Hi, I'm Yiwei Yang.

PhD, University of Washington

LLM Agents • Tool Use • Post-training

Currently seeking full-time Research Scientist or Machine Learning Engineer positions.

I am especially interested in roles focused on LLM agents for coding and tool use, as well as post-training and agent evaluation.

Yiwei Yang

I earned my PhD in Information Science from the University of Washington in August 2026, advised by Bill Howe, and my B.S. in Computer Science and Engineering from the University of Michigan.

I study how frontier models fail in practice: I identify failure modes in real user-agent trajectories, turn them into benchmarks and RL environments, and fix them with post-training. My current focus is whether coding agents recognize ambiguous tasks and ask only the clarification questions worth asking. My prior work studied spurious tool use in RL-trained agents and the robustness of vision-language models.

I have also interned at LinkedIn Core AI, Adobe, and Sony AI.

Research Highlights

From real failure modes to benchmarks, RL environments, and post-training fixes.

Ongoing project · 2026

Benchmarking Coding Agents on Underspecified Tasks

I study whether coding agents recognize ambiguous requests and ask only the clarification questions worth asking. From 4,677 real coding-agent sessions, I built 171 sandboxed tasks with executable verifiers and a simulated user who reveals the missing specification only when asked, and I score each question by the cost of not asking.

18%

Runs where Fable 5.1 asked for clarification

Its decisions to ask were weakly calibrated: 60% of its questions were low-value (vs. an 80% base rate), and the final-state pass rate was 23%.

  • Coding agents
  • Benchmark construction
  • Agent evaluation
Project Paper in preparation

arXiv 2026

Reliable Tool Use in LLM Agents

I study why RL-trained agents make unnecessary tool calls and design rewards that improve tool-use decisions.

40.1% → 0.2%

Unnecessary search calls

With Tool Necessity Reward on GSM8K with spurious cues, while preserving accuracy.

  • Reinforcement learning
  • Reward design
  • Agent evaluation
Project Paper

NeurIPS 2025

SpuriVerse: Robustness in Vision-Language Models

I built a benchmark testing whether vision-language models generalize beyond misleading visual cues.

35.2% → 78.4%

Accuracy on held-out spurious patterns

After fine-tuning Qwen2.5-VL-7B on diverse synthetic examples.

  • Multimodal learning
  • Benchmarking
  • Fine-tuning
Project Paper

Selected Publications

Representative work on LLM agents, tool use, and robustness in vision-language models.

  • Spurious Tool Use: When RL Agents Learn the Wrong Reason to Act
    Yiwei Yang*, H. Zhang*, B. Wen, Y. Lu, Y. Wu, L. Zhang, J. McAuley, P. Lu, B. Howe
    arXiv preprint, 2026 — arXiv:2609.16268
  • Escaping the SpuriVerse: Can Large Vision-Language Models Generalize Beyond Seen Spurious Correlations?
    Yiwei Yang*, C. P. Lee*, S. Feng, D. Zhao, B. Wen, A. Z. Liu, Y. Tsvetkov, B. Howe
    ICML 2025 R2-FM, NeurIPS 2025 — arXiv:2506.18322
  • Label-Efficient Group Robustness via Out-of-Distribution Concept Curation
    Yiwei Yang, A. Liu, R. Wolfe, A. Caliskan, B. Howe
    CVPR 2024 — Paper
  • Towards Zero-shot Annotation of the Built Environment with Vision-Language Models
    B. Han, Yiwei Yang, A. Caspi, B. Howe
    SIGSPATIAL 2024 — arXiv:2408.00932
  • Laboratory-scale AI: Open-Weight Models are Competitive with ChatGPT Even in Low-Resource Settings
    R. Wolfe, I. Slaughter, B. Han, B. Wen, Yiwei Yang, et al.
    FAccT 2024 — arXiv:2405.16820
  • Regularizing Model Gradients with Concepts to Improve Robustness to Spurious Correlations
    Yiwei Yang, A. Liu, R. Wolfe, A. Caliskan, B. Howe
    ICML SCIS 2023 — Workshop listing

Project pages: Underspecified Coding Tasks  |  Spurious Tool Use  |  SpuriVerse  |  Concept Correction  |  All projects