News
Lately
Log
-
BenchShield was featured in 机器之心 (Synced).
-
October 8 — Ground Truth covered a follow-up study using SkillsBench to examine when the same agent skill helps one setup and hurts another.
-
Glad to announce that SkillsBench has been accepted to the NeurIPS 2026 Evaluations and Datasets Track as a poster!
-
ScienceIDE is on arXiv — turning pinned scientific codebases into agent-learnable environments. I packaged the bioinformatics environment (Scirpy, single-cell immune-receptor repertoire analysis) and its tasks.
-
BenchShield is now on arXiv — formal model-backed instrumentation for reward integrity in LLM-agent evaluation.
-
Joined Deffai as a Research Engineer Intern — building a full-stack FDA review agent and FDABench, an end-to-end evaluation of its review workflow.
-
ClawsBench is accepted at COLM 2026, and BenchShield — model-backed instrumentation for reward integrity in agent evaluation — is under review. That closes out a summer as a Research Scientist Intern at BenchFlow.
-
SkillsMetric is on arXiv — mapping where static analysis stops being able to tell a useful agent skill from a hostile one.
-
August 10 — Muse-Glimmer-30B was released by Meta; its model card reports SkillsBench results with skills.
-
July 6 — Tencent Hy3 was released; its model-card benchmark appendix includes SkillsBench evaluations.
-
SkillsBench v4 lands on arXiv — the fourth revision since the February release.
-
June 30 — openPangu-2.0-Flash included SkillsBench results in its English model card uploaded on this date.
-
June 25 — GitHub Blog included SkillsBench in its evaluation of the GitHub Copilot agentic harness across models and tasks.
-
Two papers accepted to Agent Skills '26, the first workshop on agent skills, at ACM CAIS 2026 — which I am also helping to organize.
-
Started the summer as a Research Scientist Intern at BenchFlow in San Francisco — the eval environment, the evaluation runtime, and the benchmarks that run on it.
-
May 19 — Qwen3.7-Max included SkillsBench in its official API-model release evaluation.
-
ClawsBench preprint out: 44 tasks across five simulated productivity services, testing capability and safety on the same axis.
-
April 21 — Qwen3.6-27B was released with SkillsBench results in its model card.
-
April 17 — Qwen3.6-Max-Preview reported SkillsBench improvements in its official API-model preview announcement.
-
March 18 — OpenHands discussed SkillsBench in “How to Evaluate Agent Skills (And Why You Should).”
-
SkillsBench v1 out on arXiv (v2 and v3 follow in March). I built the agent sandbox environment and designed the statistical evaluation behind the results.
-
February 26 — Computerworld covered SkillsBench and human guidance in its 2-Minute Tech Briefing podcast.
-
February 23 — Agentman discussed SkillsBench and the performance gains from curated agent skills.
-
February 20 — InfoWorld covered SkillsBench in “AI agents still need humans to teach them.”
-
February 19 — The Register covered SkillsBench in “AI agents can’t teach themselves new tricks – only people can.”
-
February 18 — HackerNoon covered SkillsBench and the usefulness of reusable agent skills.
-
February 16 — Hacker News discussed SkillsBench: benchmarking how well agent skills work across diverse tasks.
-
Joined BenchFlow as an open-source contributor and maintainer.
-
DeepDR preprint out with Shuting Jin's group at WUST — 15+ deep learning models over a knowledge graph of 5.9M edges.
Education
Education
-
The Ohio State University Aug 2022 — present
Ph.D. Candidate · Columbus, OH
Advisor: Professor Lijun Cheng
-
Dalian University of Technology Sep 2018 — Jun 2022
B.S. in Bioinformatics · Dalian, CN