Blair Yang

I'm Head of AI at Coolwei AI Lab. My research focuses on the evaluation, safety, and reasoning capabilities of large language models.

I am interested in understanding model behavior beyond aggregate benchmark scores, improving reasoning through reinforcement learning, and evaluating agents in realistic settings.

Previously, I studied computer science at the University of Toronto and worked with Roger Grosse, Jimmy Ba, and Ashton Anderson.

Publications

2026

SafeGEO: Understanding Generative Engine Optimization Risks in Recommendation Agents

Qianfeng Wen*, Yifan Simon Liu*, Xin Liu*, Difan Jiao, Blair Yang, Junda Wu, Zhenwei Tang

EMNLP 2026

Evaluating how seller-controlled web content can steer recommendation agents, and how to defend against it.

ThinkTwice: Jointly Optimizing Large Language Models for Reasoning and Self-Refinement

Difan Jiao*, Qianfeng Wen*, Blair Yang, Zhenwei Tang, Ashton Anderson

arXiv preprint, 2026

Training models to solve problems and revise their answers with a shared correctness reward.

Grounded Chess Reasoning in Language Models via Master Distillation

Zhenwei Tang, Qianfeng Wen, Seth Grief-Albert, Yahya Elgabra, Blair Yang, Honghua Dong, Ashton Anderson

arXiv preprint, 2026

Distilling expert chess reasoning into language models that explain their decisions in natural language.

OasisSimp: An Open-source Asian-English Sentence Simplification Dataset

Hannah Liu*, Muxin Tian*, Iqra Ali, Haonan Gao, Qiaoyiwen Wu, Blair Yang, Uthayasanker Thayasivam, En-Shiun Annie Lee, Pakawat Nakwijit, Surangika Ranathunga, Ravi Shekhar

LREC 2026 · Oral

A sentence simplification dataset spanning English, Sinhala, Tamil, Pashto, and Thai.

SWE-Bench Mobile: Can Large Language Model Agents Develop Industry-Level Mobile Applications?

Muxin Tian*, Zhe Wang*, Blair Yang, Zhenwei Tang, Kunlun Zhu, Honghua Dong, Hanchen Li, Xinni Xie, Guangjing Wang, Jiaxuan You

KDD 2026 · Applied Data Science Track

Evaluating coding agents on real feature-development tasks from a production mobile application.

2025

SEAM: Semantically Equivalent Across Modalities Benchmark for Vision-Language Models

Zhenwei Tang, Difan Jiao, Blair Yang, Ashton Anderson

COLM 2025

Testing whether vision-language models reason consistently when the same information is presented as text or images.

2024

Report Cards: Qualitative Evaluation of Language Models Using Natural Language Summaries

Blair Yang*, Fuyang Cui*, Keiran Paster, Jimmy Ba, Pashootan Vaezipoor, Silviu Pitis, Michael R. Zhang

NeurIPS 2024 SoLaR Workshop · Spotlight

Human-readable summaries of model behavior that complement numerical benchmarks.

* Equal contribution.

Coolwei AI Lab

At Coolwei AI Lab, we conduct research on AI safety, coding agents, reasoning, model evaluation, and Chinese-language NLP. Our work connects empirical studies of model behavior with the development of deployed language model systems.

I welcome enquiries about research collaborations in these areas. Please contact me at blair@cs.toronto.edu.

Earlier projects

Dreamitator: Dream and Imitate to Control

Charlotte Chen*, Scott Cui*, Ray Hung*, Blair Yang*

University of Toronto · CSC413 course project, 2024

Code