PaperlayerView exact revision on arXiv

Source-linked AI summary

Finding Where the Buck Stops: An Automated Failure Attribution-Based Reflection Framework for Multi-Agent Collaboration

Xiaoqing Wang, Keman Huang, Bin Liang, Hongyu Li, Xiaoyong Du, Wuqiong Pan

arXiv:2608.28264v1cs.AI

TL;DR

Multi-agent systems often fail despite additional computational resources, while existing reflection methods can make all agents reflect on failures. DoCtOR attributes failure to a decisive error agent, corrects that step counterfactually, and targets reflection accordingly; it improves collaboration and supports efficient reflection in low-resource settings.

  • Problem

    Multi-agent systems frequently fail to outperform robust single-agent baselines, while existing reflection methods require all agents to reflect despite failures often originating from one decisive error agent.

  • Method

    DoCtOR uses automated failure attribution to identify the decisive error step and agent, counterfactual reasoning to correct the step, and targeted reflection by that agent.

  • Results

    DoCtOR achieves superior reflection quality and enhances multi-agent collaboration, while diagnose-then-correct generalizes to existing prompt-based reflection and post-error steps match complete trajectories in low-resource settings.

  • Takeaways & Limitations

    Targeting the decisive error agent prevents memory contamination of regular-behaving agents and can enable efficient reflection with reduced trajectory context.

  • Takeaways & Limitations

    The evaluation focuses on task-oriented collaborations with clear success criteria, English-language datasets, and relatively small multi-agent teams.

Abstract

from arXiv · show

Multi-agent systems (MAS) powered by large language models have shown promise for complex tasks but suffer from high failure rates. Current self-reflection methods for MAS require all agents to reflect upon failure, overlooking a critical reality: failures typically stem from a specific agent leading the task astray, namely the decisive error agent, while others merely fulfill their regular duties. Forcing regular-behaving agents to reflect contaminates their memory with wrong insights. Hence, we propose DoCtOR (Diagnose-then-Correct PPO-enhanced Reflection), a novel reflection framework that enhances multi-agent collaboration. DoCtOR first identifies the decisive error step and decisive error agent through automated failure attribution, then employs counterfactual reasoning to generate a corrected decisive error step, and finally engages only the decisive error agent to produce targeted reflections. Experimental results show DoCtOR achieves 22%, 26%, and 27% improvements over initial success rates on HotPotQA, ChartQAPro, and Mind2Web datasets, outperforming Reflexion, Retroformer, and COPPER. We further establish the generalizability of our diagnose-then-correct paradigm and demonstrate that in low-resource settings, focusing reflection on reasoning steps after the decisive error step achieves comparable quality to reflecting on the complete failure trajectory.

1 Introduction

Multi-agent systems can fail frequently, and existing reflection methods may incorrectly involve agents whose actions were sound. DoCtOR addresses this by attributing failure to the decisive error agent, correcting that step, and targeting reflection accordingly.

  • MAS can exhibit failure rates between 60% and 86.7% despite additional computational resources, raising concerns about practical application.
  • Existing COPPER-style reflection requires all participating agents to reflect after failure, assuming shared responsibility.
  • In a replayed analysis task, AgentC misinterprets minor fluctuations as a downward trend, while AgentA and AgentB perform their assigned duties correctly.
  • Forced reflection can contaminate regular-behaving agents’ memories with erroneous insights, potentially introducing new execution errors.
  • DoCtOR diagnoses the decisive error step and agent, uses counterfactual correction, and engages only that agent for targeted reflection.
  • Reflection on reasoning steps after the decisive error achieves comparable quality to reflection on complete trajectories in low-resource settings.The paper presents this as an efficient reflection-generation strategy.

2 Preliminary

The paper models cooperative multi-agent collaboration as sequential interaction over natural-language states and actions, with agents pursuing a shared objective. It frames failure attribution around locating the first incorrect action and its responsible agent.

  • 2.1 Multi-Agent Collaboration: An LLM-based MAS is represented by (N, S, A, Pξo, R), where N is agent count and S and A are joint state and action spaces.States and actions are described in natural language, while Pξo governs transitions and ξo captures transition randomness.
  • 2.1 Multi-Agent Collaboration: Agents interact sequentially with the environment to generate trajectories, and cooperative agents share an objective of maximizing collective performance.The formalization follows the standard reinforcement-learning framing of cumulative reward.
  • 2.2 Automated Failure Attribution: Automated failure attribution identifies the decisive error step and decisive error agent in a failed trajectory.
  • 2.2 Automated Failure Attribution: The decisive error step is the first incorrect action, whose downstream consequences can propagate through subsequent agents’ reasoning.The decisive error agent is the agent responsible for that erroneous action.

3 Method

DoCtOR diagnoses the earliest decisive failure, generates and verifies a counterfactual correction, then conditions targeted reflection on the responsible agent. Its reflector is trained through an RLHF pipeline and evaluated using subsequent episode returns.

  • Reflection Module: DoCtOR combines diagnosis, correction, and targeted reflection around the decisive error agent.The reflector receives the failed trajectory, agent profile, original error action, and corrected action.
  • Diagnosis Submodule: ProFA assigns each reasoning step a correctness score and uses a thresholded binary indicator to locate the earliest failure point.The correctness threshold γ is set to 0.5.
  • Correction Submodule: Counterfactual reasoning generates an alternative decisive-error action from the current state, prior trajectory, and decisive agent profile.ProFA verifies the alternative; if its score remains at or below γ, correction is repeated.
  • Reflection Module: The reflector produces a concise, high-level plan for the decisive error agent that targets root causes and mitigates similar future failures.The reflector is further fine-tuned with proximal policy optimization.
  • Rating Score of Reflections: Reflection quality is rated by the change from the current episode return to the subsequent return after injecting the decisive agent’s reflection.Frozen, low-temperature actor agents make reflection the primary source of return variation.
  • Reflector Optimization: The reflector is optimized through SFT, reward-model regression with MSE loss, and PPO with KL regularization against the supervised reference model.The KL coefficient β controls the regularization strength.

4 Experiments

Experiments evaluate DoCtOR, ProFA, generalizability, trajectory scope, and component contributions across multi-agent benchmarks. DoCtOR improves task success, while ProFA localizes decisive failures and targeted trajectory selection preserves reflection quality.

  • Main Results: DoCtOR improves initial success rates by 22%, 26%, and 27% on HotPotQA, ChartQAPro, and Mind2Web, respectively.It outperforms Reflexion, Retroformer, and COPPER across the evaluated datasets.
  • Main Results: ProFA achieves approximately 80% agent-level accuracy on held-in datasets and 50% on held-out datasets.Step-level accuracy reaches 53%, 78%, and 42% on held-in datasets, and 38% and 20% on held-out datasets.
  • ProFA Performance: ProFA improves held-in agent-level accuracy by 31% on HotPotQA and 27% on ChartQAPro compared with all-at-once.Its held-in step-level accuracy improvements are 25% on HotPotQA and 33% on ChartQAPro.
  • ProFA Performance: On held-out datasets, ProFA improves agent-level accuracy by 4% over all-at-once and step-level accuracy by 20% over random or 9% over step-by-step.The step-level comparisons apply to the Algorithm-Generated and Hand-Crafted subsets, respectively.
  • ProFA Performance: ProFA consistently achieves the smallest absolute distance between predicted and actual decisive error steps across datasets.This evaluates how closely automated predictions locate the actual decisive error step.
  • Trajectory Scope: Steps after the decisive error step yield reflection quality comparable to complete multi-agent trajectories.The result supports selecting post-error steps to reduce computational requirements while retaining reflection quality.
  • Ablation Study: Removing the correction, diagnosis and correction, or PPO components causes consistent performance drops across datasets.Removing both diagnosis and correction produces the largest degradation.

5 Conclusion

DoCtOR improves multi-agent collaboration by attributing failures to decisive error agents and targeting reflection accordingly, reducing memory contamination for regular-behaving agents.

  • DoCtOR identifies decisive error agents and uses targeted reflection to avoid contaminating regular-behaving agents’ memories.The framework combines automated failure attribution with targeted reflection for more precise feedback.
  • The framework is presented as significantly enhancing multi-agent collaboration and reflection quality.
  • DoCtOR’s targeted reflections are intended to support more effective multi-agent collaboration systems.

Limitations

The paper identifies several boundaries for DoCtOR’s evaluation and describes its experimental data, models, and training procedures. Its current scope is primarily task-oriented, English-language, and based on relatively small multi-agent teams.

  • Limitations: DoCtOR is currently focused on task-oriented collaborations with clear success criteria, leaving open-ended creative tasks and subjective evaluation settings for future work.
  • Limitations: The evaluation primarily uses English-language datasets and tasks, so multilingual and culturally diverse scenarios require further investigation.
  • Limitations: The experiments primarily involve relatively small multi-agent teams, leaving scalability to larger teams unresolved.
  • Training Details: The reflector is fine-tuned with a three-stage pipeline using LoRA, supervised fine-tuning, reward modeling, and PPO-based policy optimization.The LoRA adapters use 4-bit quantization and 0.015%-0.06% of the base model parameters.
  • Data Collection: The ProFA dataset combines failure logs from HotPotQA, ChartQAPro, and Mind2Web with automated annotation followed by review from three experienced researchers.The extracted failure logs number 667, 379, and 711 respectively.

B.3 Evaluation

The evaluation uses F1 to assess structured answer quality by balancing precision and recall, with normalized token overlap providing finer-grained feedback for Mind2Web outputs.

  • F1 evaluates generated answer quality by balancing precision and recall after normalizing predictions and ground-truth responses.
  • For standard text responses, precision measures overlapping tokens among prediction tokens, while recall measures them among ground-truth tokens.
  • Mind2Web uses structured web-navigation outputs, for which F1 provides finer-grained feedback than binary task-level execution correctness.The outputs include answer, action, and value fields.

B.4 Reproducibility

The reproducibility setup uses controlled generation settings, a hierarchical Magentic-One multi-agent architecture, and specialized agents coordinated by an orchestrator.

  • Generation Settings: The action module uses temperature 0 and top-p 1, while the reflection module uses temperature 0.8 to encourage diverse reflections.These settings separate generation randomness from reflection effects and promote more creative reflection generation.
  • Multi-Agent Architecture: Magentic-One provides a hierarchical multi-agent structure for complex task solving.
  • Agent Roles: The Orchestrator decomposes tasks, plans strategically, manages workflow, monitors progress, and delegates subtasks to specialized agents.
  • Agent Roles: WebSurfer retrieves information and interacts with webpages, while Coder performs code generation, data analysis, and artifact creation.
  • Agent Roles: ComputerTerminal executes commands, runs programs, and installs libraries to provide system-level access.

C.2 Dataset-Specific Collaboration Settings

The study configures different multi-agent teams for HotPotQA, ChartQAPro, and Mind2Web, with standardized answer formats tailored to each dataset. It also evaluates HotPotQA across test-set sizes and finds consistent performance improvements.

  • Dataset-specific teams: HotPotQA uses Magentic’s Orchestrator, WebSurfer, FileSurfer, Coder, and Terminal agents for multi-hop question answering.The ensemble supports the retrieval and reasoning chains required by complex questions.
  • Dataset-specific teams: ChartQAPro adds a specialized ImageAgent to the Orchestrator, Coder, and Terminal agents for chart analysis.The ImageAgent uses tools for image question answering and visual description generation.
  • Dataset-specific teams: Mind2Web combines a specialized WebAgent with the Orchestrator, Coder, and Terminal agents for multi-step website interaction.The WebAgent is designed for website design, navigation, and task execution.
  • Response formats: HotPotQA and ChartQAPro use structured answer templates, while Mind2Web normalizes responses into Answer, Action, and Value components.These standardized formats support systematic evaluation and analytical consistency across datasets.
  • Scalability evaluation: The method consistently improves performance across HotPotQA test-set sizes, indicating stability across evaluation scales.The reported pattern is presented in Figure 10.

D.2 Robustness Across Model Families and Sizes

The framework is evaluated across model families and sizes to assess scalability and robustness. Results validate the choice of Llama-3.1-8B and show consistent improvements across the tested models.

  • Robustness across models: The framework consistently improves performance across diverse model families and sizes.The evaluation also validates the selection of Llama-3.1-8B.
  • Robustness across models: Larger models, including 8B models, generally achieve better results than smaller 1-2B models.

D.3 Effect of Correctness Threshold γ

The correctness threshold γ is examined alongside the prompts used to deploy DoCtOR and its component modules. Results indicate that γ = 0.5 generally performs best, while improvement remains robust across threshold choices.

  • Correctness threshold: γ = 0.5 generally yields the best performance across datasets.The threshold is based on practices in PRM implementations using frameworks such as TRL and OpenRLHF.
  • Correctness threshold: The method shows robust improvement patterns regardless of the specific correctness threshold chosen.
  • DoCtOR prompts: The counterfactual reasoning prompt identifies the decisive error step, supplies preceding dialogue history, and asks the agent to generate an alternative correction.
  • DoCtOR prompts: DoCtOR’s deployment prompts assign dataset-specific roles for web retrieval, image-based chart answering, and website navigation.The prompts request external-page analysis for HotPotQA, concise image-grounded answers for ChartQAPro, and task execution from webpage HTML for Mind2Web.
  • DoCtOR prompts: The reflector prompt frames the agent as responsible for diagnosing team failures and improving problem-solving through reflection.
Loading 2608.28264v1…