Stay Focused: Problem Drift in Multi-Agent Debate

1 Introduction

• The paper addresses failures in multi-agent debate, including ineffective criticism, coordination problems, and deteriorating communication during long discussions.These limitations remain insufficiently understood in prior work.Source: §1 · p. 1
• Problem drift occurs when agents gradually diverge from the original task and cannot recover their goal during extended discussions.The study analyzes this deterioration across tasks and links it to performance degradation.Source: §1 · pp. 1–2
• Generative tasks show problem drift in 76%-89% of samples, while instruction-following tasks show it in 21%.Most drifting discussions do not recover, although recovery occurs in 9% of translation and 45% of ethical question-answering examples.Source: §1 · p. 2
• DRIFTPolicy reduces drifting discussions by 31% and improves task accuracy by up to 3.6% for weaker agents.DRIFTJudge detects drift at test time, while DRIFTPolicy supplies feedback to participating agents.Source: Figure 1 · p. 1, §1 · p. 2
• Human experts identify lack of progress as the most common drift-related issue, affecting 35% of drifting samples.Lack of progress can lead agents to overanalyze problems and provide low-quality feedback.Source: §1 · p. 2

2 Related Work

• Research on multi-agent LLMs spans simulation and problem-solving, with this paper focusing on problem-solving as a controllable setting for studying interaction.The broader literature considers multi-agent systems for human-like communication and complex task solving.Source: §2 · p. 2
• Supportive studies report gains in divergent thinking, reasoning, creative writing, dialogue generation, and theory of mind, but heterogeneous experimental designs hinder comparison.Differences include agent orchestration, decision-making, prompting, personas, and self-correction mechanisms.Source: §2 · p. 2
• Critical studies report multi-turn performance degradation, computational costs, diminished planning, irrelevant criticism, and broader coordination and robustness concerns.These findings motivate examining why prolonged debate can fail rather than only measuring its benefits.Source: §2 · p. 2

Research Gap

• The paper fills a gap by studying performance deterioration, prevalence, recovery, causes, detection, and mitigation of degrading discussions.These questions are organized as six research questions spanning empirical characterization and intervention.Source: §2 · p. 2

3 Problem Drift

• Problem drift is systematic decay in turn-based task performance as agents gradually diverge from the initial task.The definition concerns deterioration over multiple discussion turns.Source: §3 · p. 3

FOCUS

• FOCUS compares the performance of a round’s solution with the previous round using a downstream metric against the gold solution.The metric supports partial scores between 0 and 1 for non-binary solutions.Source: §3 · p. 3
• A negative FOCUS value indicates that solution quality degrades during a round, whereas a positive value indicates improvement.FOCUS is outcome-based and measures task performance rather than semantic similarity alone.Source: §3 · p. 3

Problem Drift

• Problem drift is identified when current solutions perform below an earlier debate round across multiple consecutive rounds.Its strength is represented over a sequence of rounds by aggregating FOCUS values.Source: §3 · p. 3

Recovery

• A discussion recovers when a later round returns to or exceeds the performance achieved before drift began.Recovery requires a later round whose performance meets or improves on the earlier level.Source: §3 · p. 3

4.1 Debate Setup

• The study runs three distinct-persona agents for seven conversational turns using voting and consensus protocols.Agents see one another’s messages, can disagree by proposing new solutions, and typically reach initial agreement within two turns.Source: §4.1 · p. 3
• The experiments include Llama-3.1 and Qwen2 with voting and consensus to test whether drift and FOCUS generalize across model and conversational designs.The authors make the framework and code available for reproducibility.Source: §4.1 · pp. 3–4

4.2.1 Detection with DRIFTJudge

• DRIFTJudge detects drift at test time by comparing solutions from consecutive turns without requiring gold labels.It uses the same model as the debating agents as a baseline judge, while the approach could also use lightweight classifiers.Source: §4.2.1 · p. 4

4.2.2 Mitigation Methods

• DRIFTPolicy adds a fourth agent after a drifting turn to provide situational feedback aimed at improving the conversation and resolving drift-related issues.The alternative regeneration method instead undoes and regenerates the drifting turn once.Source: §4.2.2 · p. 4

4.3 Datasets & Metrics

• The evaluation covers ten datasets across generative, reasoning, knowledge, and instruction-following tasks.Generative tasks are characterized by subjective or ambiguous answer spaces, while StrategyQA, AQUA-RAT, and MMLU-Pro require more reasoning steps.Source: §4.3 · p. 4
• Multiple-choice tasks use accuracy, generative tasks use BERTScore, and IFEval uses strict accuracy.Table 1 tracks performance and FOCUS across seven debate rounds for Llama-3.1 with voting.Source: §4.3 · p. 4, Table 1 · p. 5

RQ1: Debate length and performance

• 4.7%-69.9% of discussions suffer performance drops from negative problem focus, while 3.7%-37.9% improve with longer debates.Generative tasks show the most frequent drift, whereas reasoning and knowledge tasks drift less often but can switch from correct to completely wrong solutions.Source: §5 · p. 5

RQ2: Prevalence and task dependence

• 76-89% of generative-task samples drift, compared with 7-21% for objective, high-complexity tasks, and drift occurs across all tested tasks.Generative tasks accumulate small wording changes, while multiple-choice knowledge and reasoning tasks remain more constrained.Source: §5 · pp. 5–6

RQ3: Recovery

• 8.5%-19.8% of drifted samples recover on subjective tasks, compared with 18.5%-39.1% on more complex tasks.Recovery reaches 44.7% on IFEval and 45.4% on ETHICS, but only 18.5% on StrategyQA and 8.5% on WMT19.Source: §5 · p. 6

RQ4: Human-identified causes

• 35% of drifted samples show a lack of progress, often accompanied by low-quality feedback and lack of clarity.Experts identified eight local or temporal error types in 170 excerpts; agents may repeat arguments, overanalyze, make hypercritical changes, or violate task requirements.Source: §5 · pp. 6–7, Table 3 · p. 7

RQ5: Detection

• DRIFTJudge achieves specificity 0.92, recall 0.57, precision 0.15, and accuracy 0.91 on MMLU-Pro.The detector is presented as an initial baseline, with precision remaining a key improvement target.Source: §5 · pp. 7–8

RQ6: Mitigation

• DRIFTPolicy mitigates 30.6% of drifting examples, while combining DRIFTJudge and DRIFTPolicy reduces a -3.66% accuracy drop to -0.08% with weaker agents.The policy adds targeted feedback based on identified error types; its combined computational cost is 8.5% above default MAD.Source: §5 · p. 8

Discussion

• Problem drift remains unresolved because DRIFTJudge and Regenerate mitigate only some cases.The authors propose further mitigation and fine-tuning agents for higher-quality contributions, including use of DRIFTEval’s human-annotated error types.Source: §5 · p. 8

• 89% of debates at the high end exhibited problem drift, with rates ranging from 7% to 89% across ten tasks and higher prevalence on low-complexity generative problems.Source: §6 · p. 8
• DRIFTPolicy mitigated 30.6% of problem drift for weaker agents, while DRIFTJudge and DRIFTPolicy provided initial detection and mitigation baselines.Source: §6 · p. 9
• Eight human participants characterized eight error types, grouped into temporal errors such as lack of progress and local errors such as task non-compliance.Source: §6 · p. 9
• Future work should examine whether off-task exploration supports deeper reasoning and how agent debates differ from human debates in explorativeness, conciseness, and argument structure.Source: §6 · p. 9

Limitations

• Human annotation examined only the drifting turn and its immediate predecessor, excluding potentially informative earlier discussion history.Experiments likewise restricted agents to the current and previous turns for comparability.Source: Limitations · p. 9
• DRIFTJudge requires forwarding solution pairs from consecutive turns, and combining it with DRIFTPolicy increases computational cost by 8.5%.Source: Limitations · p. 9
• A more cost-efficient classifier such as BERT fine-tuned on drifting and non-drifting examples is suggested as a possible production alternative.Source: Limitations · p. 9

A MALLM Framework

• MALLM automatically assigns three task- and example-specific expert personas to support diverse discussion.The personas are generated by Meta-llama/Meta-Llama-3.1-70B-Instruct.Source: A MALLM Framework · p. 12
• The memory paradigm lets every agent contribute once per turn while retaining all available discussion information.Source: A MALLM Framework · p. 12
• End-of-turn voting produces a solution for evaluation after each round, with ties randomly resolved; ties occurred in 0.65% of voting procedures.Source: A MALLM Framework · p. 12

B Parameters

• The implementation used default model parameters, including temperature 1.0, top_p 1.0, and a 1024-token maximum.Source: B Parameters · p. 12

C Alternative FOCUS Metric

• FOCUSSim averages the three agents’ contextual message embeddings and compares them by cosine similarity with the task-description embedding.Source: C Alternative FOCUS Metric · p. 13
• Both FOCUS and FOCUSSim indicate the strongest drift at the beginning of the MMLU-Pro debate, with values of -2.32 and -6.86 respectively.Table 5 compares the metrics across seven turns and reports standard deviations from three runs.Source: C Alternative FOCUS Metric · p. 13, Table 5 · p. 13
• The semantic variant was excluded because semantic variability need not imply worse outcomes, and its correlation with FOCUS was weak (Spearman ρ = 0.024, p < 0.05).Source: C Alternative FOCUS Metric · p. 13
• The alternative metric can penalize temporary explorations that are later recovered, whereas a momentum term may miss drift severity at particular rounds.Source: C Alternative FOCUS Metric · p. 13

D Prompts

• Figure 3 supplies the contributing agent with the latest draft and memory, replacing them with an explicit first-solution instruction at the discussion start.Source: Figure 3 · p. 13
• Figures 4–8 provide prompts for extracting solutions, processing end-of-turn voting, assigning complementary personas, judging drift, and generating policy feedback.Source: Figure 4 · p. 13, Figure 5 · p. 13, Figure 6 · p. 14

E Datasets

• Table 6 organizes the benchmarks into three generative, three reasoning-heavy, three knowledge-intensive, and one instruction-following task.Source: Table 6 · p. 15
• The study evaluates subsets sampled with a 95% confidence interval and 5% margin of error, conservatively assuming a sample proportion of p = 0.5.Source: E Datasets · p. 15
• The resulting test sets contain several hundred samples per dataset, and experiments are repeated on randomized subsets with standard deviations reported between runs.Source: E Datasets · p. 15

F.1 Lack of Progress

• Lack of progress comprises inefficiency, redundancy, circular discussion, repetition, and unproductive disagreement, exemplified by agents repeating reasons for changing “statement” to “assertion” without introducing novel ideas.The voted solution in one turn used “statement” with BLEU 56.59, while another used “assertion” with BLEU 0.0.Source: §F.1 · pp. 16–17

F.2 Low-Quality Feedback

• Low-quality feedback includes excessive criticism, excessive agreement, self-contradiction, and unhelpful feedback that can introduce unnecessary changes.A correct answer naming Ian McKellen was expanded with irrelevant character details, producing an answer marked incorrect.Source: §F.2 · pp. 17, 19

F.3 Low-Quality Engagement

• Low-quality engagement involves poor collaboration, minimal participation, disjointed contributions, and ignoring earlier points by introducing separate perspectives or solutions.In the example, discussion initially maintained the correct answer B) Asian-Americans, but a later alternative proposing F) Hispanics was voted incorrect.Source: §F.3 · pp. 19–21

F.4 Lack of Clarity

• Lack of clarity comprises overanalysis, overgeneralization, and insignificant changes that can replace a correct answer with a less-supported alternative.For the Jane Eyre question, the initially correct answer B was later reframed as F, “more reserved in her observations,” without changing the underlying evidence.Source: §F.4 · pp. 21–23
• The lack-of-clarity example’s voted solution was incorrect.The recorded vote states that the proposed answer was false.Source: §F.4 · p. 23

F.5 Task Non-Compliance

• The task-non-compliance example involved agents inventing an answer outside the required multiple-choice options, despite initially voting for the correct option B.The debate later selected a B-variant involving indirect chromatin remodeling, which was marked incorrect.Source: §F.5 · pp. 23–25

• Knowledge gaps include assumptions, missing data, hallucinated facts, and wrongly cited facts; here, the cited China GSHS 2015 report did not exist.Source: §F.6 · p. 25
• The agents first treated E) 19% as the answer, although the task’s hidden reference was E) 19%.Source: §F.6 · p. 26
• Another agent retained E) 19% as a reasonable estimate while acknowledging that specific reliable data were unavailable.Source: §F.6 · p. 26
• The discussion introduced an allegedly credible China GSHS 2015 source and revised the answer to C) 29%.Source: §F.6 · pp. 26–27
• The revised answer was endorsed as more accurate because the discussion treated the purported report as specific, nationally representative, and WHO-conducted.Source: §F.6 · p. 27

F.7 Logical Error

• The example classifies the failure as a logical error involving misinterpretation of the relationship between accessibility and dusting frequency.The reading-comprehension specialist’s initial reasoning treats the dresser as easier to reach, while later reasoning reverses that conclusion.Source: §F.7 · pp. 27–29
• Agents’ disagreement centers on whether greater dusting frequency implies easier access to the dresser or greater difficulty reaching the fan.The discussion explicitly contrasts these two readings of the sentence.Source: §F.7 · p. 28
• The debate ultimately corrects the interpretation to answer A: the fan was harder to reach.Multiple agents initially support the dresser interpretation before reassessing the comparison of dusting frequencies.Source: §F.7 · p. 29

F.8 Linguistic Error

• Linguistic errors appear as fluency problems, grammatical errors, false pronouns, and loops of text generation.The example specifically describes agents becoming stuck in repetitive generation.Source: §F.8 · p. 29
• The debate’s final reassessments restore agreement that the original meal-allocation calculation is correct.The agents emphasize that the lunch and dinner choices remain accurately represented by 5C3 and 4C3.Source: §F.8 · pp. 30–31
• The debate repeatedly supports the meal-allocation solution A) 5C3 * 4C3 for three lunches and three dinners under the stated constraints.Several agents independently justify 5C3 for lunches and 4C3 for dinners, and the voted solution is marked correct.Source: §F.8 · pp. 30–31
• One extended disagreement introduces an alternative 4C2 * 4C3 calculation but becomes repetitive and explicitly acknowledges that its approach is incorrect.The passage contains prolonged repeated wording before another agent returns to the original calculation.Source: §F.8 · p. 31

G Task Performance

• Tables 7 and 8 compare one- and seven-turn MMLU-Pro accuracy and the turn-seven-minus-turn-one performance change across two models and decision-making approaches.Table 7 reports absolute accuracy, while Table 8 reports the corresponding performance delta.Source: Table 7 · p. 32, Table 8 · p. 32
• Table 9 tracks performance levels and positive or negative FOCUS across seven rounds for gpt-5-mini with voting.Green and red cells distinguish sustained performance thresholds and improvement or degradation relative to round one.Source: Table 9 · p. 32

H Overview of Consensus

• Consensus-based MAD continuously improves the solution draft without voting between turns.The figure describes consensus as an alternative discussion protocol.Source: Figure 9 · p. 33

I.1 Annotation Instructions

• The annotation materials include instructions for human annotation and publicly released DriftEval samples with human annotations.The instructions are presented in Figure 10, while the produced dataset is identified as DriftEval.Source: Figure 10 · p. 34, §I.2 · p. 35

I.2 Annotation Procedure

• The annotation procedure extracts drifting debates, generates candidate error labels automatically, reduces them to categories, and has eight experts annotate 170 discussions.Experts label the turns immediately before and during drift, with options for additional explanations, other causes, and false extractions.Source: §I.2 · p. 35

J Computation

• Computational cost is reported as average time per sample for 373 seven-turn MMLU-Pro discussions using Llama-3.1-70B-Instruct.Table 10 defines the scope and measurement basis of the cost analysis.Source: Table 10 · p. 35

K Usage of AI

• The paper discloses using ChatGPT-4o and Grammarly for writing, coding, and experiment-data aggregation, while noting inherent limitations of these tools.The disclosure frames the conclusions as combining human and machine insights.Source: K Usage of AI · p. 35

BibTeX Entry

• The remaining materials provide the paper’s citation sheet, BibTeX entry, and related-paper list.These records identify the publication metadata and surrounding literature rather than experimental findings.Source: CITATION SHEET · p. 36, BibTeX Entry · p. 36, Related Papers · p. 36