PaperlayerView exact revision on arXiv

Source-linked AI summary

Understanding Communication Failures in Multi-Agent LLM Systems

Priya Pitre, Jun Hou, Yi Fang, Naren Ramakrishnan, Xuan Wang

arXiv:2610.32060v1cs.AI

TL;DR

Final accuracy cannot reveal whether multi-agent failures arise from initial reasoning, peer influence, failed recovery, or aggregation. The paper builds and applies a corpus-grounded taxonomy to failed multi-round trajectories, finding that recovery and consensus problems dominate and motivating more careful disagreement handling and evidence verification.

  • Problem

    Final accuracy records that a multi-agent system failed but does not identify whether the failure came from initial reasoning, peer influence, failed recovery, or aggregation.

    Source: §1 · p. 1

  • Method

    The paper constructs a corpus-grounded taxonomy from failed multi-round trajectories and uses it to analyze errors across tasks and system designs.

    Source: §1 · p. 2, §3 · p. 3

  • Results

    Recovery/correction failures occur in 94.9% of failed trajectories and consensus/decision failures in 83.4%; recovery/correction failures are common across tasks, while consensus failures are especially frequent in QA-style tasks.

    Source: §5 · p. 5

  • Takeaways & Limitations

    The findings support better disagreement handling, evidence verification, and reconsideration of earlier decisions rather than simply generating more answers or adding debate rounds.

    Source: §5 · p. 5, §7 · p. 8

    The study focuses on incorrect multi-round trajectories and therefore does not capture all issues in correct trajectories or single-turn multi-agent settings.

    Source: Limitations · p. 9

Abstract

from arXiv · show

Multi-agent LLM systems aim to improve reasoning through debate, evidence sharing, and mutual correction, but interaction can also introduce failures that final accuracy alone cannot explain. Agents may converge too early, ignore dissent, suppress correct minority answers, or preserve a wrong framing across rounds. We present a corpus-grounded taxonomy of communication failures, built from 3,211 failed multi-round trajectories across medical QA, coding, finance, multi-hop reasoning, ethics, and embodied text tasks. Using this taxonomy, we analyze failure distributions across datasets and system designs, round-level error dynamics, recovery hardness, majority transitions, minority-signal suppression, and self-reported confidence. We find that failures are dominated by recovery/correction and consensus/decision problems rather than isolated first-round mistakes, with many emerging after agents first observe peer responses. Based on these findings, we recommend stronger disagreement handling, evidence verification, minority-signal preservation, and recovery checks before finalization.

1 Introduction

Multi-agent communication can both repair and create errors, but final accuracy alone cannot distinguish initial reasoning mistakes from failures of peer influence, recovery, or aggregation. The paper builds a corpus-grounded taxonomy to make these interaction patterns measurable and uses it to study how errors emerge and change during deliberation.

Introduction

  • Correct answers may be abandoned after peer influence, weak claims amplified through repetition, and useful corrections lost before finalization.Final accuracy indicates failure but cannot show whether it arose from initial reasoning, peer influence, failed recovery, or aggregation.Source: §1 · p. 1
  • The corpus-grounded taxonomy makes interaction patterns measurable and comparable across datasets, models, agent counts, and orchestration settings.It is used to examine when errors appear, who introduces them, whether recovery is attempted, and how group answers change across rounds.Source: §1 · p. 2
  • The taxonomy is built from trajectory evidence, with categories expanded when dense or unmapped evidence reveals coherent new failure types.Each trace is summarized into an evidence card and mapped to diagnostic categories, including an option for unsupported cases.Source: §1 · p. 2
  • Recovery/correction failures occur in 94.9% of failed trajectories and consensus/decision failures in 83.4%, outweighing isolated first-round mistakes.Many failures emerge in Round 2, after agents first observe peer responses.Source: §1 · p. 2
  • The analysis reports suppression of correct minority signals and very high recall for early-warning monitors of persistent wrong framing, premature consensus, and oscillating consensus.The paper uses these findings to motivate recommendations for more reliable multi-agent systems.Source: §1 · p. 2

2 Related Work

Related work examines multi-agent reasoning, trace-level diagnosis, attribution, and failure detection. The paper distinguishes its contribution by treating communication itself as the object of analysis and using a taxonomy to guide real-time detection.

Multi-agent systems

  • Prior multi-agent systems use debate, discussion, voting, or orchestration to expose errors, but their gains are conditional on setup and interaction.Debate can reduce accuracy when agents abandon correct answers after peer responses (Wynn et al., 2025).Source: §2 · p. 2

Communication-centered diagnosis

  • Existing taxonomies and attribution work diagnose broad failure sources or locate critical agents and steps, but leave interaction itself under-specified.The paper frames communication failures as patterns produced between agents rather than only within an agent or component.Source: §2 · p. 2

Detection and mitigation

  • The paper's taxonomy-guided trigger estimates after each round whether continued deliberation will recover or amplify an error.It is presented as real-time and propagation-aware, extending beyond final correctness or generic judge scores.Source: §2 · p. 3

3 Methodology

The methodology collects failed multi-agent trajectories, constructs a corpus-grounded communication-error taxonomy, and applies it to trigger-error detection. It combines model-assisted taxonomy building with human validation, a correct-trajectory contrast set, and an early-warning evaluation.

Multi-Agent Trajectory Collection

  • The study ran 7,300 task instances, of which 3,543 ended incorrectly, and used 3,211 incorrect multi-round traces for taxonomy construction.Traces span eight datasets and multiple open-weight model families; experiment variants used fixed 100-example samples where available.Source: §3 · p. 3

Taxonomy Construction

  • Taxonomy construction combines top-down broad dimensions with bottom-up expansion from trajectory evidence cards, candidate tags, and supporting justifications.Candidate tags are mapped and deduplicated; unmapped tags are clustered to identify coverage gaps and possible new labels.Source: §3 · pp. 3–4
  • Large taxonomy leaves are subdivided when they exceed 600 representative rows, and iterative mapping and expansion continue until no leaf exceeds that threshold.The procedure can expand taxonomy depth to depth 3 where needed while retaining row-level evidence links.Source: §3 · p. 4

Human Annotation and Taxonomy Validation

  • Three annotators assess 200 sampled failed trajectories through open-ended labeling, taxonomy mapping, and comparison of automatic labels with adjudicated human annotations.Open annotation measures taxonomy coverage, while taxonomy-visible annotation supports agreement measurement.Source: §3 · p. 4

Contrast Set

  • A contrast set of 500 correct multi-round trajectories from six non-coding datasets is annotated outcome-blind and compared with failed trajectories.The comparison adjusts for dataset-distribution differences.Source: §3 · p. 4

Trigger Model

  • The trigger model predicts whether an error will appear in the next round using answer dynamics, without message text or dataset identity.It evaluates premature consensus and two recovery failures on 304 paired debates using leave-one-pair-out validation against rule-based baselines and an LLM judge.Source: §3 · p. 4

4 AGENTINTERACTERROR Taxonomy

The AGENTINTERACTERROR taxonomy organizes communication failures into broad parent categories and finer labels. Labels depend on a behavior's role in the trajectory, and a single failed debate can receive multiple evidence-linked labels.

Taxonomy Structure

  • The taxonomy contains six broad parent categories, 24 child labels, and 11 final leaf labels.The parent categories cover reasoning/task understanding, communication/deliberation, consensus/decision, recovery/correction, coordination/orchestration, and tool/evidence use.Source: §4 · p. 4

Context-dependent Labels

  • A behavior's label depends on its effect in the trajectory; answer drift can reflect unstable consensus or loss of an available correction.The distinction turns on whether agreement becomes unstable or later discussion pulls agents away from a correct answer, correction, or useful evidence.Source: §4 · p. 4

Multi-error Annotation

  • One failed debate may receive multiple labels when its evidence shows several distinct errors, such as evidence misuse, weak dissent, and failed recovery.Each label is tied to specific evidence in the trace.Source: §4 · pp. 4–5

Evaluation

  • Taxonomy evaluation checks path granularity, sibling coherence, dimension alignment, paper relevance, and coverage.The first three semantic metrics were judged with GPT-4.1-mini; paper relevance and coverage were computed from corpus-level assignments.Source: §4 · p. 5, Table 1 · p. 5

5 Analysis and Discussion

Across tasks, recovery and correction failures are common, while failure patterns vary by dataset and system design. Failures often emerge after agents see peer responses, and later discussion does not necessarily repair them.

Human Annotation and Taxonomy Validation

  • The taxonomy’s automatic labels reach 0.91 micro-F1 at the parent level and 0.84 at the fine-grained level against adjudicated human labels.Open annotation mapped 93.4% of human-identified errors to a fine-grained label and 98.1% to a parent category; annotators reached Jaccard agreement of 0.82 at parent and 0.73 at fine-grained level.Source: §5 · p. 5

Failure Distribution

  • Recovery/correction failures are common across tasks, while QA-style datasets show more consensus failures and coding, finance, and multi-hop tasks show more reasoning and tool/evidence failures.The dataset-level patterns suggest that failure types depend partly on task, and the central problem is often failure to correct an initial mistake.Source: Figure 3 · p. 5, §5 · p. 5

What drives failures?

  • Retrieval reduces overall failure from 74.6% to 36.4% on HotpotQA and MuSiQue (95% CI: −43.8 to −32.7 pp), while increasing the team from three to seven agents does not significantly reduce failures.More agents reduce reasoning errors but increase failures after abstention and suppression of correct minority answers; hierarchy reduces oscillation but not overall failures, and explicit roles have no detectable effect.Source: §5 · p. 6

Failure vs. success contrast

  • Failed debates have more recovery, consensus, and coordination failures than successful debates, with persistent wrong framing showing the largest difference.Even among 500 successful trajectories, 91.2% contain at least one error under outcome-blind annotation; reasoning and tool/evidence errors occur at similar rates in successful and failed groups.Source: §5 · p. 6, Table 2 · p. 6

Round-level error dynamics

  • After agents see peer responses, reasoning and evidence errors decrease while consensus errors increase; later rounds contain more recovery failures.The first exchange can turn individual mistakes into group-level failures, as agents may agree on an incorrect answer and reinforce it rather than correct it.Source: §5 · p. 6, Figure 5 · p. 6

Temporal error cascades

  • Premature consensus is a common precursor to ignored tool results, weak evidence, and later corrections, while ignored evidence and corrections can also precede premature consensus.Some co-occurring errors, including premature consensus and persistent wrong framing, have lift close to 1, so co-occurrence does not always indicate a temporal relationship.Source: §5 · pp. 6–7, Table 3 · p. 7

Recovery hardness

  • Local interpretation mismatches, question misreadings, and unsupported assumptions are easier to correct than wrong framing, failed recovery, answer drift, and premature or oscillating consensus.More debate rounds do not necessarily improve recovery; the authors suggest independent reconsideration, evidence rechecking, or a separate challenger as explicit recovery mechanisms.Source: §5 · p. 7, Figure 6 · p. 7

Initial-majority transitions

  • Wrong initial majorities are more likely to remain wrong than recover, and correct minority agents are suppressed more often than they change the group’s answer.Patterns vary by task: MedQA shows stronger recovery, while Ethics has more corruption and minority suppression; correct answers also have shorter explanations in MedQA, MuSiQue, and Ethics.Source: §5 · p. 7, Table 4 · p. 8

6 Recommendations for Multi-Agent System Design

The authors recommend error-sensitive oversight rather than trusting confidence, simple majority voting, or merely extending debate. They pair task-specific verification with targeted early-warning interventions.

6 Recommendations for Multi-Agent System Design

  • Self-reported confidence should not determine trust or stopping alone, because agents can become more confident while wrong.The authors also caution that hierarchy alone does not reliably improve accuracy.Source: §6 · p. 8
  • Active supervisors should question unsupported claims, track disagreements, and verify competing answers instead of simply following the majority.The recommendation is active supervision rather than simple majority voting.Source: §6 · p. 8
  • Verification should match the task: check calculations, coding edge cases, multi-hop evidence, and protection of correct minority answers in ethics-style tasks.The authors distinguish these checks by task type.Source: §6 · p. 8
  • Interventions should match the error type, with early-warning triggers identifying when they may be needed rather than simply adding debate rounds.The recommendations pair evidence checks with premature consensus, question reconsideration with wrong framing, and conflict resolution with oscillation.Source: §6 · p. 8

7 Conclusion

The paper presents a corpus-grounded taxonomy of multi-agent communication failures, including premature consensus, ignored corrections, and unsuccessful recovery. It emphasizes evidence verification, preserved disagreement, correction during deliberation, and the taxonomy's potential for early failure detection.

7 Conclusion

  • The taxonomy identifies communication failures beyond incorrect reasoning, including premature consensus, ignored corrections, and unsuccessful recovery.The authors describe it as corpus-grounded.Source: §7 · p. 8
  • The findings support verifying evidence, preserving disagreement, and correcting errors during deliberation.Source: §7 · p. 8
  • The taxonomy supports early failure detection and provides a basis for targeted monitoring and intervention in multi-agent systems.Source: §7 · p. 8

Limitations

The study's conclusions are bounded by its trajectory and environment coverage, corpus composition, and LLM-assisted taxonomy processes. The authors frame the taxonomy as diagnostic evidence, not a guarantee of correctness or a single-cause explanation.

Limitations

  • The analysis focuses on incorrect multi-round trajectories, so it does not capture all issues in correct trajectories or single-turn multi-agent settings.Correct trajectories serve as a contrast set in selected analyses, including majority-transition dynamics and trigger-model evaluation.Source: Limitations · p. 9
  • The corpus centers on communication across rounds, leaving open-ended environments with file, memory, tool, or external-state interaction incompletely covered.Such settings may surface additional failures involving planning, execution, tool chaining, state tracking, persistent memory, and file-level coordination.Source: Limitations · p. 9
  • Because the taxonomy reflects included datasets, models, and orchestration styles, additional domains may reveal new fine-grained failure types.The authors use broad parent categories and report that high-level categories stabilize relatively quickly.Source: Limitations · p. 9
  • LLM use in taxonomy construction and evaluation may introduce model-specific bias in label generation, mapping, or semantic evaluation.Mitigations include structured trajectory cards, explicit unmapped options, evidence-backed labels, and separate structural checks.Source: Limitations · p. 9
  • Superficial monitoring without meaningful validation could make unreliable systems appear more trustworthy; the taxonomy is not a guarantee of correctness.The authors recommend human review in high-risk settings and using labels as structured debugging evidence rather than single-cause explanations.Source: Limitations · p. 9
  • Exact costs may vary across providers and time, although the paper reports approximate token usage and inference costs.Source: Limitations · p. 9

Ethical Considerations

The paper bounds its findings to analysis of system behavior and details experiments across multiple datasets, paired design comparisons, and early-warning trigger models. These analyses contrast successful and failed trajectories and report consistency for premature-consensus labels.

Ethical Considerations

  • The paper studies system behavior using existing benchmarks, not private user data, and says its findings do not establish that debate-based systems are inherently reliable.For high-stakes applications, it recommends combining taxonomy-based monitoring with external validation and human oversight.Source: Ethical Considerations · p. 9

A Experimental Setting

  • The study analyzes eight datasets and 3,211 incorrect multi-round trajectories from 7,300 task instances, using varied debate protocols, agent roles, and group sizes.Experiments used fixed 100-example samples where available and seed 42; most non-code debates used up to five rounds.Source: A Experimental Setting · pp. 11–12

C Contrast Set Details

  • Moderator/hierarchy failure (12.0% vs. 42.5%) and oscillating consensus (14.4% vs. 28.7%) were more frequent in failures, while communication errors were more frequent in successes (34.6% vs. 18.4%).The contrast set sampled 500 correct trajectories and weighted failed trajectories to match its dataset mix; communication-error gaps are interpreted cautiously because many success labels were minor issues.Source: C Contrast Set Details · pp. 12–13

D Design-Factor Analysis

  • Design-factor comparisons pair configurations differing in one factor, use pair-by-question fixed effects, and correct 252 tests; estimates cover only varied factors.Configurations with different question sets, including heterogeneous versus homogeneous teams, are not compared.Source: D Design-Factor Analysis · pp. 13–14

B Expanded Related Work

  • Table 9 situates related empirical and diagnostic studies alongside, but distinct from, resources primarily focused on taxonomies, datasets, or benchmarks.The paper says Tables 8 and 9 distinguish its work from related work.Source: B Expanded Related Work · p. 12, Table 9 · p. 14

E Trigger Model Details

  • Trigger labels define premature consensus and recovery failures from transcript events, while prediction rows use pre-onset debate states to predict an error in the next round.The model uses answer dynamics and confidence rather than message text or dataset identity, with leave-one-pair-out evaluation against baselines and an LLM judge.Source: E Trigger Model Details · pp. 14–15
  • Premature-consensus annotations agreed on 81.2% of debates (κ = 0.62) and on onset round in 80.6% of debates both runs labeled present.The two independent runs used near-identical definitions.Source: E Trigger Model Details · p. 14

F Additional Failure Distribution Results

Across failed trajectories, recovery/correction and consensus/decision problems dominate; persistent wrong framing, premature consensus, and recovery failures often overlap or vary by task and system setting.

F Additional Failure Distribution Results

  • Persistent wrong framing after correction and premature consensus appear in 62.1% and 59.6% of failed trajectories, respectively.The taxonomy is multi-label, so percentages do not sum to 100%; anchor-driven recovery failure, social convergence without verification, and unstable coordination are the most common child suberrors.Source: F Additional Failure Distribution… · p. 15, Table 11 · p. 16
  • Recovery failures overlap with consensus failures in 78.6% of failed trajectories and also commonly co-occur with coordination, tool/evidence, and reasoning failures.This pattern links recovery to how agents handle disagreement, evidence, revisions, and final decisions.Source: F Additional Failure Distribution… · p. 15, Table 12 · p. 16
  • Failure mechanisms differ by task and setting: FinQA overrepresents calculation and representation errors, while seven-agent runs overrepresent failure to recover after unknown states and low-signal contribution.LiveCodeAgent overrepresents edge-case handling and failed computational repair; MBPP shows higher failed recovery after disagreement.Source: F Additional Failure Distribution… · p. 15, Table 10 · p. 16
  • Recovery/correction and consensus/decision failures dominate the failed trajectory set, showing that failed multi-agent deliberation is driven less by isolated first-pass errors than by failures of repair, verification, and agreement.Source: F Additional Failure Distribution… · p. 15

G Additional Round-Level Error Dynamics

Localized error events peak after the first exchange, and later rounds can continue a bad state rather than repair it. Round 2 is therefore a diagnostic point for whether deliberation is correcting or stabilizing an early mistake.

G Additional Round-Level Error Dynamics

  • Errors peak in Round 2, after agents see peer responses and begin revising, aligning, or challenging one another.Figure 8 counts localized error events by round; the peak supports the finding that failures can emerge during interaction, not only in initial answers.Source: G Additional Round-Level Error Dynamics · p. 16, Figure 8 · p. 16
  • Fewer new localized errors in later rounds do not necessarily indicate repair; many trajectories enter a bad state after the first exchange and then fail to correct it.The later-round decline is consistent with a larger share of recovery-oriented failures in those rounds.Source: G Additional Round-Level Error Dynamics · p. 16

H Additional Initial-Majority Transition Results

Initial-majority transitions show that adding agents brings little change in recovery while minority signals are more often suppressed in larger groups. Model-only taxonomy labeling is unreliable for fine-grained errors, and overall accuracy varies more by task and orchestration than by group size alone.

H Additional Initial-Majority Transition Results

  • Individual rescue falls from 7.87% with 3 agents to 2.50% with 7, while individual suppression rises from 11.37% to 17.70%; recovery changes only modestly.Recovery is 19.50% with 3 agents, 22.19% with 5, and 22.06% with 7; failed recovery remains near 40%.Source: H Additional Initial-Majority… · p. 17, Table 13 · p. 17

I Model-Only Taxonomy Classification

  • Model-only taxonomy classification varies substantially across models and labels and performs especially weakly on fine-grained leaf errors.Models sometimes identify broad consensus or recovery categories but struggle to track evidence, attribute failures to turns, and determine whether later responses repaired earlier mistakes.Source: I Model-Only Taxonomy Classification · p. 17

J Overall accuracy for all runs

  • Accuracy varies more by dataset and task setting than by simply adding agents, with strong results on several datasets and substantial gains from roles or tools on HotpotQA and MuSiQue.FinQA remains near chance-level across tool and tool-role settings, and LiveCodeAgent accuracy remains low across configurations.Source: J Overall accuracy for all runs · p. 17
  • Increasing agent count does not uniformly improve accuracy; orchestration and task structure matter more than group size alone.Some settings improve with larger groups, while others are flat or decline; explicit evidence aggregation is most useful when the task benefits from it.Source: J Overall accuracy for all runs · p. 17

K Round Analysis

This section examines multi-round debate as both an opportunity for recovery and a source of error risk, alongside limits reflected in taxonomy saturation, confidence dynamics, and UNKNOWN responses. It also outlines a staged workflow for extracting communication failures from debate traces, mapping them into a taxonomy, and refining cluster-level decisions into child labels.

K Round Analysis

  • 24.3% of debates end after one round, while 75.7% require more and 36.7% reach five or more; P(correct | 1 round) = 82.1% versus P(correct |> 1 round) = 46.4%.Multi-round traces account for 39.4% of wrong outcomes and 34.1% of correct outcomes, making debate length a useful process-level signal for disagreement, answer drift, or failed correction.Source: K Round Analysis · pp. 17–19, Figure 10 · p. 20

L Taxonomy Saturation

  • 50 traces recover all parent and child categories but 80.8% of leaves; leaf coverage rises to 91.9% at 200 traces, 95.4% at 500, and 99.2% at 2,000.The taxonomy spans 6 parent categories, 11 child categories, and 26 mapped leaves; rare labels require substantially more data to capture.Source: L Taxonomy Saturation · p. 19, Figure 11 · p. 21

M Self-Reported Confidence Dynamics

  • Confidence rises for both correct and incorrect traces, more for incorrect traces (+10.49) than correct traces (+5.54), while incorrect traces still end near 90% confidence.Confidence also rises after answer changes in both outcome groups, indicating that flips can increase confidence even when they move toward a wrong answer.Source: M Self-Reported Confidence Dynamics · p. 19, Table 15 · p. 20

N UNKNOWN Answer Behavior

  • 46.3% of debates contain an UNKNOWN response; MedQA shows 89.3% persistence and only 10.7% next-round recovery, while blanks and reasoning without a final answer predominate over format failures.UNKNOWN behavior varies by dataset: question-only setups are more vulnerable to persistence, whereas tool or role structure is more likely to support recovery.Source: N UNKNOWN Answer Behavior · pp. 19–20, Table 16 · p. 20

O Description of taxonomy (detail)

  • Recovery/correction failure affects 3,046 traces and consensus/decision failure 2,677; examples include persistent wrong framing, ignored corrections, premature consensus, and answer drift.These categories describe failures to repair or preserve better-supported answers and flawed convergence, aggregation, or finalization.Source: O Description of taxonomy (detail) · pp. 20–21
  • The taxonomy also includes coordination (1,702 traces), tool/evidence use (1,430), reasoning/task understanding (1,390), and communication failures (456), including low-signal contributions and ignored evidence.Communication failures also include interpretation mismatch, poor information sharing, dogmatic disagreement, and sycophancy.Source: O Description of taxonomy (detail) · pp. 21–22

P Prompts for all setup

  • The supplied prompts specify question-only or context/tool-supported initial answers, role-based debates, follow-up rounds, consensus judgments, and moderator final decisions.Agents are instructed to give concise answers and confidence scores, review peer reasoning in follow-up rounds, and distinguish equivalent answers from meaningful disagreement.Source: P Prompts for all setup · p. 23
  • Trajectory-card prompts trace agent contributions, answer shifts, agreement, tools, and failed corrections across rounds; candidate tags then summarize evidence-linked failure modes.The card identifies concrete rounds or steps and responsible agents, and ends with an overall explanation of trajectory failure.Source: P Prompts for all setup · p. 23
  • Taxonomy prompts map each tag to one broad category and each taxonomy row to one fine-grained label, allowing unmapped decisions rather than forcing poor fits.Broad categories cover reasoning or task understanding, communication or deliberation, consensus or decision, recovery or correction, coordination or orchestration, and tool or evidence use.Source: P Prompts for all setup · p. 24
  • Unmapped rows are reviewed in semantic clusters, consolidated into final labels, then assigned one allowed child label according to the dominant failure mechanism.The prompts favor existing labels, avoid preserving tiny one-off labels, and distinguish premature-consensus mechanisms such as social convergence, anchor-driven lock-in, masked conflict, and closure on incomplete evidence.Source: P Prompts for all setup · p. 24

Related papers

Loading 2610.32060v1…