Found in Conversation: LLMs Teach Themselves to Close the Multi-Turn Gap

• URL: http://arxiv.org/abs/2605.24432v1
• Date: Sat, 23 May 2026 06:58:52 GMT
• Title: Found in Conversation: LLMs Teach Themselves to Close the Multi-Turn Gap
• Abstract summary: We introduce Found in Conversation, a training framework where a model teaches itself to find and recover its single-turn competence.<n>View-Asymmetric Self-Distillation distills across two views of the same task information--single-turn view for the teacher, multi-turn view for the student--transferring strong single-turn behavior into weak multi-turn behavior.<n>Across model families (Llama, Qwen, Phi, and OLMo), FiC recovers at least 92% of single-turn performance and reaches 100% on two Llama backbones.
• Score: 51.23637310487053
• License: http://arxiv.org/licenses/nonexclusive-distrib/1.0/
• Abstract: Large Language Model (LLM) interactions are typically underspecified, with users clarifying all necessary details across multiple conversational turns. Yet recent work shows that LLMs perform far worse in this multi-turn setting than in a single turn with same information being available at once, a phenomenon termed "Lost-in-Conversation." However, bridging this gap effectively remains an open problem. Here we introduce Found in Conversation (FiC), a training framework where a model teaches itself to find and recover its single-turn competence given underspecified multi-turn prompts. We develop View-Asymmetric Self-Distillation, which distills across two views of the same task information--single-turn view for the teacher, multi-turn view for the student--transferring strong single-turn behavior into weak multi-turn behavior. This requires no stronger external teacher, which is unavailable as even frontier LLMs exhibit this gap. Across model families (Llama, Qwen, Phi, and OLMo) and sizes (3B-14B), FiC recovers at least 92% of single-turn performance and reaches 100% on two Llama backbones, yielding more efficient and helpful multi-turn conversations with single-turn capabilities intact.

Related papers

• Beyond Continuity: Challenges of Context Switching in Multi-Turn Dialogue with LLMs [6.89598705331814]
  We stress-test the multi-turn understanding of Large Language Models (LLMs)<n>We study the following two sub-tasks: detecting whether the user pivots or refines in the current turn, and shortlisting relevant context from previous turns.<n>We evaluate the zero-shot performance of ten LLMs (open-weight, closed-source and reasoning)
  arXiv  Detail & Related papers  (2026-05-10T02:28:23Z)
• TurnWise: The Gap between Single- and Multi-turn Language Model Capabilities [57.865325277036455]
  We introduce a new benchmark, TurnWiseEval, for multi-turn capabilities that is directly comparable to single-turn chat evaluation.<n>We additionally introduce our synthetic multi-turn data pipeline TurnWiseData which allows the scalable generation of multi-turn training data.<n>Our experiments with Olmo 3 show that training with multi-turn data is vital to achieving strong multi-turn chat performance.
  arXiv  Detail & Related papers  (2026-03-17T16:40:35Z)
• Intent Mismatch Causes LLMs to Get Lost in Multi-Turn Conversation [26.91734024759386]
  We argue that the root cause lies in an intent alignment gap rather than intrinsic capability deficits.<n>We propose to decouple intent understanding from task execution through a Mediator-Assistant architecture.
  arXiv  Detail & Related papers  (2026-02-07T03:41:04Z)
• MuCo: Multi-turn Contrastive Learning for Multimodal Embedding Model [57.89395815934156]
  Multi-Turn Contrastive Learning (MuCo) is a dialogue-inspired framework that revisits this process.<n>Experiments exhibit MuCo with a newly curated 5M multimodal multi-turn dataset (M3T)
  arXiv  Detail & Related papers  (2026-02-06T05:18:33Z)
• A Simple "Try Again" Can Elicit Multi-Turn LLM Reasoning [58.80217284841095]
  Multi-turn problem solving is critical yet challenging for Large Reasoning Models (LRMs) to reflect on their reasoning and revise from feedback.<n>Existing Reinforcement Learning (RL) methods train large reasoning models on a single-turn paradigm with verifiable rewards.<n>We introduce Unary Feedback as Observation (UFO) for reinforcement learning, which uses minimal yet common unary user feedback during iterative problem solving.
  arXiv  Detail & Related papers  (2025-07-18T18:07:38Z)
• LLMs Get Lost In Multi-Turn Conversation [44.26588510453331]
  Large Language Models (LLMs) are conversational interfaces.<n>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.
  arXiv  Detail & Related papers  (2025-05-09T15:21:44Z)
• Grounded Chain-of-Thought for Multimodal Large Language Models [66.04061083611863]
  We propose a new learning task for multimodal large language models (MLLMs) called Grounded Chain-of-Thought (GCoT)<n>GCoT is keen to helping MLLMs to recognize and ground the relevant visual cues step by step, thereby predicting the correct answer with grounding coordinates as the intuitive basis.<n>To facilitate this task, we also carefully design and construct a dataset called multimodal grounded chain-of-thought (MM-GCoT) consisting of 24,022 GCoT examples for 5,033 images.
  arXiv  Detail & Related papers  (2025-03-17T04:07:47Z)
• Retrospective Learning from Interactions [18.5871047885934]
  We introduce ReSpect, a method to learn from such signals in past interactions via retrospection without additional annotations.<n>We show how ReSpect gradually improves task completion rate from 31% to 82%, all without any external annotation.
  arXiv  Detail & Related papers  (2024-10-17T17:59:03Z)
• NoteLLM-2: Multimodal Large Representation Models for Recommendation [71.87790090964734]
  Large Language Models (LLMs) have demonstrated exceptional proficiency in text understanding and embedding tasks.<n>Their potential in multimodal representation, particularly for item-to-item (I2I) recommendations, remains underexplored.<n>We propose an end-to-end fine-tuning method that customizes the integration of any existing LLMs and vision encoders for efficient multimodal representation.
  arXiv  Detail & Related papers  (2024-05-27T03:24:01Z)
• Few-shot Partial Multi-view Learning [103.33865779721458]
  We propose a new task called few-shot partial multi-view learning. It focuses on overcoming the negative impact of the view-missing issue in the low-data regime. We conduct extensive experiments to evaluate our method.
  arXiv  Detail & Related papers  (2021-05-05T13:34:43Z)

This list is automatically generated from the titles and abstracts of the papers in this site.

This site does not guarantee the quality of this site (including all information) and is not responsible for any consequences.