Our paper “Mechanistic Data Attribution: Tracing the Training Origins of Interpretable LLM Units” was selected for Oral Presentation at ICML 2026. Only 168 submissions were selected for orals this year (out of 23,918 total submissions, award rate: 0.7%).

I received the Outstanding Area Chair Award at EACL 2026 (Top 24 out of 1000 ACs, Award Rate: 2.4%).

We have 3 papers accepted by ACL 2026. Topics involve intrinsic interpretability, RL mechanisms, and reverse LLMs. We have 2 papers accepted by ICML 2026. Topics involve mechanistic data attribution and logical reasoning. One paper was selected as Spotlight Paper (Top 2.2%).

Our paper “MuSLR: Multimodal Symbolic Logical Reasoning” was awarded the Best Paper Award (Top 1 out of 75; Award Rate: 1.3%) at the AAAI 2026 Bridge Program on Logical and Symbolic Reasoning in Language Models, held in Singapore.

I am thrilled to annonce that I will be joining the School of Computer Science, Peking University as an Assistant Professor, starting from Sep 2025.

We have 2 papers accepted by EMNLP 2025. One paper is on understanding how LLMs are distracted by irrelevant context in reasoning. (Main Conference, Oral). The other paper is making the LLM reasoning more concise with RL (Findings).

We have 6 papers accepted by ACL 2025 (6 Main Conference). Topics involves logical reasoning, rule following, data contamination, uncertainty estimation, and automated LLM agents.

We have 3 papers accepted by NAACL 2025 (1 Main Conference, 2 Findings). Topics involves causal reasoning, code generation, and fact-checking of LLMs.

Our paper TART: An Open-Source Tool-Augmented Framework for Explainable Table-based Reasoning was awarded Best Paper Runner-up (Top 3 out of 59; Award Rate: 5.1%) at the 3th Table Representation Learning Workshop @ NeurIPS 2024.

5 papers accepted by EMNLP 2024 (2 Main Conference, 3 Findings). Topics involves logical reasoning, automated fact-checking, multi-agent reasoning, and knowledge editing. Our paper MMLongBench-Doc was accepted as spotlight paper in NeurIPS 2024 (Dataset and Benchmark).