Research map / Multi-agent systems and games / LLM agents
Who&When Pro: Can LLMs Really Attribute Failures in AI Agents?
Read the paperCheck your idea against 108,967 papers
AI-extracted summary. Written by AI from the paper's full text. Numbers and claims can be wrong or missing context, so check the paper before you cite or build on it.
Introduces Who&When Pro, a 12,326-trace benchmark for automated failure attribution in text, image, and video agents, built by warm-start error injection with golden decisive-step labels, and evaluates ten LLMs under three attribution protocols.
Problem
Automated failure attribution (identifying the responsible agent, decisive step, and failure mode) lacks large-scale, multimodal benchmarks with reliable golden labels, since prior benchmarks are small, text-only, or have noisy labels.
Approach
Construct failure traces by replaying a successful seed trajectory up to a chosen step, injecting an erroneous action there, and letting the agent continue, so the injected step is the decisive error by construction; failure-mode profiles from human review constrain which injections are plausible.
Key results
- The best model reaches 73.9% step accuracy on text traces but only 22.2% error-mode macro-F1 (GLM-5 / Qwen3.5 area of the table), with similar gaps on image and video.
- Step accuracy falls from 94% on traces under 3K tokens to 50% on traces over 12K tokens.
- All-at-once attribution achieves the highest Joint accuracy for eight of ten models and improves Step accuracy over step-by-step by 7.6 points.
Surprising findings
- Providing the ground-truth answer to the judge improves perception-error attribution but degrades reasoning-error attribution.
- Open-weight models match or surpass closed-source models at a fraction of the cost; Qwen3.5 achieves the highest text accuracy at about seven times lower cost than GPT-5.4.
Limitations
- Each trace has exactly one decisive error step, not compounding or causal chains of errors.
- The 18-mode taxonomy does not claim exhaustive coverage of agent failure modes.
- Injected errors are produced by frontier LLM prompts, so injected failures may not match naturally occurring failure distributions.
Experimental setup
Datasets: DA-Bench DA-Code DataBench DeepSearch QA KramaBench TableBench BrowseComp-VL MMSearch MMSearch-Plus ChartQAPro CharXiv EvoChart
Benchmarks: Who&When Pro
Metrics: Agent accuracy Step accuracy Error macro-F1 Joint accuracy Step@k
Baselines: Who&When MAST TRAIL AgentErrorBench AEGIS AgenTracer AgentRx
Builds on: Who&When decisive error formulation AEGIS AgenTracer
Closest papers by meaning
- Which Agent Causes Task Failures and When? On Automated Failure Attribution of LLM Multi-Agent Systems (ICML 2025 · similarity 0.94)
- AgenTracer: Who Is Inducing Failure in the LLM Agentic Systems? (ICLR 2026 · similarity 0.90)
- Beyond LLM-Based Reasoning: Lightweight GNNs for Agent Failure Attribution (NeurIPS 2026 · similarity 0.88)
- Why Do Multi-Agent LLM Systems Fail? (NeurIPS 2025 · similarity 0.87)
- AgentForesight: Online Auditing for Early Failure Prediction in Multi-Agent Systems (NeurIPS 2026 · similarity 0.87)
- Towards a Science of AI Agent Reliability (ICML 2026 · similarity 0.87)
Writing a paper in this area? scime draft is a free LaTeX editor with AI included, real-time co-authors and a check of your references and venue rules before you submit. It also imports OpenAI Prism projects.