Cross-Domain Limits of CoT Surface Features
A single-model measurement study showing that cheap reasoning-trace features are strong for math, partial for science, and weak for executable coding correctness.
Posts by Collection
portfolio
publications
Token-Level Verification under Controlled Evaluation
Published in ICML 2026 Workshop on AI for Math (AI4Math), 2026
A controlled audit of shallow token-level verification signals for LLM math reasoning. Accepted at the ICML 2026 Workshop on AI for Math (AI4Math).
talks
From Bandits to PPO: RL Prerequisite Seminar
Published:
A one-hour mathematical bridge from bandits to PPO…
Designed as preparation for GRPO/DAPO discussions.