LLM Agent Researcher
Tencent TEG Hunyuan · Tencent 2026 Project UP
Ph.D. candidate at East China Normal University · Shanghai, China
I am a fourth-year Ph.D. candidate advised by Prof. Yanghua Xiao. I also work with Dr. Jiaqing Liang at the Knowledge Works Lab, Fudan University.
My research focuses on language agents, reasoning models, and LLM post-training, with an emphasis on self-evolution, tool use, reasoning efficiency, and reinforcement learning.
Selected for Tencent 2026 Project UP and the ByteDance Soaring Star Talent Program.
I am currently exploring job opportunities in both industry and academia. Feel free to reach out by email.
Tencent TEG Hunyuan · Tencent 2026 Project UP
ByteDance · Soaring Star Talent Program
Knowledge Works Lab · Fudan University
Recent updates
Excited to release SKILL-USE 🚀, a benchmark that asks whether LLM agents can recognize and correctly apply skills.
Joined Tencent TEG Hunyuan as a research intern 🤖, working on LLM agent evaluation.
Two papers accepted at ACL 2026 and one at ICML 2026 🎉.
GenericAgent is open-sourced and has passed 13K stars on GitHub ⭐
Attended ICLR 2026 in Rio de Janeiro, Brazil.
Three papers accepted at ICLR 2026, a good start to the year.
CDS accepted at ACL 2025.
Released a compact GRPO implementation, now at 1.6K+ stars ⭐
First-author and co-authored papers
* Equal contribution. Only selected work is shown. See the full publication list.
Key contribution. Introduces SKILL-USE, a benchmark of 79 real skills and 177 executable tasks that evaluates Trigger, Compliance, and Boundary under progressive disclosure. PaperCode
Key contribution. Synthesizes plausible but ultimately incorrect reasoning traces as higher-quality negative data for preference optimization. Paper
Key contribution. Decouples reasoning into generation, verification, and revision, with targeted supervision that improves self-verification and reduces redundant reasoning. Paper
Key contribution. Proposes JET, which uses truncated rollouts and a quality-controlled length reward to teach reasoning models when to stop without sacrificing accuracy. PaperCode
Key contribution. Proposes PASR, enabling a model to decide whether, when, and how to refine its output during generation rather than repeatedly rewriting the full response. PaperCode
Key contribution. Introduces FineCE for continuous, fine-grained confidence estimation during text generation, including backward confidence integration at inference time. Paper
Key contribution. Builds intrinsic self-correction data with Partial Answer Masking, enabling small language models to trigger and perform correction without complex prompts. Paper
Key contribution. Introduces NoiseBench, a paired evaluation framework spanning four agent-input sources and three noise types. Experiments on APEX-Agents and Toolathlon reveal unverified error propagation and increased computational costs under noise. Paper
Key contribution. Maximizes decision-relevant information within a fixed context through compact tools, hierarchical memory, experience-to-skill evolution, and context compression. PaperCode
Key contribution. Uses abstract meta-hints and affinity-aware policy optimization to guide difficult RL rollouts while avoiding answer leakage and unstable off-policy updates. Paper
Key contribution. Proposes MENTOR, which injects expert guidance only at critical decision points to preserve both effective and diverse exploration in RLVR. Paper
Key contribution. Introduces DUCL, a curriculum framework that jointly scores sample difficulty and utility and schedules data with window ordering. Paper
Key contribution. Proposes Hindsight Policy Optimization, using Wasserstein distance in an intent space to derive low-variance learning signals for long-horizon agent training. Paper
Key contribution. Decouples efficiency and correctness rewards through a fast/slow mode-selection token, preserving reasoning quality while enabling continuous inference-time control of the efficiency–performance trade-off. Paper
Key contribution. Collects mistake-relevant knowledge through iterative evaluation and combines continual instruction tuning with pre-training to improve data efficiency and mitigate forgetting. Paper
Key contribution. Applies cognitive diagnosis to identify fine-grained model weaknesses and synthesize targeted training data for those knowledge points. Paper
Key contribution. Proposes progressive thought refinement with collaborative data selection and thought-mask fine-tuning to improve responses across diverse tasks. Paper