Yimin Liu

Model

Yimin Liu

Yimin Liu

Ph.D. Candidate Agent Infrastructure & Evaluation · AI for Science

The Ohio State University

yiminliu.career at gmail.com

About

What I build

I build the infrastructure that decides whether LLM agents actually work — the skill systems, harnesses, benchmarks, and reward-integrity tooling that turn “the agent seems good” into a number you can defend.

An eval stack fails in three places: the runtime, the tasks, and the score itself. At BenchFlow (Research Scientist Intern, summer 2026; maintainer since January) I built for all three. Its evaluation runtime executes containerized benchmark tasks across coding, terminal, and productivity domains. SkillsBench asks whether skills actually help — 87 tasks across 8 domains, where curated skills lift pass rates by 16.6 points — and ClawsBench (COLM 2026) puts capability and safety on the same axis instead of in separate benchmarks. For BenchShield, I hardened runtime evidence checks and built a resumable annotation runner.

In ScienceIDE with the AItonomy Foundation, I packaged Scirpy, a single-cell immune-receptor analysis codebase, as an agent-learnable environment with tasks that check scientific outputs.

My doctoral research focuses on deep learning for single-cell and spatial multi-omics in cancer — which is also where the hard, real scientific tasks in my evaluations come from.

  • Agent Skills & Benchmarks
  • Capability & Safety Evaluation
  • Reward Integrity
  • Single-Cell & Spatial Transcriptomics

Now

  • Deffairesearch engineer intern · FDA review agent
  • Ph.D. candidateThe Ohio State University · expected 2027
  • BenchFlowmaintainer · eval infrastructure
  • In flightAutoRSI · AgentFuzzBench · ScienceIDE

Selected work

  • SkillsBenchDo agent skills actually help? 87 tasks across 8 domains; curated skills lift pass rates by 16.6 points.
  • ClawsBenchFive simulated productivity services, 44 tasks — capability and safety on one axis. COLM 2026.
  • BenchShieldRuntime evidence integrity and reproducible annotation for agent evaluation. Now on arXiv.
  • ScienceIDEScientific codebases as agent-learnable environments. I packaged Scirpy and tasks that check scientific outputs.
  • BenchFlowThe evaluation environment and runtime executing 100+ containerized tasks.
  • Agent Skills '26Co-organizer of the first workshop on agent skills, at ACM CAIS 2026.

Research

Building the instruments

SkillsBench
Per-trial sandbox isolation and run monitoring across 18 model–harness configurations, plus paired evaluation across 87 tasks.
ClawsBench
44 tasks across five simulated productivity services, testing capability and safety on the same axis instead of in separate benchmarks. COLM 2026.
BenchShield
Runtime evidence checks, regression coverage, and a resumable annotation runner with per-call usage accounting. arXiv:2609.11028.
ScienceIDE
Packaged Scirpy, a single-cell immune-receptor analysis codebase, as an agent-learnable environment with tasks that check scientific outputs. arXiv:2609.19134.
Multi-omics
Deep learning over single-cell, spatial and knowledge-graph data in cancer biology — DG-scRNA, spatial transcriptomics, DeepDR.
The research page

Publications

Papers

SkillsBench
Benchmarking how well agent skills work across diverse tasks. Accepted to NeurIPS 2026 Evaluations and Datasets Track as a poster.
ClawsBench
Capability and safety of LLM productivity agents in simulated workspaces. COLM 2026.
BenchShield
Formal model-backed instrumentation for reward integrity in LLM-agent evaluation. arXiv:2609.11028.
ScienceIDE
Turning World’s Scientific Codebase into Agent Learnable Environments. arXiv:2609.19134.
Earlier
Back through DeepDR and DG-scRNA to a 2020 Nature Communications paper on O-GlcNAc chromatin proteomics.
All publications

News

Lately

2026

  1. BenchShield was featured in 机器之心 (Synced).

  2. October 8 — Ground Truth covered a follow-up study using SkillsBench to examine when the same agent skill helps one setup and hurts another.

  3. Glad to announce that SkillsBench has been accepted to the NeurIPS 2026 Evaluations and Datasets Track as a poster!

  4. 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.

  5. BenchShield is now on arXiv — formal model-backed instrumentation for reward integrity in LLM-agent evaluation.

  6. Joined Deffai — a full-stack FDA review agent and FDABench, its end-to-end evaluation.

  7. ClawsBench is accepted at COLM 2026, and I wrapped a summer as a Research Scientist Intern at BenchFlow — the eval environment, the eval runtime, and BenchShield, now under review.

  8. August 10 — Muse-Glimmer-30B was released by Meta; its model card reports SkillsBench results with skills.

  9. SkillsBench v4 is on arXiv — the fourth revision since the February release.

  10. Two papers accepted to Agent Skills '26, the first workshop on agent skills, at ACM CAIS 2026 — which I am also helping to organize.

The whole run

CV

The record

Now
Ph.D. candidate at The Ohio State University, advised by Professor Lijun Cheng. Expected 2027.
Before
B.S. in Bioinformatics, Dalian University of Technology, 2018 — 2022.
Also
Research Scientist Intern (summer 2026) and maintainer at BenchFlow; organizer of Agent Skills '26 at ACM CAIS; reviewer for ICLR, ICML, NeurIPS, EMNLP.
The file
Agent evaluation experience, skills and service, as a PDF you can read in the page and as a copy of it in text.
The full CV