PaperlayerView exact revision on arXiv

Source-linked AI summary

Clarification Is Not Correction: LLMs Fail to Let Go

Jianzhe Lin, Xiaolin Li, Fei Wang, Robert Douglas, Rajeshkumar Golani, Jubin Chheda

arXiv:2609.25337v1cs.AI

TL;DR

Dialogue models may retain a correction yet continue answering from an earlier interpretation, a failure the paper frames as early posterior collapse. The authors test this with controlled tasks across writing, planning, and coding; information order changes outcomes, with coding especially vulnerable. They argue for managing uncertainty and invalidating stale assumptions, while noting limits in model, task, evaluation, and intervention scope.

  • Problem

    Dialogue failures are often framed as memory failures, but the paper studies cases where a short, recent, explicit clarification is present and the final output still reflects an obsolete interpretation.

    Source: §1 · p. 2

  • Method

    The authors use controlled writing, planning, and coding dialogues to test early posterior collapse and diagnose residual traces of invalidated hypotheses in final outputs.

    Source: §1 · p. 3, §3 · p. 5, §4.1 · p. 6

  • Results

    Across thousands of trials, information order changed outcomes, coding showed especially high contamination, and common prompting or memory strategies did not reliably preserve uncertainty.

    Source: §1 · p. 3

  • Takeaways & Limitations

    Interactive systems should be evaluated for whether they discard stale hypotheses after clarification, not only for memory retention or final-task success.

    Source: §6 · p. 9

    The study evaluates two Gemini models on controlled tasks with a model-based judge, uses an imperfect contamination metric, and tests state management only through prompt-level approximations.

    Source: §5 · p. 9

Abstract

from arXiv · show

Dialogue failures in language models are usually framed as memory failures: context too long, summaries lossy, a constraint forgotten. We argue this misses a deeper problem: in many conversations the model does not forget, it commits too early. An ambiguous early turn collapses into a single hidden interpretation, and later clarification is filtered through that commitment. We call this early posterior collapse: unresolved user intent collapsing into a committed task state before ambiguity is resolved. We study it with controlled dialogue tasks in writing, planning, and coding using Gemini-2.5-Pro and Gemini-2.5-Flash. Across thousands of trials, the same information in different orders yields different outcomes, even when the final dialogue contains equivalent task-relevant information. This order effect suggests later clarification is treated as extra context rather than a corrective signal: it refines a stale task state without invalidating it. Coding tasks are especially vulnerable, suggesting early assumptions get embedded in structured artifacts such as interfaces and control flow. Standard prompting and memory strategies do not reliably help: summaries can collapse ambiguity, and chain-of-thought can reduce explicit wrong commitment in reasoning traces without improving final task success. These findings motivate uncertainty-preserving state management. If assistants cannot let go of early interpretations, robustness cannot rely on post hoc correction alone; it must keep ambiguous early turns from hardening into one task state. Assistants should hold tentative hypotheses while ambiguity remains, ask before executing when high-impact ambiguity persists, and rebuild from a revised state when later evidence invalidates an earlier reading. Rather than one prompting fix, we aim to redirect research for interactive LLMs from retaining more context toward preserving uncertainty.

1 Introduction

Dialogue failures can persist even when a short, explicit clarification is present: an early interpretation may remain embedded in the answer after it should have been invalidated. The paper studies this as early posterior collapse and tests its predictions across writing, planning, and coding tasks.

1 Introduction

  • The paper distinguishes stale-state failure from memory failure: a model may have the correction in context yet fail to remove the obsolete interpretation.The correction may be short, recent, and explicit, while the final output still carries lexical, semantic, or structural traces of the earlier reading.Source: §1 · pp. 2–3
  • Across thousands of trials, information order changed downstream outcomes, coding showed especially high contamination, and common prompting or memory strategies did not reliably preserve uncertainty.The intervention experiments are mechanism checks: rollback reduced measured contamination, while a two-phase response policy eliminated wrong commitment and contamination in a small controlled setting.Source: §1 · p. 3
  • Early posterior collapse occurs when unresolved user intent becomes a committed task state before ambiguity is resolved, leaving old-hypothesis residue in final outputs.The paper uses old-hypothesis contamination to diagnose residual traces of interpretations that later clarification should have invalidated.Source: §1 · p. 3
  • The study provides a controlled diagnostic suite across writing, planning, and coding, examining how information order, task structure, and memory compression shape contamination.Its research agenda is uncertainty-preserving state management: track whether assumptions are tentative, confirmed, unresolved, or invalidated, rather than merely retaining their text.Source: §1 · p. 3

2 Related Work

Related work examines dialogue degradation, clarification under ambiguity, memory, and reasoning-time methods. This paper focuses on whether a corrected interpretation leaves residue in the final output.

Dialogue degradation and path dependence

  • Dialogue benchmarks study extended-conversation degradation, feedback integration, and structured flows, while this work targets traces remaining after an explicit correction [Laban et al., 2025; Li et al., 2025a; Zhang et al., 2025; Li et al., 2025b; Rawal et al., 2025].The paper asks whether the final output still reflects an interpretation invalidated by the dialogue, rather than measuring only performance drops.Source: §2 · p. 3

Ambiguity, clarification, and underspecified intent

Memory, summarization, and dialogue state

Reasoning-time methods and posterior tracking

  • Reasoning-time methods can expose or revise trajectories, but hypothesis enumeration alone does not ensure invalidated task states are removed [Wei et al., 2022; Wang et al., 2023; Yao et al., 2023; Shinn et al., 2023].Uncertainty-preserving inference requires maintaining tentative hypotheses, invalidating contradicted assumptions, and regenerating from a revised belief state.Source: §2 · p. 4

3 Theoretical Framing: Conversational Posterior Collapse

The paper frames dialogue as inference over latent task hypotheses that should remain uncertain while user intent is unresolved. Early posterior collapse occurs when the model commits prematurely and later clarification fails to remove the stale hypothesis from generation.

Conversational posterior framing

  • An assistant should infer a latent task state from dialogue while maintaining a posterior over plausible hypotheses and preserving uncertainty when several remain possible.A later clarification should update that posterior, increasing support for compatible hypotheses and reducing support for contradicted ones.Source: §3 · p. 4

Early posterior collapse

  • Early posterior collapse is premature commitment to an earlier task hypothesis before ambiguity is resolved, with the later clarification failing to remove it from generation.The paper operationalizes the residual effective probability through lexical, semantic, or structural traces of the invalidated hypothesis in the final output.Source: §3 · p. 5

Distinguishing forgetting from contamination

  • Unlike forgetting, posterior-collapse failure occurs when the relevant clarification is present but the model fails to update the epistemic status of an earlier hypothesis.The output can acknowledge that the old hypothesis is invalid while still being shaped by it; this residual influence is old-hypothesis contamination.Source: §3 · p. 5

Compression and explicit state tracking

  • Memory compression can induce collapse when a summary preserves facts but converts multiple plausible interpretations into one apparent state.Chain-of-thought or an explicit single-state ledger can also leave ambiguity settled prematurely; explicit representations may become additional anchors rather than safeguards.Source: §3 · pp. 5–6

Design principle

  • The design principle is to separate hypothesis management from answer generation, clarify unresolved slots, and reconstruct the task state after evidence invalidates an earlier reading.The paper proposes updating or rebuilding from the clarified state rather than merely patching the previous output.Source: §3 · p. 6

4 Diagnostic Evidence

The diagnostic suite tests whether information order, task structure, and state representations shape old-hypothesis contamination. Its reported results show lower success after ambiguous-first interactions and greater vulnerability in coding tasks.

4 Diagnostic Evidence

  • The suite tests path dependence, structural vulnerability, state-representation failures, and whether rollback or two-phase resolution reduces contamination.The predictions include that stronger structural commitments increase fragility and that summaries, chain-of-thought, or explicit ledgers may fail to preserve epistemic status.Source: §4 · p. 6

4.1 Experimental Setup

  • Tasks use an underspecified initial request followed by a turn resolving ambiguity, with final answers evaluated for satisfying clarified intent without retaining invalidated content.The three families are writing, planning, and coding; the study evaluates task success, constraint satisfaction, revision, commitment, clarification, and contamination.Source: §4.1 · p. 6

4.2 Path Dependence: Same Information, Different Order

  • Both models had lower success in ambiguous-first than clarified-first interactions despite equivalent final task-relevant information.The authors treat this as an initial path-dependence signature, not conclusive evidence, because order effects can overlap with recency bias and contamination was not monotonic for every model.Source: §4.2 · p. 6, Table 1 · p. 7

4.3 Structural Embedding: Coding Is Most Vulnerable

  • Coding had the lowest success and highest contamination, roughly twice the contamination observed in writing.The authors connect this vulnerability to assumptions embedded in interfaces, data schemas, dependencies, and control flow, making them harder to remove through local revision.Source: §4.3 · p. 7, Table 2 · p. 7

4.5 Memory and Reasoning Baselines Do Not Preserve Uncertainty

Conventional memory and reasoning baselines do not reliably preserve uncertainty, while intervention results support rebuilding task state after clarification. A two-phase policy that resolves uncertainty before execution eliminates wrong commitment and measured contamination in a small controlled setting.

4.5 Memory and Reasoning Baselines Do Not Preserve Uncertainty

  • No baseline eliminates contamination; summary memory improves success relative to raw history but increases contamination, chain-of-thought does not reliably improve task success, and a single-state ledger increases contamination.The baselines test raw history, summary memory, chain-of-thought, and an explicit single-state ledger.Source: §4.5 · p. 7, Table 4 · p. 8

4.6 State Representations Can Collapse or Distract

  • Most summaries collapse ambiguity, while chain-of-thought rarely states wrong commitments but still underperforms in task success.Both representations can fail to preserve the epistemic status of assumptions.Source: §4.6 · p. 8

4.7 Interventions: Rollback and Two-Phase Policies

  • After clarification, explicit rollback discards prior assumptions and rebuilds from clarified requirements, improving success by 10 points and reducing measured contamination to zero in the supplementary setting.The authors treat this as support for rebuilding rather than as evidence that a rollback prompt is a general solution.Source: Table 6 · p. 8, §4.7 · p. 8
  • In a small controlled setting, the two-phase policy eliminates both wrong commitment and measured contamination.It resolves uncertainty before execution, then regenerates from the clarified state.Source: §4.7 · p. 8

5 Discussion, Alternative Views, and Limitations

The discussion frames dialogue robustness as epistemic state management: systems must track whether assumptions are tentative, confirmed, contradicted, or unresolved. Its prompt-level policy results offer initial evidence for rebuilding from clarified state, while the diagnostics and study design have important limits.

5 Discussion, Alternative Views, and Limitations

  • In this setting, the two-phase policy eliminates both wrong commitment and measured contamination, unlike policies that permit execution under uncertainty.The paper cautions that this is a small, prompt-level mechanism check, not evidence that a simple prompt solves the general problem.Source: Table 7 · p. 9, §5 · p. 9
  • Dialogue robustness requires preserving assumptions’ epistemic status, not merely retaining their content in longer context or compressed memory.The paper identifies confirmed, tentative, contradicted, and unresolved as relevant statuses.Source: §5 · pp. 8–9
  • The diagnostic experiments do not fully separate stale-state contamination from recency bias, prompt-position effects, or instruction-following brittleness.The paper treats contamination as a distinct, under-measured residue, not the sole explanation for dialogue degradation.Source: §5 · p. 9
  • Visible reasoning or a single-state ledger does not ensure uncertainty is preserved; the missing capability is updating, arbitrating, and invalidating assumptions as evidence arrives.The paper frames posterior tracking as state management rather than merely a prompting pattern.Source: §5 · p. 9
  • The study evaluates two Gemini models on controlled tasks with a model-based judge, and its contamination metric—especially for coding structure—is imperfect.Rollback and two-phase tests are small prompt-level checks; broader implementations may require explicit controllers, epistemic memory schemas, or training objectives.Source: §5 · p. 9
  • Future work should test whether this failure is amplified in tool-using or memory-augmented agents, where assumptions can propagate into external state and require rollback.Source: §5 · p. 9

6 Conclusion

Clarification can be present and acknowledged without removing the influence of an obsolete interpretation. The paper concludes that interactive systems should be evaluated for whether they preserve assumptions’ epistemic status and let go of stale hypotheses.

6 Conclusion

  • Old-hypothesis contamination is a diagnostic target because outputs can remain shaped by an interpretation that should have been invalidated.Source: §6 · p. 9
  • Robust interactive systems should be evaluated not only for memory retention or final-task success, but also for removing stale hypotheses after clarification.Source: §6 · p. 9
  • The central challenge is tracking which assumptions remain tentative, have been confirmed, or must be discarded.Source: §6 · p. 9

A Sketch of Uncertainty-Preserving State Management

The proposed system delays commitment while high-impact ambiguity remains, then invalidates contradicted hypotheses and regenerates from the revised state. Its protocol is motivated by experiments showing that clarification alone may leave stale assumptions in place, and is presented as a design primitive rather than a complete solution.

A Sketch of Uncertainty-Preserving State Management

  • When high-impact slots remain unresolved, the assistant should ask targeted clarification questions instead of producing a committed answer.Impact measures how much the answer would change with a slot’s value; the threshold is task-dependent.Source: A Sketch of Uncertainty-Preserving… · p. 11
  • Explicit rebuilding from a revised belief state improves success and eliminates measured contamination, unlike patching an earlier answer that preserves latent structure.The protocol removes contradicted assumptions before regenerating the answer.Source: A Sketch of Uncertainty-Preserving… · p. 12
  • The protocol tracks plausible task hypotheses and unresolved slots, then removes contradicted hypotheses and regenerates rather than patching the prior output.It is a system-level inference policy that can use prompting, controllers, memory schemas, or dialogue-state logic.Source: A Sketch of Uncertainty-Preserving… · pp. 11–12
  • A two-phase response policy first resolves uncertainty and then executes from the clarified state, eliminating wrong commitment and contamination in the supplementary setting.Direct answering collapses immediately; asking first alone does not necessarily enforce rebuilding.Source: A Sketch of Uncertainty-Preserving… · p. 12
  • The protocol is a design primitive, not a complete solution; experiments do not establish prompt-level uncertainty preservation as sufficient across dialogue settings.More robust implementations may require explicit state controllers, epistemic-status memory schemas, or models trained to maintain and revise uncertainty.Source: A Sketch of Uncertainty-Preserving… · pp. 12–13

B.1 Task Families

  • Tasks use an under-specified first turn followed by clarification, testing revised intent across writing, planning, and coding tasks.Coding tasks test revisions to structural assumptions such as API contracts and data structures.Source: §B.1 · p. 13

B.2 Datasets and Run Scale

The appendix describes the datasets, evaluation measures, and experiment suite used to test dialogue failures. At the model level, acknowledgement and behavioral revision do not show a universal mismatch, while coding displays the clearest gap.

B.2 Datasets and Run Scale

  • The full dataset contains 290 generated tasks, alongside 20 hand-crafted pilot tasks and 30 adversarial coding and planning tasks.Generated tasks provide coverage and scale; hand-crafted tasks stress-test adversarial ambiguity.Source: §B.2 · p. 13
  • The evaluation uses Gemini-2.5-Pro and Gemini-2.5-Flash and comprises approximately 13,900 API calls and 7,160 trials.Gemini-2.5-Pro also serves as the automatic judge at temperature 0.Source: §B.2 · p. 13

B.3 Evaluation Metrics

  • Old-hypothesis contamination measures whether a final response retains lexical, semantic, or structural traces of an interpretation later invalidated by clarification.A model may acknowledge a correction while its final answer remains contaminated.Source: §B.3 · p. 13

B.4 Complete Experiment Suite

  • The experiment suite distinguishes memory loss, insufficient reasoning, and premature state commitment through signature tests, strategy comparisons, and interventions.The tested signatures include order sensitivity, acknowledgement–behavior gaps, and task-family susceptibility.Source: §B.4 · p. 14

C.1 Model-Level Acknowledgement-Behavior Gap

  • Coding shows the strongest task-specific acknowledgement–revision gap: models acknowledge corrections more often than they behaviorally revise.The aggregate model-level gap is modest and not always positive.Source: §C.1 · p. 14

C.2 Old-Hypothesis Contamination Across Methods

The prompting and memory comparisons show that none of the tested methods eliminates old-hypothesis contamination. Summary memory, a single-state ledger, and naive multi-hypothesis prompting each show a distinct pattern of concern.

C.2 Old-Hypothesis Contamination Across Methods

  • None of the tested prompting or memory methods eliminates contamination; summary memory increases mean contamination relative to raw history.The authors suggest summarization can compress away epistemic status.Source: §C.2 · pp. 14–15, Table 14 · p. 15
  • A single-state ledger also increases contamination, consistent with forcing one explicit state prematurely settling ambiguity.Source: §C.2 · p. 15, Table 14 · p. 15
  • Naive multi-hypothesis prompting produces the highest contamination, suggesting that enumeration without arbitration or invalidation can preserve stale hypotheses.Source: §C.2 · p. 15, Table 14 · p. 15

C.3 Per-Model K-Hypothesis Ablation

The supplementary experiments show that maintaining more hypotheses has unstable effects across models, and that explicit reasoning or clarification alone does not guarantee successful revision. Rollback and two-phase response policies provide stronger intervention evidence by reducing contamination and wrong commitment.

C.3 Per-Model K-Hypothesis Ablation

  • Success rises from 4.7% at K = 1 to 19.1% at K = 5 for Gemini-2.5-Pro, but falls from 22.6% at K = 1 to 5.0% at K = 5 for Gemini-2.5-Flash.The divergent trends indicate that enumeration alone is not posterior tracking; hypotheses must be updated, arbitrated, and invalidated.Source: §C.3 · p. 15, Table 15 · p. 16

C.4 Chain-of-Thought Commitment Analysis

  • Chain-of-thought shows little explicit wrong commitment and zero final explicit contamination in this analysis, yet underperforms in the broader baseline comparison.The authors suggest attentional diversion or over-analysis rather than explicit wrong assertion may explain the result.Source: §C.4 · p. 16, Table 16 · p. 16

C.5 Clarification Policy

  • The direct policy commits to a wrong hypothesis 42.5% of the time, while the clarification policy asks in all cases and reduces wrong commitment to zero.The stronger intervention also executes from the clarified state rather than patching the earlier one.Source: §C.5 · p. 16, Table 17 · p. 16

C.6 State Rollback and Response-Policy Interventions

  • State rollback improves success and eliminates contamination; two-phase interaction eliminates both wrong commitment and contamination in the supplementary setting.Adversarial validation shows contamination remains observable on deliberately difficult tasks.Source: §C.6 · p. 16, Table 18 · p. 17

Related papers

Loading 2609.25337v1…