LLMs Get Lost in Multi-Turn Conversation
- Authors: Philippe Laban, Hiroaki Hayashi, Yingbo Zhou, Jennifer Neville
- Published: 2025-05-09
- Source: https://arxiv.org/abs/2505.06120
- Type: Academic paper
- A minor loss in aptitude — reduced raw capability to solve the task
- A significant increase in unreliability — far more variance and inconsistency in outcomes
- Multi-turn degradation is universal: All top open- and closed-weight models tested showed significant performance drops in multi-turn settings, averaging 39% across six generation tasks.
- Unreliability is the dominant factor: The performance gap stems more from increased variance and inconsistency than from a uniform loss of capability.
- Early assumptions are costly: Models lock in premature assumptions and over-rely on early partial solutions, failing to self-correct as new information arrives in later turns.
- Evaluation gap: Standard benchmarks overwhelmingly test single-turn, fully-specified instructions, underrepresenting real user behavior where underspecification and iterative refinement are common.
- A Survey on Multi-Turn Interaction Capabilities of Large Language Models (arXiv:2501.09959)
- Evaluating LLM-based Agents for Multi-Turn Conversations: A Survey (arXiv:2503.22458)
- Aligning Query Representation with Rewritten Query and Relevance Judgments (DOI: 10.1145/3627673.3679534)
- An Empirical Analysis on Multi-turn Conversational Recommender Systems (DOI: 10.1145/3626772.3657893)
- Beyond Whole Dialogue Modeling: Contextual Disentanglement (arXiv:2504.17427)
- CHIQ: Contextual History Enhancement for Improving Query Rewriting (arXiv:2406.05013)
- Original paper: LLMs Get Lost In Multi-Turn Conversation — https://arxiv.org/abs/2505.06120
Summary
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, the authors perform large-scale simulation experiments to compare LLM performance in single- and multi-turn settings. Their experiments confirm that all the top open- and closed-weight LLMs tested exhibit significantly lower performance in multi-turn conversations than in 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:
The authors find that LLMs often make assumptions in early turns and prematurely attempt to generate final solutions, on which they overly rely. In simpler terms: *when LLMs take a wrong turn in a conversation, they get lost and do not recover*.
Key Findings
Implications for Search / RAG / Personalization
1. Architecture: Cascaded retrieval + reranking + generation remains mainstream, but agentic paradigms increasingly treat retrieval count and strategy as learnable decisions. 2. Data: High-quality instruction data and session/click logs matter equally; synthetic data must guard against knowledge leakage and distribution shift. 3. Evaluation: The gap between offline metrics and online satisfaction is widening; LLM-as-judge should be cross-validated with human evaluation. 4. Product: Latency, cost, interpretability, and safety policies are hard constraints for industrial deployment — one cannot optimize only academic benchmarks.