PaperlayerView exact revision on arXiv

Source-linked AI summary

You Changed Your Mind, The Model Didn't: Demystifying Intent in Multi-Turn Dialogue

Junle Chen, Wei Chen, Zhengjun Huang, Zhoujin Tian, Yuxuan Liu, Kai Wang, Rui Chen, Xiaofang Zhou

arXiv:2610.06496v1cs.AI

TL;DR

Models can confuse mentioned requirements with those still in effect as users clarify or change tasks across turns. Intent-Eval measures this behavior, and Intent-OPSD trains on full dialogues using decision-matched Teacher supervision; results show degradation from inactive content and improved mean accuracy over base models.

  • Problem

    Prior benchmarks do not compare clarifications, accepted changes, and rejected changes on the same tasks, despite users introducing all three as intent evolves.

    Source: §1 · p. 1

  • Method

    Intent-OPSD selects a Teacher’s single-turn task based on the user’s decision and trains a Student on the full dialogue.

    Source: §1 · p. 2

  • Results

    Across eight models, Retained and Revised averaged 8.06 and 5.55 pp below Original, respectively; Intent-OPSD improved mean accuracy over corresponding base models by 10.81 pp.

    Source: §1 · pp. 1–2

  • Takeaways & Limitations

    The findings highlight the importance of distinguishing what was said from what remains in effect.

    Source: §6 · p. 10

    Intent-Eval focuses on decisions about one selected requirement; multiple interacting requirements and more varied user behavior remain future work.

    Source: F Limitations and Discussion · p. 23

Abstract

from arXiv · show

When a large language model handles a multi-turn task and a user proposes a change but ultimately rejects it, the model should continue as if nothing changed. We find a surprising failure: merely mentioning a rejected change can derail task execution, even when the user's final intent remains unchanged. To systematically study language model behavior under evolving user intent, we introduce Intent-Eval, a controlled benchmark spanning tool actions, code, databases, and mathematics. Across diverse tasks, models are vulnerable to both rejected proposals and superseded requirements, consistent with mentioned-as-in-effect confusion: conversational content is treated as active requirements even after it has been rejected or replaced. Accuracy degradation can deepen or persist as interaction continues, highlighting the need to distinguish what has been mentioned from what remains in effect. Building on this insight, we propose Intent-OPSD, a decision-conditioned on-policy self-distillation framework with Teacher and Student initialized from the same model. The frozen Teacher provides active-intent supervision from the complete task matching the user's decision, training the Student on the full dialogue to follow active requirements reflecting user intent.

1 Introduction

The paper studies whether models follow the requirements that remain active as users clarify tasks or propose changes they accept or reject. It introduces Intent-Eval to compare these interactions and Intent-OPSD to train a dialogue-reading Student from a Teacher’s decision-matched task.

1 Introduction

  • Across eight models, Original multi-turn disclosure lowers mean accuracy by 36.30 pp versus single-turn controls; Neutral adds 4.43 pp, while Retained and Revised average 8.06 and 5.55 pp below Original.The benchmark covers Actions, Code, Database, and Math; proposed changes further degrade performance whether accepted or rejected, including in single-turn controls.Source: §1 · p. 1
  • As clarification or rejected changes accumulate, accuracy can fall; new Retained and Revised errors often contain inactive content, consistent with mentioned-as-in-effect confusion.The analysis suggests models treat content that was said as still in effect.Source: §1 · p. 2
  • Intent-OPSD trains a Student on the full dialogue to match a frozen Teacher given the complete single-turn task that reflects the user’s decision.The Teacher uses the original task after rejection and the revised task after acceptance; at inference, the Student needs neither Teacher nor single-turn prompt.Source: §1 · p. 2
  • Intent-Eval compares four controlled multi-turn conditions with matched single-turn controls across four domains, while Intent-OPSD improves mean accuracy over base models by 10.81 pp across four models and four domains.The benchmark also tests whether degradation deepens or persists in longer dialogues.Source: §1 · p. 2

2 Intent-Eval: Evaluating Active Intent in Multi-Turn Dialogue

Intent-Eval tests whether models track active requirements across matched task interactions, using four multi-turn conditions, single-turn controls, and extensions to longer dialogues. Its source tasks are revised and checked so the conditions differ in controlled ways.

2 Intent-Eval: Evaluating Active Intent in Multi-Turn Dialogue

  • Intent-Eval contains 414 source tasks across four domains; each yields four multi-turn conditions and four single-turn controls, for 3,312 evaluation instances per model.The controls help separate multi-turn effects from task difficulty, and the benchmark examines interference across interaction depth.Source: §2 · p. 2, Figure 2 · p. 3

2.1 Multi-Turn Conditions

  • Original continues disclosure, Neutral adds clarifications, Retained rejects a proposal, and Revised accepts the same proposal; only Revised changes the active requirement and reference answer.Only the final response is scored; Original, Neutral, and Retained use the original answer, while Revised uses the revised answer.Source: §2.1 · p. 3

2.2 Data and Construction

  • The 414 tasks are adapted from LiC across Actions, Code, Database, and Math, with one candidate revision per task selected through constraint filtering and human review.Revisions must be unambiguous, atomic, and verifiable; they are rendered as shared proposals accepted in Revised or rejected in Retained, while Neutral adds clarifying questions.Source: §2.2 · pp. 3–4

2.3 Controls and Interaction Extensions

  • Single-turn controls distinguish revised-task difficulty from decision effects, while extended protocols test whether clarification or rejection interference accumulates or persists across added interaction blocks.Controls pair Original, Neutral, and Retained with SOrig and Revised with SDir; added-block count k ranges from 0 to 4.Source: §2.3 · p. 4

3 Intent-OPSD: Learning Active Intent from Multi-Turn Dialogue

Intent-OPSD uses a complete single-turn task matched to the user’s decision to supervise a Student that reads the full dialogue. It trains on the Student’s own answer prefixes and uses the Student alone at inference.

3 Intent-OPSD: Learning Active Intent from Multi-Turn Dialogue

  • Intent-OPSD pairs a Student receiving the full dialogue with a frozen Teacher receiving the complete single-turn task that has the same active requirements.The Teacher’s task remains original after rejection and switches to revised after acceptance.Source: §3 · p. 4, §3.1 · p. 5, Figure 3 · p. 5

3.1 Active-Task Supervision and Source Selection

  • Training sources are retained only when the Teacher solves both complete original and revised tasks and passes their verifiers; references and verifier results are not shown to either model.The Teacher supervises the Student under the matching multi-turn conditions.Source: §3.1 · p. 5

3.2 On-Policy Self Distillation of Intent

  • For each condition, the Student samples the final answer on-policy, and the frozen Teacher scores the Student-generated prefixes; the Student is updated while the Teacher remains frozen.The loss uses Jensen–Shannon divergence over the full vocabulary, excludes intermediate replies, and averages the four condition losses equally; inference needs neither Teacher nor complete-task prompt.Source: §3.2 · p. 5

4 Experiments

Across four research questions, the experiments measure multi-turn accuracy, characterize errors, test whether interference accumulates or persists, and evaluate decision-conditioned distillation. Intent-OPSD improves mean accuracy over Base and SFT, while the other analyses find degradation from turn distribution, clarification, and proposed changes.

4.1 Multi-turn accuracy across conditions (RQ1)

  • Keeping the same task across turns lowers mean accuracy by 36.30 pp; Neutral clarification adds 4.43 pp, while proposed changes lower multi-turn accuracy by 5.55 pp when accepted and 8.06 pp when rejected, relative to Original.In single-turn controls, Revised and Retained reduce accuracy by 5.74 and 7.52 pp, respectively, compared with stating the same active task directly. Accuracy remains at least 85% across all multi-turn Code conditions.Source: §4.1 · p. 7

4.2 Characterizing errors under Retained and Revised (RQ2)

  • Inactive content appears in 59.63% of new Retained errors and 30.97% of new Revised errors across ten models.In Math, inclusion reaches 67.70% under Retained and 47.94% under Revised; the paper suggests reused intermediate calculations may preserve rejected or superseded values.Source: Figure 4 · p. 7, §4.2 · p. 7

4.3 Does interference accumulate or persist? (RQ3)

  • At final depth, accuracy falls by 20.77 pp for Retained and 8.13 pp for Neutral relative to Original.Across four models, the first added block produces the largest average single-step decline, while accuracy generally continues to decline with added blocks.Source: Figure 5 · p. 8, §4.3 · p. 8
  • Four added Neutral blocks lower accuracy by 2.72 pp under Retained and 3.50 pp under Revised relative to their decision baselines.Mean accuracy stays below the Retained or Revised baseline across all tested clarification depths.Source: §4.3 · p. 8

4.4 Can decision-conditioned distillation improve accuracy? (RQ4)

  • Intent-OPSD improves mean accuracy by 10.81 pp over Base and 3.73 pp over SFT across four models.It outperforms Base in all 16 model–domain pairs and SFT in 13; gains over Base suggest active-task supervision helps the Student follow active requirements despite rejected or superseded content.Source: §4.4 · pp. 8–9, Table 2 · p. 9
  • Intent-OPSD improves Actions accuracy by 21.46 pp over Base, while Code shows smaller gains.The authors suggest the method may help when intent tracking is a bottleneck, but structural or implementation errors can still cause failure.Source: §4.4 · p. 9

5 Related Work

Prior work studies general multi-turn competence, intent alignment, and evolving requirements in different settings. Intent-Eval jointly examines clarification, rejection, and revision with matched single-turn controls and extended dialogues; Intent-OPSD conditions Teacher supervision on the user’s conversational decision.

Related multi-turn dialogue benchmarks

Related evolving-intent benchmarks

On-Policy Distillation

  • Intent-OPSD conditions Teacher supervision on the user’s conversational decision and distills the resulting active requirements into a Student operating on the full dialogue.Related approaches include on-policy self-distillation with reference solutions (Zhao et al., 2026), conversational transfer of single-turn competence (Chen et al., 2026b), and distillation from cleaned dialogue histories (Zheng et al., 2026).Source: §5 · pp. 9–10

6 Conclusion

Intent-Eval compares clarification, rejection, and revision with matched single-turn controls across four domains, revealing degradation that can deepen or persist in longer dialogues. Intent-OPSD aligns Teacher supervision with the active requirements and trains the Student on the full dialogue.

6 Conclusion

  • Intent-Eval compares clarification, rejection, and revision with matched single-turn controls across four domains, revealing degradation that can deepen or persist in longer dialogues.The analysis identifies mentioned-as-in-effect confusion, in which models treat inactive content as still operative.Source: §6 · p. 10
  • Intent-OPSD aligns the Teacher’s single-turn task with requirements that remain in effect and trains the Student on the full dialogue.The method mitigates the issue without requiring the Teacher at inference.Source: §6 · p. 10
  • The findings support distinguishing what was said from what remains in effect when evaluating evolving user intent.The paper identifies longer and more open-ended interactions with richer intent evolution as future work.Source: §6 · p. 10

Supplementary material

The supplementary material adds construction and scoring details, plus a complete worked task example.

Supplementary material

A.1 Construction details

The construction details describe how candidate task revisions are selected, checked, rendered into dialogue conditions, and scored. A worked onion-purchase example traces the resulting task states and references.

A.1 Construction details

  • Selected candidates are discarded when references are invalid or the original and revised states cannot be distinguished by the applicable verifier or output-contract checks.Source: §A.1 · p. 13
  • Domain edits target tool arguments, entry-point names, SQL requirements, or numerical premises, with reference updates and checks tailored to each task type.For Actions, edits must satisfy the tool schema and task totals; Code checks entry-point names and execution tests; Database revisions are discarded if original and revised queries return the same result; Math references are recomputed and checked for validity and difference.Source: §A.1 · pp. 13–14
  • Scoring compares generated outputs with the active reference, using domain-specific checks such as execution results, function names, or parsed tool calls.Source: §A.1 · pp. 13–14
  • Selected revisions are reviewed against source tasks, rendered with templates for proposals and decisions, and assigned to conditions using a stable record-based wording choice.All 414 selected revisions pass automated checks that Neutral reveals neither value and Retained and Revised share the proposal.Source: §A.1 · p. 15

A.2 A complete worked example

  • In the onion example, the bag count changes from four to two, while weight and price stay fixed; the reference cost is $300 originally and $150 after acceptance.Rejection keeps the original $300 reference; acceptance uses $150.Source: §A.2 · p. 15
  • The example places the proposal after the bag count is introduced and before the remaining weight and price information is disclosed.Source: §A.2 · pp. 15–16

B Evaluation Settings and User Simulation

The evaluation settings describe model generation, simulated user disclosure, and reliability checks. The interaction protocols extend clarification and rejection patterns, while Intent-OPSD conditions Teacher supervision on the user's decision.

B.1 Models and generation settings

  • Local and API model generation settings are specified, with API sampling at defaults and single- and multi-turn output limits matched.The settings table covers local models and the user simulator, with output limits measured in tokens per response.Source: §B.1 · p. 16, Table 5 · p. 16

B.2 User simulation

  • The user simulator reveals task shards over turns, then the protocol inserts fixed clarification or proposal-and-decision messages after the selected requirement is introduced.After an event, mandatory-disclosure and recovery procedures require eligible task information to resume when needed.Source: §B.2 · p. 17

B.3 User simulator reliability

  • Task-relevant information was conveyed in 2,284 of 2,317 shard-bearing messages (98.6%), and all 2,345 inspected messages (100%) were appropriately contextualized.Source: §B.3 · p. 18

C.1 Accumulation: repeated clarification or rejection

  • Accumulation protocols add repeated clarification blocks or distinct proposal-and-rejection blocks about the same requirement, preserving the original task and reference.Each event message is a separate user turn followed by an assistant reply.Source: §C.1 · p. 18

C.2 Persistence: clarification after a decision

  • Persistence protocols add neutral clarification after rejection or acceptance while keeping the corresponding active task unchanged at every depth.Source: §C.2 · pp. 18–19

D Intent-OPSD Training and Evaluation

  • Intent-OPSD selects the original Teacher task after rejection and the revised Teacher task after acceptance, while training the Student on the full dialogue.The Teacher is aligned with the requirements that remain active after the user's decision.Source: D Intent-OPSD Training and Evaluation · p. 19

D.1 Training data and conversation construction

The appendix describes training data, conversation construction, and objectives, then reports supplementary results showing lower multi-turn accuracy in two additional models, especially under Retained.

D.1 Training data and conversation construction

  • Training pools use BFCL for Actions, Spider for Database, GSM8K for Math, and MBPP, HumanEval, and LiveCodeBench for Code.Database candidates are checked for task fit and differing query results before Teacher admission.Source: §D.1 · p. 19, Table 6 · p. 19
  • Each task is split into information shards revealed at most one per turn, with condition-specific messages added to histories generated online as the Student changes.The conversation construction follows LiC's sharded setup.Source: §D.1 · p. 19

D.2 Training objectives and generation budgets

  • Intent-OPSD uses final-answer distillation, with correct and incorrect Student answers contributing to a loss averaged over conditions and sources.The comparison with matched SFT holds base model, sources, condition weights, and optimization schedule constant, contrasting on-policy distribution matching with hard-target learning under active-task supervision.Source: §D.2 · pp. 19–20
  • Training sequences allow up to 32,768 tokens for Code, Database, and Math, versus 8,192 for Actions, whose function-call format motivates a more compact budget.Intermediate replies and final answers have separate limits; evaluation uses common multi-turn generation settings.Source: §D.2 · p. 20

E.1 Overall performance (RQ1)

  • In every domain, both additional models score lower on Multi Original than Single Original, and lower again on Multi Retained.These supplementary models are Qwen3-4B-Instruct-2507 and Qwen2.5-7B-Instruct, outside the eight-model aggregate in Table 1.Source: §E.1 · p. 20, Table 8 · p. 20

E.2 Error diagnostics and examples (RQ2)

Error analysis finds inactive content in many new Retained and Revised errors. Supplementary depth results show further accuracy losses during interaction, while Intent-OPSD remains above Base across tested paths and depths.

E.2 Error diagnostics and examples (RQ2)

  • Inactive content means a rejected proposal in Retained or an original requirement replaced in Revised, detected with domain-specific checks on complete final responses.The analysis selects outputs correct in Original but incorrect in Retained or Revised; checks cover task-specific values and include explanations and code comments.Source: §E.2 · p. 21
  • 59.63% of new Retained errors and 30.97% of new Revised errors contain inactive content across ten models.Figure 6 shows model-level content-inclusion rates among Original-correct outputs that become incorrect in the named condition.Source: Figure 6 · p. 21, §E.2 · p. 21

E.3 Interaction depth in six additional models (RQ3)

  • Across six additional models and four interaction paths, the first added block causes the largest single-step accuracy drop; repeated Retained has larger average losses than Neutral, especially in Math.After either decision, further clarification leaves mean accuracy below the corresponding baseline at the final depth, with effects varying by model and domain.Source: §E.3 · p. 22

E.4 Intent-OPSD across interaction depths (RQ4)

  • 9.61–11.74 pp over Base and 2.16–3.86 pp over SFT at k = 1–4; Intent-OPSD beats Base at every depth in all four pooled paths.Table 9 reports accuracy across interaction depths, with each path pooling 1,488 accumulation or 1,656 persistence evaluations.Source: Table 9 · p. 23, §E.4 · p. 23

F Limitations and Discussion

The discussion identifies limits in interaction coverage and in transferring supervision from benchmark specifications to natural conversations.

F Limitations and Discussion

  • Intent-Eval tests decisions about one selected requirement; multiple interacting requirements and more varied user behavior remain outside its current coverage.The benchmark makes active requirements and reference answers explicit and verifiable through controlled clarification, rejection, and revision events.Source: F Limitations and Discussion · p. 23
  • Intent-OPSD improves overall accuracy over Base and matched SFT, but mean Retained accuracy is similar to SFT and rejected changes remain challenging.The authors identify handling rejected changes as a direction for further method development.Source: F Limitations and Discussion · p. 23
  • Training assumes complete active-task specifications supplied by benchmark construction; applying the approach to natural conversations requires reconstructing active requirements from messages and decisions.The paper proposes extracting and updating needs from dialogue, with consistency checks and human review for ambiguous cases, as future work.Source: F Limitations and Discussion · p. 23
Loading 2610.06496v1…