README.md

metadata

license: cc-by-4.0
pretty_name: Who&When Pro
language:
  - en
tags:
  - agents
  - multi-agent
  - failure-attribution
  - llm-evaluation
  - benchmark
size_categories:
  - 10K<n<100K
configs:
  - config_name: default
    data_files:
      - split: text
        path: data/text.jsonl
      - split: image
        path: data/image.jsonl
      - split: image_gui
        path: data/image_gui.jsonl
      - split: video
        path: data/video.jsonl

Who&When Pro

Who&When Pro is a failure attribution benchmark for LLM agent systems. Each trace is an agent trajectory, single or multi agent, with one realistic error injected at a known step. The label records who made the error, when it happened, and what kind of error it was. Given a trajectory, the evaluated model has to recover all three.

Row schema

Each row is one trace:

column	meaning
id	unique trace id
framework	agent framework of the trajectory
benchmark	source benchmark
task	the original task: query, reference answer, inputs
trajectory	the full agent rollout
ground_truth	the label: agent, step, and error mode
extras	framework specific metadata

task, trajectory, ground_truth, and extras are JSON strings, so json.loads them to consume. Image and image_gui traces embed their images in the row; video traces reference frame files under video_assets/.

Splits

• text, image, video — tool-use, chart/search, and video agent traces across single- and multi-agent frameworks.
• image_gui — GUI agent traces with full screenshot observations: coact (OSWorld, multi-agent) and openai_cua (OSWorld), plus agentoccam and gemini (WebVoyager). ground_truth.step is 1-indexed and matches the trajectory's step_number directly.

Taxonomy

17 error modes in 6 categories, defined in taxonomy.yaml:

• Perception: visual misidentification, spatial grounding
• Reasoning: hallucination, reasoning error, numerical error, task misunderstanding
• Planning: ineffective planning, goal drift
• Action: tool parameter error, output format error, premature termination, repetitive looping
• Verification: context loss, inadequate verification
• Coordination: delegation error, communication failure, over-reliance on other agents

Evaluation

The companion evaluation harness, with renderers for every framework and the scoring pipeline, is on GitHub: whowhenpro/whowhen_pro.

Citation

@article{liu2026pro,
  title={Who\&When Pro: Can LLMs Really Attribute Failures in AI Agents?},
  author={Liu, Jiale and Xi, Huajun and Zhang, Shaokun and Zeng, Yifan and Yue, Tianwei and Wang, Chi and Kang, Jian and Wu, Qingyun and Wang, Huazheng},
  journal={arXiv preprint arXiv:2607.09996},
  year={2026}
}