Research projects

Research on AI safety, mechanistic interpretability of reasoning models, and robust misinformation detection, with methods, empirical findings, and code.

CoT is Not the Chain of Truth

·ICML·2026

Research area: AI safety and mechanistic interpretability

Which attention-routing components are associated with unsafe CoT before final refusal, and can targeted updates improve reasoning-model safety?

AI safety and mechanistic interpretability of reasoning LLMs through attention-head analysis and parameter-efficient safety fine-tuning.

Research noteAI Safety and Mechanistic Interpretability of Chain-of-Thought Reasoning

Research notes

Research perspectives on reasoning-trace safety and robust fake news detection.