Publications
🤖 LLM Reasoning
- [ICML 2026] SLAT: Segment-Level Adaptive Trimming for Efficient CoT Reasoning
This work studies how to reduce structural redundancy in chain-of-thought reasoning. It proposes segment-level adaptive trimming to selectively suppress low-utility redundant reasoning segments, improving the accuracy-efficiency trade-off for large reasoning models.
- [NeurIPS 2025 Spotlight] Diversity-Aware Policy Optimization for Large Language Model Reasoning
This work studies the relationship between solution diversity and reasoning potential in LLM reasoning, and proposes a diversity-aware policy optimization method for reinforcement learning training.
- [arXiv 2025] VAR-MATH: Probing True Mathematical Reasoning in LLMs via Symbolic Multi-Instance Benchmarks
This work introduces VAR-MATH, a symbolic multi-instance benchmark that probes whether LLM mathematical gains reflect genuine reasoning rather than benchmark-specific overfitting or memorization.
🎮 Diversity in RL
- [ICLR 2025] Diverse Policies Recovering via Pointwise Mutual Information Weighted Imitation Learning
This work focuses on recovering stylistically diverse policies from expert trajectories by weighting state-action pairs with pointwise mutual information.
- [NeurIPS 2023] Policy Space Diversity for Non-Transitive Games
This work studies diversity in non-transitive games and proposes a policy-space diversity measure that is more aligned with Nash approximation quality.
- [ICLR 2023] Quality-Similar Diversity via Population Based Reinforcement Learning
This work studies how to learn user-controllable and task-relevant diverse policy sets while maintaining similar policy quality.
🧩 Others (CV & GNN)
- [NeurIPS 2022 Spotlight] Self-supervised Amodal Video Object Segmentation
This work proposes a self-supervised video segmentation framework for inferring complete object shapes in occluded scenes using temporal information.
- [NeurIPS 2021] GRIN: Generative Relation and Intention Network for Multi-agent Trajectory Prediction
This work combines conditional generative modeling with graph neural networks to model agent intentions and social relations for multi-agent trajectory prediction.