search it locally. - **Your own code:** fetch the files above directly. They change only when the site is rebuilt; the [sitemap](https://qrsntu.org/sitemap.xml) gives the date each page's content last changed. ## NeurIPS 2026 guide: API and MCP Our [NeurIPS 2026 attendee guide](https://qrsntu.org/service/neurips26/) has more for assistants than the rest of the site: a REST API and MCP servers that search the schedule, check clashes and save an itinerary at a permanent link. Setup for Claude, ChatGPT and Gemini is on its [AI assistants page](https://qrsntu.org/service/neurips26/agents/), and its own index is [service/neurips26/llms.txt](https://qrsntu.org/service/neurips26/llms.txt). The rest of qrsntu.org has no API or MCP server. It is small enough that the files above answer the same questions. ## What these files hold They hold only what the public pages show. Research Academy applications and assessments are not on this site and are not in these files; see [Privacy](https://qrsntu.org/privacy/). Assistants can make mistakes: the web page is the reference, and corrections are welcome at [contact@qrsntu.org](mailto:contact@qrsntu.org). Last updated 12 November 2026. --- # Programmes > The QRS Research Academy is QRS's research programme for students, with applications for 2026 Fall #4, and a path to QRS Researcher roles. Web page: https://qrsntu.org/programmes/ (content last changed 2026-10-23) Site index for agents: https://qrsntu.org/llms.txt Entry-stage research training at QRS@NTU, with an introductory programme under development. Active programme · 2026 Fall #4 recruiting ### QRS Research Academy From guided research to independent research leadership. Academy members lead a derived project as Slicers, under mentorship, towards a first-author workshop paper. 2026 Fall #4 begins with the December 2026 – January 2027 research cycle, aimed at relevant AAMAS 2027 workshops. 2026 Fall #3 is working towards AAAI. Completed: 2026 Fall #1 (five works accepted to NeurIPS 2026 workshops) and 2026 Fall #2 (six works accepted to ICAIF 2026 workshops). [Apply now](https://forms.gle/GGV5hH8oQqBY1ReP8) [View details](https://qrsntu.org/research-academy/) QRS@NTU · TBC ### Quant Foundations An introductory quantitative programme under consideration at QRS@NTU. Details will be announced once confirmed. Status --- # QRS Research Academy: Cohorts, Outcomes and Applications > A selective QRS research programme, from guided research to independent research leadership: selection process, how the Academy works, cohorts and applications for 2026 Fall #4. Web page: https://qrsntu.org/research-academy/ (content last changed 2026-10-24) Site index for agents: https://qrsntu.org/llms.txt 1. [Programmes](https://qrsntu.org/programmes/) 2. Research Academy Research programme From guided research to independent research leadership. A selective, research-intensive programme for academically capable students. Academy members join an active QRS research programme as Slicers: each leads a derived research project of their own, under structured mentorship, towards a first-author workshop paper, with the opportunity to continue towards archival publications. Once you are assigned a derived project and carry out the full research on it, including its extension, first authorship of at least one workshop paper from that work is guaranteed. First authorship on archival-venue papers is determined by contribution. Selection weighs research aptitude, intellectual curiosity and analytical rigour above existing credentials. The Academy is a research environment built around real projects. [Apply to QRS Research Academy](https://forms.gle/GGV5hH8oQqBY1ReP8) · [Meet Our Researchers](https://qrsntu.org/research-academy/members/) The form opens in Google Forms and asks you to sign in. Recruiting ## 2026 Fall #4 Upcoming intake Research cycle December 2026 – January 2027 Target venue Relevant AAMAS 2027 workshops Eligibility and how to apply - 2026 Fall #4 - Selection - How it works - Cohorts - Roles - Research - FAQ ## Recruitment: 2026 Fall #4 Upcoming intake Research cycle December 2026 – January 2027. > We particularly welcome early-stage undergraduate and graduate students seeking to undertake original research. Exceptional pre-university students will also be considered. Prior research or publication experience is not required; selection emphasises intellectual curiosity, analytical rigour, and the potential to develop into independent researchers. ### Who should apply - Early-stage undergraduate and graduate students who want to carry out original research. - Exceptional pre-university students, considered individually. - Students at any university; the programme is not limited to NTU. - No prior research or publication record is needed. ### Expected aptitude - Quantitative maturity: probability, statistics, linear algebra or related mathematics at the level needed to follow technical papers. - Programming ability sufficient to implement and test ideas, typically in Python. - Research habits: careful reading, precise writing, and honesty about what the evidence supports. - Project-specific prerequisites are discussed during selection. ### What participants work towards - Joining an active QRS research programme as a Slicer, under mentorship. - Leading a derived research project of your own within that programme, including its extension. - Writing at least one first-author workshop paper from that project. - Continuing towards broader responsibilities as a QRS Researcher. ### How to apply 1. Submit the application form with your résumé. 2. Shortlisted applicants are invited through the later stages described under Selection. 3. Selected applicants join the cohort and are assigned a derived project to lead as a Slicer. ### Research cycles and target venues **2026 Fall #4** begins with the December 2026 – January 2027 research cycle, targeting relevant AAMAS 2027 workshops. Each research cycle is organised around target workshops, chosen by each project's scope and the workshop's deadlines. Acceptance is decided by the workshop's reviewers. [Apply to QRS Research Academy](https://forms.gle/GGV5hH8oQqBY1ReP8) · [Questions: contact@qrsntu.org](mailto:contact@qrsntu.org) ## Selection Process Selection has four stages. Shortlisted applicants are invited to each next stage, and dates are sent directly. 1. ### Application and background review The committee reads your application form and résumé: projects, research experience, and availability for the research cycle. 2. ### Online assessment Shortlisted applicants take two short, timed online tests: quantitative reasoning, and qualitative reasoning (reading comprehension, interpretation and logic). 3. ### Research assessment and interview Invited applicants choose one research question and write a short research note on it. Correct partial results are expected. The interview then discusses that note. 4. ### Final selection and project placement The committee selects the cohort, and each selected participant is assigned a derived project in an active research programme to lead as a Slicer. ## How the Academy Works Five stages from guided contributions to independent research. 1. ### Research Orientation Literature review, identifying open questions, and evaluating potential research directions. 2. ### Research Formulation Defining a tractable question, its novelty, and a theoretical or empirical approach. 3. ### Research Execution Developing methods, performing experiments or proofs, establishing baselines, and critically evaluating evidence. 4. ### Academic Communication Preparing manuscripts, conducting internal peer review, and identifying appropriate publication venues. 5. ### Continued Research Opportunities to assume greater research responsibilities or progress to QRS Researcher roles. ## Cohorts Each cohort's research cycle, target venue and outcome. 1. ### 2026 Fall #1 Completed Inaugural cohort Target venue: NeurIPS 2026 workshops Outcome: [Five works accepted to NeurIPS 2026 workshops](https://qrsntu.org/news/neurips-2026-workshops/) Members: [Jianing Zhang](https://qrsntu.org/jn/), [Xiaotong Li](https://qrsntu.org/people/xiaotong-li/), [Yingzhi Tang](https://qrsntu.org/yz/) [Members and alumni](https://qrsntu.org/research-academy/members/#2026-fall-1) 2. ### 2026 Fall #2 Completed Target venue: ICAIF 2026 workshops Outcome: [Six works accepted to ICAIF 2026 workshops](https://qrsntu.org/news/icaif-2026-workshops/) Members: [Shufang Chen](https://qrsntu.org/people/shufang-chen/), [Wang Zimeng](https://qrsntu.org/people/wang-zimeng/), [Xinyue Wu](https://qrsntu.org/people/xinyue-wu/), [Jingshu Liu](https://qrsntu.org/people/jingshu-liu/), [Yizhe Chia](https://qrsntu.org/people/yizhe-chia/), [Annie Huaiyu Wang](https://qrsntu.org/people/annie-huaiyu-wang/) [Members and alumni](https://qrsntu.org/research-academy/members/#2026-fall-2) 3. ### 2026 Fall #3 Working Target venue: AAAI Workshops 4. ### 2026 Fall #4 Upcoming intake Research cycle December 2026 – January 2027 Target venue: Relevant AAMAS 2027 workshops [Apply now](https://forms.gle/GGV5hH8oQqBY1ReP8) [Shortened here. The full page is https://qrsntu.org/research-academy/index.md; it goes on with: Project Roles and Progression; Research Directions; Frequently Asked Questions; Apply for 2026 Fall #4.] --- # Join QRS > Apply to the QRS Research Academy (2026 Fall #4) or get involved with the QRS@NTU community. Web page: https://qrsntu.org/join_us/ (content last changed 2026-10-23) Site index for agents: https://qrsntu.org/llms.txt Ways to get involved with QRS and QRS@NTU. ## QRS@NTU Community Help run NTU-based events, local activities, and the operations and infrastructure of QRS@NTU. [Upcoming Events](https://qrsntu.org/events/) ## Quant Foundations — TBC An introductory programme at QRS@NTU is under development. Details will be announced once confirmed. [Status](https://qrsntu.org/programmes/#quant-foundations) ## QRS Research Academy A selective research programme, from guided research towards leading your own project. 2026 Fall #4 is recruiting for the December 2026 – January 2027 research cycle. [Apply to QRS Research Academy](https://forms.gle/GGV5hH8oQqBY1ReP8) [Programme detail and outcomes](https://qrsntu.org/research-academy/) [Current Projects](https://qrsntu.org/projects/) To get involved, reach out to us directly: [contact@qrsntu.org](mailto:contact@qrsntu.org) --- # Application Prerequisite Guidelines > Prerequisite guidelines for applying to QRS@NTU: mathematics, probability and statistics, and programming foundations. Web page: https://qrsntu.org/join_us/prerequisites/ (content last changed 2026-10-15) Site index for agents: https://qrsntu.org/llms.txt [← Join QRS](https://qrsntu.org/join_us/) Quantitative Research Society @NTU is an advanced, research-focused society. To ensure applicants are prepared for the rigor of our workshops, all applicants must complete foundational coursework in **mathematics, statistics, and programming**. Applicants may satisfy prerequisites through: - **Equivalent NTU undergraduate modules** (see [NTU SPMS course list](https://www.ntu.edu.sg/spms/about-us/mathematics/undergrad/course-info)), or - **Approved external MOOCs** (listed below). All prerequisites must be completed for a **letter grade (B or above)** or an official **certificate of completion**. To apply, you must complete **one course in each mandatory category (Math, Stats, Programming)**. Completion of recommended or advanced preparation courses may allow applicants to **start directly in Semester 2 or 3, or participate in research projects immediately**. ## 1. Mathematics Foundations (Linear Algebra & Analysis) — *mandatory* A solid background in linear algebra and introductory analysis is essential for functional analysis, PDEs, and optimization. - [Advanced Linear Algebra: Foundations to Frontiers (edX, US\$99)](https://www.edx.org/course/advanced-linear-algebra-foundations-to-frontiers?index=product&queryID=04c4ebbab19190b38953f77a3176861d&position=2) - [Mathematics for Engineers Specialization (Coursera, free for NTU students)](https://www.coursera.org/specializations/mathematics-engineers) ✅ NTU equivalents include **MH1201 Linear Algebra II** and **MH3100 Real Analysis I**. ## 2. Probability & Statistics — *mandatory* Applicants must have prior exposure to probability and statistics, which underpin stochastic processes, econometrics, and data-driven modeling. - [Advanced Statistics for Data Science Specialization (Coursera Plus)](https://www.coursera.org/specializations/advanced-statistics-data-science#courses) - [Econometrics (Coursera, free for NTU students)](https://www.coursera.org/learn/erasmus-econometrics) - [Data Analysis with R Specialization (Coursera, free for NTU students)](https://www.coursera.org/specializations/statistics#courses) ✅ NTU equivalents include **MH2500 Probability** and **HE2003 Econometrics I**. ## 3. Programming & Algorithms — *mandatory* Strong programming skills are required for research projects, simulations, and optimization. Applicants should be proficient in **Python or R**; **C++ is optional but recommended** for computational finance. - [Algorithms Specialization (Coursera, free for NTU students)](https://www.coursera.org/specializations/algorithms#courses) - Practice on coding platforms: [LeetCode](https://leetcode.com/) | [HackerRank](https://www.hackerrank.com/) ✅ NTU equivalents include **MH4701 Mathematical Programming**. ## 4. Recommended Preparation — *optional but can accelerate placement* Applicants completing any of these courses may be allowed to **skip Semester 1** and start in **Semester 2 or 3**. - **Numerical Methods**: [Numerical Analysis (Coursera, free for NTU students)](https://www.coursera.org/learn/computers-waves-simulations) - **Optimization**: [Convex Optimization (Stanford Online, free)](https://www.edx.org/learn/engineering/stanford-university-convex-optimization) - **Machine Learning**: [Machine Learning Specialization (Coursera Plus)](https://www.coursera.org/specializations/machine-learning-introduction) - **Finance & Economics**: - [Financial Markets (Coursera, free for NTU students)](https://www.coursera.org/learn/financial-markets-global#about) - [Principles of Economics: Macroeconomics (Coursera, free for NTU students)](https://www.coursera.org/learn/principles-of-economics-macroeconomics/) - **Finance Applications**: [Akuna Capital Options 101 (free)](https://akunacapital.teachable.com/p/options101) ✅ NTU equivalents include **MH3700 Numerical Analysis I**, **MH3701 Basic Optimization**, **MH4514 Financial Mathematics**, and **HE2002 Macroeconomics II**. ## 5. Advanced Preparation — *guaranteed accelerated placement / research readiness* Completion of any of the following programs demonstrates **advanced mastery**, allowing applicants to **enter Semester 2, Semester 3, or participate in research projects immediately**. - [MFE Preparation | Master of Financial Engineering | Berkeley Haas](https://mfe.haas.berkeley.edu/academics/mfe-preparation) (US$1,995 per course) - **Math Foundations for Financial Engineers**: 30-hour course covering essential math frameworks for financial modeling. - **Statistics for Financial Engineers**: 30-hour course covering probability, statistics, and modeling techniques. (VEE Approved) - [Pre-MFE Program | Baruch College](https://mfe.baruch.cuny.edu/pre-mfe-program/) (US$1,650 per course) - **Advanced Calculus with Financial Engineering Applications** - **Probability Theory for Financial Applications** - **Numerical Linear Algebra for Financial Engineering** - **Machine Learning for Financial Engineering Applications** Completion of these programs demonstrates strong quantitative and computational foundations and may allow direct placement into **QRS\@NTU advanced workshops or research projects**. ✨ **Summary:** - **Mandatory**: Math + Probability/Stats + Programming - **Recommended**: accelerates Semester 2/3 placement - **Advanced Preparation**: guarantees advanced workshop or research eligibility --- # Research Academy Members > Meet QRS Research Academy members and alumni by cohort, with their research areas, accepted works, and researchers promoted from the Academy. Web page: https://qrsntu.org/research-academy/members/ (content last changed 2026-10-24) Site index for agents: https://qrsntu.org/llms.txt 1. [Programmes](https://qrsntu.org/programmes/) 2. [Research Academy](https://qrsntu.org/research-academy/) 3. Members Current members and alumni, with persistent cohort history. ## QRS Research Academy, 2026 Fall #1 Completed Target venue: NeurIPS 2026 workshops. [Five works accepted to NeurIPS 2026 workshops](https://qrsntu.org/news/neurips-2026-workshops/) ### [Jianing Zhang](https://qrsntu.org/jn/) Researcher, Quantitative Research Society Research areas: [AI Native Workflow](https://qrsntu.org/projects/ai-native-workflow/), [GeoAttention](https://qrsntu.org/projects/geoattention/), [SoftRank](https://qrsntu.org/projects/softrank/) Accepted works: [AutoMLR Workshop, NeurIPS 2026](https://qrsntu.org/projects/ai-native-workflow/#research-geoaft-ai-native-workflow-automlr-2026); [NeurReps Workshop, NeurIPS 2026](https://qrsntu.org/projects/softrank/#research-approximation-rank-neurreps-ea-2026) ### [Xiaotong Li](https://qrsntu.org/people/xiaotong-li/) Research Academy Member Research areas: [AI Native Workflow](https://qrsntu.org/projects/ai-native-workflow/), [GeoAttention](https://qrsntu.org/projects/geoattention/) Accepted works: [AutoMLR Workshop, NeurIPS 2026](https://qrsntu.org/projects/ai-native-workflow/#research-geoaft-ai-native-workflow-automlr-2026) ### [Yingzhi Tang](https://qrsntu.org/yz/) Researcher, Quantitative Research Society Research areas: [PolicyAttention](https://qrsntu.org/projects/policyattention/), [RiskHedge](https://qrsntu.org/projects/riskhedge/) Slicer — [RiskHedge](https://qrsntu.org/projects/riskhedge/) Accepted works: [EvoRobust Workshop, NeurIPS 2026](https://qrsntu.org/projects/policyattention/#research-policyattention-evorobust-2026); [EvoRobust Workshop, NeurIPS 2026](https://qrsntu.org/projects/riskhedge/#research-riskhedge-evorobust-2026) ## QRS Research Academy, 2026 Fall #2 Completed Target venue: ICAIF 2026 workshops. [Six works accepted to ICAIF 2026 workshops](https://qrsntu.org/news/icaif-2026-workshops/) ### [Shufang Chen](https://qrsntu.org/people/shufang-chen/) Research Academy Member Slicer — [PolicyAttention](https://qrsntu.org/projects/policyattention/) Accepted works: [AutoMLR Workshop, NeurIPS 2026](https://qrsntu.org/projects/ai-native-workflow/#research-geoaft-ai-native-workflow-automlr-2026) ### [Wang Zimeng](https://qrsntu.org/people/wang-zimeng/) Research Academy Member · NUS Sub-branch Coordinator Slicer — [RiskHedge](https://qrsntu.org/projects/riskhedge/) Accepted works: [AutoMLR Workshop, NeurIPS 2026](https://qrsntu.org/projects/ai-native-workflow/#research-geoaft-ai-native-workflow-automlr-2026) ### [Xinyue Wu](https://qrsntu.org/people/xinyue-wu/) Research Academy Member Slicer — [PolicyAttention](https://qrsntu.org/projects/policyattention/) Accepted works: [EvoRobust Workshop, NeurIPS 2026](https://qrsntu.org/projects/policyattention/#research-policyattention-evorobust-2026) ### [Jingshu Liu](https://qrsntu.org/people/jingshu-liu/) Research Academy Member Slicer — [SoftRank](https://qrsntu.org/projects/softrank/) ### [Yizhe Chia](https://qrsntu.org/people/yizhe-chia/) Research Academy Member Slicer — [AI Native Workflow](https://qrsntu.org/projects/ai-native-workflow/) ### [Annie Huaiyu Wang](https://qrsntu.org/people/annie-huaiyu-wang/) Research Academy Member Slicer — [FRTBTrace](https://qrsntu.org/projects/frtbtrace/) ## QRS Research Academy, 2026 Fall #4 Upcoming intake Target venue: Relevant AAMAS 2027 workshops. Members will be listed here once the cohort is selected. Research cycle: December 2026 – January 2027. [Apply now](https://forms.gle/GGV5hH8oQqBY1ReP8) --- # Projects, Competitions & Recognition > Explore QRS research projects, competitions, and recognition in quantitative finance, machine learning, and mathematics. Web page: https://qrsntu.org/projects/ (content last changed 2026-11-12) Site index for agents: https://qrsntu.org/llms.txt Our portfolio of research, competitions, and infrastructure work. Research · Infrastructure · Active ### [AI Native Workflow](https://qrsntu.org/projects/ai-native-workflow/) Research on hierarchical, verification-gated and recoverable workflows for scaling long-horizon AI-assisted research and agent execution. Lead: [**Yuhe** Sui](https://qrsntu.org/yh/) Authors: [**Jianing** Zhang](https://qrsntu.org/jn/), [Shufang Chen](https://qrsntu.org/people/shufang-chen/), [**Danlin** Chen](https://qrsntu.org/cdl/), [Wang **Zimeng**](https://qrsntu.org/people/wang-zimeng/), [Xiaotong Li](https://qrsntu.org/people/xiaotong-li/) Slicers: [Yizhe Chia](https://qrsntu.org/people/yizhe-chia/) Research · Active ### [FRTBTrace](https://qrsntu.org/projects/frtbtrace/) Active research on faithful, auditable, trade-level FRTB attribution. Lead: [**Yuhe** Sui](https://qrsntu.org/yh/) Authors: [**Yingzhi** Tang](https://qrsntu.org/yz/) Slicers: [**Annie** Huaiyu Wang](https://qrsntu.org/people/annie-huaiyu-wang/) [[arXiv]](https://arxiv.org/abs/2609.32238) Research · Active ### [GeoAFT](https://qrsntu.org/projects/geoaft/) Geometric Attention Fine-Tuning: adapting the rule a pretrained Transformer uses to compare representations, not only the representations themselves. Lead: [**Yuhe** Sui](https://qrsntu.org/yh/) Research · Active ### [GeoAttention](https://qrsntu.org/projects/geoattention/) Research on attention geometry, representation-relative identifiability, and geometry-adaptive prediction. Lead: [**Yuhe** Sui](https://qrsntu.org/yh/) Authors: [**Jianing** Zhang](https://qrsntu.org/jn/), [Xiaotong Li](https://qrsntu.org/people/xiaotong-li/) Research · Active ### [PolicyAttention](https://qrsntu.org/projects/policyattention/) Research on softmax policy improvement, its robustness, and its failure modes. Lead: [**Yuhe** Sui](https://qrsntu.org/yh/) Authors: [Shufang Chen](https://qrsntu.org/people/shufang-chen/), [**Yingzhi** Tang](https://qrsntu.org/yz/), [Xinyue Wu](https://qrsntu.org/people/xinyue-wu/) Slicers: [Shufang Chen](https://qrsntu.org/people/shufang-chen/), [Xinyue Wu](https://qrsntu.org/people/xinyue-wu/) [[arXiv]](https://arxiv.org/abs/2609.30500) Research · Active ### [RiskHedge](https://qrsntu.org/projects/riskhedge/) Research on spectral-risk decisions, diverse scenarios, and selective robustness evaluation. Lead: [**Yuhe** Sui](https://qrsntu.org/yh/) Authors: [**Yingzhi** Tang](https://qrsntu.org/yz/), [**Jianing** Zhang](https://qrsntu.org/jn/), [Mi **Zeyuan**](https://qrsntu.org/people/mi-zeyuan/) Slicers: [Wang **Zimeng**](https://qrsntu.org/people/wang-zimeng/), [**Yingzhi** Tang](https://qrsntu.org/yz/) Research · Active ### [SoftRank](https://qrsntu.org/projects/softrank/) Theoretical research on softmax-attention approximation rank, support geometry, and robust interaction dimension. Lead: [**Yuhe** Sui](https://qrsntu.org/yh/) Authors: [**Jianing** Zhang](https://qrsntu.org/jn/), [**Yingzhi** Tang](https://qrsntu.org/yz/) Slicers: [Jingshu Liu](https://qrsntu.org/people/jingshu-liu/) [[arXiv]](https://arxiv.org/abs/2608.28150) Research · Early-stage / Active derived research ### [HedgeFidelity](https://qrsntu.org/projects/hedgefidelity/) An early-stage research programme with derived research in progress. Lead: [**Yuhe** Sui](https://qrsntu.org/yh/) Research · Early-stage / Active derived research ### [DecisionRank](https://qrsntu.org/projects/decisionrank/) An early-stage research programme with derived research in progress. Lead: [**Yuhe** Sui](https://qrsntu.org/yh/) Infrastructure · Active ### [ntuinfo.com](https://qrsntu.org/projects/ntuinfo-com/) A bilingual English/Chinese full-stack web application with a quantitative, multi-criteria scoring framework (mental load, physical load, difficulty) that ranks 100+ NTU student clubs and surfaces algorithmically-driven recommendations for students. Participants: [**Yuhe** Sui](https://qrsntu.org/yh/) [[Website]](https://ntuinfo.com) Research · Competition · Published ### [AlphaAllocator](https://qrsntu.org/projects/alphaallocator/) A constrained Chinese financial-news overlay on classical risk-parity allocation, presented as a poster and oral at NLPCC 2026 Shared Task 4. Participants: [**Yuhe** Sui](https://qrsntu.org/yh/) #### Recognition - NLPCC 2026 Shared Task 4 — Track 1: **1st** - NLPCC 2026 Shared Task 4 — Track 2: **6th** [[OpenReview]](https://openreview.net/forum?id=u09PzTYNHk) Competition · Research · Active ### [CFFA 2026 — China Future Financial Analyst Competition](https://qrsntu.org/projects/cffa-2026/) 2026 第五届中国未来金融分析师大赛 • Built a Greenwashing Risk Score across 1,676 Chinese green-bond issuances (2016–2025), using issue-spread regressions and portfolio sorts. • Pricing effects are weak unconditionally, but meaningful within credit-risk buckets after controls. Participants: [**Yuhe** Sui](https://qrsntu.org/yh/), [Wang **Zimeng**](https://qrsntu.org/people/wang-zimeng/), [**Luis** Ji](https://qrsntu.org/people/luis-ji/), Wang Zhoufu, Yu Jiaxue #### Recognition - CFFA 2026: **National Finalist** [[GitHub]](https://github.com/yuhesui/CFFA-2026) Research · Active ### [MathResearchOS](https://qrsntu.org/projects/mathresearchos/) A recoverable, audited command-line system for mathematical work: one objective becomes a staged, verifier-gated run across Codex, Claude and API models, with every attempt kept on disk. Lead: [**Yuhe** Sui](https://qrsntu.org/yh/) [Shortened here. The full page is https://qrsntu.org/projects/index.md; it goes on with: All projects.] --- # Publications > Browse publications and preprints by QRS researchers, with author, year, and venue filters. Web page: https://qrsntu.org/publications/ (content last changed 2026-11-07) Site index for agents: https://qrsntu.org/llms.txt A curated list of publications from our members. ## Publications - **Diverse Scenarios, Selective Evaluation: Certified Spectral-Risk Decisions under Costly Robustness Tests** (2026). [Yuhe Sui](https://qrsntu.org/yh/), [Yingzhi Tang](https://qrsntu.org/yz/). EvoRobust Workshop, NeurIPS 2026. Accepted. [OpenReview](https://openreview.net/forum?id=OzOTmkkAWW) - **Geometric Attention Fine-Tuning Is Representation-Relative: Autonomous ML Research with AI-Native Workflow** (2026). [Yuhe Sui](https://qrsntu.org/yh/), [Jianing Zhang](https://qrsntu.org/jn/), [Shufang Chen](https://qrsntu.org/people/shufang-chen/), [Xiaotong Li](https://qrsntu.org/people/xiaotong-li/), [Wang Zimeng](https://qrsntu.org/people/wang-zimeng/). AutoMLR Workshop, NeurIPS 2026. Accepted. [OpenReview](https://openreview.net/forum?id=9PwlVQ3gyj) - **NTU-QRS AlphaAllocator: Bounded Transformer News Overlays for Chinese Macro-Asset Allocation** (2026). [Yuhe Sui](https://qrsntu.org/yh/). NLPCC 2026 Shared Task 4 · Poster + Oral. [OpenReview](https://openreview.net/forum?id=u09PzTYNHk) - **PolicyAttention: Robustness and Failure Modes of Softmax Policy Improvement** (2026). [Yuhe Sui](https://qrsntu.org/yh/), [Yingzhi Tang](https://qrsntu.org/yz/), [Xinyue Wu](https://qrsntu.org/people/xinyue-wu/). EvoRobust Workshop, NeurIPS 2026. Accepted. [OpenReview](https://openreview.net/forum?id=VgswuGrqp6) - **PolicyAttention: When Softmax Attention Becomes Policy Improvement** (2026). [Yuhe Sui](https://qrsntu.org/yh/). PTA Workshop, NeurIPS 2026. Accepted. [OpenReview](https://openreview.net/forum?id=zbGWnk0kY3) - **The Approximation Rank of Softmax Attention: Sharp Geometric Laws and Robust Interaction Dimension** (2026). [Yuhe Sui](https://qrsntu.org/yh/), [Jianing Zhang](https://qrsntu.org/jn/). NeurReps Workshop, NeurIPS 2026 (extended abstract). Accepted. [OpenReview](https://openreview.net/forum?id=0zgsZ8FjVz) [arXiv](https://arxiv.org/abs/2608.28150) --- # AI Native Workflow > Research on hierarchical, verification-gated and recoverable workflows for scaling long-horizon AI-assisted research and agent execution. Web page: https://qrsntu.org/projects/ai-native-workflow/ (content last changed 2026-10-15) Site index for agents: https://qrsntu.org/llms.txt [← All projects](https://qrsntu.org/projects/) Active Research on hierarchical, verification-gated and recoverable workflows for scaling long-horizon AI-assisted research and agent execution. Lead: [Yuhe Sui](https://qrsntu.org/yh/) Authors: [Jianing Zhang](https://qrsntu.org/jn/), [Shufang Chen](https://qrsntu.org/people/shufang-chen/), [Danlin Chen](https://qrsntu.org/cdl/), [Wang Zimeng](https://qrsntu.org/people/wang-zimeng/), [Xiaotong Li](https://qrsntu.org/people/xiaotong-li/) Slicers: [Yizhe Chia](https://qrsntu.org/people/yizhe-chia/) ## Accepted Research - ### Geometric Attention Fine-Tuning Is Representation-Relative: Autonomous ML Research with AI-Native Workflow [Yuhe Sui](https://qrsntu.org/yh/), [Jianing Zhang](https://qrsntu.org/jn/), [Shufang Chen](https://qrsntu.org/people/shufang-chen/), [Xiaotong Li](https://qrsntu.org/people/xiaotong-li/), [Wang Zimeng](https://qrsntu.org/people/wang-zimeng/) AutoMLR Workshop, NeurIPS 2026 · 2026 · Accepted [[OpenReview]](https://openreview.net/forum?id=9PwlVQ3gyj) ## Related projects [GeoAFT](https://qrsntu.org/projects/geoaft/) ## Publications - **Geometric Attention Fine-Tuning Is Representation-Relative: Autonomous ML Research with AI-Native Workflow** (2026). [Yuhe Sui](https://qrsntu.org/yh/), [Jianing Zhang](https://qrsntu.org/jn/), [Shufang Chen](https://qrsntu.org/people/shufang-chen/), [Xiaotong Li](https://qrsntu.org/people/xiaotong-li/), [Wang Zimeng](https://qrsntu.org/people/wang-zimeng/). AutoMLR Workshop, NeurIPS 2026. Accepted. [OpenReview](https://openreview.net/forum?id=9PwlVQ3gyj) --- # AlphaAllocator > AlphaAllocator: a news-reading allocation agent that asks one small model which asset a story is about and another whether it is good or bad, and lets a risk model allocate. 1st in NLPCC 2026 Track 1; ICAIF 2026 coming soon. Web page: https://qrsntu.org/projects/alphaallocator/ (content last changed 2026-11-11) Site index for agents: https://qrsntu.org/llms.txt [← All projects](https://qrsntu.org/projects/) QRS research · Trading agents · Financial NLP Bounded news overlays for risk-aware asset allocation Don’t ask one language model for the trade. Ask small models narrow questions, and let a risk model allocate. A common way to build a news-reading trading agent is to hand the news to a large language model and let it produce the trade, an overlay or an event signal directly. AlphaAllocator splits that job into the two questions news can answer: **which asset is this about**, and **is it good or bad**, each with its own small model. A classical risk model decides the allocation, and the news may only nudge it within a fixed budget. [Details](https://qrsntu.org/projects/alphaallocator/details/) [Paper (OpenReview)](https://openreview.net/forum?id=u09PzTYNHk) · [Talk slides (PDF)](https://qrsntu.org/projects/alphaallocator/assets/alphaallocator-nlpcc2026-talk.pdf) · [Plain-text summary](https://qrsntu.org/projects/alphaallocator/summary.md) Published Participants: [Yuhe Sui](https://qrsntu.org/yh/) NLPCC 2026: 1st of 10 in Track 1 · ICAIF 2026: coming soon Illustrative example; the agent reads up to 20 items a day. ## AlphaAllocator at a glance The idea For news to help a portfolio, an agent has to know two things: which asset an item is about, and whether it is good or bad news. We give each question its own small model and multiply the answers: $$news score_j = relevance_j × tone_j$$ The scores, weighted for recency and certainty and centred across assets, tilt a risk-parity portfolio. The whole tilt is capped at 1.0% of the portfolio, so prices decide and news only nudges. Experiments #### NLPCC 2026 · Chinese funds tested | Track 1 | Sharpe | Ann. return | Max DD | |---|---|---|---| | 1st | 1.1144 | 13.52% | −6.23% | 2025 replay Sharpe, with news · 2.270 · same agent, news zeroed · 2.198 Track 1: 11 macro assets, 10 teams, organizer-run on a hidden 2026 period (next-best Sharpe 0.7792). Track 2 (sectors): 6th of 13. #### ICAIF 2026 · US equities Coming soon The same design, adapted to 30 US stocks and 7 decisions a day, is in the ACM ICAIF 2026 Trading Agent Competition. Results coming soon. **Every team’s official Sharpe, NLPCC 2026.** The organizers’ hidden 2026 evaluation, in rank order, ours in teal. Track 1: 1st of 10. Track 2: 6th of 13. [More results →](https://qrsntu.org/projects/alphaallocator/details/#results) [Limitations and what didn’t work →](https://qrsntu.org/projects/alphaallocator/details/#limitations) ## Where this is going proposed 1. ### Cleaner tests of the news Preregistered controls over several periods and markets, with the official order and cost rules. 2. ### Better links from news to assets Entity and event linking, so a query can tell direct, upstream, downstream and market-wide news apart. 3. ### Constraints that are enforced Pass the realised previous weights to every guard and test how often each limit binds, end to end. [What we would change, in full →](https://qrsntu.org/projects/alphaallocator/details/#next-steps) ## Papers - ### NTU-QRS AlphaAllocator: Bounded Transformer News Overlays for Chinese Macro-Asset Allocation NLPCC 2026 Shared Task 4 · poster and oral · oral at Evaluation Workshop I, 5 November 2026 [[OpenReview]](https://openreview.net/forum?id=u09PzTYNHk) [[Talk slides, PDF]](https://qrsntu.org/projects/alphaallocator/assets/alphaallocator-nlpcc2026-talk.pdf) - ### ACM ICAIF 2026 Trading Agent Competition Write-up coming soon ## Recognition - NLPCC 2026 Shared Task 4 · Track 1 · 1st - NLPCC 2026 Shared Task 4 · Track 2 · 6th Scan to open this page ## Publications - **NTU-QRS AlphaAllocator: Bounded Transformer News Overlays for Chinese Macro-Asset Allocation** (2026). [Yuhe Sui](https://qrsntu.org/yh/). NLPCC 2026 Shared Task 4 · Poster + Oral. [OpenReview](https://openreview.net/forum?id=u09PzTYNHk) --- # AlphaAllocator: How It Works > How AlphaAllocator works: why not one LLM, two small models for relevance and tone, a capped news tilt on a risk-parity base, the leakage-safe backtest, NLPCC 2026 results and ablations, and reproducibility notes. Web page: https://qrsntu.org/projects/alphaallocator/details/ (content last changed 2026-11-11) Site index for agents: https://qrsntu.org/llms.txt [← AlphaAllocator overview](https://qrsntu.org/projects/alphaallocator/) AlphaAllocator · How it works Small models answer the two questions news can answer: which asset a story is about, and whether it is good or bad. A risk model decides the allocation, and the news only nudges it. [Paper (OpenReview)](https://openreview.net/forum?id=u09PzTYNHk) · [Talk slides (PDF)](https://qrsntu.org/projects/alphaallocator/assets/alphaallocator-nlpcc2026-talk.pdf) · [Plain-text summary](https://qrsntu.org/projects/alphaallocator/summary.md) Participants: [Yuhe Sui](https://qrsntu.org/yh/) Contents 1. The idea 2. Why not just ask an LLM? 3. Two questions, two small models 4. Why the news only nudges 5. The system in one picture 6. Where it runs 7. The evidence 8. A leakage-safe backtest 9. Results: Track 1 10. News on and off 11. ICAIF 2026 12. Limitations and what didn’t work 13. What we would change 14. Technical details 15. The two quantitative bases 16. The news score 17. From tilt to orders 18. Full tables 19. Reproducibility notes 20. Summary 21. In brief 22. Sources Part 1 · The idea ## 1. Why not just ask an LLM? The obvious way to build a trading agent that reads the news is to give the headlines to a large language model and ask it for the answer: the trade itself, a tilt on top of a portfolio, or an event signal to act on. It is quick to build, and one model does everything. That is also the problem. Inside one answer the model has to decide **which asset** a story is about, **whether** it is good or bad news, and **how much** money to move. When the result is wrong, you cannot tell which of the three went wrong; the answer can shift with the wording of the prompt; and nothing in it knows about risk, turnover or the limits of the portfolio. Our project’s methodology rejected a direct LLM allocator for exactly these reasons: prompt-sensitive, hard to reproduce and weakly risk-controlled. **Figure 1. One model for everything, or one question per part.** AlphaAllocator answers the two questions news can answer with two small, dedicated models, and leaves the question of how much to hold to a risk model, with a cap on how far the news may move it. So we asked a narrower question: *how could news help a portfolio at all?* Only if the agent knows what a story is about and whether it is good or bad. Those are two different skills, so they get two different models, and the allocation stays with the part of the system that understands risk. ## 2. Two questions, two small models ### Which asset? Relevance model A small sentence encoder (4 layers, BGE family) turns each article and a short description of each asset into vectors, and their similarity says how much the article is about that asset. One article can matter to several assets. ### Good or bad? Tone model A financial sentiment classifier (FinBERT) reads the same article and gives probabilities for positive, neutral and negative. The tone is positive minus negative, between −1 and +1. ### How much? Risk model, with a cap Not a language question. A risk-parity portfolio sets the weights from prices, and the news score may only tilt it, by at most 1.0% of the portfolio in total. For each asset the agent pools the day’s articles, weighting each by its relevance to that asset and by how recent it is, and multiplies **relevance × tone**, together with how sure the tone model is and how strongly the article matches one of 8 fixed event types. The result is one news score per asset. Both models are small, run locally and give the same output for the same input, so each can be inspected and tested on its own. ## 3. Why the news only nudges Even with good relevance and tone, turning each headline straight into a large position is fragile: small day-to-day changes in the news would make the portfolio churn. So the news score enters as a **small, centred adjustment**: what one asset gains, others give up, and the total size of the adjustment is capped. **Figure 2. Try it.** A base portfolio over the Track 1 assets, a news score per asset, and the tilt the agent would apply: centred, scaled by $η = 0.010$, capped so that $‖Δw‖_1 ≤ η$, then clipped and rescaled to 97% invested. Illustrative base and scores, not model output. The cap is a design choice, not a forecast: whatever the text models conclude, they cannot move more than 1.0% of the portfolio in Track 1 (1.2% in Track 2) before normalisation. ## 4. The system in one picture **Figure 3. The pipeline.** The stages carry the names of the paper’s Fig. 1. A price branch builds the quantitative base; a text branch reads the visible news with the two local models and produces one score per asset. The centred score becomes a capped tilt, which is added to the base, normalised and handed to the order planner. - **Prices decide.** In Track 1 a shrinkage risk-parity portfolio gives every asset an equal share of portfolio risk. - **Text is local and deterministic.** Both models load from local files, with no online calls; the encoder was adapted on 2024 data only. - **Failures are explicit.** No visible news gives a zero tilt. A missing model stops the agent with a stable failure state rather than downloading a substitute. ## 5. Where it runs Nothing in the design is tied to one market: the relevance model needs a short description of each asset, the tone model needs text in the market’s language, and the risk model needs prices. ### NLPCC 2026 Shared Task 4 tested Chinese funds · daily decisions · Chinese news **Track 1**, our main system: 11 macro and broad-market assets, from the CSI 300 and ChiNext to government bonds and gold. It placed 1st of 10 teams. Track 2, a secondary extension over 16 sector funds, placed 6th of 13. ### ACM ICAIF 2026 Trading Agent Competition Coming soon US equities · 7 decisions a day · English news The same design, adapted to 30 US-listed stocks. Results coming soon. Part 2 · The evidence ### Tested so far NLPCC 2026 - The complete system on the organizers’ hidden 2026 evaluation, both tracks - A replay of the submitted agents over all 243 public-2025 decision dates - Same-stack Track 1 controls: news zeroed, and news taken from 5 or 21 dates later - The packaged agents reproduce the development agents’ weights on all 243 replay dates ### In progress Coming soon - ACM ICAIF 2026 Trading Agent Competition: US equities, live evaluation - The ICAIF write-up [Shortened here. The full page is https://qrsntu.org/projects/alphaallocator/details/index.md; it goes on with: A leakage-safe backtest; Results: Track 1; News on and off; ICAIF 2026 Coming soon; Limitations and what didn’t work; What we would change; The two quantitative bases; The news score; From tilt to orders; Full tables; Reproducibility notes; In brief; Sources; Recognition.] --- # CFFA 2026 — China Future Financial Analyst Competition > • Built a Greenwashing Risk Score across 1,676 Chinese green-bond issuances (2016–2025), using issue-spread regressions and portfolio sorts. • Pricing effects are weak unconditionally, but meaningful within credit-risk buckets after controls. Web page: https://qrsntu.org/projects/cffa-2026/ (content last changed 2026-10-15) Site index for agents: https://qrsntu.org/llms.txt [← All projects](https://qrsntu.org/projects/) Active • Built a Greenwashing Risk Score across 1,676 Chinese green-bond issuances (2016–2025), using issue-spread regressions and portfolio sorts. • Pricing effects are weak unconditionally, but meaningful within credit-risk buckets after controls. 2026 第五届中国未来金融分析师大赛 Participants: [Yuhe Sui](https://qrsntu.org/yh/), [Wang Zimeng](https://qrsntu.org/people/wang-zimeng/), [Luis Ji](https://qrsntu.org/people/luis-ji/), Wang Zhoufu, Yu Jiaxue ## Recognition - CFFA 2026 · National Finalist [[GitHub]](https://github.com/yuhesui/CFFA-2026) --- # DecisionRank > An early-stage research programme with derived research in progress. Web page: https://qrsntu.org/projects/decisionrank/ (content last changed 2026-10-15) Site index for agents: https://qrsntu.org/llms.txt [← All projects](https://qrsntu.org/projects/) Early-stage / Active derived research An early-stage research programme with derived research in progress. Lead: [Yuhe Sui](https://qrsntu.org/yh/) --- # FRTBTrace > FRTBTrace: exact trade-level attribution of FRTB-IMA market-risk capital. Five typed questions, one reverse pass through the full formula, and a ledger on (date, trade) positions that adds up to capital. Web page: https://qrsntu.org/projects/frtbtrace/ (content last changed 2026-11-12) Site index for agents: https://qrsntu.org/llms.txt [← All projects](https://qrsntu.org/projects/) QRS research · Regulatory capital · Explainable finance Exact trade-level attribution of **FRTB**-IMA capital Trade-level explanations of FRTB market-risk capital, one question at a time. The usual way to explain a bank’s FRTB market-risk capital trade by trade is to compute one attribution vector, typically a gradient or a published allocation rule, and use it for everything: desk reports, what-if screening, trade rankings. FRTBTrace separates the **five questions** that vector is asked to answer and computes each one exactly through the implemented internal-models calculation, keeping track of the **past observation date** each part of the capital comes from. [Details](https://qrsntu.org/projects/frtbtrace/details/) [Paper (arXiv)](https://arxiv.org/abs/2609.32238) · [Plain-text summary](https://qrsntu.org/projects/frtbtrace/summary.md) · [Full results](https://qrsntu.org/projects/frtbtrace/results/) Active Lead: [Yuhe Sui](https://qrsntu.org/yh/) Authors: [Yingzhi Tang](https://qrsntu.org/yz/) Slicers: [Annie Huaiyu Wang](https://qrsntu.org/people/annie-huaiyu-wang/) arXiv 2609.32238 · two ICAIF 2026 workshop papers, accepted Illustrative numbers; the benchmark’s main books have 96 trades and 60 stored dates. ## FRTBTrace at a glance The idea Write the whole capital calculation as one graph over (observation date, trade) positions. A forward pass records which branch is active at every floor, maximum and minimum; one reverse pass then carries the allocation back to every position. The resulting ledger adds up to capital: $$∑_{s,i} A_{s,i}^K = K$$ Sensitivities to today’s book miss exactly the part on past dates. Removal effects and hedge choices are answered by exact revaluation, and ties are flagged rather than hidden. Experiments #### Synthetic benchmark · 60 books tested | Capital on past dates, median | 61% | |---|---| | Range, 14 smooth main books | 21–99% | | Ledger adds up to K, max error | 5.6 × 10⁻¹⁶ | | Questions answered, of five | all · others ≤ 2 | Pre-registered: 42 small tie books and 18 main books of 96 trades, compared with six published allocators and a numerical Euler control. After the benchmark: [Shapley values over trades](https://qrsntu.org/projects/frtbtrace/details/#shapley) under two declared conventions, and every question answered on a synthetic book of 960 trades. #### Bank portfolios Coming soon Every book so far is synthetic. Real portfolios with real desk groupings, an independent implementation of capital and timing measurements come next. **Capital on past observation dates, per main book.** On the 14 smooth main books, 21–99% of capital (median 61%) sits on books held on earlier dates, so no sensitivity to today’s positions can account for it. Books at a tie have no reference share. [More results →](https://qrsntu.org/projects/frtbtrace/details/#results) [Limitations and what didn’t work →](https://qrsntu.org/projects/frtbtrace/details/#limitations) ## Where this is going proposed 1. ### An independent check of capital A second implementation of the capital calculation, and the complete set of marginals at ties. 2. ### Real books and desk views Bank-style portfolios, real management groupings, and backtests of expected-shortfall contributions over time. 3. ### Agents in the loop Typed answers behind a tool interface, tested end to end with timing and scaling measured. [What comes next, in full →](https://qrsntu.org/projects/frtbtrace/details/#next-steps) ## Papers - ### Exact Trade-Level Attribution of FRTB-IMA Capital arXiv 2609.32238 · 26 September 2026 [[arXiv]](https://arxiv.org/abs/2609.32238) - ### Auditable FRTB Capital Explanations: A Tool Contract for Financial AI Agents Yuhe Sui Workshop on Responsible AI Operations for Finance (RAIOps4Fin), ICAIF 2026 · accepted [[News]](https://qrsntu.org/news/icaif-2026-workshops/) - ### What Post-hoc Attributions of FRTB-IMA Capital Answer Yuhe Sui, Yingzhi Tang Workshop on Interpretability and Alignment in Financial Models (IAFM), ICAIF 2026 · accepted [[News]](https://qrsntu.org/news/icaif-2026-workshops/) Scan to open this page --- # FRTBTrace: How It Works > How FRTBTrace works: why one attribution vector is not enough, ties and history, the capital graph, a pre-registered benchmark of 60 synthetic books, the comparison with published allocators, and the limitations. Web page: https://qrsntu.org/projects/frtbtrace/details/ (content last changed 2026-11-12) Site index for agents: https://qrsntu.org/llms.txt [← FRTBTrace overview](https://qrsntu.org/projects/frtbtrace/) FRTBTrace · How it works Marginal impact, removal effect, accounting share, historical origin and hedge choice are different questions about the same capital number. FRTBTrace answers each one exactly against one implemented FRTB-IMA calculation, on (observation date, trade) positions. [Paper (arXiv)](https://arxiv.org/abs/2609.32238) · [Plain-text summary](https://qrsntu.org/projects/frtbtrace/summary.md) · [Full results](https://qrsntu.org/projects/frtbtrace/results/) Lead: [Yuhe Sui](https://qrsntu.org/yh/) Authors: [Yingzhi Tang](https://qrsntu.org/yz/) Slicers: [Annie Huaiyu Wang](https://qrsntu.org/people/annie-huaiyu-wang/) Contents 1. The idea 2. Why not one attribution vector? 3. Five questions 4. Ties and history 5. The method in one picture 6. Where it runs 7. The evidence 8. A pre-registered benchmark 9. Results: capital on past dates 10. Results: ties 11. What other methods answer 12. One book, audited 13. Shapley values and a larger book 14. Certified selective repricing 15. Limitations and what didn’t work 16. What comes next 17. Technical details 18. The capital formula 19. Node rules and identities 20. Full tables 21. Reproducibility notes 22. Summary 23. In brief 24. Sources Part 1 · The idea ## 1. Why not one attribution vector? Under the internal-models approach of the Fundamental Review of the Trading Book (FRTB-IMA), market-risk capital is set for a bank’s approved trading desks as a whole. The rules do not say how that number should be split across trades, yet risk managers, capital planners and validators ask for exactly that. The obvious answer is to compute one number per trade, usually a gradient (Euler) allocation or one of the published FRTB allocation rules, and to use it wherever a trade-level number is needed. That vector is then asked several different questions. A desk report needs numbers that add up to capital. A what-if screen needs the change in capital if a trade is removed. A pre-trade check needs the local sensitivity. An audit needs to know where the capital came from. These answers coincide only in special cases, and FRTB is rarely one: its expected-shortfall, stress-ratio, history and standardised-approach layers are nonsmooth, and part of today’s charge is computed from books held weeks earlier. **Figure 1. One vector for every use, or the question typed first.** FRTBTrace names the question before it computes anything, and each answer says whether it may be added up. So FRTBTrace starts from the question. It distinguishes five of them, evaluates all five against one deterministic calculation of the full capital formula, and returns with each answer what it may be used for. The ingredients are known. Euler and Aumann–Shapley allocation for homogeneous risk measures, the fact that removal effects do not add up, and allocation at points where the risk measure has a kink are standard results in capital allocation. What FRTBTrace adds is exact, node-by-node propagation through the *whole* internal-models calculation, including the history maxima and averages, the stress-ratio floor, the amber surcharge and the comparison with standardised capital; a ledger over (observation date, trade) rather than over today’s trades; the exact frozen remainder that separates the two; and answers that state their question, with an explicit outcome when there is no answer. ## 2. Five questions ### Marginal Which trades move K locally today? A gradient on today’s book. At a tie it is a set; the engine returns one flagged element of it. Adds up to K: not in general ### Removal What happens to K if trade i goes? A finite counterfactual, answered by revaluing the book without the trade. Adds up to K: not in general ### Accounting How should K be split over current trades or desks? A declared remap of the origin ledger onto today’s trades. Adds up to K: yes ### Origin Which (date, trade) positions generated K? The ledger from the reverse pass, one entry per stored date and trade. Adds up to K: yes ### Hedge choice Which candidate action gives the lowest K? A ranking of candidate actions by exact revaluation. A ranking, not a sum The questions are not five estimates of one hidden quantity. Using a marginal vector as a desk report gives numbers that need not add up; using an accounting share as a removal estimate gives a number that is not the what-if effect; and looking only at today’s book misses the capital that stored history generates. Coalition values (Shapley values over trades) are a sixth candidate, but on these books many of the sub-book capitals they need do not exist, so they have no answer. ## 3. Ties and history Two features of the formula pull the questions apart. The first is ties. Take capital $K = max(w_1 ,w_2 )$, which is itself a two-scenario expected shortfall, at $w = (1,1)$. Removing either trade leaves $K = 1$, so both removal effects are zero and their sum is zero, not one. The local marginal is the whole segment between $(1,0)$ and $(0,1)$. An accounting ledger can still add up to K, but only by declaring how the tie is split. Two trades, one tie Illustrative Move trade 2 off the tie and the marginal becomes one gradient; at the tie it is a set, the removal effects stop adding up, and the ledger needs a declared tie weight λ. FRTBTrace flags such points instead of returning one vector as if it were unique. The second is history. FRTB compares recent charges with moving averages of past charges (60 daily and 12 weekly observations). When an averaged branch is active, part of today’s capital is generated by books held on earlier dates. Differentiating only today’s positions cannot reach that part; at smooth points its size is exact, and we call it the **frozen remainder**: $$w^⊤ ∇_w K = K − ∑_{s < T − 1} x_s^⊤ ∇_{x_s } K_⏟ _{\text{frozen remainder}}$$ Here $x_s$ is the book held on stored date $s$ and $w = x_{T − 1}$ is today’s book. A sensitivity to today’s book can be numerically correct and still leave most of the capital unexplained. ## 4. The method in one picture FRTBTrace writes the full capital calculation as a graph. Expected shortfall is evaluated on blocks for each risk class, factor set and liquidity horizon, aggregated over horizons, scaled by a floored stress ratio and combined into the internal-models charge; non-modellable and default-risk charges are added; history enters through maxima against moving averages; and the result is compared with standardised-approach capital. **Figure 2. The FRTB-IMA capital graph.** Shaded nodes can be nondifferentiable. The nodes inside the dashed box are evaluated on each stored date’s book and feed the history maxima. - **One forward pass**Capital K is evaluated once, and every nonsmooth node records its active branch: whether the stress-ratio floor binds, which history branch is selected, which side of the final minimum is active. - **One reverse pass**Allocation vectors flow back through every node by local Euler rules: dual weights at each expected shortfall, the active branch at floors and maxima, the active default-quantile scenario. - **Typed answers**The origin ledger lands on (date, trade) positions and adds up to K; removal and hedge choice are exact revaluations; at a tie, one valid marginal point is returned and flagged. [Shortened here. The full page is https://qrsntu.org/projects/frtbtrace/details/index.md; it goes on with: Where it runs; A pre-registered benchmark; Results: capital on past dates; Results: ties; What other methods answer; One book, audited; Results: Shapley values and a larger book; Certified selective repricing specialist; Limitations and what didn’t work; What comes next; The capital formula; Node rules and identities; Full tables; Reproducibility notes; In brief; Sources.] --- # FRTBTrace: Full Results > FRTBTrace full results: per-book outcomes for all 60 benchmark books, per-trade and per-date answers for a 960-trade book and the audited book, the multiplier sweep and certified repricing, as CSV files with SHA-256 hashes. Web page: https://qrsntu.org/projects/frtbtrace/results/ (content last changed 2026-11-12) Site index for agents: https://qrsntu.org/llms.txt [← FRTBTrace: how it works](https://qrsntu.org/projects/frtbtrace/details/) FRTBTrace · Full results The benchmark outcome and the post-review measurements for each of the 60 synthetic books, the 960-trade book and the audited 96-trade book trade by trade and date by date, the multiplier and history sweep, the finite-difference check and certified repricing. Every table downloads as CSV; the 173 files are listed with their SHA-256 at the end. [Overview](https://qrsntu.org/projects/frtbtrace/) · [How it works](https://qrsntu.org/projects/frtbtrace/details/) · [Limitations](https://qrsntu.org/projects/frtbtrace/details/#limitations) · Files and SHA-256 Contents 1. Every benchmark book 2. The 960-trade book 3. The audited book 4. Per-trade files, every book 5. Multiplier and history sweep 6. Finite-difference check 7. Tie sets on the main books 8. Certified selective repricing 9. Files and SHA-256 [Shortened here. The full page is https://qrsntu.org/projects/frtbtrace/results/index.md; it goes on with: Every benchmark book; The 960-trade book after the benchmark; The audited book; Per-trade files, every book after the benchmark; Multiplier and history sweep after the benchmark; Finite-difference check of the origin ledger after the benchmark; Tie sets on the main books after the benchmark; Certified selective repricing; Files and SHA-256.] --- # GeoAFT > GeoAFT, Geometric Attention Fine-Tuning: adapting how a pretrained Transformer compares representations, and when that differs from ordinary weight adaptation. Web page: https://qrsntu.org/projects/geoaft/ (content last changed 2026-11-11) Site index for agents: https://qrsntu.org/llms.txt [← All projects](https://qrsntu.org/projects/) QRS research programme · Fine-tuning · Attention **Geo**metric **A**ttention **F**ine-**T**uning Adapting how representations relate to one another, not only the representations themselves. A pretrained Transformer learns **representations** and a **rule for comparing them**. Fine-tuning usually changes the first. GeoAFT adapts the second. On controlled tasks it recovers a known comparison rule almost exactly. On real tasks its margins over controls without the geometry are small, and a mechanism study found no geometry-specific improvement in the settings it tested. [Details](https://qrsntu.org/projects/geoaft/details/) [Paper (OpenReview)](https://openreview.net/forum?id=9PwlVQ3gyj) · [Explainer video](https://qrsntu.org/projects/geoaft/details/#video) · [Plain-text summary](https://qrsntu.org/projects/geoaft/summary.md) Active Lead: [Yuhe Sui](https://qrsntu.org/yh/) Workshop paper: AutoMLR Workshop, NeurIPS 2026 Illustrative 2-D toy. ## GeoAFT at a glance The idea LoRA-style methods change how queries and keys are *produced*. Geometric attention fine-tuning changes how they are *compared*: $$S_{ij} = s_{base} (q_i ,k_j ) + α_h g^¯ _ϕ (P_q q_i ,P_k k_j )$$ $g_ϕ$ is a small trainable geometry. The gate $α_h$ starts at zero, so training begins from exactly the pretrained attention. Experiments #### Synthetic tested Learned $L_p$ · 99.983% Mahalanobis · 99.991% Nonlinear · 99.934% Recovery of three known target relations, not classification accuracy. #### Real tasks tested **Remote-sensing few-shot:** a scoped result of 84.48 / 88.20 / 91.47 at 1/2/4 shots, averaged over 8 of the 10 RSFS datasets, against best published rows of 82.33 / 86.58 / 89.53; a projected-dot control also clears them. **Mechanism study:** no geometry-specific improvement was demonstrated on development data. On fixed class keys, with the strongest fitted host, a class bias alone came within 0.08 pp of the finalist overlay. Open text keys (retrieval) Coming soon What these results do and do not show: [Limitations and what didn’t work →](https://qrsntu.org/projects/geoaft/details/#limitations) ## Where this is going 1. ### Do geometries matter where the keys are open? Image-text retrieval, where every document is a new key and ranking is the score, is the cleanest remaining test against a projected dot product. 2. ### Can a graph metric reach the oracle’s headroom? On remote-sensing few-shot, oracle neighbour graphs leave several points of headroom that no label-free metric has reached yet. 3. ### Would other placements or recipes change the reading? Only the last attention layer and one inherited training recipe were tested; mid-layer placements and recipes that train the geometry from the first step were not. [Next studies in full →](https://qrsntu.org/projects/geoaft/details/#benchmarks) ## Accepted Research - ### Geometric Attention Fine-Tuning Is Representation-Relative: Autonomous ML Research with AI-Native Workflow [Yuhe Sui](https://qrsntu.org/yh/), [Jianing Zhang](https://qrsntu.org/jn/), [Shufang Chen](https://qrsntu.org/people/shufang-chen/), [Xiaotong Li](https://qrsntu.org/people/xiaotong-li/), [Wang Zimeng](https://qrsntu.org/people/wang-zimeng/) AutoMLR Workshop, NeurIPS 2026 · 2026 · Accepted [[OpenReview]](https://openreview.net/forum?id=9PwlVQ3gyj) ## Related projects [AI Native Workflow](https://qrsntu.org/projects/ai-native-workflow/), [GeoAttention](https://qrsntu.org/projects/geoattention/) ## Publications - **Geometric Attention Fine-Tuning Is Representation-Relative: Autonomous ML Research with AI-Native Workflow** (2026). [Yuhe Sui](https://qrsntu.org/yh/), [Jianing Zhang](https://qrsntu.org/jn/), [Shufang Chen](https://qrsntu.org/people/shufang-chen/), [Xiaotong Li](https://qrsntu.org/people/xiaotong-li/), [Wang Zimeng](https://qrsntu.org/people/wang-zimeng/). AutoMLR Workshop, NeurIPS 2026. Accepted. [OpenReview](https://openreview.net/forum?id=9PwlVQ3gyj) --- # GeoAFT: How It Works > How GeoAFT works: the intuition for adapting attention geometry, the method in one picture, synthetic and real-task results, a mechanism study with its controls, and the technical details. Web page: https://qrsntu.org/projects/geoaft/details/ (content last changed 2026-11-11) Site index for agents: https://qrsntu.org/llms.txt [← GeoAFT overview](https://qrsntu.org/projects/geoaft/) GeoAFT · Geometric Attention Fine-Tuning · How it works Start with the intuition and the pictures. Then the evidence: controlled synthetic tasks, real few-shot benchmarks, and a mechanism study of where the real-task gains come from. The precise definitions and full tables are at the end. [Paper (OpenReview)](https://openreview.net/forum?id=9PwlVQ3gyj) · [Plain-text summary](https://qrsntu.org/projects/geoaft/summary.md) Lead: [Yuhe Sui](https://qrsntu.org/yh/) Contents 1. The idea 2. Why adapt relations? 3. Same points, different relations 4. The method in one picture 5. The evidence 6. Synthetic experiments 7. Real tasks 8. Mechanism study 9. What the results mean 10. Limitations and what didn’t work 11. Next studies and future directions 12. Technical details 13. Notation and definition 14. Families, centring and masks 15. Q/K versus score adaptation 16. Protocol, controls and full results 17. Summary 18. In brief 19. Sources Part 1 · The idea ## 1. Why adapt relations? Inside every attention head, tokens become queries and keys, and a compatibility rule, usually a scaled dot product, decides how strongly each query attends to each key. So a pretrained model has learned two things: **representations**, and a **rule for comparing them**. Parameter-efficient fine-tuning such as LoRA adapts the maps that *produce* queries and keys. The rule that *compares* them is inherited unchanged. GeoAFT asks what happens if we adapt that rule instead. **Explainer (draft, October 2026).** About two and a half minutes, silent: the idea and the synthetic results, recorded before the real-task results below. The working hypothesis was that relations matter most when representations are constrained: frozen or compressed features, small adaptation budgets, or similarity-critical tasks. Where representations are flexible, ordinary fine-tuning may absorb the same change. The real-task evidence in Part 2 tests this. **Figure 1. Two places to adapt attention.** Conventional PEFT edits the projections that produce $q$ and $k$. GeoAFT adds a trainable term to the compatibility function and leaves $q$ and $k$ as they are. ## 2. Same points, different relations Changing the geometry changes what counts as similar. Hold one query and seven keys fixed. The **dot product** prefers the key furthest along the query’s direction. **Euclidean** distance prefers the nearest key. **Taxicab (L1)** distance prefers neighbours along the axes. An **anisotropic Mahalanobis** metric prefers neighbours along a direction it treats as cheap. The winner changes every time, and no point has moved. **Figure 2. Try it.** Each panel scores the same query $q$ against the same keys A–G with a different rule. Illustrative toy example, not trained-model behaviour. ## 3. The method in one picture GeoAFT keeps the pretrained score and adds a small, trainable geometry term on top of it: $$S_{ij} = s_{base} (q_i ,k_j ) + α_h g^¯ _ϕ (P_q q_i ,P_k k_j ),A_{ij} = softmax_{j ∈ 𝒱(i)} (S_{ij} )$$ **Figure 3. The GeoAFT score.** A structured geometry branch is centred over valid keys, scaled by a gate that starts at zero, and added before a single softmax. - **Starts from the pretrained model.** Each gate $α_h$ is initialized at zero, so training begins from exactly the pretrained attention. - **Small and structured.** What is trained is the geometry itself, a metric, an exponent or a projection, not a generic learned scorer. - **Two places to use it.** As an *attention overlay*, on the query-key scores inside a Transformer layer; or as a *score-level overlay*, on the class scores of a frozen-feature few-shot classifier, where the keys are class prototypes. Part 2 · The evidence ### Tested so far - Recovery of three known target relations on synthetic tasks, against Q/K-LoRA, a fixed geometry bank, generic pair-MLP scorers and exact controls - Remote-sensing few-shot (RSFS-8) and CLIP few-shot on the 11 CoOp datasets, against published results - GLUE MRPC and RTE, domain generalisation and biomedical few-shot - A mechanism study on development data: fixed class keys, the last attention layer of CLIP ViT-B/16 with and without Q/K-LoRA, and neighbour graphs for transductive few-shot ### Next Coming soon - Open text keys: remote-sensing image-text retrieval (RSICD, RSITMD) - LC-TIM+DINO with the published DINOv3 satellite encoder, once access to the weights is granted - Host redundancy and optimisation, on development data ## 4. Synthetic experiments controlled We first test GeoAFT where the right answer is known. A synthetic teacher generates attention from a known compatibility function over fixed representations, and every method adapts the same frozen dot-product base toward it. The score is **oracle-gap recovery $R$**: 0% means no improvement over the base, 100% means matching the oracle. It is not classification accuracy. ### Learned $L_p$ Target: an $L_p$ distance 99.983% **Result.** Near-oracle recovery. A fixed bank of ten $L_p$ geometries reached nearly the same. ### Mahalanobis Target: a Mahalanobis metric 99.991% **Result.** Near-oracle recovery. An exact bilinear-plus-key-quadratic score, its known equivalent (§12), matched or slightly exceeded it. ### Structured nonlinear Target: a nonlinear compatibility 99.934% **Result.** 46 trainable parameters fit about as well as an 8,705-parameter pair MLP (99.539%), winning 15/15 paired cells. On the $L_p$ and Mahalanobis targets, Q/K-LoRA stays 39.9–41.7 points of $R$ behind GeoAFT at the tested budgets, and the tested pair-MLP scorers 6.4–12.2 points; each range spans both targets. **Figure 4. Every tested method on the three synthetic targets.** Each point is one method: its trainable parameters against its oracle-gap recovery $R$. The vertical axis is logarithmic in the remaining gap, so 99%, 99.9% and 99.99% are evenly spaced. Diamonds are GeoAFT variants; the solid ring marks the selected one and the dotted ring the best other method shown. On the Mahalanobis target the exact bilinear-plus-key-quadratic control is not plotted; it matched or slightly exceeded GeoAFT in all 25 paired cells. Counts are trainable parameters, not inference cost. [Shortened here. The full page is https://qrsntu.org/projects/geoaft/details/index.md; it goes on with: Real tasks; Mechanism study development data; What the results mean; Limitations and what didn’t work; Next studies and future directions; Notation and definition; Families, centring and masks; Q/K versus score adaptation; Protocol, controls and full results; In brief; Sources; Accepted Research; Related projects.] --- # GeoAttention > Research on attention geometry, representation-relative identifiability, and geometry-adaptive prediction. Web page: https://qrsntu.org/projects/geoattention/ (content last changed 2026-10-15) Site index for agents: https://qrsntu.org/llms.txt [← All projects](https://qrsntu.org/projects/) Active Research on attention geometry, representation-relative identifiability, and geometry-adaptive prediction. Lead: [Yuhe Sui](https://qrsntu.org/yh/) Authors: [Jianing Zhang](https://qrsntu.org/jn/), [Xiaotong Li](https://qrsntu.org/people/xiaotong-li/) ## Related projects [GeoAFT](https://qrsntu.org/projects/geoaft/), [SoftRank](https://qrsntu.org/projects/softrank/) --- # HedgeFidelity > An early-stage research programme with derived research in progress. Web page: https://qrsntu.org/projects/hedgefidelity/ (content last changed 2026-10-15) Site index for agents: https://qrsntu.org/llms.txt [← All projects](https://qrsntu.org/projects/) Early-stage / Active derived research An early-stage research programme with derived research in progress. Lead: [Yuhe Sui](https://qrsntu.org/yh/) --- # MathResearchOS > MathResearchOS: a command-line system that turns a mathematical objective into a recoverable, verifier-gated run across Codex and Claude models, keeping every attempt on disk. Install, quick start and a one-run figure. Web page: https://qrsntu.org/projects/mathresearchos/ (content last changed 2026-11-12) Site index for agents: https://qrsntu.org/llms.txt [← All projects](https://qrsntu.org/projects/) QRS research · Mathematical research tooling A recoverable, audited operating system for mathematical work Give it a mathematical objective. Get back a result with its evidence, or the exact stage where it stopped. MathResearchOS is a command-line runtime that turns one objective into a staged run: normalise the problem, search the literature, explore proof ideas, try to refute them, integrate an argument and write it up. Models from the Codex CLI, the Claude Code CLI or an HTTP API do the work; **every stage output must pass a verifier verdict** before the next stage uses it. Each case keeps the exact objective, its frozen contract, every attempt and an event log, so a run can be paused, resumed or audited later. [Usage guide](https://qrsntu.org/projects/mathresearchos/details/) [Plain-text summary](https://qrsntu.org/projects/mathresearchos/summary.md) · [Install](https://qrsntu.org/projects/mathresearchos/details/#install) · [Troubleshooting](https://qrsntu.org/projects/mathresearchos/details/#troubleshooting) Active Lead: [Yuhe Sui](https://qrsntu.org/yh/) Version 0.1.0 · Python 3.11+ · no third-party packages Illustrative run; stage names are the workflow’s own, verdicts vary. ## MathResearchOS at a glance Quick start Install, create a case from a problem file, preview the run for free, then run it. The full guide covers each step. ``` pip install -e . mathos create --mode solve --title "First problem" --objective-file problem.md mathos run MATH-0001 --dry-run mathos run MATH-0001 --route cross mathos status MATH-0001 ``` Live runs call real models and can cost money; `--dry-run` never does. A live run needs a logged-in `codex` or `claude` CLI. [Install details →](https://qrsntu.org/projects/mathresearchos/details/#install) Evidence so far #### Pilot run tested One public-benchmark problem (IMO-ProofBench, problem PB-Basic-008) through the v2 workflow with 8 explorers: 17 calls, about 11 minutes. All explorers delivered distinct ideas and the final proof was correct, checked by hand. One problem, one run: a check that the pipeline works end to end, not a measure of accuracy. #### Held-out comparisons In progress Model and role comparisons on held-out problems, with every arm and failure reported, are planned but not run yet. ## Jump to what you need - [Solve one problem](https://qrsntu.org/projects/mathresearchos/details/#solve) · create, dry run, run - [Choose models](https://qrsntu.org/projects/mathresearchos/details/#models) · presets, mixes, pools - [Follow and resume a run](https://qrsntu.org/projects/mathresearchos/details/#follow) · status, events, usage - [Human gates](https://qrsntu.org/projects/mathresearchos/details/#gates) · continue, hand-off - [Campaigns](https://qrsntu.org/projects/mathresearchos/details/#campaigns) · many problems, many routes - [Lean check](https://qrsntu.org/projects/mathresearchos/details/#lean) · kernel errors, sorry, axioms - [Where data is stored](https://qrsntu.org/projects/mathresearchos/details/#storage) · cases, campaigns, experiments - [Troubleshooting](https://qrsntu.org/projects/mathresearchos/details/#troubleshooting) · locks, gates, missing CLIs [What an accept means, and the limitations →](https://qrsntu.org/projects/mathresearchos/details/#limitations) ## Where this is going proposed 1. ### Held-out model comparisons A benchmark harness with contamination-aware held-out problems; effort sweeps, a verifier canary with known-wrong proofs and explorer-yield counts, each planned before the run and reported in full. 2. ### Obligation ledger and independent checkers A proof-obligation ledger the final verifier must cover, plus exact-arithmetic and Lean adapters, so more claims rest on a checkable certificate and less on a model’s audit. 3. ### Longer research runs Retrieval-backed memory with a novelty ledger, so objectives that take many sessions do not repeat approaches already tried. ## Related projects [AI Native Workflow](https://qrsntu.org/projects/ai-native-workflow/) --- # MathResearchOS: Usage Guide > MathResearchOS usage guide: install, create and run a case, family presets and model mixes, status and resume, human gates, manual hand-off, campaigns, Lean checks, where data is stored, safety limits and troubleshooting. Web page: https://qrsntu.org/projects/mathresearchos/details/ (content last changed 2026-11-12) Site index for agents: https://qrsntu.org/llms.txt [← MathResearchOS overview](https://qrsntu.org/projects/mathresearchos/) MathResearchOS · Usage guide Everything you need to run MathResearchOS from the command line: install, create a case, preview and run it, pick models, follow and resume a run, handle human gates, run campaigns and Lean checks, and know where the files go. Every command on this page is checked against the program’s own `--help`. [Plain-text summary](https://qrsntu.org/projects/mathresearchos/summary.md) · Install · Troubleshooting Lead: [Yuhe Sui](https://qrsntu.org/yh/) Contents 1. Get started 2. Install 3. Solve one problem 4. What happens in a run 5. Follow, pause and resume 6. Choose models 7. Presets and mixes 8. Pools and other backends 9. Reading the selection guide 10. Beyond one run 11. Human gates 12. Manual hand-off 13. Campaigns 14. Lean check 15. Where data is stored 16. Safety and trust 17. Safety limits 18. Limitations and what an accept means 19. Troubleshooting 20. Summary 21. In brief 22. Sources Part 1 · Get started ## 1. Install You need Python 3.11 or newer. The package has no third-party dependencies. From a checkout of the repository: ``` pip install -e . mathos --help ``` This registers the `mathos` command. `python -m mathresearchos` does the same thing without installing, and the examples on this page work with either. Replace `mathos` with `python -m mathresearchos` if you did not install. For a **live** run you also need at least one model CLI, installed and logged in with your own account: - `codex`, the Codex CLI, for the `openai` family. - `claude`, the Claude Code CLI, for the `claude` family. Use both for a cross-family setup. - Optionally an OpenAI-compatible HTTP endpoint (for example DeepSeek) with its API key in an environment variable. This is configured in a pool file; see Pools. Creating a case, `validate`, `status`, `events` and every `--dry-run` need none of these. To check the install without any model, run the offline tests; they make no model calls: ``` python -m unittest discover -s tests -v python benchmarks/run_smoke.py ``` ## 2. Solve one problem A **case** is one objective plus everything the run produces. Put the problem statement, exactly as you want it solved, in a text file. The system stores it unchanged in `OBJECTIVE.md` and never rewrites it. ### Create the case The classic workflow (`v1`) has eight stages and a budget of 16 model calls: ``` mathos create --mode solve --title "First problem" --objective-file problem.md ``` The tiered workflow (`v2`) adds an explorer swarm, a separate refuter and auditors. `--explorers` sets the swarm size, from 2 to 32 (default 8), and the call budget becomes 16 plus the number of explorers. Use `--text` instead of `--objective-file` for a one-line statement, and `--mode research` for an open-ended objective. ``` mathos create --mode solve --workflow v2 --explorers 8 --title "First problem" --objective-file problem.md ``` The command prints the new case id (`MATH-0001`, then `MATH-0002`, and so on). Look at the generated `cases/MATH-0001/DIC/` and `EPS/INITIAL.md`: that is the frozen contract the run will follow, and the moment to fix any ambiguity in the statement. Then check the case is consistent: ``` mathos validate MATH-0001 ``` ### Preview with a dry run A dry run resolves the stages and the model for every role, prints the call budget, and makes **no model calls and no writes**. Always do it before a live run. This example uses the built-in `cross` preset on a v2 case with four explorers: ``` mathos run MATH-0002 --dry-run --route cross Dry run: MATH-0002; ready: specify Route: cross {"reasoning": "anthropic", "worker": "anthropic", "verifier": "openai"} specify (thinker) -> claude-default [claude-sonnet-5-5] literature (worker) -> claude-default [claude-sonnet-5-5] explore_swarm (explorer) -> claude-default [claude-sonnet-5-5] falsify (thinker) -> claude-default [claude-sonnet-5-5] triage (orchestrator) -> claude-default [claude-sonnet-5-5] integrate (thinker) -> claude-default [claude-sonnet-5-5] refute (refuter) -> openai-default [gpt-6.1-sol] verify (auditor) -> openai-default [gpt-6.1-sol] report (writer) -> claude-default [claude-sonnet-5-5] verify_final (auditor) -> openai-default [gpt-6.1-sol] Swarm: 4 explorers, 4 in parallel, needs 2 deliveries; case call budget 20 No model calls, state changes, or external writes performed ``` The output above is abridged. Read any `WARNING` lines: they name a same-family pairing or a missing API key, and are cheaper to fix now than after a run. ### Run it A live run calls real models and may incur charges on your accounts. With no model flags, a run uses the Codex CLI alone with its own configured defaults. Choose models with `--route`, `--mix` or `--pool` (see Presets and mixes): ``` mathos run MATH-0001 mathos run MATH-0001 --route cross mathos run MATH-0001 --mix "claude/openai" mathos run MATH-0001 --pool pool.json --route tiered-swarm ``` Useful limits on any `run` or `resume`: `--max-nodes N` stops after N stages, and `--timeout-seconds S` caps each model call (default 1200, that is 20 minutes). The command exits with 0 when the case is completed, ready or paused and with 2 otherwise, for example at a human gate, so a script can tell the two apart. ## 3. What happens in a run A run walks a fixed graph of stages in dependency order. The v1 stages are `specify`, `literature`, `explore`, `falsify`, `integrate`, `verify`, `report` and `verify_final`. The v2 stages replace `explore` with `explore_swarm` and add `triage` and `refute`: the swarm runs N explorers, each forced onto a different method family, and `triage` merges their ideas with exact duplicates removed. The node fails, and is retried, if fewer than half of the explorers deliver. The audited stages (`verify` and `verify_final`) end in one of four verdicts: - **accept**: the audit accepted the supplied evidence under the contract; the run moves on. - **repair**: the draft goes back for a bounded repair, at most 2 repair cycles per case. - **replan** or **escalate**: the run stops at a human gate instead of pretending success. `RESULT.md` is written only after the final verifier accepts. Failed and superseded attempts stay on disk. The design and its rationale are in the repository’s `docs/OS_DESIGN.md`; this guide stays with usage. [Shortened here. The full page is https://qrsntu.org/projects/mathresearchos/details/index.md; it goes on with: Follow, pause and resume; Presets and mixes; Pools and other backends; Reading the model selection guide; Human gates; Manual hand-off; Campaigns; Lean check; Where data is stored; Safety limits; Limitations and what an accept means; Troubleshooting; In brief; Sources.] --- # ntuinfo.com > A bilingual English/Chinese full-stack web application with a quantitative, multi-criteria scoring framework (mental load, physical load, difficulty) that ranks 100+ NTU student clubs and surfaces algorithmically-driven recommendations for students. Web page: https://qrsntu.org/projects/ntuinfo-com/ (content last changed 2026-10-15) Site index for agents: https://qrsntu.org/llms.txt [← All projects](https://qrsntu.org/projects/) Active A bilingual English/Chinese full-stack web application with a quantitative, multi-criteria scoring framework (mental load, physical load, difficulty) that ranks 100+ NTU student clubs and surfaces algorithmically-driven recommendations for students. Participants: [Yuhe Sui](https://qrsntu.org/yh/) [[Website]](https://ntuinfo.com) --- # PolicyAttention > PolicyAttention: one softmax attention row computes the policy mirror descent step exactly. A causal actor-environment-critic loop with a bound on the returned policy, learned Transformers checked against it, and where a stronger step backfires. Web page: https://qrsntu.org/projects/policyattention/ (content last changed 2026-11-11) Site index for agents: https://qrsntu.org/llms.txt [← All projects](https://qrsntu.org/projects/) QRS research · Reinforcement learning · Transformers Softmax attention as the **policy-improvement step** of an in-context controller Softmax attention already has the form of a policy-improvement update. Build the controller around it, then find where it breaks. Earlier Transformer constructions for in-context control improve a policy with linear attention, while constructions with standard softmax attention stop at evaluating one. PolicyAttention starts from a classical identity: the policy mirror descent update $π^{ + } ∝ πe^{ηQ}$ **is a softmax**, so one attention row can be the actor. We build a causal actor–environment–critic loop around it, bound the loss of the policy it returns, and map where finite attention makes a stronger update backfire. [Details](https://qrsntu.org/projects/policyattention/details/) [arXiv](https://arxiv.org/abs/2609.30500) · [PTA paper (OpenReview)](https://openreview.net/forum?id=zbGWnk0kY3) · [EvoRobust paper (OpenReview)](https://openreview.net/forum?id=VgswuGrqp6) · [ICAIF 2026: market making](https://qrsntu.org/projects/policyattention/icaif/) · [Plain-text summary](https://qrsntu.org/projects/policyattention/summary.md) Active Lead: [Yuhe Sui](https://qrsntu.org/yh/) Authors: [Shufang Chen](https://qrsntu.org/people/shufang-chen/), [Yingzhi Tang](https://qrsntu.org/yz/), [Xinyue Wu](https://qrsntu.org/people/xinyue-wu/) Slicers: [Shufang Chen](https://qrsntu.org/people/shufang-chen/), [Xinyue Wu](https://qrsntu.org/people/xinyue-wu/) NeurIPS 2026 workshops: PTA and EvoRobust · arXiv 2609.30500 Illustrative numbers. ## PolicyAttention at a glance The idea Negative-entropy policy mirror descent (PMD) moves a policy $π$ towards actions with high value $Q$, with step size $η$, state by state: $$π_{k + 1} = softmax( log π_k + ηQ_k )$$ An attention row whose scores are $log π + ηQ$ computes this exactly. The environment then supplies fresh samples under the new policy, a softmax critic evaluates it, and the errors of both add up to a bound on the loss of the policy the loop returns. Experiments #### Exact decoder tested Ideal / implemented gain, η = 0.8 · 0.950 / +0.916 · Ideal / implemented gain, η = 2.4 · 2.084 / −2.009 A larger step makes ideal PMD better but the implemented update worse once $ηc > (1 − γ)κ$; a state-local mask removes this. With coverage relaxed, one on-policy stream raises the returned-policy loss from 0.009 to 0.029. #### Learned Transformer tested Row error to PMD, held out · 0.0357 · Repeated control, loss ratio to exact PMD · 1.052 The next-closest fixed rule is at 0.112. Ratio: median of 5 trained actors after 20 rounds, all within a 1.5× band. #### Without the oracle Not yet tested The guarantee assumes fresh samples for every state–action pair each round. A learned critic on ordinary trajectory data, and learned closed-loop rollouts with matched errors, are not yet tested. #### Finance ICAIF 2026 Two ICAIF 2026 workshop papers study the same update rule in synthetic allocation and market-making settings: [market-making pages](https://qrsntu.org/projects/policyattention/icaif/). **A stronger step can reverse improvement.** Exact value gain in a two-state family through the fully specified decoder (margin κ = 8, γ = 0.8). Ideal PMD keeps improving as the step grows; without a mask the implemented gain crosses zero at η = 1.6, as predicted; with a state-local mask it equals PMD. [More results →](https://qrsntu.org/projects/policyattention/details/#failure) [Limitations and what didn’t work →](https://qrsntu.org/projects/policyattention/details/#limitations) ## Where this is going open 1. ### Learning without the oracle A learned critic run on ordinary trajectory data, where some pairs go unvisited and samples are dependent. 2. ### Closed-loop audits of learned models Learned rollouts whose actor and critic errors are measured against the policy they actually execute. 3. ### Routing failures in general models Whether the reversal seen in the exact decoder occurs in learned Transformers that see other states’ memories. [The open questions in full →](https://qrsntu.org/projects/policyattention/details/#next-steps) ## Accepted Research - ### PolicyAttention: Robustness and Failure Modes of Softmax Policy Improvement [Yuhe Sui](https://qrsntu.org/yh/), [Yingzhi Tang](https://qrsntu.org/yz/), [Xinyue Wu](https://qrsntu.org/people/xinyue-wu/) EvoRobust Workshop, NeurIPS 2026 · 2026 · Accepted [[OpenReview]](https://openreview.net/forum?id=VgswuGrqp6) - ### PolicyAttention: When Softmax Attention Becomes Policy Improvement [Yuhe Sui](https://qrsntu.org/yh/) PTA Workshop, NeurIPS 2026 · 2026 · Accepted [[OpenReview]](https://openreview.net/forum?id=zbGWnk0kY3) ## Preprint and ICAIF 2026 - ### PolicyAttention: Softmax Attention Implements Policy Mirror Descent for Closed-Loop Control Yuhe Sui, Yingzhi Tang, Shufang Chen arXiv 2609.30500 · 2026 [[arXiv]](https://arxiv.org/abs/2609.30500) - ### PolicyAttention: A Certifiable and Testable Update Rule for Attention-Based Allocation Controllers IAFM'26 workshop, ICAIF 2026 · 2026 · Accepted - ### Where Does Learned In-Context Policy Mirror Descent Pay Off in Market Making? A Component Diagnosis ICAIF 2026 Workshop on AI-Driven Market Microstructure · 2026 [[Project pages]](https://qrsntu.org/projects/policyattention/icaif/) Affiliations: Quantitative Research Society, Singapore · Nanyang Technological University. Scan to open this page ## Publications - **PolicyAttention: Robustness and Failure Modes of Softmax Policy Improvement** (2026). [Yuhe Sui](https://qrsntu.org/yh/), [Yingzhi Tang](https://qrsntu.org/yz/), [Xinyue Wu](https://qrsntu.org/people/xinyue-wu/). EvoRobust Workshop, NeurIPS 2026. Accepted. [OpenReview](https://openreview.net/forum?id=VgswuGrqp6) - **PolicyAttention: When Softmax Attention Becomes Policy Improvement** (2026). [Yuhe Sui](https://qrsntu.org/yh/). PTA Workshop, NeurIPS 2026. Accepted. [OpenReview](https://openreview.net/forum?id=zbGWnk0kY3) --- # PolicyAttention: How It Works > How PolicyAttention works: why softmax rather than linear attention, the actor, environment and critic errors, the exact reversal boundary and state-local mask, coverage relaxed, learned Transformers in repeated control, and the limitations. Web page: https://qrsntu.org/projects/policyattention/details/ (content last changed 2026-11-11) Site index for agents: https://qrsntu.org/llms.txt [← PolicyAttention overview](https://qrsntu.org/projects/policyattention/) PolicyAttention · How it works One softmax attention row computes the policy mirror descent step exactly. Around it we build a causal actor–environment–critic loop, bound the loss of the policy it returns, and test both the exact construction and learned Transformers against it. [arXiv](https://arxiv.org/abs/2609.30500) · [PTA paper (OpenReview)](https://openreview.net/forum?id=zbGWnk0kY3) · [EvoRobust paper (OpenReview)](https://openreview.net/forum?id=VgswuGrqp6) · [ICAIF 2026: market making](https://qrsntu.org/projects/policyattention/icaif/) · [Plain-text summary](https://qrsntu.org/projects/policyattention/summary.md) Lead: [Yuhe Sui](https://qrsntu.org/yh/) Authors: [Shufang Chen](https://qrsntu.org/people/shufang-chen/), [Yingzhi Tang](https://qrsntu.org/yz/), [Xinyue Wu](https://qrsntu.org/people/xinyue-wu/) Slicers: [Shufang Chen](https://qrsntu.org/people/shufang-chen/), [Xinyue Wu](https://qrsntu.org/people/xinyue-wu/) Contents 1. The idea 2. Why softmax, not linear attention? 3. Three calls, three errors 4. Why a stronger step can backfire 5. The decoder in one picture 6. Where it runs 7. The evidence 8. How we test it 9. Results: the exact decoder 10. Relaxing the coverage oracle 11. Results: learned Transformers 12. Stress test and matched audit 13. Limitations and what didn’t work 14. Open questions 15. Technical details 16. Definitions and the bound 17. The two-state family 18. Full tables 19. Reproducibility notes 20. Summary 21. In brief 22. Sources Part 1 · The idea ## 1. Why softmax, not linear attention? Transformers can run learning algorithms in context. For reinforcement learning, the constructions that improve a policy, and not only evaluate it, have used **linear attention**: its unnormalized bilinear products can carry a semi-gradient or actor–critic update. Standard **softmax attention** behaves differently. Its outputs are positive, normalized mixtures, and the constructive results for it have stopped at policy evaluation (temporal-difference learning). For one well-known control update that normalization is not an obstacle but the update itself. Negative-entropy policy mirror descent (PMD) reweights each action by the exponential of its value and renormalizes: $$π_{k + 1} (a∣s) = \frac{π_k (a∣s)e^{η_k Q_k (s,a)} }{∑_b π_k (b∣s)e^{η_k Q_k (s,b)} }$$ State by state, that is exactly a softmax of the old log-probabilities plus the scaled values: $$π_{k + 1} = softmax( log π_k + η_k Q_k ).$$ **Figure 1. Earlier constructions, and softmax as the improvement step.** PolicyAttention keeps standard softmax attention and gives it the actor’s job: one row with scores $log π + ηQ$. The environment and a softmax critic close the loop, and the actor’s and critic’s errors bound the loss of the policy the loop returns. The identity is classical. What is new is treating it as the native actor of a causal Transformer and asking which control guarantees follow, and where they stop. ## 2. Three calls, three errors Each round runs in a fixed order: the actor must form a policy before the environment can sample under it, and the critic may only evaluate that policy after the samples exist. The order separates three kinds of error. ### How should the policy change? Actor · softmax routing error $ζ_k$ Memories carry each state–action pair’s $log π$ and $Q$; a query routes to its own state’s memories and the softmax takes the PMD step. With a finite margin $κ$, other states’ memories keep a small weight; $ζ_k$ measures how far the row ends up from exact PMD. ### What happens under the new policy? Environment · sampling The environment draws fresh transitions, with successor actions taken from the policy the actor just produced. The guarantee assumes a generative oracle: a fresh batch for every state–action pair, every round. ### How good is the new policy? Critic · TD error $δ_k$ The critic first looks up the next state’s value for each transition and forms a TD target, then averages the targets for each state–action pair. $δ_k$ measures its distance from an exact one-step evaluation of the produced policy. A greedy-defect recursion then carries $ζ_k$ and $δ_k$ through the rounds into a bound on $‖Q^⋆ − Q^{π^^ _T } ‖_∞$, the loss of the policy the loop finally returns (§13). The last critic call does not enter, because the returned policy is formed before it. ## 3. Why a stronger step can backfire A larger step $η$ makes ideal PMD greedier, which helps. It also spreads the scores further apart, and a finite margin has to keep the other states’ memories down against that spread. In a two-state family where every quantity is exact, the competition reduces to one ratio, $ρ = ηc/((1 − γ)κ)$ for reward scale $c$: the implemented update improves the policy when $ρ<1$ and makes it worse when $ρ > 1$. **Figure 2. Try it.** Two self-looping states with rewards $( − c,0)$ and $(c,0)$, a uniform prior and its exact $Q$, $c = 1$, $γ = 0.8$; the curve is the value gain at the worse state. Exact computation for this family, not an illustration. Within one state, the common level of the $Q$-values cancels out of PMD. Across states it does not: it shifts the two states’ softmax normalizers in opposite directions, by $ηc/(1 − γ)$ in total. So a larger step uses up the margin $κ$ faster than it sharpens the intended row. The critic is exact throughout: the wrong action is chosen from a perfect value table. ## 4. The decoder in one picture **Figure 3. Two calls per round.** One fixed decoder reads an actor packet, then, after the environment has sampled under the new policy, a critic packet. Causal masking means no critic token can see a transition before the policy that generated it exists. - **Small and fixed.** 2 blocks, 2 heads, head width S+A+3 and model width 6S+6A+14, with a gated ReLU feed-forward map of 8 active units. The same weights serve both calls; the step size $η_k$ arrives as token content. - **Normalization-free.** No layer normalization: the value channels must keep controlled amplitudes while the key channels carry score-dependent routing, and normalizing tokens would change both at once. - **Declared before freezing.** Step-size caps, packet sizes and coverage floors are fixed before the margins are chosen and the weights frozen; the bound holds for that declared horizon. [Shortened here. The full page is https://qrsntu.org/projects/policyattention/details/index.md; it goes on with: Where it runs; How we test it; Results: the exact decoder; Relaxing the coverage oracle; Results: learned Transformers; Stress test and matched audit; Limitations and what didn’t work; Open questions; Definitions and the bound; The two-state family; Full tables; Reproducibility notes; In brief; Sources; Accepted Research.] --- # In-Context PMD in Market Making: A Component Diagnosis > Where does learned in-context policy mirror descent pay off in market making? An ICAIF 2026 workshop paper that swaps the learned actor and critic for their exact counterparts in an exactly solvable dealer market. Web page: https://qrsntu.org/projects/policyattention/icaif/ (content last changed 2026-11-11) Site index for agents: https://qrsntu.org/llms.txt [← PolicyAttention](https://qrsntu.org/projects/policyattention/) PolicyAttention · ICAIF 2026 workshop paper · Market making A **component diagnosis** of the actor and the critic under fee, volatility, liquidity and toxicity stress When an in-context market maker does badly after a regime shift, its returns cannot tell you which learned part failed. In a market we can solve exactly, we swap the learned actor and the learned critic, one at a time, for their exact counterparts. After 20 rounds the learned actor adds little; **the critic stage carries the excess normalized loss**, and in a follow-up an untrained average of the critic's own six samples loses about as much. [Details](https://qrsntu.org/projects/policyattention/icaif/details/) [Workshop listing](https://icaif2026.org/workshop-papers/#paper-ai-market-microstructure-1245) · [Workshop page](https://icaif-ai-driven.github.io/) · [Plain-text summary](https://qrsntu.org/projects/policyattention/icaif/summary.md) Authors: [Xinyue Wu](https://qrsntu.org/people/xinyue-wu/), [Yuhe Sui](https://qrsntu.org/yh/), [Shufang Chen](https://qrsntu.org/people/shufang-chen/), [Yingzhi Tang](https://qrsntu.org/yz/) Poster 03 · ICAIF 2026 Workshop on AI-Driven Market Microstructure · Milan, 14 November 2026 Paired excess normalized loss over exact TD-PMD after 20 rounds on nominal markets (lower is better): the median of 320 paired values, 5 runs × 64 markets. The learned actor's 95% interval is [−0.0002, +0.0043]. ## The paper at a glance The idea Policy mirror descent (PMD) improves a policy $π$ against action values $Q$ with a step $η$: $$π_{k + 1} = softmax( log π_k + ηQ_k )$$ One frozen Transformer runs both halves in context: an **actor** that takes this step and a **critic** that estimates the next $Q$ from 6 sampled transitions per state–action pair. Because our dealer market is exactly solvable, each half has an exact counterpart, and a failure becomes a set of measured component costs. Experiments #### Synthetic market tested Learned actor · +0.0020 Learned critic · +0.0219 Both learned · +0.0235 Excess normalized loss over exact TD-PMD on nominal markets. In a follow-up, an untrained average of the critic's six samples loses 0.70–1.10 times as much as the learned critic, across all 9 families. #### Against desk rules The loop beats HBEST, the best nominal rule, under severe fee (+0.1661 in value capture) and toxicity (+0.1247) stress. #### Real markets Not tested [Limitations and what didn't work →](https://qrsntu.org/projects/policyattention/icaif/details/#limitations) ## Where this is going from the paper 1. ### A better critic target A critic that fitted its six-sample target perfectly would keep at least 70% of the cost, so the next step is a target better than the six-sample average, tried at other sample budgets. 2. ### A converged starting policy Run the adaptation tests from a converged nominal policy and for more than 20 rounds. 3. ### Why the critic and not the actor Find a theory for why the critic, not the actor, limits control. [The directions in full →](https://qrsntu.org/projects/policyattention/icaif/details/#open) ## Paper - ### Where Does Learned In-Context Policy Mirror Descent Pay Off in Market Making? A Component Diagnosis [Xinyue Wu](https://qrsntu.org/people/xinyue-wu/), [Yuhe Sui](https://qrsntu.org/yh/), [Shufang Chen](https://qrsntu.org/people/shufang-chen/), [Yingzhi Tang](https://qrsntu.org/yz/) ICAIF 2026 Workshop on AI-Driven Market Microstructure · 2026 · Accepted, Poster 03 [[Workshop listing]](https://icaif2026.org/workshop-papers/#paper-ai-market-microstructure-1245) This work is part of [PolicyAttention](https://qrsntu.org/projects/policyattention/). --- # In-Context PMD in Market Making: How It Works > How the ICAIF 2026 component diagnosis works: an exactly solvable market-making family, the in-context PMD loop and its exact swap, what the actor and the critic cost, the payoff against desk rules, and the limitations. Web page: https://qrsntu.org/projects/policyattention/icaif/details/ (content last changed 2026-11-11) Site index for agents: https://qrsntu.org/llms.txt [← ICAIF 2026 overview](https://qrsntu.org/projects/policyattention/icaif/) PolicyAttention · ICAIF 2026 · How it works A learned in-context market maker is taken apart in a market that can be solved exactly. The learned actor and the learned critic are replaced, one at a time, by their exact counterparts, and the cost of each is measured under fee, volatility, liquidity and toxicity stress. [Workshop listing](https://icaif2026.org/workshop-papers/#paper-ai-market-microstructure-1245) · [Workshop page](https://icaif-ai-driven.github.io/) · [Plain-text summary](https://qrsntu.org/projects/policyattention/icaif/summary.md) Authors: [Xinyue Wu](https://qrsntu.org/people/xinyue-wu/), [Yuhe Sui](https://qrsntu.org/yh/), [Shufang Chen](https://qrsntu.org/people/shufang-chen/), [Yingzhi Tang](https://qrsntu.org/yz/) Contents 1. The idea 2. Why take the controller apart? 3. A market we can solve exactly 4. The loop and the swap 5. The evidence 6. What each part costs 7. By round and by family 8. Against desk rules 9. Adapting after a shift 10. What the results mean 11. Limitations and what didn't work 12. Where this is going 13. Technical details 14. Notation and metrics 15. Protocol and controls 16. Summary 17. In brief 18. Sources Part 1 · The idea ## 1. Why take the controller apart? A market maker keeps re-setting its quote skew as fees, volatility, liquidity and order-flow toxicity change. In-context reinforcement learning offers a controller that adapts without being re-engineered for each regime: a sequence model with frozen weights adapts at inference time by conditioning on recent experience. When such a controller does badly after a regime shift, the practitioner needs to know which learned component failed before deciding what to fix. End-to-end scores mix three things: how well the policy-improvement step is learned, how well the value estimate it relies on is learned, and how hard the new regime is. A fixed rule also looks best in the regime it was designed for. In an exactly solvable market, each learned component can be swapped for its exact counterpart on the same market and the same initial policy. A failure then becomes a set of measured component costs under each stress. ## 2. A market we can solve exactly The dealer holds inventory $q$ from −2 to 2 and observes order-flow imbalance $z ∈ { − 1,0, + 1}$, giving 15 states. Each step it sets a quote skew $u ∈ { − 1,0, + 1}$; at most one fill occurs. Buy pressure raises both the chance that the dealer's ask is hit and the chance that the price then rises: the adverse-selection channel, carried here by an observed state. The reward is spread income minus fees, plus mark-to-market, minus a mean–variance inventory penalty, with discount 0.8. There are 9 families of 64 paired markets: nominal (G0) and fee, volatility, liquidity and toxicity stress (F, V, L, I), each at a moderate (1) and a severe (2) dose. Every nominal market has a twin in each stress family that re-draws one group of parameters and keeps the same random initial policy, so every comparison is paired market by market. At the severe dose, fees rise from [0, 0.2] to [0.6, 0.8], volatility from [0.8, 1.2] to [1.8, 2.2], and liquidity falls from [1, 2] to [0.3, 0.6]. The family is not trivial. On nominal markets (G0) the inventory rule HINV reaches a median value capture of 0.678 and exact PMD from the initial policy reaches 0.883 after 20 rounds, and severe stress changes at least one optimal action in 95–100% of markets. ## 3. The loop and the swap in one picture **The in-context loop and the exact swap.** The formulas are written out under Notation and metrics. - **Actor.** Reads the policy $π_k$, the action values $Q_k$ and the step $η$, and returns $π_{k + 1}$. It is trained toward the PMD step, $softmax( log π_k + ηQ_k )$. - **Critic.** Reads $Q_k$, the reward table and 6 sampled successors per state–action pair, and returns $Q_{k + 1}$. It is trained toward the average backup of those six samples. - **The arms.** Exact runs exact PMD with the exact backup (exact TD-PMD, the reference). ActorOnly learns only the actor, CriticOnly only the critic, FullLoop both. The follow-up arm SampleAvg replaces the critic network by the plain six-sample average it is trained to output, computed from the same samples. The controller is [PolicyAttention](https://qrsntu.org/projects/policyattention/): one Transformer (4 layers, 4 heads, width 64) trained on nominal markets only, in 5 runs. Its weights stay frozen across rounds, so it adapts only through its context. Every arm starts from the same random policy and its exact action values, and runs 20 rounds. Part 2 · The evidence ### Tested synthetic, pre-registered - Each component's cost on all 9 families: 5 trained runs × 64 paired markets. - The full loop against desk rules and against the carried old-regime optimum. - Holds rules committed before any learned result; two-sided tests and statistics code before any locked output was read. - Follow-up, registered after those results: the untrained SampleAvg arm. ### Not tested open - Real markets and trading P&L. - A converged starting policy, or more than 20 rounds. - Any sample budget other than six successors per state–action pair. ## 4. What each part costs after 20 rounds Each cost is a paired difference in normalized loss against exact TD-PMD on the same market: the median of 320 paired values (5 runs × 64 markets) with a 95% hierarchical-bootstrap interval. Lower is better. ### Learned actor Pre-registered · nominal markets +0.0020 interval [−0.0002, +0.0043], which includes 0. Within the holds margin of exact TD-PMD on all nine families (ratios 0.990–1.086). ### Learned critic Pre-registered · nominal markets +0.0219 interval [+0.0186, +0.0248]. Positive on every family, from +0.0191 (F1) to +0.0580 (L2). The full loop's excess, +0.0235, is almost all the critic's. ### Untrained average Follow-up · all nine families 0.70–1.10× the learned critic's excess, for SampleAvg, the plain average of the same six samples (+0.0197 to +0.0450 over exact TD-PMD). So a critic that fitted its six-sample target perfectly would keep at least 70% of the cost: improving the critic stage means improving what it is trained toward, not only how well it fits. [Shortened here. The full page is https://qrsntu.org/projects/policyattention/icaif/details/index.md; it goes on with: By round and by family; Where the loop pays off against desk rules; Adapting after a shift; What the results mean; Limitations and what didn't work; Where this is going; Notation and metrics; Protocol and controls; In brief; Sources.] --- # RiskHedge > RiskHedge: a fail-closed gate for financial risk decisions that accepts a cheap hedge ranking only when a certificate shows exhaustive evaluation would choose the same hedges; otherwise it buys exact prices or falls back. ICAIF 2026 FinAI-SPS workshop paper. Web page: https://qrsntu.org/projects/riskhedge/ (content last changed 2026-11-11) Site index for agents: https://qrsntu.org/llms.txt [← All projects](https://qrsntu.org/projects/) QRS research · Financial risk · Certified decisions Certify the Decision, Not the Shortcut: Fail-Closed Screening for Financial Risk Decisions Accept a cheap risk computation only when a certificate shows it picks the same hedges as exhaustive evaluation. The usual shortcut ranks hedges from cheap approximate prices and accepts the result, so near the cut-off a small error can **silently change which hedges are chosen**. RiskHedge keeps the cheap computation but accepts its answer only when a certificate proves that exhaustive evaluation would choose the same hedges. Otherwise it pays for exact prices or falls back to exact evaluation. [Details](https://qrsntu.org/projects/riskhedge/details/) [Companion paper (OpenReview)](https://openreview.net/forum?id=OzOTmkkAWW) · [Plain-text summary](https://qrsntu.org/projects/riskhedge/summary.md) Active Paper by [Yingzhi Tang](https://qrsntu.org/yz/) and [Yuhe Sui](https://qrsntu.org/yh/) · FinAI-SPS workshop, ICAIF 2026 Illustrative toy, not paper data. ## RiskHedge at a glance The idea Each candidate hedge $j$ has a risk interval $[ℓ_j ,u_j ]$ computed from certified price bounds. A proposed top-$k$ set $D$ is accepted only if $$(u_d ,id_d ) < (ℓ_e ,id_e )$$ for every chosen $d ∈ D$ and every rejected $e ∉ D$. Every chosen hedge is better at its worst than every rejected hedge at its best, so exact repricing cannot change the set; IDs break ties. Experiments #### Synthetic tested Gate wrong or unsupported · 0/200 Midpoint shortcut silently wrong · 10/200 Exact work, median · 17.2% 200 fresh instances from one synthetic Merton option-hedging pipeline, enriched near the decision boundary. Work is the share of exhaustive exact pricing paid after the initial bounds. #### Real portfolios Not yet tested No real portfolio data has been used yet. [Scope, caveats and negative results: Limitations and what didn’t work →](https://qrsntu.org/projects/riskhedge/details/#limitations) ## Open questions from the paper’s limitations 1. ### What does certification cost end to end? The work count leaves out building the cheap certified price bounds, and no wall-clock time was measured in the paper. 2. ### Does it hold beyond one synthetic book? Every result comes from one synthetic Merton pipeline with 24 hedges, 64 scenarios and a top-3 choice. 3. ### Can learned models supply valid bounds? Statistical sets from a learned surrogate missed some prices, so the guarantee does not transfer to them. [The questions in full →](https://qrsntu.org/projects/riskhedge/details/#open) ## Workshop paper - ### Certify the Decision, Not the Shortcut: Fail-Closed Screening for Financial Risk Decisions [Yingzhi Tang](https://qrsntu.org/yz/) (Quantitative Research Society, Singapore; The Chinese University of Hong Kong, Shenzhen), [Yuhe Sui](https://qrsntu.org/yh/) (Quantitative Research Society, Singapore; Nanyang Technological University) Financial AI Security, Privacy, and Safety Workshop (FinAI-SPS) at ICAIF 2026 · poster · non-archival ## Accepted Research - ### Diverse Scenarios, Selective Evaluation: Certified Spectral-Risk Decisions under Costly Robustness Tests [Yuhe Sui](https://qrsntu.org/yh/), [Yingzhi Tang](https://qrsntu.org/yz/) EvoRobust Workshop, NeurIPS 2026 · 2026 · Accepted [[OpenReview]](https://openreview.net/forum?id=OzOTmkkAWW) Project team Lead: [Yuhe Sui](https://qrsntu.org/yh/) Authors: [Yingzhi Tang](https://qrsntu.org/yz/), [Jianing Zhang](https://qrsntu.org/jn/), [Mi Zeyuan](https://qrsntu.org/people/mi-zeyuan/) Slicers: [Wang Zimeng](https://qrsntu.org/people/wang-zimeng/), [Yingzhi Tang](https://qrsntu.org/yz/) Scan to open this page ## Publications - **Diverse Scenarios, Selective Evaluation: Certified Spectral-Risk Decisions under Costly Robustness Tests** (2026). [Yuhe Sui](https://qrsntu.org/yh/), [Yingzhi Tang](https://qrsntu.org/yz/). EvoRobust Workshop, NeurIPS 2026. Accepted. [OpenReview](https://openreview.net/forum?id=OzOTmkkAWW) --- # RiskHedge: How It Works > How RiskHedge works: why not accept the cheap ranking, buying exact prices until the choice is fixed, the decision flow, results on 200 boundary-enriched synthetic instances, limitations, and the certificate in full. Web page: https://qrsntu.org/projects/riskhedge/details/ (content last changed 2026-11-11) Site index for agents: https://qrsntu.org/llms.txt [← RiskHedge overview](https://qrsntu.org/projects/riskhedge/) RiskHedge · How it works The intuition and the figures come first. Protocol, results, limitations, definitions and the full table follow. [Companion paper (OpenReview)](https://openreview.net/forum?id=OzOTmkkAWW) · [Plain-text summary](https://qrsntu.org/projects/riskhedge/summary.md) Paper by [Yingzhi Tang](https://qrsntu.org/yz/) (Quantitative Research Society, Singapore; The Chinese University of Hong Kong, Shenzhen) and [Yuhe Sui](https://qrsntu.org/yh/) (Quantitative Research Society, Singapore; Nanyang Technological University). Contents 1. The idea 2. Why not accept the cheap answer? 3. Buy exact prices until it is fixed 4. The gate in one picture 5. The evidence 6. How it was tested 7. Results 8. Where the shortcut fails 9. What the results mean 10. The companion paper 11. Limitations and what didn’t work 12. Open questions 13. Technical details 14. Notation and the certificate 15. Full results 16. Summary 17. In brief 18. Sources Part 1 · The idea ## 1. Why not accept the cheap answer? A risk engine compares candidate hedges by the tail loss of the hedged book over many market scenarios and keeps the least risky few. Exact evaluation reprices every option in every scenario, so systems use price bounds, caches, proxies or learned surrogates instead. These can be accurate on average and still swap the last chosen hedge for the first rejected one when the two are close. The obvious approach is to rank by the approximate numbers and accept the result. Model validation describes how good an approximation is in aggregate, but it does not by itself establish that *this* hedge decision matches exact evaluation, and a wrong ranking looks exactly like a right one. RiskHedge asks a narrower question: when may an approximate computation be allowed to decide which hedges are chosen? The answer is fail-closed. Accept only when a certificate proves exact-set equivalence; otherwise buy exact information or fall back. Uncertainty produces more computation, not an unqualified decision. ## 2. Buy exact prices until the decision is fixed Each bar below is a hedge's risk interval: its exact risk lies somewhere inside. Buy exact prices and the bars narrow. The choice is certified once the three chosen bars sit entirely to the left of the other three. Switch on the midpoint shortcut to see a cheap ranking pick the wrong third hedge with nothing to warn you. **Illustrative toy, not paper data.** Six hedges, top three chosen; each batch is 8 exact primitives, the batch size the paper's gate uses. ## 3. The gate in one picture **Decision flow.** Redrawn from Fig. 1 of the FinAI-SPS paper; the refinement box says “primitives” where the paper's figure says “legs”, matching the paper's text. Teal is the certified path; the dashed ochre box is the uncertified midpoint baseline. - **Are the bounds usable?** The price-bound provider must be available and valid, the bounds must belong to the current input, and the call must be in the registered scope. A stale input is recomputed; any other failure falls back to exact evaluation. - **Can any exact value change the choice?** If every chosen hedge is better at its worst than every rejected hedge at its best, the gate returns CERTIFIED with the exact top-3 set. - **What if it can?** Buy up to 8 more exact primitives and test again; if the budget runs out, return EXACT_FALLBACK. Every run writes a safety record a separate checker can replay. Part 2 · The evidence ### Tested so far synthetic - Pre-specified confirmation on 200 fresh, boundary-enriched instances: 0 wrong or unsupported decisions. - 60 injected guard faults, each rejected by its intended guard. - All 260 confirmation records replayed by the standalone checker. - Post hoc: work-matched baselines, corrupted bounds, a sampled audit, a refusal path, a learned surrogate. ### Not yet tested open - Real portfolio data. - Wall-clock time and the cost of building the price bounds. - Other pricers, decisions and books. ## 4. How it was tested one synthetic Merton pipeline The decision chooses the 3 least risky of 24 hedge packages by expected shortfall over 64 market scenarios, on a book that is long one stock. Exhaustive evaluation needs 320 shared exact option-pricing primitives; work is the fraction of those paid after the initial bounds. **Baseline.** The midpoint baseline gets the same input, bounds, risk measure and tie rule, replaces each interval by its midpoint, ranks and accepts. It pays no exact prices, so it is not work matched. **Two stages.** A discovery study (120 instances) required midpoint errors in at least two strata, found only 3 in one, and stopped. A new registration then fixed fresh seeds, references, quotas, bins and the single confirmatory test before any confirmation outcome was computed. The two studies are never pooled. The 200 confirmation instances were enriched near the decision boundary: 50 near-tie, wide-bound cases (B1) and 150 others (B2–B5). ## 5. Results ### Confirmation Pre-specified · 200 fresh, boundary-enriched instances 0/200 wrong or unsupported decisions by the gate, at a median 17.2% of exhaustive work (max 0.350). The midpoint baseline was silently wrong on 10/200. ### Guard faults Pre-specified · 60 injected states 60/60 rejected by the intended guard: 40 missing-provider and out-of-scope runs fell back exactly, 20 stale-input runs recomputed and certified. ### Work-matched Post hoc · same refinement order 11 wrong for an uncertified rule given about the same total work as the gate (8,600 vs 8,518 primitives). Among the budgets tried, zero errors first appeared at 1.88× the gate's work. ## 6. Where the shortcut fails, and what certainty costs **By boundary risk.** The 200 confirmation instances in five equal groups of boundary risk (initial uncertainty relative to the exact margin at the cut-off), frozen before reporting. The midpoint baseline fails only in the top group (10/40); the gate is never wrong, and its median work rises from 0 to 0.21–0.23 of exhaustive, then levels off. Descriptive. Work-matched comparison post hoc **Same order, no certificate.** Uncertified midpoint rules follow the gate's refinement order, pay a fixed budget of exact primitives per instance, then accept. Selected budgets from the paper's table; totals over the 200 instances. ## 7. What the results mean - **Errors are silent.** The baseline's output looks the same whether it is right or wrong. Every one of its 10 errors swapped only the third chosen hedge, all in the near-tie, wide-bound group (pre-specified Fisher test, p = 0.016). - **Ambiguity becomes work.** 35 instances were certified from the initial bounds and 165 after refinement; none needed exhaustive pricing. - **The certificate adds stopping.** A fixed budget either wastes work or leaves errors; the certificate says when each instance can stop. [Shortened here. The full page is https://qrsntu.org/projects/riskhedge/details/index.md; it goes on with: The companion paper EvoRobust, NeurIPS 2026; Limitations and what didn’t work; Open questions; Notation and the certificate; Full results; In brief; Sources; Accepted Research.] --- # SoftRank > SoftRank: the approximation rank of softmax attention, measured on the operator rather than on one choice of query and key coordinates. A gauge-invariant bound from the softmax-visible interaction, a sharp r/2 exponent, and certified brackets on trained heads. NeurReps 2026. Web page: https://qrsntu.org/projects/softrank/ (content last changed 2026-11-11) Site index for agents: https://qrsntu.org/llms.txt [← All projects](https://qrsntu.org/projects/) QRS research · Attention theory · Representation geometry The approximation **rank** of **soft**max attention How much rank does an attention head need? Measure the operator, not one choice of query and key coordinates. A natural way to judge how complex a head is, is to look at its queries and keys: their radii, their spectra, the rank of the logits. But one attention matrix has many query–key factorizations, and those numbers change from one to the next. SoftRank asks how much rank is needed to reproduce the head’s output on **every** bounded set of values, and bounds it with the geometry that softmax itself can see, which is **the same in every coordinate frame**. [Details](https://qrsntu.org/projects/softrank/details/) [Paper (arXiv)](https://arxiv.org/abs/2608.28150) · [Extended abstract (OpenReview)](https://openreview.net/forum?id=0zgsZ8FjVz) · [Plain-text summary](https://qrsntu.org/projects/softrank/summary.md) Active Lead: [Yuhe Sui](https://qrsntu.org/yh/) Authors: [Jianing Zhang](https://qrsntu.org/jn/), [Yingzhi Tang](https://qrsntu.org/yz/) Slicers: [Jingshu Liu](https://qrsntu.org/people/jingshu-liu/) Quantitative Research Society, Singapore · Nanyang Technological University NeurReps 2026: extended abstract accepted · arXiv 2608.28150 (v2) The rank-two example from the paper, computed exactly. ## SoftRank at a glance The idea Softmax cannot see anything that adds the same number to a whole row of logits, and the logits cannot see how they are split between queries and keys. What is left is the doubly centred log-attention $C = P( log A)P$. If a rank-$r$ part $T$ of $C$ leaves a residual $τ(T) < ε$, then $$r_ε (A) ≤ C_r (1 + \frac{κ(T)}{(ε − τ(T))^2 })^{r/2} ,$$ where $κ(T)$ is the smallest radius product over all factorizations of $T$. Nothing in the bound depends on the coordinate frame, and no smaller exponent than $r/2$ works in the worst case. Experiments #### Learned attention heads tested | Certified rank brackets | 9,978 | |---|---| | Median upper ÷ lower, five model arms | 2.6–3.8 | | Best construction ÷ lower certificate | 2.6–3.7× | BERT, GPT-2, Qwen2.5-1.5B and two ViT-Base checkpoints at ε = 0.25; the construction gap is on a separate held-out panel of 1,152 operators. On Qwen2.5-1.5B and, in a blind run on fresh documents, Llama-3.2-1B, fixed rank budgets of 8 and 16 are shown feasible less often as the block widens. #### Validated numerics In progress Interval-checked versions of the certificates, and brackets for κ(T) itself on learned heads, are not finished. Until then the certificates are float64 computations without interval enclosure. **2,302 learned attention operators, each with a certified lower rank and a checked upper rank (ε = 0.25).** Teal operators are resolved: the two ends are within a factor of two or a few ranks of each other. [More results →](https://qrsntu.org/projects/softrank/details/#brackets) [Limitations and what didn’t work →](https://qrsntu.org/projects/softrank/details/#limitations) ## Where this is going proposed 1. ### Validated certificates Rounding-error and residual bounds for every stored witness, so each verdict holds beyond float64. 2. ### The intrinsic size on real heads Primal and dual brackets for κ(T) on a representative set of heads, in place of the balanced-factor surrogate. 3. ### Causal masks A theory for row-dependent supports, so decoder self-attention is covered without common-prefix blocks. [What comes next, in full →](https://qrsntu.org/projects/softrank/details/#next-steps) ## Papers - ### Intrinsic Interaction Geometry Controls the Low-Rank Complexity of Softmax Attention Yuhe Sui, Jianing Zhang, Yingzhi Tang arXiv 2608.28150 · v1 28 August 2026, v2 22 September 2026 [[arXiv]](https://arxiv.org/abs/2608.28150) ## Accepted Research - ### The Approximation Rank of Softmax Attention: Sharp Geometric Laws and Robust Interaction Dimension [Yuhe Sui](https://qrsntu.org/yh/), [Jianing Zhang](https://qrsntu.org/jn/) NeurReps Workshop, NeurIPS 2026 (extended abstract) · 2026 · Accepted [[OpenReview]](https://openreview.net/forum?id=0zgsZ8FjVz) Scan to open this page ## Publications - **The Approximation Rank of Softmax Attention: Sharp Geometric Laws and Robust Interaction Dimension** (2026). [Yuhe Sui](https://qrsntu.org/yh/), [Jianing Zhang](https://qrsntu.org/jn/). NeurReps Workshop, NeurIPS 2026 (extended abstract). Accepted. [OpenReview](https://openreview.net/forum?id=0zgsZ8FjVz) [arXiv](https://arxiv.org/abs/2608.28150) --- # SoftRank: How It Works > How SoftRank works: why query-key radii depend on the coordinate frame, the visible interaction C = P(log A)P and the intrinsic size kappa, where the r/2 exponent comes from, certified brackets on BERT, GPT-2, Qwen2.5, ViT and Llama heads, and the limitations. Web page: https://qrsntu.org/projects/softrank/details/ (content last changed 2026-11-11) Site index for agents: https://qrsntu.org/llms.txt [← SoftRank overview](https://qrsntu.org/projects/softrank/) SoftRank · How it works The rank an attention head needs is a property of the attention operator, not of the query and key vectors that happen to produce it. SoftRank defines that rank by the outputs it must preserve, bounds it with geometry that survives every change of coordinates, and brackets it directly on trained heads. [Paper (arXiv)](https://arxiv.org/abs/2608.28150) · [Extended abstract (OpenReview)](https://openreview.net/forum?id=0zgsZ8FjVz) · [Plain-text summary](https://qrsntu.org/projects/softrank/summary.md) Lead: [Yuhe Sui](https://qrsntu.org/yh/) Authors: [Jianing Zhang](https://qrsntu.org/jn/), [Yingzhi Tang](https://qrsntu.org/yz/) Slicers: [Jingshu Liu](https://qrsntu.org/people/jingshu-liu/) Quantitative Research Society, Singapore · Nanyang Technological University Contents 1. The idea 2. Why not measure the queries and keys? 3. Three questions, three objects 4. Coordinates against the operator 5. Where the exponent comes from 6. Measuring the rank directly 7. The evidence 8. Results: brackets on learned heads 9. Constructions against certificates 10. Rank budgets as the block widens 11. Other attention regimes 12. Native coordinates as a predictor 13. Limitations and what didn’t work 14. Where this is going 15. Technical details 16. Definitions and the theorem 17. Support-geometry laws 18. Full tables 19. Reproducibility notes 20. Summary 21. In brief 22. Sources Part 1 · The idea ## 1. Why not measure the queries and keys? An attention head computes logits $H = βQK^⊤$ and normalises each row with softmax. The obvious way to say how complex the head is, is to measure the two point clouds that produce the logits: how far the queries and keys spread (their radii, combined into $Γ = βρ_Q ρ_K$), or the spectrum of $W_q^⊤ W_k$. A bound written in those terms is correct for the factorization it is given. But the factorization is not unique. For any invertible matrix $S$, replacing $Q$ by $QS$ and $K$ by $KS^{ − ⊤}$ leaves every logit, and so the whole attention matrix, unchanged. The point clouds do change. In the paper’s rank-two example, four queries and keys at $±e_1 ,±e_2$ give $Γ = 1$; after the shear $S_t$ with $t = 2,5,10$ the same attention matrix has $Γ = 5,26,101$. A measure that can be made as large as we like without touching the operator describes the coordinates, not the head. **Two ways to measure.** Above, quantities of one query–key factorization, which move under a change of coordinates. Below, quantities of the normalised operator, which do not. SoftRank changes both ends. The target is the operator’s own approximation rank. The geometry is read off the part of the logits that softmax can see, and its size is minimised over every factorization at once, so that no frame is privileged. ## 2. Three questions, three objects ### What must be preserved? The approximation rank $r_ε (A)$ The least rank of any real matrix $B$ within maximum-row-$ℓ_1$ error $ε$ of $A$. That error is exactly the worst output error over all value vectors of norm at most one, in any normed space, so $r_ε (A)$ is the rank needed to preserve every bounded attention output. ### What can softmax see? The visible interaction $C = P_m ( log A)P_N$ Row softmax ignores any constant added to a row of logits, and terms that depend only on the key act as common base weights. Centring $log A$ on both sides removes the first and keeps the second apart. $C$ is a function of $A$ alone, and $R = rankC$ is the exact visible interaction dimension. ### How large is it? Residual $τ(T)$ and size $κ(T)$ Keep a rank-$r$ part $T$ of $C$. The residual $τ(T)$ is the output error that dropping the rest can cause; the size $κ(T)$ is the smallest product of the two enclosing radii over every factorization $T = XY^⊤$. Both are fixed by $A$ and $T$, whatever coordinates produced them. Put together, they give the central theorem: whenever $τ(T) < ε$, $$r_ε (A) ≤ min{p,C_r (1 + \frac{κ(T)}{(ε − τ(T))^2 })^{r/2} },$$ Here $p = min(m,N)$ is the trivial cap. The bound is never worse than the native bound with the same retained dimension, and it does not change under any invertible refactorization of the queries and keys. ## 3. Coordinates against the operator The two sliders separate what changes the coordinates from what changes the head. The shear $t$ moves the queries and keys; the matrix below is recomputed from the moved vectors each time and does not change. The logit scale $β$ changes the head itself: the matrix sharpens, and the intrinsic size grows with it. **Explorer: one operator, many frames.** Here $κ(C) = 1$ exactly at $β = 1$: the original factorization reaches a radius product of 1, and the rectangular contrast $C_{11} − C_{12} − C_{21} + C_{22} = 4$ shows that no factorization can do better. ## 4. Where the exponent comes from The bound rests on a cover of Gibbs rows: every softmax row is close, in $ℓ_1$, to one of a small set of rows, and the size of that set is a question about convex bodies in fixed dimension. Its exponent is set by the geometry of the support. For self-attention on the unit sphere in $ℝ^d$, the worst-case rank grows like $(1 + β)^{(d − 1)/2}$ with the temperature; on the full ball, with enough tokens, it grows like $β^{d/2}$. One radial degree of freedom adds exactly one half to the exponent. The intrinsic theorem inherits the same exponent, now with the visible dimension $r$ in place of the ambient one, and it cannot be lowered. On a finite grid of $L^r$ points with logits $2sa^⊤ b − s‖b‖^2$, the exact approximation rank is $L^r$ while $κ(C) = sr(L − 1)^2 /2$, so $r_ε (A) = Θ(κ(C)^{r/2} )$. Low interaction rank alone is not enough either: rank-one families with no residual still need growing rank, so a size term like $κ$ cannot be dropped. **The $r/2$ exponent is reached.** The grid family at $r = 2$, $ε = 0.25$: exact rank $L^2$ against $κ(C)$, for $L = 3,5,8,12,20$. Exact values from the proposition. [Shortened here. The full page is https://qrsntu.org/projects/softrank/details/index.md; it goes on with: Measuring the rank directly; Results: brackets on learned heads; Constructions against certificates; Rank budgets as the block widens; Other attention regimes; Native coordinates as a predictor; Limitations and what didn’t work; Where this is going proposed; Definitions and the theorem; Support-geometry laws; Full tables; Reproducibility notes; In brief; Sources; Accepted Research.] --- # People > Meet QRS@NTU officers and QRS researchers, and explore their profiles, projects, and publications. Web page: https://qrsntu.org/people/ (content last changed 2026-11-07) Site index for agents: https://qrsntu.org/llms.txt QRS@NTU officers, QRS researchers, and coordinators. ## Officers [**Yuhe** Sui President & Founder, QRS@NTU · Principal Researcher, Quantitative Research Society](https://qrsntu.org/yh/) [Wang **Zimeng** NUS Sub-branch Coordinator](https://qrsntu.org/people/wang-zimeng/) [**Luis** Ji QRS@NTU](https://qrsntu.org/people/luis-ji/) [Mi **Zeyuan** Vice President, QRS@NTU](https://qrsntu.org/people/mi-zeyuan/) [Xinyue Wu QRS@NTU](https://qrsntu.org/people/xinyue-wu/) ## Researchers [**Jianing** Zhang Researcher, Quantitative Research Society](https://qrsntu.org/jn/) [**Yingzhi** Tang Researcher, Quantitative Research Society](https://qrsntu.org/yz/) [QRS Research Academy](https://qrsntu.org/research-academy/members/) ## Other School Branches [QRS@NUS](https://www.qrs-nus.org/) [QRS@NTU (Taiwan)](https://qrsntu.org/tw/) [QRS@CMU — Coming soon](https://qrsntu.org/cmu/) [QRS@CUHKSZ 量化研究协会@港中深 — Coming soon](https://qrsntu.org/cuhk/) [QRS@FDU 量化研究协会@复旦 — Coming soon](https://qrsntu.org/fdu/) [QRS@SJTU 量化研究协会@交大 — Coming soon](https://qrsntu.org/sjtu/) ## People - [Yuhe Sui](https://qrsntu.org/yh/): President & Founder, QRS@NTU; Principal Researcher, Quantitative Research Society - [Jianing Zhang](https://qrsntu.org/jn/): Researcher, Quantitative Research Society - [Danlin Chen](https://qrsntu.org/cdl/) - [Shufang Chen](https://qrsntu.org/people/shufang-chen/): QRS Research Academy Member - [Xiaotong Li](https://qrsntu.org/people/xiaotong-li/): QRS Research Academy Member - [Wang Zimeng](https://qrsntu.org/people/wang-zimeng/): NUS Sub-branch Coordinator; QRS Research Academy Member - [Duy Anh Nguyen](https://qrsntu.org/people/duy-anh-nguyen/) - [Luis Ji](https://qrsntu.org/people/luis-ji/). QRS@NTU - [Mi Zeyuan](https://qrsntu.org/people/mi-zeyuan/): Vice President, QRS@NTU - [Yingzhi Tang](https://qrsntu.org/yz/): Researcher, Quantitative Research Society - [Xinyue Wu](https://qrsntu.org/people/xinyue-wu/): QRS Research Academy Member. QRS@NTU - [Jingshu Liu](https://qrsntu.org/people/jingshu-liu/): QRS Research Academy Member - [Yizhe Chia](https://qrsntu.org/people/yizhe-chia/): QRS Research Academy Member - [Annie Huaiyu Wang](https://qrsntu.org/people/annie-huaiyu-wang/): QRS Research Academy Member --- # Annie Huaiyu Wang > QRS Research Academy Member Web page: https://qrsntu.org/people/annie-huaiyu-wang/ (content last changed 2026-10-23) Site index for agents: https://qrsntu.org/llms.txt [← Back to People](https://qrsntu.org/people/) BS, Tsinghua University, *Mathematics and Applied Mathematics* [Member, QRS Research Academy, 2026 Fall #2](https://qrsntu.org/research-academy/members/) ## About Annie is a first-year undergraduate in Mathematics and Applied Mathematics at Tsinghua University, in Qiuzhen College and the Yau Mathematical Sciences Leaders Program. Her interests include quantitative finance, risk and mathematical machine learning. She has previously worked on machine learning for mining signal peptides in plant genomes and on agent tools for meteorological data. [Projects](https://qrsntu.org/projects/) · [Publications](https://qrsntu.org/publications/) ## Projects - [FRTBTrace](https://qrsntu.org/projects/frtbtrace/): Active research on faithful, auditable, trade-level FRTB attribution. Status: active. People: [Yuhe Sui](https://qrsntu.org/yh/), [Yingzhi Tang](https://qrsntu.org/yz/). ## Links - OpenReview: https://openreview.net/profile?id=~Annie_Huaiyu_Wang1 - LinkedIn: https://www.linkedin.com/in/wang-annie-huaiyu-621b0a445 --- # Danlin Chen > BS, Western University Web page: https://qrsntu.org/cdl/ (content last changed 2026-10-15) Site index for agents: https://qrsntu.org/llms.txt [← Back to People](https://qrsntu.org/people/) BS, Western University [Projects](https://qrsntu.org/projects/) · [Publications](https://qrsntu.org/publications/) ## Projects - [AI Native Workflow](https://qrsntu.org/projects/ai-native-workflow/): Research on hierarchical, verification-gated and recoverable workflows for scaling long-horizon AI-assisted research and agent execution. Status: active. People: [Yuhe Sui](https://qrsntu.org/yh/), [Jianing Zhang](https://qrsntu.org/jn/), [Shufang Chen](https://qrsntu.org/people/shufang-chen/), [Danlin Chen](https://qrsntu.org/cdl/), [Wang Zimeng](https://qrsntu.org/people/wang-zimeng/), [Xiaotong Li](https://qrsntu.org/people/xiaotong-li/). ## Links - OpenReview: https://openreview.net/profile?id=~Danlin_Chen1 --- # Duy Anh Nguyen > BS, Nanyang Technological University Web page: https://qrsntu.org/people/duy-anh-nguyen/ (content last changed 2026-10-15) Site index for agents: https://qrsntu.org/llms.txt [← Back to People](https://qrsntu.org/people/) BS, Nanyang Technological University [Projects](https://qrsntu.org/projects/) · [Publications](https://qrsntu.org/publications/) ## Links - GitHub: https://github.com/IMainMordekaiser --- # Jianing Zhang > Researcher, Quantitative Research Society Web page: https://qrsntu.org/jn/ (content last changed 2026-10-15) Site index for agents: https://qrsntu.org/llms.txt [← Back to People](https://qrsntu.org/people/) Researcher, Quantitative Research Society MS, Carnegie Mellon University, *Computational Data Science* · Aug 2025 – May 2027 BS, New York University, *Data Science & Mathematics* · Sep 2021 – May 2025 Quant Intern, UBS · Algo · May 2026 – Aug 2026 [Graduate, QRS Research Academy, 2026 Fall #1](https://qrsntu.org/research-academy/members/) ## About Jianing is pursuing a master's in Computational Data Science at Carnegie Mellon University. Previous experience at Alphabet and iFLYTEK includes fraud detection and multilingual AI pipelines. Jianing's interests span applied AI, machine learning engineering, data science, and data engineering, with an emphasis on scalable, trustworthy systems for healthcare and other practical applications. [Projects](https://qrsntu.org/projects/) · [Publications](https://qrsntu.org/publications/) ## Projects - [AI Native Workflow](https://qrsntu.org/projects/ai-native-workflow/): Research on hierarchical, verification-gated and recoverable workflows for scaling long-horizon AI-assisted research and agent execution. Status: active. People: [Yuhe Sui](https://qrsntu.org/yh/), [Jianing Zhang](https://qrsntu.org/jn/), [Shufang Chen](https://qrsntu.org/people/shufang-chen/), [Danlin Chen](https://qrsntu.org/cdl/), [Wang Zimeng](https://qrsntu.org/people/wang-zimeng/), [Xiaotong Li](https://qrsntu.org/people/xiaotong-li/). - [GeoAttention](https://qrsntu.org/projects/geoattention/): Research on attention geometry, representation-relative identifiability, and geometry-adaptive prediction. Status: active. People: [Yuhe Sui](https://qrsntu.org/yh/), [Jianing Zhang](https://qrsntu.org/jn/), [Xiaotong Li](https://qrsntu.org/people/xiaotong-li/). - [RiskHedge](https://qrsntu.org/projects/riskhedge/): Research on spectral-risk decisions, diverse scenarios, and selective robustness evaluation. Status: active. People: [Yuhe Sui](https://qrsntu.org/yh/), [Yingzhi Tang](https://qrsntu.org/yz/), [Jianing Zhang](https://qrsntu.org/jn/), [Mi Zeyuan](https://qrsntu.org/people/mi-zeyuan/). - [SoftRank](https://qrsntu.org/projects/softrank/): Theoretical research on softmax-attention approximation rank, support geometry, and robust interaction dimension. Status: active. People: [Yuhe Sui](https://qrsntu.org/yh/), [Jianing Zhang](https://qrsntu.org/jn/), [Yingzhi Tang](https://qrsntu.org/yz/). ## Publications - **Geometric Attention Fine-Tuning Is Representation-Relative: Autonomous ML Research with AI-Native Workflow** (2026). [Yuhe Sui](https://qrsntu.org/yh/), [Jianing Zhang](https://qrsntu.org/jn/), [Shufang Chen](https://qrsntu.org/people/shufang-chen/), [Xiaotong Li](https://qrsntu.org/people/xiaotong-li/), [Wang Zimeng](https://qrsntu.org/people/wang-zimeng/). AutoMLR Workshop, NeurIPS 2026. Accepted. [OpenReview](https://openreview.net/forum?id=9PwlVQ3gyj) - **The Approximation Rank of Softmax Attention: Sharp Geometric Laws and Robust Interaction Dimension** (2026). [Yuhe Sui](https://qrsntu.org/yh/), [Jianing Zhang](https://qrsntu.org/jn/). NeurReps Workshop, NeurIPS 2026 (extended abstract). Accepted. [OpenReview](https://openreview.net/forum?id=0zgsZ8FjVz) [arXiv](https://arxiv.org/abs/2608.28150) ## Links - OpenReview: https://openreview.net/profile?id=~Jianing_Zhang17 - ORCID: https://orcid.org/0009-0004-3810-9316 - GitHub: https://github.com/Catherineya - LinkedIn: https://www.linkedin.com/in/jianing-zhang-831446223/ --- # Jingshu Liu > QRS Research Academy Member Web page: https://qrsntu.org/people/jingshu-liu/ (content last changed 2026-10-15) Site index for agents: https://qrsntu.org/llms.txt [← Back to People](https://qrsntu.org/people/) BS, Tsinghua University, *Mathematics* · Sep 2025 – Jun 2029 [Member, QRS Research Academy, 2026 Fall #2](https://qrsntu.org/research-academy/members/) [Projects](https://qrsntu.org/projects/) · [Publications](https://qrsntu.org/publications/) ## Projects - [SoftRank](https://qrsntu.org/projects/softrank/): Theoretical research on softmax-attention approximation rank, support geometry, and robust interaction dimension. Status: active. People: [Yuhe Sui](https://qrsntu.org/yh/), [Jianing Zhang](https://qrsntu.org/jn/), [Yingzhi Tang](https://qrsntu.org/yz/). ## Links - OpenReview: https://openreview.net/profile?id=~Jingshu_Liu1 - LinkedIn: https://www.linkedin.com/in/%E5%A9%A7%E5%A7%9D-%E5%88%98-7b034b384/ --- # Luis Ji > QRS@NTU Web page: https://qrsntu.org/people/luis-ji/ (content last changed 2026-10-15) Site index for agents: https://qrsntu.org/llms.txt [← Back to People](https://qrsntu.org/people/) QRS@NTU BS, Nanyang Technological University, *Economics* · Aug 2025 – Present ## About Happy to connect. [Projects](https://qrsntu.org/projects/) · [Publications](https://qrsntu.org/publications/) ## Projects - [CFFA 2026 — China Future Financial Analyst Competition](https://qrsntu.org/projects/cffa-2026/): • Built a Greenwashing Risk Score across 1,676 Chinese green-bond issuances (2016–2025), using issue-spread regressions and portfolio sorts. • Pricing effects are weak unconditionally, but meaningful within credit-risk buckets after controls. Status: active. ## Links - LinkedIn: https://www.linkedin.com/in/luisjicy/ --- # Mi Zeyuan > Vice President, QRS@NTU Web page: https://qrsntu.org/people/mi-zeyuan/ (content last changed 2026-10-15) Site index for agents: https://qrsntu.org/llms.txt [← Back to People](https://qrsntu.org/people/) Vice President, QRS@NTU BS, Nanyang Technological University, *Accountancy and Data Science & Artificial Intelligence* · Aug 2024 – May 2028 ## About Mi studies Accountancy and Data Science & Artificial Intelligence at Nanyang Technological University, with an interest in quantitative finance. Through URECA, Mi researches volatility pricing and leverage-decay risk in leveraged ETF options, and has worked on systematic investment strategies with the Nanyang Student Investment Fund. Technical interests include statistical modelling and using mathematical and computational methods to study financial markets. [Projects](https://qrsntu.org/projects/) · [Publications](https://qrsntu.org/publications/) ## Projects - [RiskHedge](https://qrsntu.org/projects/riskhedge/): Research on spectral-risk decisions, diverse scenarios, and selective robustness evaluation. Status: active. People: [Yuhe Sui](https://qrsntu.org/yh/), [Yingzhi Tang](https://qrsntu.org/yz/), [Jianing Zhang](https://qrsntu.org/jn/), [Mi Zeyuan](https://qrsntu.org/people/mi-zeyuan/). ## Links - LinkedIn: https://www.linkedin.com/in/mizeyuan/ --- # Shufang Chen > QRS Research Academy Member Web page: https://qrsntu.org/people/shufang-chen/ (content last changed 2026-10-15) Site index for agents: https://qrsntu.org/llms.txt [← Back to People](https://qrsntu.org/people/) MS, The University of Hong Kong, *Computer Science* BS, Beijing University of Posts and Telecommunications, *Computer Science* Quant Intern, UBS · May 2026 – Aug 2026 [Member, QRS Research Academy, 2026 Fall #2](https://qrsntu.org/research-academy/members/) [Projects](https://qrsntu.org/projects/) · [Publications](https://qrsntu.org/publications/) ## Projects - [AI Native Workflow](https://qrsntu.org/projects/ai-native-workflow/): Research on hierarchical, verification-gated and recoverable workflows for scaling long-horizon AI-assisted research and agent execution. Status: active. People: [Yuhe Sui](https://qrsntu.org/yh/), [Jianing Zhang](https://qrsntu.org/jn/), [Shufang Chen](https://qrsntu.org/people/shufang-chen/), [Danlin Chen](https://qrsntu.org/cdl/), [Wang Zimeng](https://qrsntu.org/people/wang-zimeng/), [Xiaotong Li](https://qrsntu.org/people/xiaotong-li/). - [PolicyAttention](https://qrsntu.org/projects/policyattention/): Research on softmax policy improvement, its robustness, and its failure modes. Status: active. People: [Yuhe Sui](https://qrsntu.org/yh/), [Shufang Chen](https://qrsntu.org/people/shufang-chen/), [Yingzhi Tang](https://qrsntu.org/yz/), [Xinyue Wu](https://qrsntu.org/people/xinyue-wu/). ## Publications - **Geometric Attention Fine-Tuning Is Representation-Relative: Autonomous ML Research with AI-Native Workflow** (2026). [Yuhe Sui](https://qrsntu.org/yh/), [Jianing Zhang](https://qrsntu.org/jn/), [Shufang Chen](https://qrsntu.org/people/shufang-chen/), [Xiaotong Li](https://qrsntu.org/people/xiaotong-li/), [Wang Zimeng](https://qrsntu.org/people/wang-zimeng/). AutoMLR Workshop, NeurIPS 2026. Accepted. [OpenReview](https://openreview.net/forum?id=9PwlVQ3gyj) ## Links - OpenReview: https://openreview.net/profile?id=~SHUFANG_CHEN1 - ORCID: https://orcid.org/0009-0009-3627-7726 - GitHub: https://github.com/zhaohuisuo1111-droid --- # Wang Zimeng > NUS Sub-branch Coordinator; QRS Research Academy Member Web page: https://qrsntu.org/people/wang-zimeng/ (content last changed 2026-10-15) Site index for agents: https://qrsntu.org/llms.txt [← Back to People](https://qrsntu.org/people/) NUS Sub-branch Coordinator BS, National University of Singapore, *Computer Science* · 2023 – 2027 Quant Intern, UBS · PQA · May 2026 – Aug 2026 [Member, QRS Research Academy, 2026 Fall #2](https://qrsntu.org/research-academy/members/) [Projects](https://qrsntu.org/projects/) · [Publications](https://qrsntu.org/publications/) ## Projects - [AI Native Workflow](https://qrsntu.org/projects/ai-native-workflow/): Research on hierarchical, verification-gated and recoverable workflows for scaling long-horizon AI-assisted research and agent execution. Status: active. People: [Yuhe Sui](https://qrsntu.org/yh/), [Jianing Zhang](https://qrsntu.org/jn/), [Shufang Chen](https://qrsntu.org/people/shufang-chen/), [Danlin Chen](https://qrsntu.org/cdl/), [Wang Zimeng](https://qrsntu.org/people/wang-zimeng/), [Xiaotong Li](https://qrsntu.org/people/xiaotong-li/). - [RiskHedge](https://qrsntu.org/projects/riskhedge/): Research on spectral-risk decisions, diverse scenarios, and selective robustness evaluation. Status: active. People: [Yuhe Sui](https://qrsntu.org/yh/), [Yingzhi Tang](https://qrsntu.org/yz/), [Jianing Zhang](https://qrsntu.org/jn/), [Mi Zeyuan](https://qrsntu.org/people/mi-zeyuan/). - [CFFA 2026 — China Future Financial Analyst Competition](https://qrsntu.org/projects/cffa-2026/): • Built a Greenwashing Risk Score across 1,676 Chinese green-bond issuances (2016–2025), using issue-spread regressions and portfolio sorts. • Pricing effects are weak unconditionally, but meaningful within credit-risk buckets after controls. Status: active. ## Publications - **Geometric Attention Fine-Tuning Is Representation-Relative: Autonomous ML Research with AI-Native Workflow** (2026). [Yuhe Sui](https://qrsntu.org/yh/), [Jianing Zhang](https://qrsntu.org/jn/), [Shufang Chen](https://qrsntu.org/people/shufang-chen/), [Xiaotong Li](https://qrsntu.org/people/xiaotong-li/), [Wang Zimeng](https://qrsntu.org/people/wang-zimeng/). AutoMLR Workshop, NeurIPS 2026. Accepted. [OpenReview](https://openreview.net/forum?id=9PwlVQ3gyj) ## Links - OpenReview: https://openreview.net/profile?id=~Wang_Zimeng1 - GitHub: https://github.com/jiangsuwangjing - LinkedIn: https://www.linkedin.com/in/zimeng-wang-2b3919285/ --- # Xiaotong Li > QRS Research Academy Member Web page: https://qrsntu.org/people/xiaotong-li/ (content last changed 2026-10-15) Site index for agents: https://qrsntu.org/llms.txt [← Back to People](https://qrsntu.org/people/) BS, Dalian University of Technology, *Economics & Management* [Member, QRS Research Academy, 2026 Fall #1](https://qrsntu.org/research-academy/members/) [Projects](https://qrsntu.org/projects/) · [Publications](https://qrsntu.org/publications/) ## Projects - [AI Native Workflow](https://qrsntu.org/projects/ai-native-workflow/): Research on hierarchical, verification-gated and recoverable workflows for scaling long-horizon AI-assisted research and agent execution. Status: active. People: [Yuhe Sui](https://qrsntu.org/yh/), [Jianing Zhang](https://qrsntu.org/jn/), [Shufang Chen](https://qrsntu.org/people/shufang-chen/), [Danlin Chen](https://qrsntu.org/cdl/), [Wang Zimeng](https://qrsntu.org/people/wang-zimeng/), [Xiaotong Li](https://qrsntu.org/people/xiaotong-li/). - [GeoAttention](https://qrsntu.org/projects/geoattention/): Research on attention geometry, representation-relative identifiability, and geometry-adaptive prediction. Status: active. People: [Yuhe Sui](https://qrsntu.org/yh/), [Jianing Zhang](https://qrsntu.org/jn/), [Xiaotong Li](https://qrsntu.org/people/xiaotong-li/). ## Publications - **Geometric Attention Fine-Tuning Is Representation-Relative: Autonomous ML Research with AI-Native Workflow** (2026). [Yuhe Sui](https://qrsntu.org/yh/), [Jianing Zhang](https://qrsntu.org/jn/), [Shufang Chen](https://qrsntu.org/people/shufang-chen/), [Xiaotong Li](https://qrsntu.org/people/xiaotong-li/), [Wang Zimeng](https://qrsntu.org/people/wang-zimeng/). AutoMLR Workshop, NeurIPS 2026. Accepted. [OpenReview](https://openreview.net/forum?id=9PwlVQ3gyj) ## Links - OpenReview: https://openreview.net/profile?id=~李晓彤1 - ORCID: https://orcid.org/0009-0005-2622-8120 - GitHub: https://github.com/ClzHhhhhhh --- # Xinyue Wu > QRS Research Academy Member; QRS@NTU Web page: https://qrsntu.org/people/xinyue-wu/ (content last changed 2026-10-15) Site index for agents: https://qrsntu.org/llms.txt [← Back to People](https://qrsntu.org/people/) QRS@NTU BS, Nanyang Technological University, *Mathematical Sciences and Economics* · Aug 2024 – Aug 2027 [Member, QRS Research Academy, 2026 Fall #2](https://qrsntu.org/research-academy/members/) [Projects](https://qrsntu.org/projects/) · [Publications](https://qrsntu.org/publications/) ## Projects - [PolicyAttention](https://qrsntu.org/projects/policyattention/): Research on softmax policy improvement, its robustness, and its failure modes. Status: active. People: [Yuhe Sui](https://qrsntu.org/yh/), [Shufang Chen](https://qrsntu.org/people/shufang-chen/), [Yingzhi Tang](https://qrsntu.org/yz/), [Xinyue Wu](https://qrsntu.org/people/xinyue-wu/). ## Publications - **PolicyAttention: Robustness and Failure Modes of Softmax Policy Improvement** (2026). [Yuhe Sui](https://qrsntu.org/yh/), [Yingzhi Tang](https://qrsntu.org/yz/), [Xinyue Wu](https://qrsntu.org/people/xinyue-wu/). EvoRobust Workshop, NeurIPS 2026. Accepted. [OpenReview](https://openreview.net/forum?id=VgswuGrqp6) ## Links - OpenReview: https://openreview.net/profile?id=~Xinyue_Wu5 - LinkedIn: https://www.linkedin.com/in/eric-xinyue-wu/ --- # Yingzhi Tang > Researcher, Quantitative Research Society Web page: https://qrsntu.org/yz/ (content last changed 2026-10-15) Site index for agents: https://qrsntu.org/llms.txt [← Back to People](https://qrsntu.org/people/) Researcher, Quantitative Research Society BS, The Chinese University of Hong Kong, Shenzhen, *Financial Engineering (minor in Artificial Intelligence)* · 2023 – 2027 Quantitative Developer Intern, UBS · Jul 2026 – Present [Graduate, QRS Research Academy, 2026 Fall #1](https://qrsntu.org/research-academy/members/) ## About Yingzhi studies Financial Engineering with a minor in Artificial Intelligence at The Chinese University of Hong Kong, Shenzhen. Her interests span quantitative finance, machine learning, and robust decision-making. Her experience includes quantitative factor research, factor timing, and machine-learning-based trading strategies, alongside work on AI agent systems as a quantitative developer intern at UBS. At QRS, she contributes to PolicyAttention and RiskHedge, exploring the reliability of attention-based policy improvement and decision-making under uncertainty. [Projects](https://qrsntu.org/projects/) · [Publications](https://qrsntu.org/publications/) ## Projects - [FRTBTrace](https://qrsntu.org/projects/frtbtrace/): Active research on faithful, auditable, trade-level FRTB attribution. Status: active. People: [Yuhe Sui](https://qrsntu.org/yh/), [Yingzhi Tang](https://qrsntu.org/yz/). - [PolicyAttention](https://qrsntu.org/projects/policyattention/): Research on softmax policy improvement, its robustness, and its failure modes. Status: active. People: [Yuhe Sui](https://qrsntu.org/yh/), [Shufang Chen](https://qrsntu.org/people/shufang-chen/), [Yingzhi Tang](https://qrsntu.org/yz/), [Xinyue Wu](https://qrsntu.org/people/xinyue-wu/). - [RiskHedge](https://qrsntu.org/projects/riskhedge/): Research on spectral-risk decisions, diverse scenarios, and selective robustness evaluation. Status: active. People: [Yuhe Sui](https://qrsntu.org/yh/), [Yingzhi Tang](https://qrsntu.org/yz/), [Jianing Zhang](https://qrsntu.org/jn/), [Mi Zeyuan](https://qrsntu.org/people/mi-zeyuan/). - [SoftRank](https://qrsntu.org/projects/softrank/): Theoretical research on softmax-attention approximation rank, support geometry, and robust interaction dimension. Status: active. People: [Yuhe Sui](https://qrsntu.org/yh/), [Jianing Zhang](https://qrsntu.org/jn/), [Yingzhi Tang](https://qrsntu.org/yz/). ## Publications - **Diverse Scenarios, Selective Evaluation: Certified Spectral-Risk Decisions under Costly Robustness Tests** (2026). [Yuhe Sui](https://qrsntu.org/yh/), [Yingzhi Tang](https://qrsntu.org/yz/). EvoRobust Workshop, NeurIPS 2026. Accepted. [OpenReview](https://openreview.net/forum?id=OzOTmkkAWW) - **PolicyAttention: Robustness and Failure Modes of Softmax Policy Improvement** (2026). [Yuhe Sui](https://qrsntu.org/yh/), [Yingzhi Tang](https://qrsntu.org/yz/), [Xinyue Wu](https://qrsntu.org/people/xinyue-wu/). EvoRobust Workshop, NeurIPS 2026. Accepted. [OpenReview](https://openreview.net/forum?id=VgswuGrqp6) ## Links - OpenReview: https://openreview.net/profile?id=~Yingzhi_Tang2 - ORCID: https://orcid.org/0009-0004-7571-0100 - LinkedIn: https://www.linkedin.com/in/yingzhi-tang-735603360/ --- # Yizhe Chia > QRS Research Academy Member Web page: https://qrsntu.org/people/yizhe-chia/ (content last changed 2026-10-15) Site index for agents: https://qrsntu.org/llms.txt [← Back to People](https://qrsntu.org/people/) BS, Nanyang Technological University, *Mathematics* · Aug 2025 – Dec 2028 [Member, QRS Research Academy, 2026 Fall #2](https://qrsntu.org/research-academy/members/) ## About An early fascination with the Pythagorean theorem inspired Yizhe to study Mathematical Sciences. Yizhe is interested in exploring connections across mathematics and applying mathematical ideas to practical problems. [Projects](https://qrsntu.org/projects/) · [Publications](https://qrsntu.org/publications/) ## Projects - [AI Native Workflow](https://qrsntu.org/projects/ai-native-workflow/): Research on hierarchical, verification-gated and recoverable workflows for scaling long-horizon AI-assisted research and agent execution. Status: active. People: [Yuhe Sui](https://qrsntu.org/yh/), [Jianing Zhang](https://qrsntu.org/jn/), [Shufang Chen](https://qrsntu.org/people/shufang-chen/), [Danlin Chen](https://qrsntu.org/cdl/), [Wang Zimeng](https://qrsntu.org/people/wang-zimeng/), [Xiaotong Li](https://qrsntu.org/people/xiaotong-li/). ## Links - LinkedIn: https://www.linkedin.com/in/yizhechia/ --- # Yuhe Sui > President & Founder, QRS@NTU; Principal Researcher, Quantitative Research Society Web page: https://qrsntu.org/yh/ (content last changed 2026-10-15) Site index for agents: https://qrsntu.org/llms.txt [← Back to People](https://qrsntu.org/people/) President & Founder, QRS@NTU Principal Researcher, Quantitative Research Society BS, Nanyang Technological University, *Mathematical Sciences and Economics* · Aug 2025 – Dec 2028 Quant Intern, UBS · PQA · May 2026 – Aug 2026 ## About Yuhe studies attention mechanisms, focusing on their geometry, approximation behaviour, and statistical efficiency. Research interests include adapting attention through post-training and fine-tuning, and applying these methods to quantitative finance, reinforcement learning, and sequential decision-making. [Projects](https://qrsntu.org/projects/) · [Publications](https://qrsntu.org/publications/) ## Projects - [AI Native Workflow](https://qrsntu.org/projects/ai-native-workflow/): Research on hierarchical, verification-gated and recoverable workflows for scaling long-horizon AI-assisted research and agent execution. Status: active. People: [Yuhe Sui](https://qrsntu.org/yh/), [Jianing Zhang](https://qrsntu.org/jn/), [Shufang Chen](https://qrsntu.org/people/shufang-chen/), [Danlin Chen](https://qrsntu.org/cdl/), [Wang Zimeng](https://qrsntu.org/people/wang-zimeng/), [Xiaotong Li](https://qrsntu.org/people/xiaotong-li/). - [FRTBTrace](https://qrsntu.org/projects/frtbtrace/): Active research on faithful, auditable, trade-level FRTB attribution. Status: active. People: [Yuhe Sui](https://qrsntu.org/yh/), [Yingzhi Tang](https://qrsntu.org/yz/). - [GeoAFT](https://qrsntu.org/projects/geoaft/): Geometric Attention Fine-Tuning: adapting the rule a pretrained Transformer uses to compare representations, not only the representations themselves. Status: active. People: [Yuhe Sui](https://qrsntu.org/yh/). - [GeoAttention](https://qrsntu.org/projects/geoattention/): Research on attention geometry, representation-relative identifiability, and geometry-adaptive prediction. Status: active. People: [Yuhe Sui](https://qrsntu.org/yh/), [Jianing Zhang](https://qrsntu.org/jn/), [Xiaotong Li](https://qrsntu.org/people/xiaotong-li/). - [PolicyAttention](https://qrsntu.org/projects/policyattention/): Research on softmax policy improvement, its robustness, and its failure modes. Status: active. People: [Yuhe Sui](https://qrsntu.org/yh/), [Shufang Chen](https://qrsntu.org/people/shufang-chen/), [Yingzhi Tang](https://qrsntu.org/yz/), [Xinyue Wu](https://qrsntu.org/people/xinyue-wu/). - [RiskHedge](https://qrsntu.org/projects/riskhedge/): Research on spectral-risk decisions, diverse scenarios, and selective robustness evaluation. Status: active. People: [Yuhe Sui](https://qrsntu.org/yh/), [Yingzhi Tang](https://qrsntu.org/yz/), [Jianing Zhang](https://qrsntu.org/jn/), [Mi Zeyuan](https://qrsntu.org/people/mi-zeyuan/). - [SoftRank](https://qrsntu.org/projects/softrank/): Theoretical research on softmax-attention approximation rank, support geometry, and robust interaction dimension. Status: active. People: [Yuhe Sui](https://qrsntu.org/yh/), [Jianing Zhang](https://qrsntu.org/jn/), [Yingzhi Tang](https://qrsntu.org/yz/). - [HedgeFidelity](https://qrsntu.org/projects/hedgefidelity/): An early-stage research programme with derived research in progress. Status: active. People: [Yuhe Sui](https://qrsntu.org/yh/). - [DecisionRank](https://qrsntu.org/projects/decisionrank/): An early-stage research programme with derived research in progress. Status: active. People: [Yuhe Sui](https://qrsntu.org/yh/). - [ntuinfo.com](https://qrsntu.org/projects/ntuinfo-com/): A bilingual English/Chinese full-stack web application with a quantitative, multi-criteria scoring framework (mental load, physical load, difficulty) that ranks 100+ NTU student clubs and surfaces algorithmically-driven recommendations for students. Status: active. - [AlphaAllocator](https://qrsntu.org/projects/alphaallocator/): A constrained Chinese financial-news overlay on classical risk-parity allocation, presented as a poster and oral at NLPCC 2026 Shared Task 4. Status: published. - [CFFA 2026 — China Future Financial Analyst Competition](https://qrsntu.org/projects/cffa-2026/): • Built a Greenwashing Risk Score across 1,676 Chinese green-bond issuances (2016–2025), using issue-spread regressions and portfolio sorts. • Pricing effects are weak unconditionally, but meaningful within credit-risk buckets after controls. Status: active. - [MathResearchOS](https://qrsntu.org/projects/mathresearchos/): A recoverable, audited command-line system for mathematical work: one objective becomes a staged, verifier-gated run across Codex, Claude and API models, with every attempt kept on disk. Status: active. People: [Yuhe Sui](https://qrsntu.org/yh/). ## Publications - **Diverse Scenarios, Selective Evaluation: Certified Spectral-Risk Decisions under Costly Robustness Tests** (2026). [Yuhe Sui](https://qrsntu.org/yh/), [Yingzhi Tang](https://qrsntu.org/yz/). EvoRobust Workshop, NeurIPS 2026. Accepted. [OpenReview](https://openreview.net/forum?id=OzOTmkkAWW) - **Geometric Attention Fine-Tuning Is Representation-Relative: Autonomous ML Research with AI-Native Workflow** (2026). [Yuhe Sui](https://qrsntu.org/yh/), [Jianing Zhang](https://qrsntu.org/jn/), [Shufang Chen](https://qrsntu.org/people/shufang-chen/), [Xiaotong Li](https://qrsntu.org/people/xiaotong-li/), [Wang Zimeng](https://qrsntu.org/people/wang-zimeng/). AutoMLR Workshop, NeurIPS 2026. Accepted. [OpenReview](https://openreview.net/forum?id=9PwlVQ3gyj) - **NTU-QRS AlphaAllocator: Bounded Transformer News Overlays for Chinese Macro-Asset Allocation** (2026). [Yuhe Sui](https://qrsntu.org/yh/). NLPCC 2026 Shared Task 4 · Poster + Oral. [OpenReview](https://openreview.net/forum?id=u09PzTYNHk) - **PolicyAttention: Robustness and Failure Modes of Softmax Policy Improvement** (2026). [Yuhe Sui](https://qrsntu.org/yh/), [Yingzhi Tang](https://qrsntu.org/yz/), [Xinyue Wu](https://qrsntu.org/people/xinyue-wu/). EvoRobust Workshop, NeurIPS 2026. Accepted. [OpenReview](https://openreview.net/forum?id=VgswuGrqp6) - **PolicyAttention: When Softmax Attention Becomes Policy Improvement** (2026). [Yuhe Sui](https://qrsntu.org/yh/). PTA Workshop, NeurIPS 2026. Accepted. [OpenReview](https://openreview.net/forum?id=zbGWnk0kY3) - **The Approximation Rank of Softmax Attention: Sharp Geometric Laws and Robust Interaction Dimension** (2026). [Yuhe Sui](https://qrsntu.org/yh/), [Jianing Zhang](https://qrsntu.org/jn/). NeurReps Workshop, NeurIPS 2026 (extended abstract). Accepted. [OpenReview](https://openreview.net/forum?id=0zgsZ8FjVz) [arXiv](https://arxiv.org/abs/2608.28150) ## Links - OpenReview: https://openreview.net/profile?id=~Yuhe_Sui1 - ORCID: https://orcid.org/0009-0002-7456-0804 - GitHub: https://github.com/yuhesui - LinkedIn: https://www.linkedin.com/in/yuhesui/ --- # Math Notes > Browse NTU mathematics and economics past-year papers, solutions, revision notes, and course resources. Web page: https://qrsntu.org/math_notes/ (content last changed 2026-11-12) Site index for agents: https://qrsntu.org/llms.txt Browse past year papers, solutions, and revision notes [Shortened here. The full page is https://qrsntu.org/math_notes/index.md; it goes on with: Courses.] --- # Post-Training, Explained > An illustrated explainer on LLM post-training: how supervised fine-tuning, reward models, RLHF with PPO, DPO and GRPO turn a pre-trained language model into an assistant. Web page: https://qrsntu.org/learn/post-training/ (content last changed 2026-11-11) Site index for agents: https://qrsntu.org/llms.txt QRS explainer · Large language models How a model that predicts the next word becomes an assistant that answers, follows instructions and reasons. Pre-training teaches a language model what text looks like. Post-training reshapes what it chooses to say: first by imitating good answers, then by learning which of its own answers people prefer, and most recently by rewarding answers that can be checked. [Read the guide](https://qrsntu.org/learn/post-training/guide/) The pipeline at a glance · [中文版指南](https://qrsntu.org/learn/post-training/zh/) A long read in seven parts, from the history to the derivations and the practice, with ten interactive figures. By **Yuhe Sui** · Last reviewed: 14 October 2026 Illustrative probabilities, not measurements from a real model. ## The pipeline at a glance The classic recipe runs left to right. The last column shows the two main departures from it. 1. Starting point ### Pre-training Predict the next token on a very large amount of text. Knowledge and skills, but no notion of answering. [History →](https://qrsntu.org/learn/post-training/guide/#history) 2. Imitate ### Supervised fine-tuning Train on prompt–response demonstrations with the same next-token loss, counted on the response only. [SFT →](https://qrsntu.org/learn/post-training/guide/#sft) 3. Compare ### Reward model Fit a scalar score to human comparisons with the Bradley–Terry model: the preferred reply should score higher. [Reward model →](https://qrsntu.org/learn/post-training/guide/#reward-model) 4. Optimise ### RL with PPO Sample replies, score them, and push the policy towards high reward while a KL penalty keeps it near the SFT model. [PPO →](https://qrsntu.org/learn/post-training/guide/#ppo) 5. Departures ### DPO Skip the reward model and the RL loop: a classification-style loss on preference pairs, derived from the same objective. ### RLVR and GRPO Replace learned rewards with checkable ones (a correct maths answer, passing tests) and the critic with a group of samples. [DPO →](https://qrsntu.org/learn/post-training/guide/#dpo)   [GRPO →](https://qrsntu.org/learn/post-training/guide/#grpo) ## What you will learn 1. [From task models to pre-train and adapt](https://qrsntu.org/learn/post-training/guide/#history) ULMFiT, GPT-1, BERT, GPT-2, T5 and GPT-3, and the question they leave open. 2. [Supervised fine-tuning](https://qrsntu.org/learn/post-training/guide/#sft) The objective, derived; chat templates and loss masks; why a little data goes a long way. 3. [Learning from preferences](https://qrsntu.org/learn/post-training/guide/#preference) Why comparisons, the Bradley–Terry reward model, and reward hacking. 4. [Reinforcement learning, from scratch](https://qrsntu.org/learn/post-training/guide/#rl-setup) MDPs, returns, policy gradients, value functions, Bellman equations, TD and GAE. 5. [RLHF in practice](https://qrsntu.org/learn/post-training/guide/#trpo-ppo) Trust regions, PPO’s clipped objective, the InstructGPT pipeline and its limits. 6. [Beyond PPO](https://qrsntu.org/learn/post-training/guide/#dpo) DPO, verifiable rewards, GRPO and DeepSeek-R1. 7. [In practice](https://qrsntu.org/learn/post-training/guide/#compare) The methods side by side, reported settings, memory costs, failure diagnosis, evaluation and open questions. An explainer, not a QRS research project. The toy figures are illustrative; every other claim cites its source in the guide’s [references](https://qrsntu.org/learn/post-training/guide/#references). --- # Post-Training: The Full Guide > The full guide to LLM post-training: history from ULMFiT to GPT-3, SFT, Bradley-Terry reward models, policy gradients, GAE, PPO, InstructGPT, DPO, RLVR, GRPO and DeepSeek-R1, with derivations and interactive figures. Web page: https://qrsntu.org/learn/post-training/guide/ (content last changed 2026-11-11) Site index for agents: https://qrsntu.org/llms.txt [← Post-training overview](https://qrsntu.org/learn/post-training/) QRS explainer · Post-training · The full guide The history first, then supervised fine-tuning, preferences, the reinforcement learning they rely on, the methods that came after PPO, and how all of this is run and measured in practice. Each idea comes with its intuition before its maths. By **Yuhe Sui** · Last reviewed: 14 October 2026 [阅读中文版 →](https://qrsntu.org/learn/post-training/zh/) Contents 1. From pre-training to post-training · about 10 min 2. One model per task 3. Pre-train, then adapt 4. Scale and prompting 5. Supervised fine-tuning · about 11 min 6. What SFT optimises 7. Data, templates and loss masks 8. Why a little data goes far 9. Strengths and limits 10. Learning from preferences · about 7 min 11. From demonstrations to comparisons 12. The Bradley–Terry reward model 13. Reinforcement learning · about 25 min 14. Agents, MDPs and returns 15. Policy gradients 16. Values and Bellman equations 17. Monte Carlo and TD learning 18. Baselines and advantages 19. Actor–critic and GAE 20. RLHF in practice · about 11 min 21. From TRPO to PPO 22. The InstructGPT pipeline 23. The limits of RLHF 24. Beyond PPO · about 13 min 25. Direct preference optimisation 26. Verifiable rewards and GRPO 27. DeepSeek-R1 and reasoning 28. In practice · about 10 min 29. Methods side by side 30. Running it: settings, cost, failures 31. How post-trained models are measured 32. Open questions 33. Summary 34. In brief 35. References Part 1 about 10 min**From pre-training to post-training** ## 1. One model per task Until the late 2010s, natural language processing mostly built one model per problem. A sentiment classifier, a question-answering system and a translation model each had their own architecture, their own labelled dataset and their own training run. If you had trained a good model and then needed to solve a different problem, you trained another model. What was shared was the bottom layer. Word embeddings such as word2vec (Mikolov et al., 2013) and GloVe (Pennington et al., 2014) learned one vector per word from unlabelled text, and task models used those vectors as input features. This was already transfer learning, but a shallow kind: each word got a single vector regardless of context, so *bank* in “river bank” and “bank account” looked the same, and everything above the embeddings still had to be learned from scratch for every task. Between 2015 and 2020 that changed. The field converged on a different recipe: train one large language model on a very large amount of unlabelled text, keep it, and adapt it to each task. That recipe, “train once broadly, adapt many times cheaply”, is the embryo of today’s split between **pre-training** and **post-training**. [Shortened here. The full page is https://qrsntu.org/learn/post-training/guide/index.md; it goes on with: Pre-train, then adapt; Scale and prompting; What SFT optimises; Data, templates and loss masks; Why a little data goes far; Strengths and limits; From demonstrations to comparisons; The Bradley–Terry reward model; Agents, MDPs and returns; Policy gradients; Values and Bellman equations; Monte Carlo and TD learning; Baselines and advantages; Actor–critic and GAE; From TRPO to PPO; The InstructGPT pipeline; The limits of RLHF; Direct preference optimisation; Verifiable rewards and GRPO; DeepSeek-R1 and reasoning; Methods side by side; Running it: settings, cost, failures; How post-trained models are measured; Open questions; In brief; References.] --- # News > News from QRS: research, releases, events and guides, with search and filters and an RSS feed. Web page: https://qrsntu.org/news/ (content last changed 2026-11-12) Site index for agents: https://qrsntu.org/llms.txt Research, releases, events and guides from QRS. [RSS feed](https://qrsntu.org/news/feed.xml) 11 Nov 2026 · Research ### [FRTBTrace: new project page](https://qrsntu.org/projects/frtbtrace/) FRTBTrace, our work on exact trade-level attribution of FRTB-IMA capital, now has its own project page: why one attribution vector cannot answer every question about capital, how one reverse pass through the full formula gives a ledger on (date, trade) positions, and the evidence from a pre-registered synthetic benchmark. Two FRTBTrace workshop papers were accepted at ICAIF 2026. Project: [FRTBTrace](https://qrsntu.org/projects/frtbtrace/) 11 Nov 2026 · Research ### [MathResearchOS: new project page and usage guide](https://qrsntu.org/projects/mathresearchos/) A command-line system that turns a mathematical objective into a recoverable, verifier-gated run across Codex and Claude models. The new project page has a quick start and a usage guide covering install, model choice, human gates, campaigns and troubleshooting. Project: [MathResearchOS](https://qrsntu.org/projects/mathresearchos/) 11 Nov 2026 · Release ### [NeurIPS 2026 Attendee Guide for Sydney, Atlanta and Paris](https://qrsntu.org/news/neurips-2026-attendee-guide/) A free, unofficial guide for everyone attending NeurIPS 2026 at any of its three sites: the full schedule, a guide to each city, a personal plan, hotels compared by price and commute, deadlines and templates for presenters, and the data for AI assistants. 11 Nov 2026 · Research ### [PolicyAttention: new project page](https://qrsntu.org/projects/policyattention/) PolicyAttention now has its own project page: why one softmax attention row can be the policy-improvement step of an in-context controller, the causal actor-environment-critic loop and its guarantee, where a stronger step backfires, and how learned Transformers compare, from our two NeurIPS 2026 workshop papers. Project: [PolicyAttention](https://qrsntu.org/projects/policyattention/) 11 Nov 2026 · Research ### [SoftRank: new project page](https://qrsntu.org/projects/softrank/) Our work on the approximation rank of softmax attention, accepted as an extended abstract at the NeurReps Workshop, NeurIPS 2026, now has its own project page: why query–key geometry depends on the coordinate frame, the softmax-visible interaction that does not, an explorer for the rank-two example, and certified rank brackets on trained heads. Project: [SoftRank](https://qrsntu.org/projects/softrank/) 5 Nov 2026 · Paper ### [NTU-QRS AlphaAllocator: Bounded Transformer News Overlays for Chinese Macro-Asset Allocation](https://openreview.net/forum?id=u09PzTYNHk) Presented at NLPCC 2026 Shared Task 4 as a poster and an oral (Evaluation Workshop I, 5 November 2026); first of ten teams in Track 1. Project: [AlphaAllocator](https://qrsntu.org/projects/alphaallocator/) 29 Oct 2026 · Release ### [Workshop Explorer: every AAAI 2027 workshop in one place](https://qrsntu.org/service/workshops/aaai-2027/) A new, independent QRS resource for researchers deciding where to submit. The Workshop Explorer lists the 45 AAAI-27 workshops by topic, with accepted paper types, page limits, deadlines, review and publication policies and official links, checked against the organisers' own pages. Filter by paper type, page limit, deadline or submission rules, shortlist and compare workshops, and add their deadlines to your calendar. 23 Oct 2026 · Announcement ### [QRS Research Academy: 2026 Fall #3 under way, 2026 Fall #4 recruiting](https://qrsntu.org/research-academy/#apply) 2026 Fall #3 is working towards AAAI. 2026 Fall #4 is recruiting for the December 2026 – January 2027 research cycle, aimed at relevant AAMAS 2027 workshops. Academy members join active rese [Shortened here. The full page is https://qrsntu.org/news/index.md.] --- # Events > Find QRS and QRS@NTU talks, workshops, competitions, and community events. Web page: https://qrsntu.org/events/ (content last changed 2026-11-07) Site index for agents: https://qrsntu.org/llms.txt Industry talks, external workshops, office visits, and industry-linked competitions. ## Events - [Global Market Making Games 2026 — Singapore Qualifier](https://qrsntu.org/gmmg/). 2026-10-16, 19:30 to 21:00 (Asia/Singapore). Singapore (venue to be announced). An individual global market-making competition open to university students across Singapore. QRS role: Singapore partner. Organisers: Amsterdam Investment Club. Partners: Trading Interview, Flow Traders (Super Day partner). - [IMC Tech Race @ NTU](https://builtin.com/job/imc-tech-race-singapore/10595673). 2026-09-02, 11:00 to 13:00 (Asia/Singapore). Nanyang Technological University. An on-campus coding race by IMC featuring 20 rapid computer-science problems, a live leaderboard, networking with IMC technologists, and prizes. Organisers: IMC Trading. - [Singapore vs Hong Kong Quant Trading Hackathon](https://luma.com/tqx5xvcy). 2026-02-23 to 2026-04-21. Singapore & Hong Kong / online. A cross-border university quantitative-trading competition in which Singapore and Hong Kong teams built and deployed live crypto trading strategies, with city qualifiers followed by a SG-vs-HK final. QRS role: Co-organizer. Organisers: HKU Data Science Association, NUS Investment Society, HKU Quantitative Research Society, Traders@UST, SMU FinTech Club, NTU Quantitative Research Society, CUHK Strategic Investment Society, SMU Quantitative Finance Society, CUHK Financial Technology Society, CUHK Quant Trading Society. Partners: IMC Trading, Optiver, Cubist Systematic Strategies (Point72), Gondor Capital. --- # Global Market Making Games 2026 Singapore > Compete individually in the Global Market Making Games 2026 Singapore qualifier on 16 October. Open to all university students in Singapore. Web page: https://qrsntu.org/gmmg/ (content last changed 2026-10-15) Site index for agents: https://qrsntu.org/llms.txt Singapore · 2026 Compete locally. Rank globally. An individual market-making competition open to university students across Singapore. Register [Official Games Website](https://www.amsterdaminvestmentclub.com/globalmarketmakinggames) ## Register Free, individual entry. No team required. Round 1 · Registration open ### Singapore Qualifier Everyone competes individually in one Singapore sitting. Your result decides who moves on to the Singapore Final. Date Friday 16 October 2026 Time 19:30–21:00 SGT Venue Singapore · Venue to be announced Entry Free · Individual Please complete both signups to register. [Singapore Signup](https://forms.cloud.microsoft/r/jJxpURLPbi) [Official Games Signup](https://www.amsterdaminvestmentclub.com/globalmarketmakinggames/s/ntu_quant_research_society) Eligibility ### Who can join? Open to university students across Singapore. Students from NTU, NUS, SMU, SUTD, SIT, SUSS and other universities in Singapore are welcome to register. You do not need to be an NTU student or a QRS member, and you do not need to form a team. ## From Singapore to the World Final Part of the official four-round format: society qualifier, campus final, final qualifier and world final. 01 · Qualifier ### Singapore Qualifier Everyone competes individually. Results count towards the ranking and decide who reaches the Singapore Final. Date Fri 16 Oct 2026 Time 19:30–21:00 SGT 02 · Campus final ### Singapore Final Leading qualifiers return for the Singapore Final, which decides the Singapore champion. Date Fri 13 Nov 2026 Time 16:00–19:00 SGT 03 · World final day ### World Final Society champions meet online. A final qualifier fills the ten seats at the final table, followed by the livestreamed World Final. Date Sat 12 Dec 2026 Where Online · Livestreamed ## What Will You Play? The Games test the kinds of quantitative reasoning used in market-making and trading interviews: pricing, probability, estimation, mental arithmetic, risk-taking and decision-making under pressure. - Make Me a Market - Market of Cards - Probability Betting - Fermi Questions Challenge - ETF Trading Challenge [Explore all games on the official website](https://www.amsterdaminvestmentclub.com/globalmarketmakinggames) ## FAQ Who can participate? The Singapore competition is open to university students across Singapore. You do not need to be an NTU student or an existing QRS member. Is this a team competition? No. Participants compete individually. You do not need to form a team. Do I need prior trading experience? The official Games materials do not list a prior-experience requirement. The Games focus on pricing, probability, estimation and mental arithmetic, and the official website shows each game in detail. How much does it cost? Entry is free. What will I play? The competition uses the Market Making Games platform from Trading Interview, with exercises in market making, probability, estimation, mental arithmetic and related quantitative decision-making. See the [official Games page](https://www.amsterdaminvestmentclub.com/globalmarketmakinggames) for full details. When is the Singapore qualifier? Friday 16 October 2026, 19:30–21:00 SGT. What happens if I qualify? Leading qualifiers progress to the Singapore Final on Friday 13 November. The Singapore champion then moves on to World Final day on 12 December, where a final qualifier fills the ten seats at the final table. Where will the event take place? Singapore. Venue details will be announced to registered participants once confirmed. What should I bring? Setup requirements have not been published yet. Registered participants will receive joining details, including anything to bring, once confirmed. Who organises the competition? The Global Market Making Games are hosted by Amsterdam Investment Club and powered by Trading Interview, with Flow Traders as Super Day partner. QRS@NTU is coordinating the Singapore participation as local partner. How do I register? Complete both steps: fill in the [Singapore Qualifier registration form](https://forms.cloud.microsoft/r/jJxpURLPbi), then also [sign up on Amsterdam Investment Club’s official Games page](https://www.amsterdaminvestmentclub.com/globalmarketmakinggames/s/ntu_quant_research_society). Registration is individual. ## Organisers Hosted by Powered by Super Day partner Singapore partner [Quantitative Research Society @ NTU](https://qrsntu.org/) [View all events](https://qrsntu.org/events/) --- # QRS@NTU Major Update > The December 2025 QRS@NTU update: website fixes to folders, search and ranking, and final examination papers for over 70 SPMS and NBS courses. Web page: https://qrsntu.org/news/dec2025-major-update/ (content last changed 2026-10-15) Site index for agents: https://qrsntu.org/llms.txt [← All news](https://qrsntu.org/news/) **Published:** December 10, 2025 **Maintained by:** QRS@NTU This is one of the most comprehensive upgrades to the QRS@NTU platform to date, featuring major improvements to the website infrastructure, a significantly expanded finals archive, and enhanced revision notes and solutions for multiple SPMS and NBS modules. # 1. Website Functionality Improvements A series of core structural issues on the platform have been fixed: ## ✔ Folder Button (Expandable Sections) — Fixed Previously, some folders failed to expand or load their contents. This has now been resolved, and all nested course directories open correctly without refresh errors. ## ✔ Search Function — Fully Restored Search now properly indexes every file across: - Finals archives - Revision notes - Problem sheets - Solutions (typed and handwritten) Users should experience significantly more reliable and accurate search results. ## ✔ Ranking Logic — Redesigned Courses are now automatically ranked based on **number of files available**, allowing students to quickly identify: - Modules with comprehensive resources - Newly updated courses - Courses requiring further expansion This upgrade also ensures consistency as new materials are added. # 2. Finals Archive Expansion (AY22/23 – AY24/25) Final examination papers have been added for **over 70 SPMS and NBS courses**, including multiple academic years for each module. Some of the major courses receiving full finals coverage include: ### **SPMS Mathematics** - MH1201, MH1301, MH1403 - MH2500, MH3100, MH3300, MH3500 - MH3700, MH3701 - MH4100, MH4110, MH4200, MH4300 - MH4311, MH4501, MH4511, MH4512, MH4518 - MH4600, MH4601, MH4700, MH4701 ### **NBS (Risk Management / Quantitative Finance)** - BR2207 - BR2210 - BR2211 - BR3213 - BR3208 Each module now contains at least one finals paper across the three academic years, with many having the full set from **AY22/23**, **AY23/24**, and **AY24/25**. # 3. Expanded Notes & New Releases ## 📘 MH1300 — Enhanced Revision Notes The MH1300 Foundations of Mathematics notes have been substantially expanded: - More step-by-step proofs - Additional examples - Highlighted standard methods (truth tables, modular arithmetic, induction, Euclid algorithm, etc.) - Improved visual structure for clarity - Added unofficial Mark Scheme ## 📘 MH4514 — New Revision Notes Complete revision notes for **Financial Mathematics** (MH4514) have been added: - Coverage of all major topics - Clean structure with definitions, examples, and key formulas - Consistent layout following QRS@NTU’s latest formatting standards ## 📘 MH5200 — New Revision Notes Full revision notes for **Advanced Investigations in Linear Algebra I** (MH5200) are now available: - Expanded explanations - Intuitive interpretations of abstract concepts - Additional theorems, examples, and diagrams - Polished LaTeX with unified formatting ## 📘 MH5200 Problem Sheet Solutions Enhanced - **At least two solution methods** for each problem (e.g., algebraic vs geometric, or direct vs spectral approach) - Clarified reasoning steps - More intermediate derivations - Added alternative insights, interpretations, and common pitfalls - Updated notation to match advanced linear algebra conventions This makes MH5200’s solution set one of the most comprehensive in QRS@NTU. # 4. Marking Schemes Included in New Solutions Where feasible, **mark distributions** are now included in: - Finals solutions - Problem sheet solutions - Selected revision notes This provides clearer expectations on exam-style writing and helps guide self-assessment. # 5. Updated QRS@NTU Formatting Standards All new materials uploaded after December 2025 follow a unified standard: - Clean LaTeX typography - Numbered sub-questions and structured headings - Final results boxed for visibility - Concise explanations with minimal redundancy - Consistent mathematical notation across courses These refinements improve readability and maintain professional quality across the entire library. # 6. Accessing the Updated Materials All updated files are available on: 👉 **https://qrsntu.org** Navigate to your course under the **Math Notes** section. For MH1300, MH4514, and MH5200, the latest notes and solutions are listed at the top of each module. # 8. Feedback & Future Plans We welcome suggestions from the NTU community. If you notice: - Missing files - Broken links - Suggested improvements - Courses you want prioritised Please reach out via the Telegram group or GitHub Issues. Upcoming planned updates: - Expansion of midterm archives - Typed solutions for earlier handwritten years - Additional cheat-sheets for selected modules - Automated metadata generation updates **Thank you for supporting QRS@NTU.** We hope these improvements help you revise more effectively and navigate your courses with confidence. --- # Quant & FinTech Events Guide > Quant and fintech events for students in January 2026: campus events, recruiting programmes, conferences and industry expos. Web page: https://qrsntu.org/news/events-guide-2026-jan/ (content last changed 2026-10-15) Site index for agents: https://qrsntu.org/llms.txt [← All news](https://qrsntu.org/news/) **Comprehensive Student Resource | Last Updated: January 22, 2026** ## 🗓️ Quick Reference | Date | Event | Location | Registration | ROI | |---|---|---|---|---| | **Jan 28** | 🏢 Jane Street @ NUS Career Fair | 🇸🇬 Singapore (NUS) | 📝 Sign up required | ⭐⭐ | | **Feb 6** | 💻 Optiver Tech Talk (Virtual) | 🌐 Zoom | 🔗 Open | ⭐⭐ | | **Feb 11** | 🧮 Jane Street Estimathon® | 🇸🇬 NTU, Singapore | **📝 Due Feb 9 EOD** | 🔥🔥🔥 | | **Feb 27** | 🎯 Citadel Discover | 🇸🇬 Singapore | **⏰ Due Feb 1** | 🔥🔥🔥 | | **Mar 20-21** | 🎯 Citadel Discover | 🇭🇰 Hong Kong | **⏰ Due Mar 1** | 🔥🔥🔥 | | **Mar 29-Apr 1** | 💰 Jane Street FTTP | 🇭🇰 Hong Kong | **⏰ Due Feb 15** | 🔥🔥🔥 | | **Jun 14-15** | 🏆 5th HK FinTech & AI Conf | 🇭🇰 CityU, Hong Kong | 📜 Papers due Mar 1 | ⭐⭐ | | **Jul 23-24** | 🌐 WikiEXPO Hong Kong | 🇭🇰 Hong Kong | 🆓 Open | ⭐⭐ | | **Nov 2-6** | 👑 HK FinTech Week 2026 | 🇭🇰 Hong Kong | 📋 Pre-reg open | 🔥🔥🔥 | ## 📋 SECTION 1: CAMPUS EVENTS ### Event 1.1: Jane Street @ NUS Computing Career Fair - **Organizer**: Jane Street + NUS School of Computing - **Date**: Tuesday, January 28, 2026 [careerfair.comp.nus.edu](https://careerfair.comp.nus.edu.sg) - **Time & Time Zone**: 12:00 PM – 5:00 PM SGT - **Location**: Physical | COM3 Multi-Purpose Hall & Foyer, National University of Singapore [careerfair.comp.nus.edu](https://careerfair.comp.nus.edu.sg) - **What It's About**: Campus fair booth with Jane Street recruiters. Opportunity to collect company materials, ask questions about roles, and network with representatives. Part of a larger computing career fair with 50+ companies. [careerfair.comp.nus.edu](https://careerfair.comp.nus.edu.sg) - **Target Audience**: All NUS computing students; especially those interested in quantitative trading and software engineering - **Ranking**: Medium ROI | Casual networking opportunity, not a structured recruitment event - **Registration Link**: https://app2.greenhouse.io/e/avbxcc - **Registration Deadline**: Not applicable - **Current Status**: ✅ Applications Open - **What to Prepare**: Resume (1 page), elevator pitch about your interests, questions about Jane Street's culture/roles - **One-Line Takeaway**: Low-barrier first contact with Jane Street recruiters; ideal for exploring without commitment ### Event 1.2: Jane Street Estimathon® @ Nanyang Technological University - **Organizer**: Jane Street - **Date**: Wednesday, February 11, 2026 - **Time & Time Zone**: 12:00 PM SGT (duration: 30 minutes competition + networking) - **Location**: Physical | Nanyang Technological University, Singapore (exact venue to be confirmed via email) - **What It's About**: Team-based estimation competition with 13 Fermi problems in 30 minutes. Teams create confidence intervals for questions ranging from trivia to advanced mathematical problems. Individual registration; teams formed on-site. - **Target Audience**: Undergraduates interested in quantitative reasoning, probability, and mental math; all majors welcome - **Ranking**: High ROI | Direct Jane Street exposure, tests core quant skills, highly competitive - **Registration Link**: https://docs.google.com/forms/d/e/1FAIpQLSe8QIfNhCJ7YpQxCvb8Z8QjDDTC8p9jYmOMR4-DIevLpUMuLg/viewform - **Registration Deadline**: End of day Monday, February 9, 2026 - **Current Status**: ✅ Applications Open | Confirmation sent by EOD Tuesday, February 10, 2026 - **What to Prepare**: Practice Fermi estimation problems (e.g., "How many piano tuners in Chicago?"), quick mental math, teamwork mindset. No calculators typically allowed. - **One-Line Takeaway**: Premier mental math competition hosted by top trading firm; winning/participating adds strong signal to quant applications [Shortened here. The full page is https://qrsntu.org/news/events-guide-2026-jan/index.md; it goes on with: 📋 SECTION 2: RECRUITING PROGRAMS; 📋 SECTION 3: ACADEMIC & RESEARCH CONFERENCES; 📋 SECTION 4: INDUSTRY EXPOS & NETWORKING EVENTS.] --- # Six works accepted to ICAIF 2026 workshops > Six works from QRS research programmes were accepted to ICAIF 2026 workshops, including the first first-author workshop papers of three 2026 Fall #2 Research Academy members. Web page: https://qrsntu.org/news/icaif-2026-workshops/ (content last changed 2026-10-24) Site index for agents: https://qrsntu.org/llms.txt [← All news](https://qrsntu.org/news/) **Published:** 23 October 2026 We are pleased to share that six works from QRS research programmes have been accepted to ICAIF 2026 workshops. Three of the six are the first first-author workshop papers of members of [QRS Research Academy](https://qrsntu.org/research-academy/), 2026 Fall #2: [Shufang Chen](https://qrsntu.org/people/shufang-chen/), [Xinyue Wu](https://qrsntu.org/people/xinyue-wu/) and [Annie Huaiyu Wang](https://qrsntu.org/people/annie-huaiyu-wang/). **Related projects:** [FRTBTrace](https://qrsntu.org/projects/frtbtrace/), [PolicyAttention](https://qrsntu.org/projects/policyattention/) and [RiskHedge](https://qrsntu.org/projects/riskhedge/). Titles, workshops and links for the individual works are not yet listed on this site. Further details on our research and programmes are available on the [Research Academy](https://qrsntu.org/research-academy/) page and in [Publications](https://qrsntu.org/publications/). --- # NeurIPS 2026 Attendee Guide for Sydney, Atlanta and Paris > The Quantitative Research Society, Singapore, has published a free, unofficial guide for NeurIPS 2026 in Sydney, Atlanta and Paris: schedule, city guides, hotels, presenting and data for AI assistants. Web page: https://qrsntu.org/news/neurips-2026-attendee-guide/ (content last changed 2026-11-11) Site index for agents: https://qrsntu.org/llms.txt [← All news](https://qrsntu.org/news/) **Published:** 11 November 2026 The Quantitative Research Society, Singapore, has published the [NeurIPS 2026 Attendee Guide](https://qrsntu.org/service/neurips26/), a free guide for everyone attending NeurIPS 2026 at any of its three sites: Sydney (6 to 12 December) and Atlanta and Paris (9 to 13 December). It is an unofficial guide. Each fact links to the official NeurIPS page it comes from, and a time, room or deadline that NeurIPS has not published is shown as not published rather than estimated. ## What the guide covers - **[Schedule](https://qrsntu.org/service/neurips26/schedule/):** every session, paper and event at the three sites, read from the official NeurIPS pages and updated every six hours, and every two hours during the conference. - **City guides for [Sydney](https://qrsntu.org/service/neurips26/guide/sydney/), [Atlanta](https://qrsntu.org/service/neurips26/guide/atlanta/) and [Paris](https://qrsntu.org/service/neurips26/guide/paris/):** the key facts at a glance, each day of the conference with its main sessions, times and rooms, getting to the venue, entry rules, food, and what is still to be published. - **[My plan](https://qrsntu.org/service/neurips26/itinerary/):** a personal schedule that shows clashes and can be added to a calendar, so changes to rooms and times reach your phone. - **[Hotels](https://qrsntu.org/service/neurips26/hotels/):** hotels in each city compared by price, door-to-door commute and breakfast, with a filter for the official NeurIPS hotels. - **[Presenting](https://qrsntu.org/service/neurips26/authors/):** deadlines for poster and oral presenters, printing in each city, and free LaTeX and PowerPoint templates for posters and talks. - **[Tickets](https://qrsntu.org/service/neurips26/tickets/), [visas](https://qrsntu.org/service/neurips26/visa/) and [questions and answers](https://qrsntu.org/service/neurips26/faq/)**, and a [news page](https://qrsntu.org/service/neurips26/news/) that lists official announcements and every change to the schedule. - **[For AI assistants](https://qrsntu.org/service/neurips26/agents/):** the schedule and the guide as data for AI assistants, through an API, an MCP server and plain-text files. The city guides and the Presenting pages also have prompts to copy, which ask an assistant to plan your days or help prepare a poster or talk. The guide remembers the city you choose, so the city guide and hotels open on your site. It works on a phone and keeps a copy for use offline. Open the guide at [qrsntu.org/service/neurips26](https://qrsntu.org/service/neurips26/). --- # Five works accepted to NeurIPS 2026 workshops > Five works from the QRS Research Academy's inaugural cohort were accepted to NeurIPS 2026 workshops. Web page: https://qrsntu.org/news/neurips-2026-workshops/ (content last changed 2026-10-15) Site index for agents: https://qrsntu.org/llms.txt [← All news](https://qrsntu.org/news/) **Published:** 4 October 2026 We are pleased to share that five of our works have been accepted to NeurIPS 2026 workshops. The accepted works span several of our current research directions, including the geometry of softmax attention, policy improvement, risk-sensitive decision-making, and AI-native research workflows. These results mark the successful completion of [QRS Research Academy](https://qrsntu.org/research-academy/), 2026 Fall #1, our inaugural cohort, in which members worked towards developing and submitting their first workshop papers. Contributors across the five accepted works include Yuhe S., Jianing Zhang, Yingzhi Tang, Shufang Chen, Xiaotong Li, Zimeng Wang, and Xinyue WU. We look forward to presenting these works at NeurIPS 2026 and to advancing the research directions they represent. Further details on our research and programmes are available at [qrsntu.org](https://qrsntu.org/). ## The accepted works - [Geometric Attention Fine-Tuning Is Representation-Relative: Autonomous ML Research with AI-Native Workflow](https://openreview.net/forum?id=9PwlVQ3gyj) AutoMLR Workshop, NeurIPS 2026 · [AI Native Workflow](https://qrsntu.org/projects/ai-native-workflow/) - [PolicyAttention: Robustness and Failure Modes of Softmax Policy Improvement](https://openreview.net/forum?id=VgswuGrqp6) EvoRobust Workshop, NeurIPS 2026 · [PolicyAttention](https://qrsntu.org/projects/policyattention/) - [PolicyAttention: When Softmax Attention Becomes Policy Improvement](https://openreview.net/forum?id=zbGWnk0kY3) PTA Workshop, NeurIPS 2026 · [PolicyAttention](https://qrsntu.org/projects/policyattention/) - [Diverse Scenarios, Selective Evaluation: Certified Spectral-Risk Decisions under Costly Robustness Tests](https://openreview.net/forum?id=OzOTmkkAWW) EvoRobust Workshop, NeurIPS 2026 · [RiskHedge](https://qrsntu.org/projects/riskhedge/) - [The Approximation Rank of Softmax Attention: Sharp Geometric Laws and Robust Interaction Dimension](https://openreview.net/forum?id=0zgsZ8FjVz) NeurReps Workshop, NeurIPS 2026 (extended abstract) · [SoftRank](https://qrsntu.org/projects/softrank/) --- # Service > QRS supports academic workshops and research communities with web, publicity, outreach, and organisational infrastructure. Web page: https://qrsntu.org/service/ (content last changed 2026-11-02) Site index for agents: https://qrsntu.org/llms.txt ## Workshop & Research Community Support QRS supports academic workshops, research communities, and student-led scientific initiatives with lightweight web, publicity, outreach, and organisational infrastructure. ## What We Can Help With Support may include: - Workshop website creation and maintenance - Calls for papers and submission-information pages - Speaker, programme, and accepted-paper pages - Publicity graphics and digital materials - Schedule and attendee-information pages - Outreach to relevant student and research communities - Lightweight coordination support where appropriate ## Workshop Explorer An independent QRS directory of conference workshops for researchers deciding where to submit. Search calls for papers by research area, paper type, page limit, deadline and publication policy, with submission requirements and official links for each workshop. Unconfirmed or conflicting details are marked as such. [AAAI 2027 Workshop Explorer](https://qrsntu.org/service/workshops/aaai-2027/) · [All conferences](https://qrsntu.org/service/workshops/) ## NeurIPS 2026 Planner An independent QRS planner for NeurIPS 2026 in Sydney, Atlanta and Paris: every session in venue time, search across authors, organisations and papers, a personal plan with clash warnings, hotels and a venue guide. Details the programme has not published are marked as such. [NeurIPS 2026 Attendee Guide](https://qrsntu.org/service/neurips26/) · [Schedule](https://qrsntu.org/service/neurips26/schedule/) ## Our Approach Our aim is to reduce the operational overhead of running research activities while helping organisers maintain a clear and professional public presence. Support is collaborative and scoped according to the needs of each workshop or research community. ## Work With Us Researchers, workshop organisers, and academic communities interested in working with QRS are welcome to contact us. [contact@qrsntu.org](mailto:contact@qrsntu.org) --- # Workshop Explorer > An independent QRS directory of conference workshops: search calls for papers by research area, paper type, page limit, deadline and publication policy. Web page: https://qrsntu.org/service/workshops/ (content last changed 2026-11-06) Site index for agents: https://qrsntu.org/llms.txt Workshop Explorer Independent QRS resource Conference workshops, searchable by what you want to submit. For each conference, the explorer lists every workshop with its research scope, accepted paper types, page limits, deadlines, publication policy and official links. Search and filter them, shortlist and compare candidates, and add their deadlines to your calendar. Details are checked against the official sources; anything unconfirmed or contradictory is labelled as such. Compiled independently by QRS. QRS is not affiliated with the conferences or workshop organisers listed here. Always confirm requirements on the official call for papers. - AAAI 2027 ## [AAAI 2027 Workshop Explorer](https://qrsntu.org/service/workshops/aaai-2027/) 45 named AAAI-27 workshops (Montréal, February 2027): submission categories, page limits, deadlines, publication policies and official links. [Open the explorer](https://qrsntu.org/service/workshops/aaai-2027/) --- # AAAI 2027 Workshop Explorer > Search and filter the AAAI 2027 workshops by research area, paper type, page limit, deadline and archival status, with submission rules and official CFP links. Web page: https://qrsntu.org/service/workshops/aaai-2027/ (content last changed 2026-11-12) Site index for agents: https://qrsntu.org/llms.txt [Workshop Explorer](https://qrsntu.org/service/workshops/) AAAI 2027 · Workshops · [Updates](https://qrsntu.org/service/workshops/aaai-2027/updates/) · About the data Independent QRS resource · AAAI 2027 Find the workshop that fits your paper. Search all 45 AAAI 2027 workshops by research area, paper type, page limit, deadline and publication policy. Open any workshop for its research scope, accepted contributions, submission requirements, key dates and official links. Anything a source does not state, or where official sources disagree, is marked as such. Workshops listed 45 named in the AAAI directory (AAAI's overview states 47) Default deadlines Submission 20 Nov 2026, notification 2 Dec 2026 (AAAI defaults; 45 workshops have confirmed their own submission date) Workshops held 22–23 Feb 2027, Montréal, Canada Submission sites 43 of 45 published Sources last checked 12 Nov 2026 ([what changed](https://qrsntu.org/service/workshops/aaai-2027/updates/)) [Shortened here. The full page is https://qrsntu.org/service/workshops/aaai-2027/index.md; it goes on with: Browse by topic; About this directory.] --- # AAAI 2027 Workshop Updates > What changed in the AAAI 2027 workshop data: moved deadlines, OpenReview due times that differ from the published deadline, and points where official sources disagree. Web page: https://qrsntu.org/service/workshops/aaai-2027/updates/ (content last changed 2026-11-12) Site index for agents: https://qrsntu.org/llms.txt [Workshop Explorer](https://qrsntu.org/service/workshops/) [AAAI 2027 workshops](https://qrsntu.org/service/workshops/aaai-2027/) · Updates · Deadline changes · OpenReview · Log Independent QRS resource · AAAI 2027 What has changed in the AAAI 2027 workshop data since it was first compiled: deadlines that moved, times and zones published later, OpenReview due times that differ from the published deadline, and points where official sources disagree. Checks 28 Oct, 1 Nov, 5 Nov, 6 Nov and 12 Nov 2026 Deadline changes 3: SAFE: Synthetic and Adversarial Forensics, Emerging AI Technologies for Music, EMC² OpenReview sites 33 checked; 11 due an hour or more before the published deadline Sources disagree 21 of 45 workshops Log entries 79 Compiled independently by QRS from the AAAI workshop directory, organisers' own pages and their submission sites. QRS is not affiliated with AAAI or the workshop organisers. Always confirm deadlines on the official call for papers before you submit. ## Deadline changes Submission deadlines that organisers have moved since the directory was first compiled. The explorer keeps the old date struck through beside the new one. | Workshop | Previous deadline | Now | Change | Recorded | |---|---|---|---|---| | [SAFE: Synthetic and Adversarial Forensics](https://qrsntu.org/service/workshops/aaai-2027/?w=SAFE) | 2 Nov 2026 23:59 AoE | 4 Nov 2026 23:59 ET | extended | 5 Nov 2026 | | [Emerging AI Technologies for Music](https://qrsntu.org/service/workshops/aaai-2027/?w=EAIM) | 6 Nov 2026 23:59 AoE | 13 Nov 2026 23:59 AoE | extended | 12 Nov 2026 | | [EMC²](https://qrsntu.org/service/workshops/aaai-2027/?w=EMC2-27) | 8 Jan 2027 23:59 PST | 25 Dec 2026 23:59 PST | moved earlier | 1 Nov 2026 | [Shortened here. The full page is https://qrsntu.org/service/workshops/aaai-2027/updates/index.md; it goes on with: OpenReview due times; Log of findings.] --- # Site Map > Find all public QRS pages, including programmes, events, research publications, mathematics notes, and member profiles. Web page: https://qrsntu.org/site-map/ (content last changed 2026-11-12) Site index for agents: https://qrsntu.org/llms.txt Explore our public pages and member profiles. - [About | QRS](https://qrsntu.org/about/) - [Using This Site with AI Assistants | QRS](https://qrsntu.org/agents/) - [Danlin Chen | QRS](https://qrsntu.org/cdl/) - [Contact | QRS](https://qrsntu.org/contact/) - [Events | QRS](https://qrsntu.org/events/) - [Global Market Making Games 2026 Singapore | QRS@NTU](https://qrsntu.org/gmmg/) - [Quantitative Research Society (QRS)](https://qrsntu.org/) - [Jianing Zhang | QRS](https://qrsntu.org/jn/) - [Join QRS | QRS](https://qrsntu.org/join_us/) - [Application Prerequisite Guidelines | QRS](https://qrsntu.org/join_us/prerequisites/) - [Post-Training: The Full Guide | QRS](https://qrsntu.org/learn/post-training/guide/) - [Post-Training, Explained | QRS](https://qrsntu.org/learn/post-training/) - [后训练：完整指南 | QRS](https://qrsntu.org/learn/post-training/zh/) - [BR2207 Quantitative Analysis: NTU Past-Year Papers | QRS](https://qrsntu.org/math_notes/br2207/) - [BR2210 Financial Markets and Products: NTU Past-Year Papers | QRS](https://qrsntu.org/math_notes/br2210/) - [BR2211 Financial and Risk Analytics I: NTU Past-Year Papers | QRS](https://qrsntu.org/math_notes/br2211/) - [BR3208 Operational Risk Management: NTU Past-Year Papers | QRS](https://qrsntu.org/math_notes/br3208/) - [BR3213 Valuation and Risk Models: NTU Past-Year Papers | QRS](https://qrsntu.org/math_notes/br3213/) - [HE1001 Microeconomics I: NTU Past-Year Papers and Solutions | QRS](https://qrsntu.org/math_notes/he1001/) - [HE1002 Macroeconomics I: NTU Notes and Solutions | QRS](https://qrsntu.org/math_notes/he1002/) - [HE2001 Microeconomics II: NTU Notes and Solutions | QRS](https://qrsntu.org/math_notes/he2001/) - [HE2002 Macroeconomics II: NTU Notes and Solutions | QRS](https://qrsntu.org/math_notes/he2002/) - [HE2003 Econometrics I: NTU Notes and Solutions | QRS](https://qrsntu.org/math_notes/he2003/) - [Math Notes | QRS](https://qrsntu.org/math_notes/) - [MH1100 Calculus I: NTU Past-Year Papers and Solutions | QRS](https://qrsntu.org/math_notes/mh1100/) - [MH1101 Calculus II: NTU Past-Year Papers and Solutions | QRS](https://qrsntu.org/math_notes/mh1101/) - [MH1200 Linear Algebra I: NTU Past-Year Papers and Solutions | QRS](https://qrsntu.org/math_notes/mh1200/) - [MH1201 Linear Algebra II: NTU Past-Year Papers and Solutions | QRS](https://qrsntu.org/math_notes/mh1201/) - [MH1300 Foundations of Mathematics: NTU Past-Year Papers and Solutions | QRS](https://qrsntu.org/math_notes/mh1300/) - [MH1301 Discrete Mathematics: NTU Past-Year Papers and Solutions | QRS](https://qrsntu.org/math_notes/mh1301/) - [MH1403 Algorithms and Computing: NTU Past-Year Papers | QRS](https://qrsntu.org/math_notes/mh1403/) - [MH2100 Calculus III: NTU Past-Year Papers and Solutions | QRS](https://qrsntu.org/math_notes/mh2100/) - [MH2220 Algebra I: NTU Past-Year Papers | QRS](https://qrsntu.org/math_notes/mh2220/) - [MH2500 Probability: NTU Past-Year Papers and Solutions | QRS](https://qrsntu.org/math_notes/mh2500/) - [MH3100 Real Analysis I: NTU Past-Year Papers and Solutions | QRS](https://qrsntu.org/math_notes/mh3100/) - [MH3110 Ordinary Differential Equations: NTU Past-Year Papers | QRS](https://qrsntu.org/math_notes/mh3110/) - [MH3200 Abstract Algebra I: NTU Past-Year Papers | QRS](https://qrsntu.org/math_notes/mh3200/) - [MH3210 Number Theory: NTU Past-Year Papers | QRS](https://qrsntu.org/math_notes/mh3210/) - [MH3220 Algebra II: NTU Past-Year Papers | QRS](https://qrsntu.org/math_notes/mh3220/) - [MH3300 Graph Theory: NTU Past-Year Papers | QRS](https://qrsntu.org/math_notes/mh3300/) - [MH3500 Statistics: NTU Past-Year Papers | QRS](https://qrsntu.org/math_notes/mh3500/) - [MH3510 Regression Analysis: NTU Past-Year Papers | QRS](https://q [Shortened here. The full page is https://qrsntu.org/site-map/index.md.] --- # Privacy > How the Quantitative Research Society handles personal data: the website, Research Academy applications and Google account data. Web page: https://qrsntu.org/privacy/ (content last changed 2026-11-12) Site index for agents: https://qrsntu.org/llms.txt How the Quantitative Research Society handles personal data. ## This website qrsntu.org sets no tracking cookies and runs no analytics or advertising scripts. Our hosting provider, Cloudflare, processes standard request data such as IP addresses to serve and protect the site. ## Research Academy applications When you apply to the QRS Research Academy we collect what you submit: your name, email addresses, university, programme and year of study, availability, the answers on the application form, your résumé, and, if you are invited, your Online Assessment answers and timings and your Research Assessment submission. The Online Assessment also records interactions with the assessment page, such as leaving the page or pasting text, to help us review attempts fairly. We use this information only to run the selection and to contact you about it. It is seen by the selection committee and is never sold or shared for marketing. We may use AI tools to help read and organise applications; selection decisions are made by the committee. Personal data is deleted once the selection it belongs to has concluded, apart from the records of members who join and of applicants who agree to our keeping their details so that we may contact them about future rounds. Applicants may ask us to delete these at any time. ## Google account data Our internal tools sign in with the Society's own Google account, with read-only access, to fetch application form responses and the résumés uploaded with them. Committee members may also sign in with Google to reach the committee site, which uses only their email address to confirm access. We do not access applicants' Google accounts. Our use of information received from Google APIs adheres to the [Google API Services User Data Policy](https://developers.google.com/terms/api-services-user-data-policy), including the Limited Use requirements. This data is used only for the selection described above and is not transferred to others or used for advertising. ## Your choices You may ask to see, correct or delete your data, or withdraw an application, at any time by writing to [contact@qrsntu.org](mailto:contact@qrsntu.org). Last updated 12 November 2026.