Tencent
· 青云计划RL and OPD (multimodal agent): Working on post-training for visual agents: evidence diagnosis, trajectory reflection, targeted token-level supervision, and efficient tool use through ReVuE.
I am currently a Ph.D. student at the Institute of Computational Linguistics (计算语言学研究所), School of Computer Science, 
Peking University, advised by Prof. Houfeng Wang. My research focuses on 1) post-train: data-efficient and stable reinforcement learning across domains (RL, OPD, for agentic AI and reasoning); 2) data and eval: data synthesis research and reliable evaluation.
Before PKU, I received my B.Eng. degree in Artificial Intelligence from 
Beihang University in 2024.
I have industrial internship experience at Tencent (2026, 青云计划), 
Qwen Pilot, Alibaba · Tongyi Lab (2025), and 
PixVerse (2024), focusing on post-train, data synthesis and agentic ai & reasoning.
Industrial Experience
· 青云计划
Algorithm Research: Worked on RLVR for mathematical reasoning and instruction following: token-level distribution shifts, cross-task learnability, and capability trade-offs. Infra & RL Framework: Built experience-guided rollouts in VeRL for sample efficiency and training stability.
Worked on data synthesis and evaluation for video generation: VLM-based video captioning, caption quality analysis, and side-by-side annotation and evaluation workflows.
News
- Our new preprint on on-policy distillation (OPD) for multimodal agents, On-Policy Visual Evidence Distillation, is now available.
- 1 paper was accepted to the NeurIPS 2026 Position Track.
- Our new preprint on continual reinforcement learning for language models, Verifier-Induced Support Reshaping in On-Policy Optimization, is now available.
Earlier updates
- 3 papers were accepted by ICML 2026.
- 1 paper was accepted by ACL 2026 Main.
- 1 paper was accepted by ICLR 2026.
- TIME was accepted as a NeurIPS 2025 Spotlight in the D&B Track.
- 2 papers were accepted by ACL 2025 (Main and Findings).
- 1 paper was accepted by Science China Information Sciences.
- 1 paper was accepted by NAACL 2025 Main.
Open Source Projects & Academic Service
Open source. Co-contributor to BloodArena, an AI-powered Blood on the Clocktower environment for human–agent dialogue and LLM gameplay evaluation. Co-author of LeafyLingo, an interactive plant recognition system with a knowledge graph. Author of 24-Game-Reasoning, a minimal tutorial on LLM reasoning with Zero-RL, SFT, and SFT+RL.
Academic service. Reviewer for AAAI 2027, ICLR 2027, ICML 2026 (Gold Reviewer Award), and the EMNLP 2025 HCI-NLP Workshop.
Publications
16 publications and preprints · * Equal contribution
On-Policy Visual Evidence Distillation
Preprint (under review)
TL;DR ReVuE diagnoses where visual agents lose evidence and turns those reflections into targeted on-policy distillation, improving visual reasoning while reducing redundant reasoning and tool use.
Verifier-Induced Support Reshaping in On-Policy Optimization
Preprint (under review)
TL;DR RLVR can raise current-task scores while making successful behaviors for other tasks harder to sample, exposing a hidden trade-off between immediate gains and future learnability.
TimE: A Multi-level Benchmark for Temporal Reasoning of LLMs in Real-World Scenarios
The 39th Annual Conference on Neural Information Processing Systems (NeurIPS 2025) · Spotlight (Datasets and Benchmarks Track)
TL;DR TimE evaluates temporal reasoning across Wikipedia, news, and dialogue, testing how models handle dense timelines, changing events, and time-dependent social interactions.
SWE-Ext: Scaling and Extending Augmented Data for Mid-Training of Repository-Level Coding Tasks
The 42nd International Conference on Software Maintenance and Evolution (ICSME 2026)
TL;DR SWE-Ext turns GitHub pull requests into multilingual coding and completion data, improving repository-level mid-training and providing stronger initialization for downstream post-training.
Experience Augmented Policy Optimization for LLM Reasoning
The 43rd International Conference on Machine Learning (ICML 2026)
TL;DR EAPO reuses a trained policy's experience at critical token decisions, with importance correction to improve reasoning while reducing the inefficiency of learning from scratch.
Sparse but Critical: A Token-Level Analysis of Distributional Shifts in RLVR Fine-Tuning of LLMs
The 14th International Conference on Learning Representations (ICLR 2026)
TL;DR RLVR changes surprisingly few token decisions; controlled token swaps show that these sparse changes account for much of the observed reasoning improvement.
Probability-Entropy Calibration: An Elastic Indicator for Adaptive Fine-tuning
The 43rd International Conference on Machine Learning (ICML 2026)
TL;DR RankTuner combines token probability and uncertainty to focus fine-tuning on under-learned tokens, improving reasoning and transfer without overemphasizing inherently uncertain targets.
One-Way Policy Optimization for Self-Evolving LLMs
The 43rd International Conference on Machine Learning (ICML 2026)
TL;DR OWPO separates verifier-guided update direction from reference-based update strength, preserving improvements beyond the reference policy and supporting continued self-improvement.
Only Say What You Know: Calibration-Aware Generation for Long-Form Factuality
arXiv 2026
TL;DR Calibration-Aware Generation separates knowledge exploration from final answers, using reliability estimates to reduce hallucinations and improve factuality in long-form generation.
Two Pathways to Truthfulness: On the Intrinsic Encoding of LLM Hallucinations
The 64th Annual Meeting of the Association for Computational Linguistics (ACL 2026) · Main
TL;DR Truthfulness signals arise through distinct question- and answer-based pathways inside LLMs; disentangling them helps explain knowledge boundaries and improve hallucination detection.
Mitigating Overthinking through Reasoning Shaping
arXiv 2025
TL;DR GRSP regularizes reasoning at the step level, reducing excessive token use while retaining accuracy and improving reinforcement-learning stability.
MindScore: Quantifying Human Preference for Text-to-image Generation Through Multi-view Lens
Science China Information Sciences, 2025
TL;DR MindScore evaluates generated images through semantic matching, faithfulness, quality, and realism, bringing automatic evaluation closer to human preferences.
Odysseus Navigates the Sirens' Song: Dynamic Focus Decoding for Factual and Diverse Open-Ended Text Generation
The 63rd Annual Meeting of the Association for Computational Linguistics (ACL 2025) · Main
TL;DR Dynamic Focus Decoding uses differences between model layers to balance factuality and diversity during generation, without extra training data or auxiliary models.
Well Begun is Half Done: Low-resource Preference Alignment by Weak-to-Strong Decoding
Findings of the Association for Computational Linguistics (ACL 2025)
TL;DR A small aligned model drafts the opening, then a larger base model takes over, improving preference alignment while preserving downstream capabilities in evaluated settings.
MeNTi: Bridging Medical Calculator and LLM Agent with Nested Tool Calling
The 2025 Annual Conference of the Nations of the Americas Chapter of the Association for Computational Linguistics (NAACL 2025) · Main
TL;DR MeNTi uses nested tool calls to select medical calculators, fill inputs, and convert units, with CalcQA providing a physician-built benchmark for evaluating these abilities.
CiteCheck: Towards Accurate Citation Faithfulness Detection
arXiv 2025
TL;DR CiteCheck combines cost-conscious annotation with synthetic negative examples to build Chinese citation-faithfulness data and train smaller, effective evaluators for retrieval-augmented generation.