mlsanityData to 7 Oct 2026

Stay Focused: Problem Drift in Multi-Agent Debate

arXiv 2502.19559 ↗First posted 26 Feb 2025cs.CLcited by 23

On the map: LLM agents and tool use → Social simulation and multi-agent debate

Abstract (arXiv)

Multi-agent debate - multiple instances of large language models discussing problems in turn-based interaction - has shown promise for solving knowledge and reasoning tasks. However, these methods show limitations when solving complex problems that require longer reasoning chains. We analyze how multi-agent debate drifts away from the initial problem over multiple turns, thus harming task performance. We define this phenomenon as problem drift and quantify its presence across ten tasks (i.e., three generative, three knowledge, three reasoning, and one instruction-following task). We find that generative tasks drift often due to the subjectivity of the answer space (76-89%), compared to high-complexity tasks (7-21%). To identify the reasons, eight human experts analyze 170 multi-agent debates suffering from problem drift. We find the most common issues related to this drift are the lack of progress (35% of cases), low-quality feedback (26% of cases), and a lack of clarity (25% of cases). We propose DRIFTJudge, an LLM-as-a-judge method, as a first baseline to detect problem drift. We also propose DRIFTPolicy, which mitigates 31% of problem drift cases. Our study is a step toward understanding a key limitation of multi-agent debate, highlighting why longer debates can harm task performance and how problem drift could be addressed.

Key facts. Cited by 23 papers in the map (data through 7 Oct 2026). Placed in Social simulation and multi-agent debate, one of 16 directions in LLM agents and tool use. Builds on 37 of its 78 references that are in the map. First posted on arXiv 26 Feb 2025.

Builds on

37 of 78 references.

Closest ancestors · 1

Same direction, newest first

  1. 2023Encouraging Divergent Thinking in Large Language Models through Multi-Agent Debate30 May

Other references · 36

  1. 2018BERT: Pre-training of Deep Bidirectional Transformers for Language Understandingcited by 40,204
  2. 2024The Llama 3 Herd of Modelscited by 14,192
  3. 2023Judging LLM-as-a-Judge with MT-Bench and Chatbot Arenacited by 8,393
  4. 2020Measuring Massive Multitask Language Understandingcited by 7,768
  5. 2019BERTScore: Evaluating Text Generation with BERTcited by 5,462
  6. 2023GPQA: A Graduate-Level Google-Proof Q&A Benchmarkcited by 2,748
  7. 2024Scaling LLM Test-Time Compute Optimally can be More Effective than Scaling Model Parameterscited by 1,928
  8. 2024MMLU-Pro: A More Robust and Challenging Multi-Task Language Understanding Benchmarkcited by 1,758
  9. 2023Improving Factuality and Reasoning in Language Models through Multiagent Debatecited by 1,739
  10. 2018Don't Give Me the Details, Just the Summary! Topic-Aware Convolutional Neural Networks for Extreme Summarizationcited by 1,308
  11. 2023ExpeL: LLM Agents Are Experiential Learnerscited by 814
  12. 2021Did Aristotle Use a Laptop? A Question Answering Benchmark with Implicit Reasoning Strategiescited by 736
24 more
  1. 2020Aligning AI With Shared Human Valuescited by 706
  2. 2017Program Induction by Rationale Generation : Learning to Solve and Explain Algebraic Word Problemscited by 642
  3. 2019How to Fine-Tune BERT for Text Classification?cited by 409
  4. 2019Emergent Tool Use From Multi-Agent Autocurriculacited by 308
  5. 2023ReConcile: Round-Table Conference Improves Reasoning via Consensus among Diverse LLMscited by 297
  6. 2021Understanding Factuality in Abstractive Summarization with FRANK: A Benchmark for Factuality Metricscited by 285
  7. 2023ToolQA: A Dataset for LLM Question Answering with External Toolscited by 273
  8. 2023Unleashing the Emergent Cognitive Synergy in Large Language Models: A Task-Solving Agent through Multi-Persona Self-Collaborationcited by 246
  9. 2023Improving Language Model Negotiation with Self-Play and In-Context Learning from AI Feedbackcited by 202
  10. 2023ExpertPrompting: Instructing Large Language Models to be Distinguished Expertscited by 140
  11. 2024Evaluation of OpenAI o1: Opportunities and Challenges of AGIcited by 127
  12. 2024Meta-Prompting: Enhancing Language Models with Task-Agnostic Scaffoldingcited by 115
  13. 2023Can Large Language Models Really Improve by Self-critiquing Their Own Plans?cited by 107
  14. 2023GPT-4 Doesn't Know It's Wrong: An Analysis of Iterative Prompting for Reasoning Problemscited by 88
  15. 2023Exchange-of-Thought: Enhancing Large Language Model Capabilities through Cross-Model Communicationcited by 78
  16. 2023Corex: Pushing the Boundaries of Complex Reasoning through Multi-Model Collaborationcited by 70
  17. 2024Learning From Failure: Integrating Negative Examples when Fine-tuning Large Language Models as Agentscited by 49
  18. 2024LLM Voting: Human Choices and AI Collective Decision Makingcited by 47
  19. 2024MEGAnno+: A Human-LLM Collaborative Annotation Systemcited by 47
  20. 2023CGMI: Configurable General Multi-Agent Interaction Frameworkcited by 41
  21. 2024Multi-Agent Large Language Models for Conversational Task-Solvingcited by 24
  22. 2024Text Generation: A Systematic Literature Review of Tasks, Evaluation, and Challengescited by 18
  23. 2020Neural Machine Translation: Challenges, Progress and Futurecited by 10
  24. 2024What's Wrong? Refining Meeting Summaries with LLM Feedbackcited by 8

Built on by

Cited by year (2026 to 7 Oct)

Citing papers per year: 2025: 8; 2026: 15.

23 papers.

Sort:

By area:

  1. 2025The Illusion of Diminishing Returns: Measuring Long Horizon Execution in LLMscited by 62
  2. 2025Literature Review Of Multi-Agent Debate For Problem-Solvingcited by 15
  3. 2025MALLM: Multi-Agent Large Language Models Frameworkcited by 15
  4. 2025Amplified Vulnerabilities: Structured Jailbreak Attacks on LLM-based Multi-Agent Debatecited by 8
  5. 2025Tree-of-Reasoning: Towards Complex Medical Diagnosis via Multi-Agent Reasoning with Evidence Treecited by 7
  6. 2025SPaRC: A Spatial Pathfinding Reasoning Challengecited by 5
  7. 2026Stop Hand-Holding Your Coding Agent: Engineering the Loops that Replace Step-by-Step Promptingcited by 5
  8. 2025Reallocating Attention Across Layers to Reduce Multimodal Hallucinationcited by 3
  9. 2026TrinityGuard: A Unified Framework for Safeguarding Multi-Agent Systemscited by 3
  10. 2026Multi-Agent Reasoning Improves Compute Efficiency: Pareto-Optimal Test-Time Scalingcited by 3
  11. 2026The Reasoning Trap: An Information-Theoretic Bound on Closed-System Multi-Step LLM Reasoningcited by 2
  12. 2026Misinformation Propagation in Benign Multi-Agent Systemscited by 2
11 more
  1. 2026Diagnosing and Mitigating Compounding Failures in Agentic Persuasion via Taxonomic Strategy Retrievalcited by 1
  2. 2026From Language Models to World-Acting Systems: Progress and Limits of Agentic AI across Digital, Social, Virtual, and Physical Environmentscited by 1
  3. 2026Loop-Back Authority in LLM Agent Teams: A Paired Experiment on Flat and Hierarchical Coordinationcited by 1
  4. 2025The Prompt Engineering Report Distilled: Quick Start Guide for Life Sciencescited by 0
  5. 2026Internal vs. External: Comparing Deliberation and Evolution for Multi-Agent Constitutional Designcited by 0
  6. 2026Enhancing Multi-Agent Communication through Attention Steering with Context Relevancecited by 0
  7. 2026Do LLM Debates Repeat Arguments Differently Across Languages?cited by 0
  8. 2026Scaling LLM-Driven Multi-Agent Systems: Design Principles and Architectural Scalability Analysiscited by 0
  9. 2026Where Reasoning Diverges: Localized Multi-Agent Debate for Multi-Hop Question Answeringcited by 0
  10. 2026What Does Multi-Agent LLM Debate Actually Change? A Layered Analysis of Disagreement and Answer Qualitycited by 0
  11. 2026From Process Loss to Assembly Bonus: Human-Grounded Diagnosis of Multi-Agent LLM Collaborationcited by 0

Similar, not cited

Close in wording, no citation either way

  1. 2025Talk Isn't Always Cheap: Understanding Failure Modes in Multi-Agent Debate
  2. 2026Agent Drift: Quantifying Behavioral Degradation in Multi-Agent LLM Systems Over Extended Interactions
  3. 2026The Cost of Consensus: Isolated Self-Correction Prevails Over Unguided Homogeneous Multi-Agent Debate
  4. 2026Not All Flips Are Conformity: Decomposing Stance Convergence in Multi-Agent LLM Debate
  5. 2025Can LLM Agents Really Debate? A Controlled Study of Multi-Agent Debate in Logical Reasoning
  6. 2025Revisiting Multi-Agent Debate as Test-Time Scaling: A Systematic Study of Conditional Effectiveness
  7. 2026Remember and Reweight: Enhancing Multi-Agent Debate with Experience Memory and Confidence Estimation
  8. 2023Should we be going MAD? A Look at Multi-Agent Debate Strategies for LLMs