PaperlayerView exact revision on arXiv

Source-linked AI summary

Confidence Estimation for LLMs in Multi-turn Interactions

Caiqi Zhang, Ruihan Yang, Xiaochen Zhu, Chengzu Li, Tiancheng Hu, Yijiang River Dong, Deqing Yang, Nigel Collier

arXiv:2601.02179v2cs.CL

TL;DR

Confidence estimation has been studied mainly in single-turn settings, leaving its dynamics in multi-turn conversations insufficiently examined. The paper introduces a calibration-and-monotonicity framework with new metrics and controlled datasets, finding that common methods struggle while P(SUFFICIENT) performs comparatively better. The results provide a foundation for evaluating confidence in interactive dialogue, although the datasets and tasks simplify real-world conversations.

  • Problem

    Prior confidence-estimation research largely focuses on single-turn question answering, leaving confidence dynamics as information accumulates in multi-turn conversations insufficiently studied.

  • Method

    The paper evaluates per-turn calibration and confidence monotonicity using InfoECE, Kendall’s τ, Hinter–Guesser dialogues, and incremental QA benchmarks.

  • Results

    Widely used methods struggle with calibration and monotonicity, while P(SUFFICIENT) performs comparatively better and more effectively distinguishes meaningful information gains from conversational filler.

  • Takeaways & Limitations

    The framework establishes a foundation for developing confidence estimators that track evidence accumulation across multi-turn and single-turn interactions.

  • Takeaways & Limitations

    The progressive datasets and Hinter–Guesser protocol simplify real conversations, omitting topic shifts, repairs, and mixed intents that may limit transfer to messy open-world dialogue.

Abstract

from arXiv · show

While confidence estimation is a promising direction for mitigating hallucinations in Large Language Models (LLMs), current research overwhelmingly focuses on single-turn settings. The dynamics of model confidence in multi-turn conversations, where context accumulates and ambiguity is progressively resolved, remain largely unexplored. This work presents the first systematic study of confidence estimation in multi-turn interactions, establishing a formal evaluation framework grounded in two key desiderata: per-turn calibration and monotonicity of confidence as more information becomes available. To facilitate this, we introduce novel metrics, including a length-normalized Expected Calibration Error (InfoECE), and a new "Hinter-Guesser" paradigm for generating controlled evaluation datasets. Our experiments reveal that widely-used confidence techniques struggle with calibration and monotonicity in multi-turn dialogues. In contrast, a novel logit-based probe we introduce, P(Sufficient), proves comparatively more effective, robustly tracking evidence accumulation and distinguishing it from conversational filler. Our work provides a foundational methodology for developing more reliable and trustworthy conversational agents.

1 Introduction

This work addresses the largely unexplored problem of estimating confidence as information accumulates across multi-turn conversations. It proposes evaluating confidence by per-turn calibration and monotonicity, using new metrics, datasets, and controlled dialogues.

  • Motivation: Multi-turn confidence estimation matters because information arrives incrementally while users refine queries, models clarify, and evidence accumulates.Reliable confidence can guide decisions about asking clarifying questions, invoking tools, or committing to actions.
  • Contribution: The paper presents the first systematic study of confidence estimation in multi-turn conversations.
  • Evaluation desiderata: Reliable confidence should satisfy calibration at each turn and monotonicity as useful information increases.Calibration measures correspondence with empirical correctness, while monotonicity requires confidence to rise as information becomes available.
  • Metrics and datasets: InfoECE measures length-normalized calibration across dialogues, while Kendall’s τ quantifies monotonicity.
  • Metrics and datasets: The evaluation covers under-specified queries through Hinter–Guesser dialogues and difficult fully specified queries through incremental QA benchmarks.Both settings provide progressively revealed clues or hints.
  • Findings: Widely used methods struggle with calibration or consistent monotonicity, whereas P(SUFFICIENT) performs comparatively better but leaves substantial room for improvement.P(SUFFICIENT) also more effectively distinguishes meaningful information gains from conversational filler.

2 Related Work

Prior work studies confidence and uncertainty estimation in LLMs and increasingly examines multi-turn dialogue, but this paper focuses on confidence dynamics under progressively increasing information. Its evaluation compares confidence methods across controlled variants of contextual complexity.

  • Confidence estimation: Confidence concerns certainty about a specific generated response, whereas uncertainty concerns variability in predictions given an input query.
  • Confidence estimation: Existing confidence estimation approaches include prompting-based verbalized methods and other mainstream techniques.
  • Multi-turn interactions: Recent multi-turn studies report that LLMs can perform worse on the same tasks in multi-turn contexts than in single-turn settings.
  • Multi-turn interactions: The paper creates question variants with increasing contextual information across turns to compare confidence methods under different complexity levels.
  • Multi-turn interactions: This work studies the cooperative regime, where confidence should rise as information accumulates, complementing work on adversarial persuasion.

3 Methodology

The methodology models confidence as a per-turn signal in dialogues whose information accumulates, evaluating calibration at each information level and monotonicity across turns. It compares prompting, self-consistency, and logit-based confidence methods, including probes for answer correctness and evidence sufficiency.

  • Dialogue formulation: Each dialogue turn produces an answer and confidence score, while correctness is recorded against a single gold label.
  • Task characteristics: The controlled task requires progressive information acquisition, step-wise answerability, and confidence that increases with turn index and tracks accuracy.
  • Evaluation regimes: The evaluation covers under-specified questions whose candidate sets shrink with hints and fully specified but difficult questions made easier by additional information.
  • Metrics: InfoECE groups predictions by normalized information exposure so calibration can be compared across dialogues of different lengths.
  • Metrics: Kendall’s τ measures pairwise confidence ordering across turns, with 1 indicating strictly increasing confidence and 0 indicating no overall trend.
  • Confidence estimation methods: The study compares verbalized confidence, self-consistency, P(TRUE), and P(SUFFICIENT), while excluding post-hoc calibration methods as orthogonal rescaling approaches.
  • Confidence estimation methods: P(SUFFICIENT) asks whether accumulated information entails that the proposed answer is uniquely correct, distinguishing evidence-supported identifiability from incidental correctness.

4 Dataset Construction

The datasets cover under-specified interactions built with 20Q and GUESS-style settings and fully specified but difficult quizbowl-style questions. The Hinter–Guesser pipeline generates fixed, incrementally informative dialogues and retains trajectories that eventually reach a correct, unique answer.

  • Dataset regimes: Under-specified datasets use 20Q and GUESS settings, while GRACE and TRICKME provide fully specified but difficult evaluation data.
  • Under-specified datasets: In 20Q and GUESS, a questioner incrementally seeks information about a secret entity, with 20Q restricted to yes/no questions and GUESS allowing open-ended questions.
  • Hinter–Guesser paradigm: The Hinter–Guesser pipeline assigns a secret entity to a Hinter, which supplies helpful non-trivial hints while a Guesser answers and probes whether multiple candidates remain.
  • Dataset examples: Table 1 presents 20Q and GUESS examples, using fixed Hinter–Guesser histories for confidence evaluation without strategic information gathering during evaluation.
  • Hinter–Guesser paradigm: Dialogues continue until the Guesser answers correctly and certifies uniqueness; unsuccessful trajectories are discarded, yielding 1,848 collected dialogues.
  • Fully specified datasets: GRACE and TRICKME are incremental quizbowl-style datasets in which clues become increasingly specific and each item has a unique gold answer.

5 Experiments

Experiments compare confidence estimation methods across models and multi-turn settings, evaluating calibration, monotonicity, information sensitivity, and single-turn versus multi-turn performance. P(SUFFICIENT) is generally strongest for tracking evidence progression, while other methods show calibration errors or sensitivity to conversational format.

  • Experimental setup: The experiments evaluate confidence estimation across four models and datasets using calibration, monotonicity, and information-versus-length comparisons.The evaluation includes model confidence on generated answers and confidence relative to ground-truth answers.
  • Monotonicity: P(SUFFICIENT) most consistently tracks increasing confidence as clues accumulate, reaching τ = 83.76 on GUESS with Qwen2.5-72B and τ = 71.38 on TRICKME with Llama3.1-70B.SC often has weak monotonicity in under-specified settings, while verbalized methods occasionally lead on particular datasets despite poor calibration.
  • Monotonicity: Ground-truth evaluation substantially improves monotonicity for all methods, with P(SUFFICIENT) reaching τ = 93.91 on GUESS with Qwen2.5-72B.For Llama3.1-70B, P(SUFFICIENT) reaches τ = 91.62, 86.55, and 85.90 on 20Q, GUESS, and GRACE, respectively.
  • Information sensitivity: Across 40 comparisons, informative turns produce more significant changes than placebo turns, 27 versus 18, and P(SUFFICIENT) most cleanly separates information gain from turn accumulation.It can decrease confidence after an uninformative hint, including a drop from 14.27 to 2.97 for Llama3.1-70B on GUESS.
  • Single-turn versus multi-turn: Multi-turn and single-turn accuracy differ by less than 1 point on average, but confidence shifts markedly with prompt format, especially for P(SUFFICIENT).P(TRUE) can rise under placebo in GUESS, while SC is generally more robust and verbalized confidence remains unstable across conditions.

6 Conclusion

This work presents a foundational framework for evaluating confidence estimation in multi-turn conversations. Its findings show that widely used methods remain poorly calibrated and non-monotonic, while P(SUFFICIENT) performs comparatively better but leaves substantial room for improvement.

  • The paper presents the first systematic study of confidence estimation for LLMs in multi-turn conversations.
  • The evaluation framework combines calibration and monotonicity with novel metrics and datasets for dynamic dialogues.
  • Widely used confidence estimation techniques struggle to maintain calibration and monotonicity as conversations progress.
  • P(SUFFICIENT) achieves comparatively better performance than widely used techniques, although the task remains significantly under-resolved.
  • Future research should target confidence signals that satisfy both desiderata, distinguish task-relevant information from conversational filler, and remain robust across interaction settings.

Limitation

The study provides a foundational framework but evaluates simplified settings and leaves several important dimensions of conversational confidence for future work.

  • Progressive datasets and the Hinter–Guesser protocol omit topic shifts, repairs, and mixed intents, potentially limiting transfer to messy, open-world dialogue.
  • The study focuses on information-seeking tasks, leaving confidence dynamics in open-ended, creative, and collaborative conversations unresolved.
  • The evaluation emphasizes calibration and rank monotonicity, while downstream effects on user utility and human trust require user studies and field deployments.
  • The framework studies confidence rather than uncertainty, leaving their relationship in multi-turn settings for future investigation.

Ethics Statement

The study reports adherence to standard ethical guidelines and describes its software, dataset, participant, privacy, and risk practices. It also acknowledges potential Western-centric bias and limited multilingual coverage.

  • The research follows standard ethical guidelines and verifies the licenses of all software and datasets used.
  • The datasets may contain Western-centric biases, and the study does not address multilingual coverage, which may limit generality across languages and cultures.
  • The experiments are fully automated, with no new participants or annotators recruited.
  • The study reports no privacy concerns because it does not collect, store, or release personally identifiable information.
  • The authors report no additional anticipated risks and used an AI assistant only for grammar checking.

B Incremental QA datasets

The incremental QA datasets evaluate confidence as clues unfold, covering both fully specified questions and adversarially authored cases. Their per-turn accuracy and confidence reports align with the proposed multi-turn evaluation framework.

  • GRACE contains quizbowl-style questions with increasingly specific, self-contained clues and evaluates when, how accurately, and how confidently models answer.
  • TrickMe uses human-in-the-loop adversarial authorship to create incremental clues that elicit confident model mistakes while remaining solvable by humans.
  • GRACE and TrickMe provide fully specified settings with unique gold answers while models decide when to commit as evidence accumulates.
  • The study reports per-turn accuracy and confidence on both datasets without altering their incremental-clue protocols.

C Calibration shifts reveal a scaling-dependent format effect.

Calibration quality changes with presentation format in a model-scale-dependent way, despite comparable accuracy. Smaller models often calibrate worse after summarization, whereas larger models can calibrate substantially better.

  • C Calibration shifts reveal a scaling-dependent format effect.: The reported accuracy remains comparable across formats, while InfoECE responds divergently by model scale.Thus, presentation format changes calibration quality without a corresponding accuracy shift.
  • C Calibration shifts reveal a scaling-dependent format effect.: P(SUFFICIENT) worsens with summarization for smaller models: Llama3.1-8B rises from 6.99 to 24.57 on 20Q and 3.82 to 9.41 on GUESS.The pattern suggests reliance on turn-by-turn conversational structure.
  • C Calibration shifts reveal a scaling-dependent format effect.: P(SUFFICIENT) improves with summarization for larger models: Llama3.1-70B falls from 40.29 to 9.81 on 20Q and 34.70 to 6.16 on GUESS.Qwen2.5-72B similarly falls from 27.51 to 2.55 on GUESS.
  • C Calibration shifts reveal a scaling-dependent format effect.: P(TRUE) improves markedly for Llama3.1-70B, with InfoECE falling from 68.00 to 53.28 on 20Q and 66.07 to 36.87 on GUESS.Smaller models show minimal change.
  • C Calibration shifts reveal a scaling-dependent format effect.: SC and verbalized methods remain largely format-invariant, with shifts typically below 5 InfoECE points.They nevertheless remain consistently poorly calibrated, exceeding 50 InfoECE.

E Placebo QA Examples

The appendix provides question-format examples and placebo controls for the 20Q and GUESS datasets. The placebo questions are designed to add neither useful information nor contradiction.

  • E Placebo QA Examples: Figure 5 compares InfoECE across five confidence methods and Accuracy for 20Q and GUESS.It contrasts multi-turn with summarized presentation, with lower InfoECE indicating better calibration.
  • E Placebo QA Examples: Table 7 shows examples from four datasets in question-answer dialogue format.
  • E Placebo QA Examples: Table 8 lists placebo question-answer pairs for Guess My City and 20Q.The control set contains 30 Guess My City questions and 26 20Q questions.
  • E Placebo QA Examples: The placebo questions provide no useful information for answering the target question.
  • E Placebo QA Examples: The placebo questions also do not contradict the conversation history.
Loading 2601.02179v2…