Confidence Estimation for LLMs in Multi-turn Interactions

AI summary

Overview

Research area: Natural Language Processing — confidence/calibration estimation for large language models, extended from single-turn question answering to multi-turn dialogue.

Technical level: Intermediate to Advanced. The core intuitions are accessible (does a model's stated confidence rise as it gets better information?), but the paper formalizes them with calibration metrics used in the uncertainty-estimation literature, so some background in LLM evaluation helps.

Scope: The paper defines an evaluation framework, metrics and datasets for measuring whether LLM confidence signals are calibrated and monotonically increasing across turns of an information-seeking dialogue, and benchmarks five confidence methods on four open-source models.

What This Paper Is About

Almost all prior work on LLM confidence estimation studies a single, static question-and-answer exchange. But real use of LLMs is conversational: users add information turn by turn, ambiguity is progressively resolved, and the model's confidence should rise as the evidence accumulates. This paper asks whether existing confidence methods actually behave that way. The authors build the first systematic testbed for multi-turn confidence, define what "good" multi-turn confidence should look like, and show that most widely used methods fail both to stay calibrated and to rise monotonically — while proposing one new probe that comes closer.

Key Contributions

1. First systematic study of confidence estimation in multi-turn interactions. The authors frame multi-turn confidence as a distinct research target and argue it is critical for agentic and human-in-the-loop systems.

2. A formal evaluation framework with two desiderata and new metrics. They specify per-turn/per-level calibration and monotonicity, introduce a length-normalized InfoECE (information-level Expected Calibration Error) that groups predictions by normalized information exposure, and adopt Kendall's τ to measure whether confidence trends upward across turns.

3. Two new dataset regimes plus a new generation paradigm. For under-specified questions they introduce the "Hinter–Guesser" paradigm (with a uniqueness-probing step and filtering of non-converging dialogues), producing 1,848 dialogue turns across 226 entities for 20Q and 1,625 turns across 223 entities for Guess. For fully specified but difficult questions they adapt the existing Grace and TrickMe incremental quizbowl-style QA datasets.

4. A new confidence method, P(Sufficient), a logit-based probe that asks whether the accumulated information is sufficient to make the current answer the only correct one — as opposed to P(True), which only asks whether the answer is correct.

Main Findings

• Calibration is generally poor across the board. Verbalized confidence (Vanilla-Verb, CoT-Verb) and the logit-based P(True) are poorly calibrated, with InfoECE values typically between 40 and 80. Self-consistency (SC) is usually the most calibrated method on the fully-specified incremental QA datasets (Grace and TrickMe).

• Sufficiency probing is strikingly better calibrated on under-specified games. For Llama3.1-70B, P(Sufficient) achieves InfoECE of 13.05 on 20Q and 5.27 on Guess, while remaining competitive on Grace and TrickMe.

• P(Sufficient) shows the most consistent monotonicity on the model's current answer. Examples reported include τ = 83.76 on Guess with Qwen2.5-72B and τ = 71.38 on TrickMe with Llama3.1-70B. By contrast, SC often shows weak monotonicity in under-specified settings (single digits on Guess). Qwen2.5 models are an exception, where verbalized confidence occasionally attains the highest τ on 20Q and Grace despite poor calibration.

• Monotonicity improves dramatically when confidence is scored against the ground-truth answer. P(Sufficient) dominates here: τ = 93.91 on Guess with Qwen2.5-72B, and 91.62, 86.55, 85.90 on 20Q, Guess and Grace respectively with Llama3.1-70B. The authors note the ground truth is unavailable in real applications.

• P(Sufficient) penalizes uninformative turns. Under the placebo condition, it frequently shows a statistically significant decrease in confidence after a useless hint (for example, a drop from 14.27 to 2.97 for Llama3.1-70B on Guess). Across all 40 comparisons, informative turns yield more significant changes than placebos (27 vs. 18 at p < 0.05).

• P(True) is confounded by turn count, especially on Guess. It rises even under placebo (mean Δ_placebo = +5.64; significant in 2/4 model pairs). Llama3.1-8B jumps +11.75 under placebo (p < 10⁻¹²) and Qwen2.5-72B jumps +14.61 (p < 10⁻⁶). Its mean informative gain on Guess is +14.07, but the placebo lift undermines its usefulness. On 20Q, placebo effects tend to be negative (mean −3.91).

• Self-consistency is moderately robust but not immune. It shows small placebo movements and sizable informative gains (mean Δ_info = +9.83 on Guess), yet Llama3.1-8B on 20Q increases under placebo by +4.82 (p = 0.0025).

• Verbalized confidence is unstable. It shows small average placebo shifts (often non-significant) and modest informative gains, and in some cases moves counterintuitively — e.g., Qwen2.5-7B on Guess decreases by −9.29 with an informative hint (p = 0.005).

• Multi-turn and single-turn-summary accuracy are comparable. The mean absolute accuracy gap is less than 1 across models and datasets, unlike the "getting lost" effect reported by Laban et al. (2025). Example: on 20Q, Llama3.1-8B goes 24.95 → 27.16 with summaries, while Llama3.1-70B goes 33.87 → 32.31.

• Confidence, unlike accuracy, shifts strongly with prompt format. P(Sufficient) consistently drops under single-turn summaries (on 20Q: Qwen2.5-7B 63.13 → 13.23; Llama3.1-70B 34.80 → 15.30), suggesting it exploits turn-structure cues. P(True) and verbalized confidence often decrease for smaller models but can increase for larger ones without accuracy gains — e.g., Llama3.1-70B on Guess: Vanilla-Verb 80.63 → 87.65, CoT-Verb 84.43 → 90.72.

• Scaling helps ranking more than absolute calibration. Larger models show consistently higher accuracy and better τ, particularly for P(Sufficient) (Qwen2.5-72B reaches τ = 83.76% on Guess vs. its 7B counterpart), but the effect on InfoECE is more nuanced — smaller models occasionally show lower absolute calibration error in specific configurations.

Methodology in Plain English

The setup. The researchers build dialogues where information arrives a piece at a time. At every turn the model must produce a concrete answer and a confidence score, so accuracy and calibration can be assessed turn by turn. Later turns always contain at least as much information as earlier ones, so a well-behaved confidence signal should climb.

Two kinds of task. In under-specified tasks the opening question admits many answers, and hints progressively eliminate candidates (e.g., "guess the city" with clues about continent, climate, region). In fully-specified but difficult tasks a unique answer exists from the start but is hard to pin down; hints make it progressively easier.

Hinter–Guesser. Naively having two LLMs play questioner and answerer can produce irrelevant or misleading questions. So the authors restructure the interaction: a "Hinter" supplies a useful but non-trivial clue each turn, and a "Guesser" makes a best guess and separately flags whether other candidates remain plausible. Dialogues run until the Guesser is both correct and certifies uniqueness; only convergent dialogues are kept. During confidence evaluation the conversations are fixed — models do not choose what to ask.

Comparing fairness. Each model answers once to produce an answer, which is compared to the gold answer. Then each confidence method scores that answer; confidence is also computed against the ground-truth answer in parallel. Sampling temperature is 1, otherwise 0.

Methods tested. (a) Verbalized: Vanilla-Verb asks the model to self-report a score in [0,100], rescaled to [0,1]; CoT-Verb adds step-by-step reasoning first. (b) Self-consistency: sample m (e.g., 20) answers independently and take the fraction agreeing with the chosen answer. (c) Logit probes: P(True) forces a constrained A/B choice and reads off the softmax probability of "A"; P(Sufficient) is the authors' new variant, asking whether the current information is sufficient to entail that the answer is the only correct one.

Two controls. First, a placebo condition replaces the real hint with a conversational filler (e.g., Q: "Is this a valid hint?" A: "Yes."), letting the authors separate genuine information gain from mere dialogue length. Second, a single-turn summary condition presents a synthesized version of all clues so far in one prompt, to see whether incremental interaction helps or hurts.

Metrics. InfoECE averages the absolute gap between mean accuracy and mean confidence across information-level bins. Kendall's τ compares all turn pairs within a dialogue and counts concordant minus discordant pairs, averaged across dialogues; 1 means strictly increasing confidence, 0 means no trend.

Models and data. Llama3.1 Instruct (8B and 70B) and Qwen2.5 Instruct (8B and 72B) — the tables label the smaller Qwen2.5 variant as 7b. Reported accuracies range from 12.92 (Qwen2.5-7b on Guess) to 53.88 (Qwen2.5-72b on TrickMe).

Why This Matters

Impact on research. The paper reframes confidence as a dynamic signal tied to dialogue progression rather than a static property of one response. It supplies reusable metrics (InfoECE), a dataset-generation paradigm (Hinter–Guesser), and a strong baseline (P(Sufficient)) that others can build on, and it explicitly positions multi-turn confidence as under-resolved rather than solved.

Real-world applications:

• Autonomous agents that need a reliable internal signal for when to ask a clarifying question, call a tool, or commit to an action.
• Human-in-the-loop systems where a trustworthy, rising confidence signal helps a person decide when to accept, override, or further supervise a model.
• Information-seeking assistants (help desks, diagnosticians, troubleshooting agents) where partial evidence accumulates over a conversation before any answer is certain.
• Model monitoring and risk control in high-stakes deployments, where the ability to distinguish "I got lucky" from "the evidence now supports this" matters more than raw accuracy.

Industry relevance. Any product built around multi-turn chat with decision-making consequences — where the model must know whether it knows — stands to benefit. The paper's placebo test is also a practical evaluation trick: it directly detects whether a confidence score is inflating simply because the conversation got longer.

Future Directions

1. Methods that satisfy calibration and monotonicity simultaneously. P(Sufficient) does comparatively well, but the authors state the task remains significantly under-resolved.

2. Better separation of information from conversational filler. The placebo experiments reveal that P(True) and verbalized confidence are still swayed by turn count in some settings.

3. Robustness across prompt formats. P(Sufficient) degrades sharply under single-turn summaries (63.13 → 13.23 on 20Q with Qwen2.5-7B), raising the question of how to keep confidence stable when the same evidence is presented differently.

4. Extending beyond the current setting. The limitations section flags transferring to messier dialogue (topic shifts, repairs, mixed intents), open-ended or creative conversations, measuring downstream impact on user utility and human trust via controlled user studies, and moving from confidence to uncertainty quantification in multi-turn settings. The ethics statement also notes the data may carry Western-centric biases and that multilingual coverage was not addressed.

Target Audience

Researchers and practitioners working on LLM reliability, hallucination mitigation, uncertainty quantification, and conversational/agentic systems. It is most useful for those designing evaluation protocols for multi-turn assistants or building confidence-aware decision logic. Readers focused purely on single-turn QA calibration will find the framework and the placebo-control design directly transferable, while those interested in dataset construction will find the Hinter–Guesser paradigm the most reusable artifact.

Note on reporting: the paper does not report human user studies, field deployments, or downstream trust measurements; the ethics statement confirms no new participants or annotators were recruited. Several details (full dataset protocols for Grace and TrickMe, the remainder of the prompt templates) are deferred to appendices that are not included in the provided content.

Authors’ abstract

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.