LLMs Get Lost In Multi-Turn Conversation

2025/05/09 by Philippe Laban, Laban, Philippe, Hiroaki Hayashi +5 · 41 voices · 109 citations
Computer Science · #Natural Language Processing Techniques #Text Readability and Simplification #Topic Modeling #cs.CL #cs.HC

paper · pdf · doi:10.48550/arxiv.2505.06120

openalex publication_date 2025/05/09 · openalex created_date 2025/10/10 · openalex updated_date 2026/07/28

Abstract

Large Language Models (LLMs) are conversational interfaces. As such, LLMs have the potential to assist their users not only when they can fully specify the task at hand, but also to help them define, explore, and refine what they need through multi-turn conversational exchange. Although analysis of LLM conversation logs has confirmed that underspecification occurs frequently in user instructions, LLM evaluation has predominantly focused on the single-turn, fully-specified instruction setting. In this work, we perform large-scale simulation experiments to compare LLM performance in single- and multi-turn settings. Our experiments confirm that all the top open- and closed-weight LLMs we test exhibit significantly lower performance in multi-turn conversations than single-turn, with an average drop of 39% across six generation tasks. Analysis of 200,000+ simulated conversations decomposes the performance degradation into two components: a minor loss in aptitude and a significant increase in unreliability. We find that LLMs often make assumptions in early turns and prematurely attempt to generate final solutions, on which they overly rely. In simpler terms, we discover that *when LLMs take a wrong turn in a conversation, they get lost and do not recover*.

Citations

• DeepSeek-R1 incentivizes reasoning in LLMs through reinforcement learning
• 2 OLMo 2 Furious
• FABLES: Evaluating faithfulness and content selection in book-length summarization
• Art or Artifice? Large Language Models and the False Promise of Creativity
• LMSYS-Chat-1M: A Large-Scale Real-World LLM Conversation Dataset
• Lost in the Middle: How Language Models Use Long Contexts
• We're Afraid Language Models Aren't Modeling Ambiguity

Cited by

• Beyond Text-to-SQL: Can LLMs Really Debug Enterprise ETL SQL?
• Enjoy Your Talk: A Human-Centered Benchmark for Multi-Turn Dialogue with Decoupled User Simulation, Target Modeling, and Judging
• The Prompt Is Not the Query: How Request State Evolves Across Multi-Turn AI Conversations
• LLMs Get Lost in Evolving User Intent
• Reclaim Evaluation: A Lossy Memory Is Worse Than an Empty One
• SWE-Milestone: Evaluating AI Agents on Continuous Software Evolution
• Selective State-Space Adaptation and Retrieval for Language Model Reasoning
• SWE-Pruner Pro: The Coder LLM Already Knows What to Prune
• MemTrace: Tracing and Attributing Errors in Large Language Model Memory Systems
• Beyond Generalist LLMs: Specialist Agentic Systems for Structured Code Workflow Execution
• How Inference Compute Shapes Frontier LLM Evaluation
• Do Modules Stay in Their Lane? Role Drift in Compound LLM Systems
• Self-Compacting Language Model Agents
• Mitigating Scaffolding Collapse in Socratic Tutors via Representation Alignment
• δ-mem: Efficient Online Memory for Large Language Models
• Just Keep Prompting: Evaluating Repetitive Socratic Prompting in VLMs
• "AI Psychosis" in Context: How Conversation History Shapes LLM Responses to Delusional Beliefs
• MT-PingEval: Evaluating Multi-Turn Collaboration with Private Information Games
• Mitigating Conversational Inertia in Multi-Turn Agents
• Do LLMs Benefit From Their Own Words?
• Echoing: Identity Failures when LLM Agents Talk to Each Other
• BED-LLM: Intelligent Information Gathering with LLMs and Bayesian Experimental Design
• Style Amnesia: Investigating Speaking Style Degradation and Mitigation in Multi-Turn Spoken Language Models
• Topics as Proxies for Sociodemographics: How Conversational Context Affects LLM Answers
• ContextEcho: A Benchmark for Persona Drift in Long Agentic-Coding Sessions
• Eliciting Behaviors in Multi-Turn Conversations
• MusiChat: Vibe Composing for Music Creation
• Tokengeist: Multi-Turn Attribution Tracing in Agentic Conversations
• MedLoCoMo: A Long-Context Multi-Session Medical Dialogue Benchmark for Large Language Models
• ClarifyMT-Bench: Benchmarking and Improving Multi-Turn Clarification for Conversational Large Language Models
• ShareChat: A Dataset of Chatbot Conversations in the Wild
• Audio MultiChallenge: A Multi-Turn Evaluation of Spoken Dialogue Systems on Natural Human Interaction
• Towards Effective Model Editing for LLM Personalization
• Non-Resolution Reasoning (NRR): A Computational Framework for Contextual Identity and Ambiguity Preservation
• Context Branching for LLM Conversations: A Version Control Approach to Exploratory Programming
• Asynchronous Reasoning: Training-Free Interactive Thinking LLMs
• Decoding Human-LLM Collaboration in Coding: An Empirical Study of Multi-Turn Conversations in the Wild
• Reliable agent engineering should integrate machine-compatible organizational principles
• Rhea: Role-aware Heuristic Episodic Attention for Conversational LLMs
• Everything is Context: Agentic File System Abstraction for Context Engineering
• Invasive Context Engineering to Control Large Language Models
• The Trojan Knowledge: Bypassing Commercial LLM Guardrails via Harmless Prompt Weaving and Adaptive Tree Search
• MultiBanana: A Challenging Benchmark for Multi-Reference Text-to-Image Generation
• SABER: Small Actions, Big Errors -- Safeguarding Mutating Steps in LLM Agents
• SciEducator: Scientific Video Understanding and Educating via Deming-Cycle Multi-Agent System
• O-Mem: Omni Memory System for Personalized, Long Horizon, Self-Evolving Agents
• LLM Reinforcement in Context
• Human or LLM as Standardized Patients? A Comparative Study for Medical Education
• Measuring and Mitigating the Distributional Gap Between Real and Simulated User Behaviors
• Unsupervised Evaluation of Multi-Turn Objective-Driven Interactions
• Unlocking the Power of Multi-Agent LLM for Reasoning: From Lazy Agents to Deliberation
• From Passive to Proactive: A Hierarchical Multi-Agent Framework for Automated Medical Pre-Consultation
• CRAG-MM: Multi-modal Multi-turn Comprehensive RAG Benchmark
• Completion ≠ Collaboration: Scaling Collaborative Effort with Agents
• CMT-RAG: Complementary Memory Traces for Multi-turn Multi-hop RAG
• AtmosERC: Modeling Dialogue-Level Affective Atmosphere for Emotion Recognition in Conversation
• Evaluating Temporal Consistency in Multi-Turn Language Models
• Ming-Flash-Omni: A Sparse, Unified Architecture for Multimodal Perception and Generation
• Verifiable Accuracy and Abstention Rewards in Curriculum RL to Alleviate Lost-in-Conversation
• StreamingThinker: Large Language Models Can Think While Reading
• Offline Policy Evaluation of Multi-Turn LLM Health Coaching with Real Users
• ERGO: Entropy-guided Resetting for Generation Optimization in Multi-turn Language Models
• D-SMART: Enhancing LLM Dialogue Consistency via Dynamic Structured Memory And Reasoning Tree
• Reducing Belief Deviation in Reinforcement Learning for Active Reasoning
• SAGE: A Top-Down Bottom-Up Knowledge-Grounded User Simulator for Multi-turn AGent Evaluation
• Simulating Viva Voce Examinations to Evaluate Clinical Reasoning in Large Language Models
• PsychoPass: Geometric Profiling of Multi-Turn Adversarial LLM Conversations
• Format Inertia: A Failure Mechanism of LLMs in Medical Pre-Consultation
• CoBia: Constructed Conversations Can Trigger Otherwise Concealed Societal Biases in LLMs
• GTAlign: Game-Theoretic Alignment of LLM Assistants for Social Welfare
• Drift No More? Context Equilibria in Multi-Turn LLM Interactions
• ReviewerToo: Should AI Join The Program Committee? A Look At The Future of Peer Review
• Haystack Engineering: Context Engineering for Heterogeneous and Agentic Long-Context Evaluation
• Flipping the Dialogue: Training and Evaluating User Language Models
• COMPASS: Benchmarking Constrained Optimization in LLM Agents
• Taxonomy of User Needs and Actions
• Prompt reinforcing for long-term planning of large language models
• Bloom: Designing for LLM-Augmented Behavior Change Interactions
• Time-To-Inconsistency: A Survival Analysis of Large Language Model Robustness to Adversarial Attacks
• In-Place Feedback: Reliable Refinement for Multi-Turn Expert-LLM Collaboration
• MADS: Multi-Agent Dialogue Simulation for Diverse Persuasion Data Generation
• Which course? Discourse! Teaching Discourse and Generation in the Era of LLMs
• Improving the Efficiency of LLM Agent Systems through Trajectory Reduction
• Test-Time Policy Adaptation for Enhanced Multi-Turn Interactions with LLMs
• Non-Collaborative User Simulators for Tool Agents
• You Can't Steal Nothing: Mitigating Prompt Leakages in LLMs via System Vectors
• KnowMT-Bench: Benchmarking Knowledge-Intensive Long-Form Question Answering in Multi-Turn Dialogues
• TRUEBench: Can LLM Response Meet Real-world Constraints as Productivity Assistant?
• CIFLEX: Contextual Instruction Flow for Sub-task Execution in Multi-Turn Interactions with a Single On-Device LLM
• Cognitive Load Limits in Large Language Models: Benchmarking Multi-Hop Reasoning
• MIND: Lightweight and Effective Memory Injection Defense for LLM Agents via Intent-Aware Information Bottleneck
• AgentRadio: Passive Awareness for Long-Horizon Multi-Agent Collaboration
• The Topological Trouble With Transformers
• Co-FactChecker: A Framework for Human-AI Collaborative Claim Verification Using Large Reasoning Models
• A State-Update Prompting Strategy for Efficient and Robust Multi-turn Dialogue
• It Depends: Resolving Referential Ambiguity in Minimal Contexts with Commonsense Knowledge
• The Prompt Engineering Report Distilled: Quick Start Guide for Life Sciences
• Towards Better Health Conversations: The Benefits of Context-seeking
• Another Turn, Better Output? A Turn-Wise Analysis of Iterative LLM Prompting
• Measuring How (Not Just Whether) VLMs Build Common Ground
• CelloAI: Leveraging Large Language Models for HPC Software Development in High Energy Physics
• CyberSleuth: Autonomous Blue-Team LLM Agent for Web Attack Forensics
• Open-Universe Assistance Games
• User-Assistant Bias in LLMs
• Text-to-SQL Task-oriented Dialogue Ontology Construction
• User Feedback in Human-LLM Dialogues: A Lens to Understand Users But Noisy as a Learning Signal
• MemTool: Optimizing Short-Term Memory Management for Dynamic Tool Calling in LLM Agent Multi-Turn Conversations
• Goal Alignment in LLM-Based User Simulators for Conversational AI
• Debating Truth: Debate-driven Claim Verification with Multiple Large Language Model Agents

Discussions

• LLMs get lost in multi-turn conversation [hn, 374 points, 259 comments]
• “Our experiments confirm that all the top open- and closed-weight LLMs we test exhibit significantly lower performance in multi-turn conversations than single-turn, with an average drop of 39% across [bsky, 47 points, 5 comments]
• Recommended reading. An anecdotal observation many regular LLM users share, quantified. arxiv.org/pdf/2505.06120 [bsky, 22 points, 3 comments]
• arxiv.org/abs/2505.06120 [bsky, 6 points, 0 comments]
• LLMs in general are actually sort of bad at multi turn conversation, in that they have trouble distinguishing correct from incorrect: arxiv.org/abs/2505.06120 [bsky, 5 points, 1 comments]
• oops, forgot to add the link to the paper - arxiv.org/abs/2505.06120 [bsky, 4 points, 0 comments]
• "when LLMs take a wrong turn in a conversation, they get lost and do not recover" arxiv.org/abs/2505.06120 [bsky, 2 points, 0 comments]
• I saw this paper shared recently which quantifies the degradation when specifying everything up-front to an LLM vs taking multiple turns to specify the problem arxiv.org/pdf/2505.06120 [bsky, 2 points, 0 comments]
• Tip: If you use current LLMs for problem-solving or with a goal in mind, you need to be as comprehensive as possible in the first prompt to get the best performance. arxiv.org/abs/2505.06120 [bsky, 1 points, 1 comments]
• YES I HAVE NOTICED arxiv.org/abs/2505.06120 [bsky, 1 points, 0 comments]
• Large language models (LLMs) see a 39% drop in effectiveness in multi-turn dialogues versus single-turn tasks due to their tendency for hasty assumptions and premature response finalization, leading t [bsky, 1 points, 0 comments]
• Every major AI model gets dramatically worse the longer you talk to it. There are big losses in aptitude and even bigger increases in unreliability. [bsky, 1 points, 0 comments]
• This is a interesting microsoft paper arxiv.org/abs/2505.06120 [bsky, 1 points, 0 comments]
• LLMs performance degradation in multi-turn conversations [bsky, 1 points, 0 comments]
• LLMs Get Lost in Multi-Turn Conversation [bsky, 1 points, 0 comments]
• LLMs get lost in multi-turn conversation View Article | Join the HN Conversation Summary of HN discussion 🧵👇 #hacker-news [bsky, 0 points, 1 comments]
• (비전문가 대충 읽음) LLM 하곤 대화를 길게, 순차적으로 이어나갈 수록 오류가 폭증한다고. 반드시 질의를 한 번에 입력해야 한다는데 그래서 XML 형태로 입력하는게 더 효율적인가 싶고... 기존 AI 벤치마크는 전부 질답 한 번 기준이라고 한다. arxiv.org/abs/2505.06120 [bsky, 0 points, 0 comments]
• LLMs get lost in multi-turn conversation https://arxiv.org/abs/2505.06120 [bsky, 0 points, 0 comments]
• ⚡ Hackernews Top story: LLMs Get Lost in Multi-Turn Conversation [bsky, 0 points, 0 comments]
• LLMs get lost in multi-turn conversation https://arxiv.org/abs/2505.06120 (http://news.ycombinator.com/item?id=43991256) [bsky, 0 points, 0 comments]
• LLMs get lost in multi-turn conversation https://arxiv.org/abs/2505.06120 (http://news.ycombinator.com/item?id=43991256) [bsky, 0 points, 0 comments]
• LLMs Get Lost In Multi-Turn Conversation https://arxiv.org/abs/2505.06120 [bsky, 0 points, 0 comments]
• https://arxiv.org/abs/2505.06120 大規模言語モデル(LLM)は、複数ターンの対話で文脈を把握し、一貫性を保つのが苦手であるという研究。 特に、会話が長くなるにつれて、LLMは以前の発言を忘れ、矛盾した応答を生成する傾向がある。 この問題を解決するために、研究者たちは、LLMが会話履歴をより効果的に利用できるようにする新しい手法を提案している。 [bsky, 0 points, 0 comments]
• LLM get wronger the more they talk to people arxiv.org/abs/2505.06120 There’s something fundamentally broken about both #AI #metacognition and its #ActiveListening capability [bsky, 0 points, 0 comments]
• LLMs Get Lost in Multi-Turn Conversation view on hacker news [bsky, 0 points, 0 comments]
• LLMs get lost in multi-turn conversation https:// arxiv.org/abs/2505.06120 # arxiv # llm # llms [mastodon, 0 points, 0 comments]
• arxiv.org/pdf/2505.0612 Laban, et al, 2025, Arxiv, “LLMs get lost in Multi-Turn Conversations” Consistent with: Shumailov, et al, “AI models collapse when trained on recursively generated data,” Natur [bsky, 0 points, 0 comments]
• LLMs Get Lost In Multi-Turn Conversation arxiv.org/abs/2505.06120 [bsky, 0 points, 0 comments]
• LLMs Get Lost in Multi-Turn Conversation https://arxiv.org/abs/2505.06120 (https://news.ycombinator.com/item?id=43991256) [bsky, 0 points, 0 comments]
• LLMs get lost in multi-turn conversation https://arxiv.org/abs/2505.06120 (https://news.ycombinator.com/item?id=43991256) [bsky, 0 points, 0 comments]
• LLMs Get Lost in Multi-Turn Conversation https://arxiv.org/abs/2505.06120 (https://news.ycombinator.com/item?id=43991256) [bsky, 0 points, 0 comments]
• LLMs Get Lost In Multi-Turn Conversation arxiv.org/abs/2505.06120 [bsky, 0 points, 0 comments]
• arxiv.org/abs/2505.06120 [bsky, 0 points, 0 comments]
• arxiv.org/abs/2505.06120 [bsky, 0 points, 0 comments]
• "LLMs tend to (1) generate overly verbose responses, leading them to (2) propose final solutions prematurely in conversation, (3) make incorrect assumptions about underspecified details, and (4) rely [bsky, 0 points, 1 comments]
• "LLMs get lost in multi-turn conversation" LLMs struggle with keeping track of conversations. Without clear context, they get confused and give poor quality answers. Managing context is really importa [bsky, 0 points, 0 comments]
• https://bsky.app/profile/hackernews.com.web.brid.gy/post/3lp72k3mgfox2 [bsky, 0 points, 0 comments]
• LLMs Get Lost in Multi-Turn Conversation #HackerNews https://arxiv.org/abs/2505.06120 [bsky, 0 points, 0 comments]
• LLMs Get Lost in Multi-Turn Conversation https://arxiv.org/abs/2505.06120 https://news.ycombinator.com/item?id=43991256 [bsky, 0 points, 0 comments]
• LLMs get lost in multi-turn conversation https://arxiv.org/abs/2505.06120 arxiv.org [bsky, 0 points, 0 comments]
• LLMs Get Lost in Multi-Turn Conversation https://arxiv.org/abs/2505.06120 [bsky, 0 points, 0 comments]

Related

• Are LLMs All You Need for Task-Oriented Dialogue?
• MT-Eval: A Multi-Turn Capabilities Evaluation Benchmark for Large Language Models
• Flipping the Dialogue: Training and Evaluating User Language Models
• MultiChallenge: A Realistic Multi-Turn Conversation Evaluation Benchmark Challenging to Frontier LLMs
• Evaluating the Sensitivity of LLMs to Prior Context