You Changed Your Mind, The Model Didn't: Demystifying Intent in Multi-Turn Dialogue
Summary
该论文提出了 Intent-Eval 基准来系统研究大语言模型在多轮对话中追踪用户意图变化的能力，发现模型会将已被拒绝或替换的提议误当作有效要求（mentioned-as-in-effect 混淆），并提出 Intent-OPSD（基于决策的在策略自蒸馏框架）来提升模型对活跃意图的遵循能力。
View Cached Full Text
Cached at: 10/09/26, 02:22 PM
Paper page - You Changed Your Mind, The Model Didn’t: Demystifying Intent in Multi-Turn Dialogue
Source: https://huggingface.co/papers/2610.06496
Abstract
Whenalargelanguagemodelhandlesamulti-turntaskandauserproposesachangebutultimatelyrejectsit,themodelshouldcontinueasifnothingchanged.Wefindasurprisingfailure:merelymentioningarejectedchangecanderailtaskexecution,evenwhentheuser’sfinalintentremainsunchanged.Tosystematicallystudylanguagemodelbehaviorunderevolvinguserintent,weintroduceIntent-Eval,acontrolledbenchmarkspanningtoolactions,code,databases,andmathematics.Acrossdiversetasks,modelsarevulnerabletobothrejectedproposalsandsupersededrequirements,consistentwithmentioned-as-in-effectconfusion:conversationalcontentistreatedasactiverequirementsevenafterithasbeenrejectedorreplaced.Accuracydegradationcandeepenorpersistasinteractioncontinues,highlightingtheneedtodistinguishwhathasbeenmentionedfromwhatremainsineffect.Buildingonthisinsight,weproposeIntent-OPSD,adecision-conditionedon-policyself-distillationframeworkwithTeacherandStudentinitializedfromthesamemodel.ThefrozenTeacherprovidesactive-intentsupervisionfromthecompletetaskmatchingtheuser’sdecision,trainingtheStudentonthefulldialoguetofollowactiverequirementsreflectinguserintent.
View arXiv pageView PDFProject pageGitHub2Add to collection
Get this paper in your agent:
hf papers read 2610\.06496
Don’t have the latest CLI?curl \-LsSf https://hf\.co/cli/install\.sh \| bash
Models citing this paper0
No model linking this paper
Cite arxiv.org/abs/2610.06496 in a model README.md to link it from this page.
Datasets citing this paper0
No dataset linking this paper
Cite arxiv.org/abs/2610.06496 in a dataset README.md to link it from this page.
Spaces citing this paper0
No Space linking this paper
Cite arxiv.org/abs/2610.06496 in a Space README.md to link it from this page.
Collections including this paper0
No Collection including this paper
Add this paper to acollectionto link it from this page.
Similar Articles
When Users Change Their Minds: Measuring and Repairing Intent Drift in LLM Agents
This paper introduces IntentFlux, an executable benchmark for measuring intent drift in multi-turn LLM agents, and StateForge, an explicit active-requirement maintenance method that partially improves task scores when users change their minds mid-dialogue.
They Infer What You Meant: Models Represent Communicative Intent More Reliably Than They Act On It
This paper studies language models' failure to act on communicative intent despite robust internal representations. Using linear probes, the authors show intent is decodable from hidden states but often not reflected in outputs, and steering a late-layer direction can recover the intended behavior.
Enjoy Your Talk: A Human-Centered Benchmark for Multi-Turn Dialogue with Decoupled User Simulation, Target Modeling, and Judging
This paper introduces EYT-Bench, a human-centered benchmark for evaluating LLMs in multi-turn dialogues with a decoupled user simulation, target modeling, and judging design. It reveals that closed- and open-source models differ significantly on objective intent-tracking but are similar on subjective dimensions, and that reasoning improves objective tracking while persona format strongly affects trajectory spread.
Multi-Objective Aligned Small Language Model Framework for SUD Patient Dialogue Generation
该论文提出一个多目标对齐的小语言模型框架，用于生成物质使用障碍（SUD）咨询中的患者对话，结合知识蒸馏、人类标注偏好优化和注意力引导的奖励塑形，以提升认知一致性与临床真实性。
Enhancing Target-Guided Proactive Dialogue Systems via Conversational Scenario Modeling and Intent-Keyword Bridging
This paper proposes a method to enhance target-guided proactive dialogue systems by jointly modeling user profiles and domain knowledge as conversational scenarios and employing intent-keyword bridging to predict future dialogue turns.