I work on foundation models and agentic intelligence at ByteDance. My research focuses on agentic reinforcement learning, agent architectures, embodied intelligence, and physical intelligence.
I am especially interested in real-world, multi-turn agent tasks: how agents plan, use tools, adapt to feedback, and carry a task through to completion. I contribute to SkillsBench 1.9k+ and Harbor Adapters & Harbor-Index, built on the Harbor framework 6.8k+.
I completed my master’s research at VIPL, Institute of Computing Technology, Chinese Academy of Sciences, under the supervision of Prof. Dr. Qingming Huang (Professor at UCAS, IEEE Fellow) and Prof. Dr. Qianqian Xu (Professor at ICT, CAS). I received my bachelor’s degree in Computer Science and Technology from Harbin Institute of Technology.
I enjoy exchanging ideas and working with people across research and engineering. I’m especially excited about foundation models, agentic RL, embodied and physical intelligence, and agents that learn through real-world interaction. Whether you’re working in these areas or simply curious, I’d love to hear from you.
Say hello by email ↗
SkillsBench passed 1.9k stars; the Harbor framework passed 6.8k stars.
AcceptedSkillsBench was accepted to NeurIPS 2026, E&B Track!
AcceptedHarbor Adapters & Harbor-Index was accepted to NeurIPS 2026, E&B Track!
CoRT was featured on Hugging Face Daily Papers. Our counterfactual replay method studies token-level credit for rubric-guided RL.
SkillsBench was featured on Hugging Face Daily Papers, bringing reusable agent skills into systematic evaluation.
Joined ByteDance as an LLM Researcher.
EDGE has been accepted to AAAI 2025!
Our paper on AI-generated video quality assessment has been accepted to ICASSP 2025 as an oral presentation!
ADA-GAD has been accepted to AAAI 2024!
HGOE has been accepted to ACM MM 2024!
RHKH has been accepted to ACM MM 2024!
* Equal contribution.
A benchmark for testing whether procedural knowledge packaged as reusable skills translates into real agent capability. Curated tasks and deterministic verifiers compare agents with and without skills, separating skill quality from an agent’s ability to discover and apply them. Used in frontier-model evaluations.
Unified adapters for 80+ agentic benchmarks, with evaluation across 54 benchmarks. Harbor-Index curates 82 tasks from 29 benchmarks into a compact meta-dataset, making diverse agent evaluation easier to reproduce and compare.
Turning successful interaction trajectories into step-level credit for training LLM agents.
Protocol-level rubric aggregation mitigates reward hacking while preserving criterion coverage.
Specializing reward-specific teachers, then merging their capabilities through online policy distillation.
Counterfactual replay assigns token-level credit from rubric feedback without an auxiliary scorer.
Anomaly-denoised pretraining that reduces anomaly overfitting in graph autoencoders.
ByteDance
LLM algorithms · Agents & reinforcement learning
Alibaba Token Hub (ATH)
Research intern · Multimodal LLMs
Tencent Applied Research Center (ARC)
Research intern · Document understanding
M.S. in Computer Science and Technology
Visual Information Processing and Learning (VIPL)
Advised by Qingming Huang, IEEE Fellow
CyclingSolo rides around Hainan Island (1,000+ km) and Taihu Lake (400+ km).
Swimming
ByteDance. My research focuses on agentic reinforcement learning, agent architectures, embodied intelligence, and physical intelligence.
Hainan Island (1,000+ km) and 
Taihu Lake (400+ km).