Source-linked AI summary
Automatic Failure Attribution and Critical Step Prediction Method for Multi-Agent Systems Based on Causal Inference
Guoqing Ma, Jia Zhu, Hanghui Guo, Weijie Shi, Jiawei Shen, Jingjiang Liu, Yidan Liang
TL;DR
MAS failure attribution remains difficult: correlation-based approaches achieve less than 20% accuracy by confusing symptoms with root causes. This paper introduces a multi-granularity causal framework, reporting up to 36.2% step-level accuracy and 22.4% task-success gains from targeted optimizations.
Problem
Correlation-based MAS failure attribution approaches achieve less than 20% accuracy, confusing symptoms with root causes.
Source: Conclusion · p. 7Method
A multi-granularity causal framework combines performance causal inversion and Shapley-based agent attribution with CDC-MAS for non-stationary interaction data.
Source: Introduction · p. 2, Conclusion · p. 7Results
Up to 36.2% step-level accuracy, significantly outperforming baselines; targeted optimizations boost task success rates by 22.4%.
Source: Conclusion · p. 7Takeaways & Limitations
The framework is intended to provide precise, interpretable, and actionable diagnostics for multi-agent system failures.
Source: Introduction · p. 2Performance causal inversion assumes an acyclic performance graph and has limitations with complex feedback loops or collaborative failures.
Source: Introduction · p. 4
Abstract
from arXiv · showhide
Multi-agent systems (MAS) are critical for automating complex tasks, yet their practical deployment is severely hampered by the challenge of failure attribution. Current diagnostic tools, which rely on statistical correlations, are fundamentally inadequate; on challenging benchmarks like Who\&When, state-of-the-art methods achieve less than 15\% accuracy in locating the root-cause step of a failure. To address this critical gap, we introduce the first failure attribution framework for MAS grounded in multi-granularity causal inference. Our approach makes two key technical contributions: (1) a performance causal inversion principle, which correctly models performance dependencies by reversing the data flow in execution logs, combined with Shapley values to accurately assign agent-level blame; (2) a novel causal discovery algorithm, CDC-MAS, that robustly identifies critical failure steps by tackling the non-stationary nature of MAS interaction data. The framework's attribution results directly fuel an automated optimization loop, generating targeted suggestions whose efficacy is validated via counterfactual simulations. Evaluations on the Who\&When and TRAIL benchmarks demonstrate a significant leap in performance. Our method achieves up to 36.2\% step-level accuracy. Crucially, the generated optimizations boost overall task success rates by an average of 22.4\%. This work provides a principled and effective solution for debugging complex agent interactions, paving the way for more reliable and interpretable multi-agent systems.
Abstract · Introduction
The paper frames multi-agent system (MAS) failure attribution as a gap in reliability and scalability, and proposes causal inference at agent and step levels to identify causes and guide counterfactual diagnosis.
Introduction
- MAS failure rates reach 40–60%, while debugging and optimization consume up to 70% of maintenance efforts; performance can also fail to surpass single-agent baselines.The failures are described as systemic and emergent, arising from inter-agent interactions, coordination breakdowns, and error propagation.Source: Introduction · p. 1
- Best prior methods reach only 17.1% critical-step accuracy on Who&When and 18.3% joint accuracy on TRAIL, while correlation diagnostics mistake downstream symptoms for root causes.The introduction also notes that manual diagnosis is costly and difficult to scale.Source: Introduction · p. 2
- The paper proposes a multi-granularity causal framework that combines agent-level attribution, step-level prediction, and counterfactual reasoning to diagnose failures and support actionable optimization.Its contributions include causal attribution, hierarchical modeling, performance causal inversion with Shapley values, and validation of diagnostics-guided optimization suggestions.Source: Introduction · p. 2
- MAS attribution must account for observable and latent confounders that can create false associations in correlation-based analyses.Examples include task complexity, agent configuration, external API performance fluctuations, and emergent collaboration patterns.Source: Introduction · pp. 2–3
- Performance Causal Inversion reverses execution-log data dependencies so graph edges point from an agent’s performance deficiency to its downstream performance effect.The principle assumes an acyclic performance graph and may be limited by complex feedback loops or collaborative failures.Source: Introduction · pp. 3–4
- SBSLocator combines Shapley-based system importance with counterfactual improvement potential, identifying agents as bottlenecks only when both are high.It asks whether the task would succeed if a suspect agent performed normally; Shapley values are approximated by Monte Carlo sampling.Source: Introduction · p. 4
- CPIdentifier uses technical, interaction, temporal, and semantic features with CDC-MAS to discover temporal causal structures and rank critical error steps under non-stationarity and confounding.CDC-MAS prepares context-aware features, discovers a temporal graph skeleton conditioned on context, orients edges using context as a confounding proxy, and ranks steps by causal effects.Source: Introduction · pp. 4–5
- CPIdentifier validates candidate steps through counterfactual interventions and ranks them using causal effect, expected improvement, and bootstrap-estimated confidence.The resulting ranking is intended to guide subsequent MAS optimization.Source: Introduction · p. 5
Experiments
Experiments evaluate the framework’s attribution accuracy, closed-loop repair utility, and component contributions across Who&When and TRAIL. The results test its causal diagnosis on varied multi-agent trajectories and its ability to guide targeted repairs.
Datasets
- Who&When contains 184 failures from 127 LLM-based multi-agent systems, while TRAIL contains 118 GAIA-based information-retrieval traces averaging 8.28 spans.Who&When includes short Algo-Generated and longer Hand-Crafted trajectories; its evaluation uses an 80/10/10 train-validation-test split.Source: Datasets · p. 5
Baseline Methods
- Baselines included Random and GPT-4o LLM-as-Judge variants: All-at-Once, Step-by-Step, and Binary Search.Methods used identical hardware and parameter settings for a fair comparison.Source: Baseline Methods · p. 5
Evaluation Metrics
- Evaluation used agent-level and step-level accuracy, with statistical significance tested by t-test at p < 0.05.Reported results are means ± standard deviations across five independent runs.Source: Evaluation Metrics · pp. 5–6
Main Results: Attribution Accuracy
- On Hand-Crafted Who&When, Step-Level Accuracy reached 18.2%, over 159% above Step-by-Step’s 7.02%, with all key improvements significant (p < 0.05).The prose attributes this result to tracing upstream root causes rather than downstream symptoms.Source: Main Results: Attribution Accuracy · p. 6, Table 1 · p. 6
- On TRAIL, step-level accuracy was 44.6% on GAIA and 14.3% on SWE Bench, versus average baseline performances of 24.8% and 10.5%, respectively.The framework’s bars exceed competing models across both subsets.Source: Main Results: Attribution Accuracy · p. 6, Figure 3 · p. 6
Operability: Closed-Loop Optimization
- Causal-guided repairs raised overall task success from 15.4% to 37.8%, a 22.4-point gain, versus 18.3% with Random Repair.The experiment diagnosed 50 failed Who&When trajectories, generated targeted revisions, then re-executed tasks from the repaired step.Source: Operability: Closed-Loop Optimization · pp. 6–7, Table 2 · p. 6
Ablation Study
- Removing Context Conditioning reduced Step-Level Accuracy by 13.9 points, while removing Shapley Values reduced it by 11.7 points; removing Causal Inversion lowered Agent-Level Accuracy by 7.8 points.Removing Counterfactual Reasoning also caused a notable decline; the study reports that all components contribute to robust attribution.Source: Ablation Study · p. 7, Table 3 · p. 7
Related Work
LLM-powered multi-agent systems coordinate specialized agents for complex problem-solving but remain failure-prone, and existing attribution methods lack robust causal diagnosis.
Multi-Agent Systems
- LLM-powered MAS coordinate specialized agents for complex problem-solving, yet exhibit 40–60% failure rates and benchmarks such as AgentBench (Liu et al. 2023a) lack causal diagnostic tools.The cited systems use role assignment and Standard Operating Procedures for coordination; MAST classifies failures including inter-agent misalignments and verification errors.Source: Multi-Agent Systems · p. 7
LLM-as-Judge for Failure Attribution
- Who&When and TRAIL attribution strategies (All-at-Once, Step-by-Step, Binary Search) have low accuracy because correlation-based analyses confuse symptoms with causes.The paper instead combines Shapley-based responsibility allocation with performance causal inversion to identify root causes in non-stationary MAS interactions.Source: LLM-as-Judge for Failure Attribution · p. 7
Conclusion
The paper presents a multi-granularity causal inference framework for MAS failure attribution, addressing limitations of manual, correlation-based, and LLM-as-Judge methods. Evaluations show improved step-level accuracy and task success, while the paper identifies directions for extending the framework.
Conclusion
- Up to 36.2% step-level accuracy and 22.4% higher task success rates from targeted optimizations demonstrate the framework’s performance and actionability.The conclusion reports that the method significantly outperforms baselines and that closed-loop experiments validate the optimizations.Source: Conclusion · p. 7
- The framework combines performance causal inversion, Shapley values for agent contribution quantification, and CDC-MAS for non-stationary data.It is presented as the first multi-granularity causal inference framework for MAS failure attribution.Source: Conclusion · p. 7
- Future directions include multi-modal interactions, online causal discovery for real-time attribution, and integration with automated self-healing mechanisms.Source: Conclusion · p. 7
- Manual analysis is unscalable, correlation-based approaches achieve less than 20% accuracy by confusing symptoms with root causes, and LLM-as-Judge methods cannot model causal dependencies.Source: Conclusion · p. 7