Publications

Towards a Mechanistic Understanding of Propositional Logical Reasoning in Large Language Models

Danchun Chen^*, Qiyao Yan^*, Liangming Pan

arXiv preprint 2026

• Mechanistic analysis on four models (Qwen3-8B/14B, Llama-3.1-8B-Instruct, Mistral-7B-Instruct) using PropLogic-MI across 11 rule categories and one-/two-hop reasoning.
• Uncovered four interlocking mechanisms: Staged Computation, Information Transmission, Fact Retrospection, and Specialized Attention Heads.
• Prompt-order control analysis confirms the staged organization is model-internal.

[Paper (earlier version)] [Code]

SWRV: Empowering Self-Verification of Small Language Models through Step-wise Reasoning and Verification

Danchun Chen, Yixin Cao, Liangming Pan

XLLM @ ACL 2025 · Certificate of 2nd Place Award

[Paper] [Code]