How do LLMs Compute Verbal Confidence

• URL: http://arxiv.org/abs/2603.17839v1
• Date: Wed, 18 Mar 2026 15:31:43 GMT
• Title: How do LLMs Compute Verbal Confidence
• Abstract summary: We show that verbal confidence reflects automatic, sophisticated self-evaluation -- not post-hoc reconstruction.<n>These findings have implications for understanding metacognition in LLMs and improving calibration.
• Score: 20.456986780602964
• License: http://creativecommons.org/licenses/by/4.0/
• Abstract: Verbal confidence -- prompting LLMs to state their confidence as a number or category -- is widely used to extract uncertainty estimates from black-box models. However, how LLMs internally generate such scores remains unknown. We address two questions: first, when confidence is computed - just-in-time when requested, or automatically during answer generation and cached for later retrieval; and second, what verbal confidence represents - token log-probabilities, or a richer evaluation of answer quality? Focusing on Gemma 3 27B and Qwen 2.5 7B, we provide convergent evidence for cached retrieval. Activation steering, patching, noising, and swap experiments reveal that confidence representations emerge at answer-adjacent positions before appearing at the verbalization site. Attention blocking pinpoints the information flow: confidence is gathered from answer tokens, cached at the first post-answer position, then retrieved for output. Critically, linear probing and variance partitioning reveal that these cached representations explain substantial variance in verbal confidence beyond token log-probabilities, suggesting a richer answer-quality evaluation rather than a simple fluency readout. These findings demonstrate that verbal confidence reflects automatic, sophisticated self-evaluation -- not post-hoc reconstruction -- with implications for understanding metacognition in LLMs and improving calibration.

Related papers

• ORCE: Order-Aware Alignment of Verbalized Confidence in Large Language Models [17.761630515662947]
  We propose a decoupled and order-aware framework for verbalized confidence calibration.<n>Our method first generates an answer and then estimates confidence conditioned on the fixed question--answer pair.<n>Experiments show that our method improves calibration and failure prediction performance while largely preserving answer accuracy.
  arXiv  Detail & Related papers  (2026-05-12T17:39:43Z)
• Identifying Influential N-grams in Confidence Calibration via Regression Analysis [53.95767806659635]
  We identify what linguistic expressions are related to confidence by applying the regression method.<n>We show that large language models (LLMs) remain overconfident when reasoning is involved.<n>We reveal that confidence calibration is possible by simply suppressing those overconfident expressions without drops in performance.
  arXiv  Detail & Related papers  (2026-04-07T11:58:49Z)
• BrowseConf: Confidence-Guided Test-Time Scaling for Web Agents [58.05949210993854]
  We investigate whether search agents have the ability to communicate their own confidence through verbalized confidence scores after long sequences of actions.<n>We propose Test-Time Scaling (TTS) methods that use confidence scores to determine answer quality, encourage the model to try again until reaching a satisfactory confidence level.
  arXiv  Detail & Related papers  (2025-10-27T15:58:51Z)
• Mind the Generation Process: Fine-Grained Confidence Estimation During LLM Generation [63.49409574310576]
  Large language models (LLMs) exhibit overconfidence, assigning high confidence scores to incorrect predictions.<n>We introduce FineCE, a novel confidence estimation method that delivers accurate, fine-grained confidence scores during text generation.<n>Our code and all baselines used in the paper are available on GitHub.
  arXiv  Detail & Related papers  (2025-08-16T13:29:35Z)
• Read Your Own Mind: Reasoning Helps Surface Self-Confidence Signals in LLMs [2.4892313127400962]
  We study the source of uncertainty in DeepSeek R1-32B by analyzing its self-reported verbal confidence on question answering (QA) tasks.<n>We show that granting DeepSeek the budget to explore its distribution by forcing a long chain-of-thought before the final answer greatly improves its verbal score effectiveness.
  arXiv  Detail & Related papers  (2025-05-28T17:01:30Z)
• On Verbalized Confidence Scores for LLMs [25.160810008907397]
  Uncertainty quantification for large language models (LLMs) can establish more human trust into their responses.<n>This work focuses on asking the LLM itself to verbalize its uncertainty with a confidence score as part of its output tokens.<n>We assess the reliability of verbalized confidence scores with respect to different datasets, models, and prompt methods.
  arXiv  Detail & Related papers  (2024-12-19T11:10:36Z)
• Confidence in the Reasoning of Large Language Models [0.0]
  Confidence is measured in terms of persistence in keeping their answer when prompted to reconsider.<n> Confidence is only partially explained by the underlying token-level probability.
  arXiv  Detail & Related papers  (2024-12-19T10:04:29Z)
• Fact-Level Confidence Calibration and Self-Correction [64.40105513819272]
  We propose a Fact-Level framework that calibrates confidence to relevance-weighted correctness at the fact level. We also develop Confidence-Guided Fact-level Self-Correction ($textbfConFix$), which uses high-confidence facts within a response as additional knowledge to improve low-confidence ones.
  arXiv  Detail & Related papers  (2024-11-20T14:15:18Z)
• Learning to Route LLMs with Confidence Tokens [43.63392143501435]
  Large language models (LLMs) have demonstrated impressive performance on several tasks and are increasingly deployed in real-world applications.<n>In high-stakes settings, it becomes vital to know when the output of an LLM may be unreliable.<n>We study the extent to which LLMs can reliably indicate confidence in their answers, and how this notion of confidence can translate into downstream accuracy gains.
  arXiv  Detail & Related papers  (2024-10-17T07:28:18Z)
• When to Trust LLMs: Aligning Confidence with Response Quality [49.371218210305656]
  We propose CONfidence-Quality-ORDer-preserving alignment approach (CONQORD) It integrates quality reward and order-preserving alignment reward functions. Experiments demonstrate that CONQORD significantly improves the alignment performance between confidence and response accuracy.
  arXiv  Detail & Related papers  (2024-04-26T09:42:46Z)
• Think Twice Before Trusting: Self-Detection for Large Language Models through Comprehensive Answer Reflection [90.71323430635593]
  We propose a novel self-detection paradigm that considers the comprehensive answer space beyond LLM-generated answers. Building upon this paradigm, we introduce a two-step framework, which firstly instructs LLM to reflect and provide justifications for each candidate answer. This framework can be seamlessly integrated with existing approaches for superior self-detection.
  arXiv  Detail & Related papers  (2024-03-15T02:38:26Z)
• Reconfidencing LLMs from the Grouping Loss Perspective [56.801251926946485]
  Large Language Models (LLMs) are susceptible to generating hallucinated answers in a confident tone. Recent findings show that controlling uncertainty must go beyond calibration. We construct a new evaluation dataset derived from a knowledge base to assess confidence scores given to answers of Mistral and LLaMA.
  arXiv  Detail & Related papers  (2024-02-07T15:40:22Z)
• An evaluation of word-level confidence estimation for end-to-end automatic speech recognition [70.61280174637913]
  We investigate confidence estimation for end-to-end automatic speech recognition (ASR) We provide an extensive benchmark of popular confidence methods on four well-known speech datasets. Our results suggest a strong baseline can be obtained by scaling the logits by a learnt temperature.
  arXiv  Detail & Related papers  (2021-01-14T09:51:59Z)

This list is automatically generated from the titles and abstracts of the papers in this site.

This site does not guarantee the quality of this site (including all information) and is not responsible for any consequences.