My recent work focuses on LLM agents that can reason over multi-step tasks, use tools, interact with task environments, and improve through evaluation, data feedback, and reinforcement learning.
ByteDanceLLM algorithms
Alibaba ATHInternmultimodal LLMs
Tencent ARCInterndocument AI
News
SkillsBench became the fastest benchmark repository to reach 1k GitHub stars, with 1.1k stars within two months of release
SkillsBench appeared in recent model-card and release discussions, including Qwen 3.6 Plus and HY-3
SkillsBench now covers 86 tasks, 11 domains, 7,308 trajectories, and 40 indexed benchmarks
Harbor v0.6.5 was released for agent evaluation, task environments, and RL-ready rollout workflows
Started working on LLM algorithms at ByteDance
Selected Publications
arXiv 2026 · Agent Skills · 1.1k stars
A benchmark for evaluating whether agent skills actually work across diverse tasks, separating skill quality from an agent’s ability to discover and use the right skill.
AAAI 2024 · CCF A · 99 citations
A two-stage graph anomaly detection framework that pretrains graph autoencoders on anomaly-denoised graphs at node, edge, and subgraph levels, then retrains the decoder on the original graph to mitigate anomaly overfitting and homophily traps.
ACM MM 2024 · CCF A
Studies graph out-of-distribution detection with hybrid synthetic and internal outlier exposure, improving robustness when abnormal patterns are scarce or shift across graphs.
ACM MM 2024 · CCF A
Models n-ary relational facts as hypergraphs for knowledge reasoning, moving beyond binary triples and capturing richer relational structures.
Open Source
A benchmark for testing whether agent skills actually work across tasks, models, and environments.
An evaluation and environment framework for agent tasks, rollouts, and RL-oriented data generation.
Stars
Visitor Map
This map starts collecting public, approximate visitor regions after it is loaded on the live site.