PaperlayerView exact revision on arXiv

Source-linked AI summary

DCFA: Dual-view Causal-inspired Attribution for Failure Reasoning in LLM-based Multi-agent Systems

Zehao Wang, Lanjun Wang, Shilong Jin, Junjie Chen, Yanghua Xiao

arXiv:2609.04749v1cs.AI

TL;DR

LLM-based multi-agent systems require failure attribution that can distinguish decisive errors from recoverable deviations while handling long traces. DCFA combines global causal-inspired dependency-graph reasoning with local counterfactual-inspired refinement, improving step-level accuracy by up to 8.27% over state-of-the-art baselines. Its applicability remains constrained by extremely long or highly branching traces and by the reasoning capabilities of the underlying models.

  • Problem

    Failure attribution must identify decisive errors in fragile multi-agent systems, but existing methods can focus on recoverable deviations and degrade on long traces.

  • Method

    DCFA is a training-free framework combining global causal-inspired dependency-graph analysis with local counterfactual-inspired reasoning.

  • Results

    DCFA improves step-level attribution accuracy by up to 8.27% over state-of-the-art baselines across six LLMs on the Who&When benchmark.

  • Takeaways & Limitations

    Combining global dependency analysis with local refinement supports decisive-error attribution across traces of varying lengths.

  • Takeaways & Limitations

    Extremely long or highly branching interactions can reduce the reliability of both global attribution and local counterfactual refinement.

Abstract

from arXiv · show

Large language model (LLM)-based multi-agent systems have experienced rapid growth in recent years. Despite their promise, such systems remain fragile, frequently exhibiting reasoning and coordination errors that can lead to system-level failures. Failure attribution in such systems relies on tracing natural language interactions among agents to identify the decisive error, which refers to the earliest action whose correction can reverse system failure. There are two key challenges: 1) Shallow attribution: Existing methods often capture only minor deviations, such as incomplete retrievals or formatting errors, which verification mechanisms can correct, while missing the decisive cause of system failure. 2) Contextual degradation: As the length of the system traces increases, the model's reasoning ability rapidly deteriorates. To address these challenges, we propose DCFA, a training-free framework for failure attribution. DCFA integrates a global module that constructs structured causal-inspired dependency graphs from system traces to identify the initial decisive error, and a local module that applies local counterfactual-inspired reasoning to refine causal-inspired attribution. Experiments on the Who&When benchmark across six LLMs show that DCFA improves step-level accuracy by up to 8.27% over state-of-the-art baselines.

1 Introduction

LLM-based multi-agent systems are fragile because coordination and reasoning errors can cascade through interaction traces. DCFA addresses shallow attribution and long-context degradation with global dependency-graph reasoning and local counterfactual-inspired refinement.

  • Multi-agent systems support autonomous collaboration but remain vulnerable to coordination and reasoning failures that can derail workflows.
  • Existing approaches trade costly annotation and retraining for training-free inference that often struggles with long, multi-step, unstructured traces.
  • Attribution can be shallow when methods identify recoverable deviations instead of the earliest action whose correction reverses system failure.
  • Step-level attribution accuracy averages around 15% for traces of length 10 but drops to nearly 0% for traces approaching length 100.
  • DCFA combines global causal-inspired dependency-graph reasoning with local counterfactual-inspired refinement to identify decisive errors in long traces.
  • Experiments across six LLMs show up to 8.27% higher step-level accuracy than state-of-the-art baselines.

2 Related Works

Related work shows that multi-agent systems suffer cascading failures and that existing attribution methods rely heavily on LLM reasoning. DCFA builds on causal-inspired graph structure and localized counterfactual validation to address these limitations.

  • LLM-based multi-agent systems exhibit coordination failures and cascading reasoning errors, especially in long or highly interactive traces.
  • Failure attribution targets the earliest mistake whose correction could reverse overall system failure, called the decisive error.
  • Fine-tuned diagnostic models provide control but require costly annotation and retraining, whereas instruction-based methods analyze traces through prompting.
  • Existing methods often produce shallow attribution and degraded performance on long traces because they rely heavily on intrinsic LLM reasoning.
  • Prior research indicates that LLMs can extract events, propose causal relations, and simulate hypothetical interventions from unstructured text.
  • These developments motivate causal-inspired scaffolding and localized counterfactual-inspired validation for failure attribution.

3 Problem Formulation

The problem formulation represents a multi-agent execution as an interaction trace with an expected final outcome and defines failure when the generated output differs. Failure attribution seeks the decisive interaction whose correction recovers or most reduces the failure.

  • A system trace is a sequence of interactions, each pairing an agent identifier with its corresponding content.
  • A system failure occurs when the generated final output deviates from the expected ground-truth outcome.
  • When one erroneous interaction causes failure, its correction should recover the correct outcome.
  • When multiple errors jointly contribute, the decisive error is the interaction whose correction yields the largest reduction in system failure.
  • DCFA identifies and refines decisive errors by combining global causal-inspired attribution with local counterfactual-inspired reasoning over a dependency graph.

4 Methodology

DCFA combines global causal-inspired graph reasoning with local counterfactual-inspired refinement to identify decisive errors in long, error-prone multi-agent traces. GCA preserves candidate deviations, models their dependencies, and proposes an initial error hypothesis; LCE then searches its neighborhood and evaluates corrections by outcome alignment.

  • Framework Overview: DCFA uses GCA for global trace reasoning and LCE for local counterfactual-inspired refinement of decisive-error attribution.GCA addresses shallow attribution and LCE addresses contextual degradation from long traces.
  • Minor Deviation Identification: Minor deviations such as malformed outputs or incomplete retrievals are often recoverable and may instead be early symptoms of deeper failures.GCA therefore retains potentially relevant deviations rather than immediately selecting one as decisive.
  • Global Causal-Inspired Analysis: GCA detects candidate deviations with short textual reasons, then represents trace interactions as vertices and putative dependencies as directed edges.The resulting graph is centered on detected deviations and their reasoning cues while remaining connected to the full trace.
  • Dependency Graph Construction: Edges are admitted only when temporality, necessity, and sufficiency all hold, providing a conservative filter for dependency relationships.Temporality preserves trace order; necessity tests removal of an upstream step, while sufficiency tests whether enforcing it can induce the downstream outcome.
  • Global Hypothesis Generation: GCA uses the trace, candidate deviations, textual reasons, graph, and ground-truth outcome to generate an initial decisive-error hypothesis and reason.The output is the hypothesized interaction τ⋆ and its corresponding reason.
  • Local Counterfactual-Inspired Refinement: LCE refines the global hypothesis by bidirectionally tracing upstream and downstream dependencies and greedily selecting interactions through counterfactual-inspired evaluation.The forward search captures propagation, while the backward search examines prior misreasoning or missing premises.
  • Counterfactual Evaluation: LCE evaluates each correction using traces with corrected prefixes and original suffixes, then measures incremental outcome-alignment improvement after preceding corrections.A corrected trace is built from the ground-truth outcome, original trace, erroneous interactions, reasons, and dependency graph; semantic alignment uses a pretrained encoder and cosine similarity.

5 Experimental Design

DCFA is evaluated on Who&When using multiple datasets, models, baselines, ablations, trace lengths, and a case study. Results show improved attribution accuracy, robustness on long traces, and a runtime–accuracy trade-off from LCE.

  • Experimental Setup: DCFA is evaluated on 184 Who&When traces spanning Algorithm-generated and Hand-crafted settings, against five LLM-based baselines under identical inference settings.The evaluation uses six open-source and commercial LLMs and Step-level Accuracy for precise failure localization.
  • Overall Performance: DCFA improves average step-level accuracy over the strongest baseline by 12.51% on Algorithm-generated data and 4.02% on Hand-crafted data.The comparison covers both dataset settings reported in Table 1.
  • Ablation on LCE: Adding LCE improves average step-level accuracy by 3.45% on Hand-crafted data and 1.93% on Algorithm-generated data.The ablation compares DCFA with and without LCE while retaining the GCA graph and hypothesis.
  • Performance on Varying Trace Lengths: DCFA outperforms baselines across nearly all five trace-length levels, with gains largest at Level 1 and maintained through Level 5.The sole reported exception is Level 2 with DeepSeek-R1-32B, where DCFA slightly underperforms ECHO.
  • Case Study: In the trace-49 case study, LCE identifies τ12 as decisive because correcting τ8 or τ9 mitigates propagation, whereas correcting τ12 yields the largest failure reduction.τ8 is a minor deviation, τ9 is the GCA-identified error, and τ12 is the decisive error.

6 Conclusion

DCFA is a training-free framework that combines global causal-inspired dependency graph analysis with local counterfactual-inspired reasoning for failure attribution. It identifies decisive errors and remains effective across varying trace lengths, improving step-level attribution accuracy by up to 8.27% over state-of-the-art baselines.

  • DCFA combines global causal-inspired dependency graph analysis with local counterfactual-inspired reasoning to identify decisive errors in multi-agent system traces.
  • DCFA remains effective across traces of varying lengths and improves step-level attribution accuracy by up to 8.27% over state-of-the-art baselines.

Limitations

DCFA remains constrained by the capabilities of its underlying LLMs, especially on long or highly branching traces and failures arising from intrinsic model reasoning errors.

  • Dependence on LLM Capability: Long traces can reduce the accuracy of global causal-inspired dependency graph construction when LLMs cannot maintain coherent dependency representations.The limitation is tied to the reasoning capability of the underlying model.
  • Dependence on LLM Capability: Smaller or less capable models may reduce the precision of decisive error identification during local counterfactual-inspired refinement.
  • Scalability to Long or Complex Traces: Extremely long or highly branching interactions can make both global attribution and local refinement less reliable.This limits applicability to systems with deeply nested or prolonged interaction patterns.
  • Attribution of Intrinsic LLM Reasoning Errors: Errors caused by intrinsic LLM hallucinations may be difficult to attribute accurately compared with failures within the multi-agent workflow.

Ethical Considerations

DCFA is intended to support failure analysis rather than make automated judgments or assign accountability, so its outputs require cautious human interpretation.

  • Ethical Considerations: DCFA is a diagnostic tool for failure analysis, not an automated mechanism for judgment or accountability.Its attribution results may be imperfect, particularly for intrinsic model errors or ambiguous interactions.
  • Ethical Considerations: The framework should assist rather than replace human analysis, especially in high-stakes settings.Over-reliance on automated explanations could misinterpret responsibility.
  • Ethical Considerations: DCFA outputs depend on the reasoning behavior of the underlying language models and may therefore be imperfect.

A Dataset

The Who&When benchmark combines algorithm-generated and hand-crafted multi-agent traces with varied lengths, tasks, annotations, evaluation baselines, and reproducibility procedures.

  • Dataset Composition: The benchmark contains 184 executions: 126 algorithmically generated CaptainAgent traces and 58 manually curated traces from systems such as Magnetic-One.
  • Dataset Composition: The dataset covers web-navigation and general-assistant reasoning tasks, with each trace annotated for the responsible agent, decisive error, and failure explanation.
  • Trace Characteristics: Algorithm-generated traces average 8.7 steps, whereas Hand-crafted traces average 51.6 steps and range up to 130 steps.Hand-crafted traces also show greater length variability.
  • Evaluation Setup: DCFA is evaluated against five baselines representing distinct strategies for localizing the failure-responsible step and agent.The baselines include All-at-Once, Step-by-Step, Binary-Search, A2P, and ECHO.
  • Evaluation Setup: Step-level Accuracy measures whether the exact erroneous step is identified, and is preferred over Agent-level Accuracy because agent-level scores can be inflated.
  • Reproducibility: Three-run robustness evaluations report mean and standard deviation to assess small fluctuations in LLM inference outputs.The reported variance is extremely small, and DCFA consistently achieves the highest performance across settings.
  • Runtime: DCFA averages about 300 seconds per trajectory, including approximately 32 seconds for GCA and 270 seconds for local LCE processing.The LCE stage processes roughly 6k tokens per inference and averages 4.5 calls.

E.2 Trade-off Analysis

DCFA trades additional computation for improved attribution, with GCA providing substantial gains and LCE adding further benefits on longer traces; its current design remains post-hoc rather than real-time.

  • Accuracy–Cost Trade-off: GCA alone improves over ECHO by 10.58% on the Algorithm-Generated dataset and 0.57% on the Hand-Crafted dataset.Both ECHO and GCA require two full-context LLM calls.
  • Accuracy–Cost Trade-off: LCE further improves detection accuracy on longer trajectories, including a 3.45% gain on the Hand-Crafted dataset.The additional local reasoning incurs extra computational cost.
  • Accuracy–Cost Trade-off: The primary performance gains are attributed to enhanced reasoning over the trajectory, while LCE adds local refinement at additional cost.
  • Practicality: DCFA analyzes complete traces after task failure and is therefore designed for post-hoc attribution rather than real-time debugging.Its outputs can guide offline changes to agent structures, context management, or tool invocation policies.
  • Practicality: Online deployment would require incremental reasoning with cached dependency subgraphs and local counterfactual evaluations to reduce repeated full-trace processing.
  • Graph Quality: Higher-capability models produce edge counts closer to the global average, whereas less capable models generate overly dense or sparse dependency graphs.The comparison concerns graphs produced during GCA on Algorithm-generated and Hand-crafted datasets.

F.2 Implications for Decisive Error Localization.

The analysis examines how dependency-graph structure and localized refinement affect decisive-error localization. DCFA’s graph grows nearly linearly with trace length, while LCE reduces the reasoning scope for long traces.

  • Gemini-2.5-Pro and GPT-5 capture the overall dependency structure but differ substantially in the critical failure region.Gemini-2.5-Pro produces a more entangled τ9–τ12 subgraph, whereas GPT-5 constructs a more compact and hierarchical structure.
  • GPT-5’s compact dependency structure enables GCA to identify τ12 as the decisive error, while Gemini-2.5-Pro initially hypothesizes τ9.
  • LCE recovers τ12 for Gemini-2.5-Pro by evaluating and eliminating less relevant upstream candidates within the local dependency subgraph.
  • 8.7 steps correspond to 8.95 edges in Algorithm-generated traces, while 51.6 steps correspond to 56.61 edges in Hand-crafted traces.These figures support nearly linear rather than quadratic edge growth.
  • LCE addresses long-trace reasoning challenges by exploring locally within the dependency graph, keeping additional search overhead roughly linear in trace length.

H Details for Case Study on Causal-inspired Dependency Graph Search

The case study shows DCFA separating recoverable deviations from the decisive fabrication error in a fifteen-interaction Unlambda debugging trace. GCA maps dependencies broadly, while LCE uses localized counterfactual evaluation to isolate the failure source.

  • The three-agent trace produces “k” instead of the correct missing backtick “`” in the Unlambda debugging task.
  • Inference Phase of GCA: GCA detects incomplete retrieval at τ8 and propagation without verification at τ9, then builds a dependency graph over these interactions.
  • Refinement Phase of LCE: LCE evaluates the neighborhood around τ9 and identifies τ12 as decisive because correcting it prevents the incorrect outcome, unlike correcting τ8.
  • Decisive Error and DCFA Attribution: τ8 is a recoverable minor deviation, while τ9 propagates partial context without introducing a fundamentally new error.
  • Decisive Error and DCFA Attribution: τ12 becomes decisive when the Assistant fabricates operator details, turning recoverable incompleteness into an irreversible factual error that drives the wrong outcome.
  • DCFA combines global dependency reasoning with local counterfactual validation to distinguish contributory deviations from decisive errors.
  • Figure 8 reports more centralized deviation distributions around zero for DCFA than ECHO across DeepSeek-R1-671B and Gemini-2.5-pro.The result indicates smaller attribution offsets and less biased localization.

J.3 Causal Reasoning and Counterfactual Analysis via LLM

This section positions DCFA as a causal-inspired, rather than formal causal-inference, framework that combines dependency-graph reasoning with counterfactual-inspired evaluation. Its procedures organize candidate deviations globally and refine them through localized bidirectional search.

  • Prior work uses LLMs to extract causal structures, aggregate causal hypotheses, and simulate interventions for reasoning about outcome changes.
  • DCFA explicitly uses causal-inspired dependency graphs and LLM-mediated approximate interventions without claiming formal causal inference.
  • Global Causal-inspired Attribution: GCA identifies minor deviations, constructs a directed dependency graph, and generates an initial decisive-error hypothesis from the full trace.
  • Local Counterfactual-inspired Enhancement: LCE refines GCA through bidirectional greedy search over upstream and downstream graph neighbors.
  • Local Counterfactual-inspired Enhancement: Counterfactual-inspired evaluation corrects selected events, simulates downstream outcomes, and estimates how each correction affects the final result.
  • Local Counterfactual-inspired Enhancement: A final semantic decision selects the candidate interaction that most plausibly constitutes the decisive error.
Loading 2609.04749v1…