Cyrus Zikai Zhou 周子恺 (zih-KAI joe)
Accountable AI · PhD Student, Stanford Computer Science
About
I build accountable AI systems whose decisions can be verified, challenged, and changed. Most recently, I led the Stanford OVAL Clinical Trial Matcher — an open-source system that matches patients to trials with accountable decisions. I am a Computer Science PhD student (2024–) at Stanford University, affiliated with the Stanford NLP Group. Previously, I received my undergraduate degree from the University of Chicago (2020–2024).
I'm actively looking for collaborators and research assistants to work on accountable/faithful AI and its applications to healthcare. Feel free to reach out if you're interested!
News
- Oct 2026 🏆 VERDICT received an Outstanding Paper Award at EMNLP 2026 (one of 23 out of 17,669 submissions)!
- Oct 2026 New preprint: CREDIT, counterfactual training that teaches language models to disclose when their answers depend on external information.
- Oct 2026 Chairing a virtual session at EMNLP 2026.
- Sep 2026 VERDICT accepted as Oral at EMNLP 2026 Main (to be presented virtually).
- Sep 2026 Starting as Head TA for CS224V: Agentic AI at Stanford.
- Jul 2026 SatIR accepted at COLM 2026.
Accountable AI
An AI system is accountable when its decisions can be examined, justified, and contested: we can see what decided it, check that its explanation is actually true, and know what would change the outcome. I pursue this along three routes: by construction, where language models translate rules into formal constraints and a solver decides (VERDICT, SatIR); by training, teaching models to report what actually changed their answers (CREDIT); and by measurement, scoring what models say about themselves against what happens when their inputs change.
Selected Publications
- Did the Model Use That? Counterfactual Training Boosts Self-Attribution in Language Models Preprint, 2026 · Code
- Accountable AI with Grounded, Faithful, Consistent, Actionable Rationales: A Case Study in Clinical Trial Matching with VERDICT EMNLP Main (Oral), 2026 · 🏆 Outstanding Paper Award
- SatIR: Scalable High-Recall Constraint-Satisfaction-Based Information Retrieval for Clinical Trials Matching COLM, 2026
- LowRA: Accurate and Efficient LoRA Fine-Tuning of LLMs under 2 Bits ICML, 2025
- If At First You Don't Succeed, Try, Try, Again...? Insights and LLM-informed Tooling for Detecting Retry Bugs in Software Systems SOSP, 2024
- Helping Users Debug Trigger-Action Programs UbiComp, 2023
Awards
- EMNLP 2026 Outstanding Paper Award
- Enlight Foundation Graduate Fellowship ($90K), Stanford University, 2024
- CRA Outstanding Undergraduate Researcher Award, Honorable Mention, 2023
- UChicago Quad Undergraduate Research Scholars (×3), 2022–2024
- Valedictorian, Ningbo Xiaoshi High School, 2020
- John Locke Institute Essay Competition, Psychology Prize Winner, 2019
Services
- Reviewer: ACL Rolling Review, NeurIPS
- Session Chair, EMNLP 2026 (virtual)
- UChicago Student Summer CS Research Fellowship Program, Student Lead, Summer 2023
- Kinzon Capital, Assistant to the General Partner, Summer 2021
- CreditEase GenZ Forum, Co-organizer, Mar – Sep 2021
Teaching
I am currently the head TA for CS224V: Agentic AI at Stanford (Fall 2026). I was also a project mentor for the same course in Fall 2025.
Experience
Early in my PhD, I got startup experience by interning at an AI infrastructure startup (2025) building memory processing technology for AI systems, where I worked on next-generation AI chip pathfinding and LLM inference cost modeling. Earlier, I worked at Moody's (2022) building data-management tooling, and at Vision Knight Capital (2021) on investment research.