Research map / Multi-agent systems and games / LLM agents
Which Agent Causes Task Failures and When? On Automated Failure Attribution of LLM Multi-Agent Systems
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 automated failure attribution for LLM multi-agent systems, releasing the Who&When dataset (127 systems, 184 annotated failure tasks) and evaluating three LLM-judge methods (all-at-once, step-by-step, binary search) that find agent-level accuracy is reasonable but step-level accuracy is very low.
Problem
Identifying which agent and which decisive step cause task failures in LLM multi-agent systems is manual, labor-intensive, and underexplored.
Approach
Formulates automated failure attribution as identifying the failure-responsible agent and earliest decisive error step from logs, builds an annotated benchmark, and compares LLM-as-judge strategies with different context receptive fields (full log, incremental, binary search), plus a hybrid of all-at-once and step-by-step.
Key results
- Best method (all-at-once, GPT-4o) reaches about 53.5% agent-level accuracy on Who&When, while step-level accuracy stays low at 14.2% for the best method.
- Step-by-step attains the highest step-level accuracy (e.g., 25.51% on algorithm-generated systems with ground truth), whereas all-at-once is often below random for step-level accuracy.
- Hybrid all-at-once + step-by-step raises hand-crafted step-level accuracy to 12.28% but at the highest token cost (149,177 tokens).
Measured effects reported in the paper
| What was changed | Where | Measured | Effect |
|---|---|---|---|
| Use the complete failure log at once instead of processing it incrementally | GPT-4o failure attribution on Who&When · Who&When, agent-level attribution | Agent-Level Accuracy: 54.33 vs 35.20 (algorithm-generated, with ground truth); 55.17 vs 34.48 (hand-crafted, with ground truth); 51.12 vs 26.0 | Helped |
| Process the failure log incrementally instead of all at once | GPT-4o failure attribution on Who&When · Who&When, step-level attribution | Step-Level Accuracy: 25.51 vs 12.50 (algorithm-generated, with ground truth); 7.02 vs 5.26 (hand-crafted, with ground truth); 8.77 vs 3.51 (h | Helped |
| Provide ground-truth answers to the attribution model | All three attribution methods on Who&When · Who&When | Accuracy | Helped |
| Increase failure-log context length | All three attribution methods on hand-crafted systems · Who&When, hand-crafted systems | Agent-Level Accuracy and Step-Level Accuracy | Hurt |
| Allow a wider tolerance around the predicted error step | All-at-Once versus Step-by-Step on hand-crafted systems · Who&When, hand-crafted systems | Step-Level Accuracy: ± 1: 12.07 vs 14.66; ± 2: 19.83 vs 16.38; ± 3: 30.17 vs 18.10; ± 4: 37.07 vs 31.90; ± 5: 43.10 vs 33.62 | Mixed |
| Combine All-at-Once agent prediction with Step-by-Step step detection | GPT-4o failure attribution on hand-crafted systems · Who&When, hand-crafted systems | Step-Level Accuracy: 7.90 -> 12.28 | Helped |
| Replace GPT-4o with a stronger reasoning model | All-at-Once and Step-by-Step attribution on hand-crafted systems · Who&When, hand-crafted systems | Agent-Level Accuracy and Step-Level Accuracy: DeepSeek R1 All-at-Once agent-level: 54.31 -> 56.90; OpenAI o1 All-at-Once step-level: 4.39 -> 10.34; OpenAI o1 Step-by- | Mixed |
| Remove explicit reasoning prompts (ablation) | All-at-Once on algorithm-generated systems · Who&When, algorithm-generated systems | Agent-Level Accuracy: -7.4% | Hurt |
| Remove explicit reasoning prompts (ablation) | Step-by-Step attribution · Who&When | Step-Level Accuracy: -4.4% | Hurt |
Surprising findings
- Some methods perform below the random baseline on step-level accuracy.
- OpenAI o1 and DeepSeek R1 do not consistently outperform GPT-4o on failure attribution.
Limitations
- Small dataset (184 annotated tasks) and only 30 GAIA logs for hand-crafted systems.
- Step-level accuracy remains very low, so instance-level attribution is not practically reliable.
- Algorithm-generated systems are all built with GPT-4o, limiting diversity.
Experimental setup
Datasets: GAIA AssistantBench
Benchmarks: Who&When
Metrics: agent-level accuracy step-level accuracy step-level accuracy with tolerance
Baselines: Random All-at-once Step-by-step Binary search OpenAI o1 DeepSeek R1
Builds on: CaptainAgent Magentic-One LLM-as-a-judge ProcessBench Math-Shepherd
Closest papers by meaning
- Who&When Pro: Can LLMs Really Attribute Failures in AI Agents? (NeurIPS 2026 · similarity 0.94)
- AgenTracer: Who Is Inducing Failure in the LLM Agentic Systems? (ICLR 2026 · similarity 0.90)
- DoVer: Intervention-Driven Auto Debugging for LLM Multi-Agent Systems (ICLR 2026 · similarity 0.90)
- Beyond LLM-Based Reasoning: Lightweight GNNs for Agent Failure Attribution (NeurIPS 2026 · similarity 0.90)
- Why Do Multi-Agent LLM Systems Fail? (NeurIPS 2025 · similarity 0.90)
- CORRECT: COndensed eRror RECognition via knowledge Transfer in multi-agent systems (ICML 2026 · similarity 0.89)
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.