Yimin Liu

Experience

• Research Scientist Intern & Open-Source Contributor Jan 2026 — present

  BenchFlow · San Francisco, CA

  Research Scientist Intern: May — Aug 2026

  • Evaluation infrastructure — built the BenchFlow evaluation environment and runtime powering benchmark suites including SkillsBench, AgentFuzzBench, and FrontierPhysics (100+ containerized tasks) across coding, terminal, and productivity domains.
  • BenchGuard (under review, USENIX Security 2027) — designed a model-backed instrumentation layer combining phase-aware static taint analysis with runtime infrastructure-side evidence to detect reward hacking; built BenchGuard Trajectories, a human-labeled corpus of 456 adjudicated trajectories from 31K+ public agent runs, reaching 96% detection accuracy and lifting full-chain recall from 23–94% to 77–100%.
  • ClawsBench (COLM 2026) — built a benchmark simulating five productivity services (Gmail, Slack, Google Calendar, Google Docs, Google Drive) with 44 tasks testing both capability and safety of LLM agents.
  • SkillsBench (87 tasks / 8 domains / 18 model–harness configs; under review, NeurIPS 2026) — implemented the native skill system for Terminus2, where curated skills lift mean pass rate from 33.9% to 50.5% (+16.6 pp); developed the mHC layer task, a credentialed A100 GPU task implementing manifold-constrained hyper-connections in nanoGPT bundling 3 reusable skills; built the analysis pipeline from raw trajectories to paper figures and led code review.
  • FrontierPhysics (ongoing) — building the contribution review pipeline and task infrastructure (Docker environments, automated verifiers, rubric-based LLM-as-judge reviewer agents) for an open benchmark of PhD-level physics research tasks.
• Independent Research Collaborator Jun 2026 — present

  AutoRSI & AgentFuzzBench · Remote

  Collaborators: Yuxuan Jiang (AutoRSI); Security Lab, UC Davis (AgentFuzzBench)

  • AutoRSI (manuscript in preparation) — a general framework that automatically strengthens a naive base method across domains via iterative propose–implement–evaluate–reflect cycles; building the RSI loop pipeline and the automatic construction and refinement of the improvement-plan pool.
  • AgentFuzzBench — a benchmark for end-to-end agentic fuzzing (build, harnessing, fuzzing, crash triage); built the uniform task-package pipeline across 8 exact-pinned OSS targets and packaged it as an externally hosted BenchFlow benchmark with committed AFL++ oracle evidence.
• Research Collaborator Sep 2025 — Nov 2025

  Wuhan University of Science and Technology · Wuhan, CN (remote)

  Collaborator: Shuting Jin, School of Computer Science and Technology

  • DeepDR — built the drug-repositioning web platform: 15+ deep learning models over a knowledge graph of 5.9M edges and 107 relationship types, with visualizations explaining why each drug is recommended.
• Graduate Research Assistant Mar 2023 — present

  The Ohio State University · Columbus, OH

  Advisor: Professor Lijun Cheng

  • Developed DG-scRNA, a graph-convolutional network annotating cell types from scRNA-seq data — 97.2% accuracy on 150K+ cells (first-author manuscript).
  • Integrated scRNA-seq and spatial transcriptomics across 25 thyroid cancer samples; identified 50+ ligand–receptor interactions and two novel cell subtypes.
  • Benchmarked CNN, RNN, and Transformer architectures for CRISPR gRNA on-target prediction over 50K+ curated sequences.
• Research Assistant Aug 2021 — Oct 2021

  Harvard Medical School · Boston, MA

  Advisor: Martin Hemberg

  • Built an automated mutation-calling pipeline (GATK, R) over 50K+ cells from public scRNA-seq datasets; clustered on SNV information and expression matrices separately to predict variant–phenotype relationships in cancer.
• Research Assistant Mar 2021 — Aug 2021

  Beijing Institute of Genomics, Chinese Academy of Sciences · Beijing, CN

  Advisor: Mingkun Li

  • Built a pipeline to optimize COVID-19 hybrid capture probes; located low-efficiency regions and redesigned them, improving sensitivity by 25%.
• Research Assistant Nov 2019 — Feb 2021

  Dalian University of Technology · Dalian, CN

  Advisor: Yubo Liu

  • Undergraduate research on O-GlcNAc chromatin regulation in breast cancer (proteomics, ChIP-seq, RNA-seq); co-authored Nature Communications 2020, Cell Biology International 2021, and BBA — General Subjects 2021.