PaperlayerView exact revision on arXiv

Source-linked AI summary

Rethinking Failure Attribution in Multi-Agent Systems: A Multi-Perspective Benchmark and Evaluation

Yeonjun In, Mehrab Tanjim, Jayakumar Subramanian, Sungchul Kim, Uttaran Bhattacharya, Wonjoong Kim, Sangwu Park, Somdeb Sarkhel, Chanyoung Park

arXiv:2603.25001v1cs.AI

TL;DR

Prior MAS failure-attribution work assumes each failure has one deterministic cause, although multiple attributions may be plausible. This paper proposes a multi-perspective approach and introduces MP-Bench with a tailored evaluation protocol; experiments find that modern LLMs can produce diverse, human-aligned attributions under appropriate evaluation.

  • Problem

    Prior work assumes a single deterministic failure-inducing step, although MAS failures can admit multiple plausible attributions because of inter-agent dependencies and differing judgments of valid trajectories.

    Source: §1 · p. 1

  • Method

    The paper defines multi-perspective attribution as identifying plausible failure-inducing steps with rationales, and introduces MP-Bench and an evaluation protocol tailored to this setting.

    Source: §1 · p. 2, §3.2 · p. 3

  • Results

    Modern LLMs produce diverse, human-aligned failure attributions when evaluated under the proposed protocol, while prior claims of poor performance largely stem from restrictive benchmark assumptions.

    Source: §1 · p. 2, §7 · p. 8

  • Takeaways & Limitations

    The findings underscore the need for benchmarks and evaluation protocols that represent multiple valid attributions in MAS debugging.

    Source: §1 · p. 2, §7 · p. 8

    MP-Bench focuses on general-purpose assistant tasks, and its quality-first approach limits benchmark scale and covered multi-agent framework diversity.

    Source: §8 · p. 8

Abstract

from arXiv · show

Failure attribution is essential for diagnosing and improving multi-agent systems (MAS), yet existing benchmarks and methods largely assume a single deterministic root cause for each failure. In practice, MAS failures often admit multiple plausible attributions due to complex inter-agent dependencies and ambiguous execution trajectories. We revisit MAS failure attribution from a multi-perspective standpoint and propose multi-perspective failure attribution, a practical paradigm that explicitly accounts for attribution ambiguity. To support this setting, we introduce MP-Bench, the first benchmark designed for multi-perspective failure attribution in MAS, along with a new evaluation protocol tailored to this paradigm. Through extensive experiments, we find that prior conclusions suggesting LLMs struggle with failure attribution are largely driven by limitations in existing benchmark designs. Our results highlight the necessity of multi-perspective benchmarks and evaluation protocols for realistic and reliable MAS debugging.

1 Introduction

The paper argues that MAS failure attribution can admit multiple reasonable explanations, and proposes a benchmark and evaluation protocol designed to assess those perspectives and their reasoning.

1 Introduction

  • Prior benchmarks, beginning with Who&When, assume one deterministic failure-inducing step, a formulation later studies also adopt [11] [31].The paper argues this assumption can misjudge reasonable predictions and undermine evaluation reliability.Source: §1 · pp. 1–2
  • Multiple plausible attributions arise because interdependent agents can follow diverse valid trajectories, making judgments depend on an analyst’s view of a correct trajectory.Expert annotators also show perspective differences on failure-inducing steps in up to about 60% of cases.Source: §1 · p. 1
  • The proposed multi-perspective paradigm pairs failure-inducing steps with explicit reasoning from each viewpoint to support systematic diagnosis and debugging.It replaces the rigid deterministic-label formulation with multiple reasoned attributions and actionable signals.Source: §1 · p. 2
  • MP-Bench contains execution logs from diverse tasks and MAS configurations, with multiple screened expert annotators providing attributions and explicit reasoning.Its evaluation protocol measures identifying failure-inducing steps from multiple perspectives and providing reliable, well-grounded justifications.Source: §1 · p. 2
  • Repeatedly sampled LLMs identify diverse attributions that align closely with human experts’ reasoning, contrasting with near-random results on deterministic benchmarks.The authors attribute the gap to deterministic evaluation assumptions rather than inherent model limitations.Source: §1 · p. 2
  • The authors recommend diverse sampling and complementary perspectives across models to capture varied causal failure patterns and support MAS debugging.Their experiments report that sampled attributions and associated reasoning align closely with human experts’ internal reasoning.Source: §1 · p. 2

2 Related Works

Inter-agent dependencies make MAS failures difficult to diagnose, while existing attribution benchmarks assume a single deterministic failure step. This work frames attribution as multi-perspective and evaluates both attributions and their reasoning.

2.1 LLM-based Multi-Agent Systems

  • LLM-based MAS collaborate through specialized agents [10] [22] [29] [33], but inter-agent dependencies propagate errors and compound reliability problems, especially in long-horizon reasoning [3].These challenges motivate systematic analysis of where and why MAS fail, the focus of this work.Source: §2.1 · pp. 2–3

2.2 Benchmarks for Failure Attribution of MAS

  • Existing benchmarks range from execution-level MAST [3] to finer-grained TracerTraj [31], Aegis [11], and expert-built Who&When, yet assume one deterministic attribution and omit reasoning evaluation.The proposed benchmark uses expert recruitment and labeling, provides annotator rationales, and evaluates attribution quality and underlying reasoning under ambiguity.Source: §2.2 · p. 3

2.3 Annotator Disagreement

  • This is the first study of annotator disagreement in MAS failure attribution, extending perspectivism from social subjectivity to technical diagnostic ambiguity.Prior perspectivist work treats disagreement as meaningful [6] and explores ways to preserve diverse viewpoints [4] [9] [27], largely for socially subjective tasks [4] [5] [20] [21].Source: §2.3 · p. 3

3 Task Formulation

The formulation moves MAS failure attribution beyond a single deterministic label: a failed execution may support multiple plausible causal explanations, each with its own rationale, and the goal is to recover explanations aligned with human expert judgments.

3.1 Preliminaries

  • The modeled MAS has N agents acting in discrete turns, with an active-agent rule and state transitions conditioned on the current state and active agent’s action.Each agent has an allowable action subset, and an execution is a finite trajectory.Source: §3.1 · p. 3
  • Existing benchmarks formulate attribution as a fixed set of failure-inducing steps for each failed trace, often simplifying it to one step [31].This treats attribution as an objective labeling problem with a single correct set.Source: §3.1 · p. 3

3.2 Multi-Perspective Failure Attribution

  • A failed execution can admit multiple plausible attributions because valid execution trajectories and expectations of correct strategy vary across analytical perspectives.Source: §3.2 · p. 3
  • The proposed output is a set of failure-inducing steps, each paired with an explicit rationale for why that action is responsible under a particular perspective.Source: §3.2 · p. 3
  • The number of attributions can vary, and simultaneously valid explanations may identify different steps and therefore different agents as responsible.Source: §3.2 · pp. 3–4
  • Given a failed trace, the objective is to predict plausible attributions and rationales that align with human expert judgments, rather than match one oracle label.Source: §3.2 · p. 4
  • The formulation reframes attribution from finding a single failure-inducing step to characterizing a space of plausible causal explanations.Source: §3.2 · p. 4

4 MP-Bench: Multi-Perspective Failure Attribution Benchmark

MP-Bench captures multiple expert perspectives on MAS failure attribution through curated execution logs, step-level annotations, and consensus-based rankings. Its analysis finds substantial disagreement among annotators, while its design prioritizes annotation quality and reasoning over dataset scale.

4.1 Benchmark Dataset Construction

  • MP-Bench contains 289 logs from 121 MAS configurations, combining MAgenticOne [7] and CaptainAgent [22] with GAIA [14] and AssistantBench [30] tasks.The benchmark is designed to cover diverse failure patterns across systems, tasks, and annotator perspectives, with structured rationales for step-level actions.Source: §4.1.1 · p. 4, §4.1 · p. 4

4.1.1 Multi-Agent Systems and Execution Log Collection.

  • MP-Bench prioritizes expert annotation reliability over scale, using multiple annotators per instance; it has 57% more instances than Who&When (289 vs. 184).The authors contrast its manual curation with fully automated benchmarks such as TracerTraj (2,500 instances) and Aegis (9,500), which they say cannot ensure annotation reliability for complex MAS failure reasoning.Source: §4.1.1 · p. 4

4.1.2 Expert Annotator Recruitment.

  • Three technically experienced annotators passed interviews and test-case reviews before annotating complete MAS traces and task descriptions.Authors manually reviewed test annotations for correctness, consistency, and clarity of reasoning.Source: §4.1.2 · p. 4, §4.1.3 · p. 4

4.1.3 Annotation Task.

  • For each step, annotators independently record a failure-inducing label, failure reason, and ideal action, without being given guiding perspectives.This supports evaluation of both identifying failure-inducing steps across perspectives and generating reasoning aligned with expert judgment, rather than deterministic classification alone.Source: §4.1.3 · p. 4

4.1.4 Annotation Consolidation.

  • MP-Bench ranks each step labeled as failure by at least one annotator according to its failure-label consensus rate, prioritizing unanimous failures over perspective-dependent ones.Free-text reasons and ideal actions are consolidated by an LLM prompted to cover annotators’ rationales without adding information; rankings and performance patterns remain consistent across GPT-5.1, GPT-4.1, and Claude-Sonnet-4.5.Source: §4.1.4 · pp. 4–5

4.2 Annotation Quality and Effort

  • The 289 instances comprise 169 hand-crafted executions averaging 33 interaction steps and 120 automatically generated executions averaging 8 steps, with triple expert annotation.Annotation requires about 31 minutes per annotator for each hand-crafted execution and 14 minutes for each automatic execution; the authors describe this quality-first approach as limiting immediate scalability while establishing a gold-standard benchmark.Source: §4.2 · p. 5

4.3 Data Analysis

  • 16.2% of steps were labeled failures by all three annotators, 27.8% by exactly two, and 56.1% by one; disagreement reached 60%.The authors define disagreement as cases where one annotator marks a step as a failure while another does not.Source: §4.3 · p. 5

5 Evaluation Protocol

The protocol evaluates whether LLM-based failure attribution captures multiple perspectives and provides reliable reasoning. It uses ranking-based assessment for failure steps and judge-based assessment for attribution reasoning.

5.1 Failure Attribution Evaluation

  • Because one execution can support multiple plausible attributions, MP-Bench aggregates N LLM runs into a consensus ranking and compares it with human rankings using nDCG@K.nDCG@K emphasizes prioritizing highly salient failure steps, aligning evaluation with debugging workflows.Source: §5.1 · p. 6

5.2 Attribution Reasoning Evaluation

  • Consolidated LLM reasoning is judged against human ground truths for consistency, execution-context grounding, explanatory adequacy, and action reasonableness, with 1–10 scores for reasons, actions, and overall quality.The reported overall reasoning score averages across steps and executions.Source: §5.2 · p. 6
  • Failure-attribution model rankings remain highly consistent when GPT-4.1 and Claude-Sonnet-4.5 replace GPT-5.1 as judge, indicating limited sensitivity to judge choice.The study reports this robustness across all three judges.Source: §5.2 · p. 6

6 Experiment

On MP-Bench, LLMs can capture multiple plausible failure attributions when evaluated with a multi-perspective protocol, challenging conclusions from deterministic benchmarks. The experiments also find that nonzero-temperature sampling, more samples, and cross-family model combinations improve attribution, informing practical system design.

6.1.1 Failure Attribution Systems Details.

  • For implementation, the experiments use zero-shot All-at-Once attribution, generally with N=3, τ=1.0, and a 4,196-token output limit, and evaluate attribution with multiple nDCG settings.Attribution reasoning is judged by GPT-5.1 at τ=0.0; manual inspection found the judgments reliable.Source: §6.1.1 · p. 6, §6.1.2 · p. 6

6.2 Experimental Results

  • Claude-Sonnet-4.5 performs strongest and all evaluated models exceed random-baseline nDCG, while Qwen3-8B is limited; this contrasts with near-random results on Who&When.The comparison uses three samples at temperature 1; stronger reasoning-oriented or newer models do not necessarily perform better.Source: §6.2 · pp. 6–7, Table 2 · p. 7
  • LLM-generated attributions receive high judge scores against human annotators, indicating that their identified failure steps are supported by reasoning aligned with expert judgments.The authors describe these attributions as meaningful diagnostic signals for MAS debugging.Source: Table 2 · p. 7, §6.2 · p. 7
  • Different samples produce substantial disagreement in predicted failure steps, mirroring human-annotator disagreement and indicating multiple plausible perspectives rather than one deterministic root cause.The analysis reports pairwise disagreement across GPT-5.1 samplings and across different LLMs.Source: §6.2 · p. 6, Figure 4 · p. 7

6.2.1 Sensitivity Analyses on Sampling Temperature.

  • Temperature 0 performs consistently worse than higher temperatures; performance improves substantially at τ≥0.3, with no monotonic gains from 0.3 to 1.0.The temperature analysis reports nDCG@5 with exponential gain and excludes o3-mini, which only supports τ=1.Source: Table 3 · p. 7, §6.2.1 · p. 7

6.2.2 Sensitivity Analyses on Number of LLM Sampling.

  • Increasing LLM runs from N=3 to N=10 consistently improves nDCG@5 with exponential gain across models, yielding more stable and reliable rankings.The authors attribute this trend to mitigating limited-sampling variability and aggregating perspective-dependent signals.Source: §6.2.2 · p. 7, Table 4 · p. 8

6.2.3 Analyses on Multi-LLM Failure Attribution.

  • Multi-LLM systems outperform single-LLM counterparts, with the largest gains from cross-family combinations; disagreement across models exceeds disagreement across repeated samples of one model.The findings attribute gains to complementary internal reasoning patterns, rather than increased sampling alone.Source: Figure 4 · p. 7, Table 5 · p. 8, §6.2.3 · p. 8

6.3 Practical Guidance for Practitioners

  • Practitioners should aggregate multiple perspectives, use temperature above zero, combine models from different families, and increase sample counts when greater reliability is needed.These recommendations aim to capture diverse causal patterns and produce more informative, stable failure diagnosis.Source: §6.3 · p. 8

7 Conclusion

The paper argues that deterministic failure attribution misses the multi-perspective nature of MAS failures and introduces a benchmark and evaluation framework for assessing them. Its experiments find that modern LLMs can produce diverse, human-aligned attributions under this evaluation.

7 Conclusion

  • MP-Bench combines expert annotations and multi-perspective evaluation to assess failure-attribution quality and reasoning.It is presented as the first benchmark and evaluation framework designed for multi-perspective failure attribution in MAS.Source: §7 · p. 8
  • Prior claims about LLM limitations largely stem from restrictive benchmark assumptions, while modern LLMs can produce diverse, human-aligned attributions when evaluated appropriately.The paper concludes that deterministic formulations fail to capture the inherently multi-perspective nature of real-world MAS failures.Source: §7 · p. 8
  • The work provides a practical foundation for reliable MAS diagnosis and debugging and calls for rethinking failure-attribution benchmarks and evaluation practices.Source: §7 · p. 8

8 Limitations and Future Work · A Use of Generative AI Tools · B Ethics Statement

MP-Bench is presented as a gold-standard foundation for multi-perspective failure attribution, while its current domain coverage, scale, and framework diversity leave directions for future work. The paper also states that generative AI was used only for minor editing and that the dataset raises no known ethical concerns.

8 Limitations and Future Work

  • MP-Bench currently focuses on general-purpose assistant tasks, leaving specialized domains such as scientific research, software engineering, and creative tasks for future expansion.Broader domain coverage would enable a more comprehensive evaluation of failure attribution methods.Source: §8 · p. 8
  • The quality-first annotation approach supports reliable, high-fidelity annotations but limits benchmark scale.Future work could investigate hybrid expert–automated annotation pipelines to improve scalability while maintaining annotation quality.Source: §8 · p. 8
  • MP-Bench covers a limited diversity of multi-agent system frameworks, and future work could incorporate a wider range.This limitation accompanies the benchmark’s current scale constraints.Source: §8 · p. 8
  • MP-Bench provides a gold-standard foundation for multi-perspective failure attribution research.The paper describes it as a valuable resource for advancing this line of work.Source: §8 · p. 9

A Use of Generative AI Tools

  • Generative AI tools were used only for minor manuscript editing, including improving sentence fluency and correcting typographical errors.They were not used to generate new content, analyses, or experimental results.Source: A Use of Generative AI Tools · p. 10

B Ethics Statement

  • The dataset uses only publicly available tasks and publicly accessible multi-agent systems, and contains no personally identifiable information.The authors state that, to the best of their knowledge, it raises no ethical concerns.Source: B Ethics Statement · p. 10

C Additional Experimental Results · C.1 Sensitivity Analyses on LLM choice for Consolidation and LLM-as-a-Judge

Across nine combinations of consolidation and judge models, attribution reasoning scores exceed 7.0 and system rankings remain highly consistent in both hand-crafted and automatic MAS executions. GPT-4.1 judges assign higher absolute scores, while GPT-5.1 and Claude-Sonnet-4.5 likely score more rigorously.

C.1 Sensitivity Analyses on LLM choice for Consolidation and LLM-as-a-Judge

  • Across all nine configurations, attribution reasoning scores exceed 7.0, indicating high-quality reasoning regardless of judge choice.GPT-4.1 consistently assigns higher scores than GPT-5.1 and Claude-Sonnet-4.5, which the authors suggest may evaluate more rigorously and conservatively.Source: §C.1 · p. 10
  • The sensitivity analysis tests all nine combinations of GPT-5.1, GPT-4.1, and Claude-Sonnet-4.5 as consolidation and judge models.GPT-5.1 is used for both roles in the main experiments; the alternatives are GPT-4.1 and Claude-Sonnet-4.5.Source: §C.1 · p. 10
  • For hand-crafted MAS executions, average Kendall’s τ is 0.896; 31 of 36 configuration pairs are significant at p < 0.05 and all at p < 0.1.The minimum τ is 0.733 and the maximum is 1.00.Source: §C.1 · p. 10
  • For hand-crafted executions, the reported ranking correlations substantially exceed conventional thresholds for strong agreement.The analysis compares 36 configuration pairs.Source: §C.1 · p. 10
  • For automatic MAS executions, average Kendall’s τ is 0.904; 30 of 36 pairs are significant at p < 0.05 and all at p < 0.1.The minimum τ is 0.733 and the maximum is 1.00.Source: §C.1 · p. 10
  • The average τ of approximately 0.90 in both settings indicates that rankings reflect differences in attribution quality rather than consolidation or judge choices.The authors conclude that practitioners can reliably identify top-performing systems across evaluation setups.Source: §C.1 · p. 10

C.2 Scaling 𝑁in Multi-LLM Failure Attribution

In multi-LLM failure attribution, increasing the sample count steadily improves performance, while relative model rankings remain stable across judge configurations despite variation in absolute scores.

C.2 Scaling 𝑁in Multi-LLM Failure Attribution

  • Average nDCG@5 with exponential gain improves steadily as N rises from 1 to 10 across hand-crafted and automatically generated MAS datasets.With three collaborating LLMs, this spans 3 to 30 total runs (3 × N).Source: Figure 5 · p. 10, §C.2 · p. 10
  • Absolute scores vary with judge model choice, but relative rankings remain highly stable across configurations.Source: §C.2 · p. 10

Related papers

Loading 2603.25001v1…