Verbalized Confidence Estimation in Language Models

97 concepts79 papers

Large Language Models (LLMs) are notoriously "confident liars," often presenting hallucinations with the same authoritative tone as factual truths. Traditionally, researchers measured a model's uncertainty by looking at its internal "white-box" states—specifically the log-probabilities (logits) of the tokens it generates. However, as models become increasingly closed-source and accessible only via APIs, these internal signals are often hidden.

Verbalized confidence estimation offers a "black-box" alternative: asking the model to explicitly state how sure it is in natural language. By treating uncertainty as a linguistic output rather than a hidden statistical property, researchers aim to bridge the gap between a model's internal knowledge and its external reliability.

The Mechanism of Self-Reported Certainty

The core intuition behind verbalized confidence is that a sufficiently capable model can act as its own evaluator. Instead of relying on the softmax distribution of the next token, the system prompts the model to generate a meta-cognitive signal.

As Makoto Yamada et al. demonstrate, this can be as simple as asking a model for a numerical score (e.g., "Rate your confidence from 0 to 100"). While smaller models often struggle with this and remain hopelessly overconfident, larger models exhibit a surprising degree of self-awareness. Tiancheng Hu et al. found that when advanced models express high verbalized confidence, their accuracy significantly increases, whereas low confidence scores correlate with performance that approaches random chance.

However, direct elicitation has limits. Jiuhai Chen argues that asking for a continuous numeric score often triggers the model's tendency toward extremes. Instead, forcing the model into a discrete choice—such as selecting from a set of confidence intervals—tends to produce better-calibrated results.

Building Intuition through Self-Reflection

Why does a model "know" it might be wrong? The effectiveness of verbalized confidence often depends on the model's ability to reason about its own reasoning process. This is frequently implemented through a "model cascade" or a self-reflective loop.

Lisa Alazraki et al. show that the most effective way to ensemble vision-language models is not through simple voting, but by having a model evaluate the reasoning steps of its peers. Similarly, Dongkeun Yoon et al. suggest that "reasoning models" (like OpenAI’s o1) are better at expressing confidence because they perform a self-critique before committing to a final probability. By forcing the model to identify potential flaws in its logic, the system elicits latent uncertainty that a single-pass answer might hide.

This self-probing can also be structured as a consistency check. The BSDETECTOR framework, proposed by Jiuhai Chen, combines sampled consistency (checking if the model gives the same answer multiple times) with self-reflection to flag hallucinations in black-box environments.

The Honesty Gap and Calibration

A recurring challenge in this field is the "honesty gap." Shiyu Ni et al. found that while internal token probabilities are generally more reliable indicators of truth, models struggle to translate these internal signals into accurate natural language. Models are often "better at feeling their knowledge boundaries than saying them."

To fix this, researchers use calibration-tuning. Sanyam Kapoor et al. demonstrate that fine-tuning a model on a small dataset of its own correct and incorrect answers can teach it to recognize its own knowledge gaps. This transforms uncertainty estimation into a learned classification task. Similarly, Yu-Neng Chuang et al. developed Self-REF, which teaches models to append specific "confidence tokens" (like <CN> for certain or <UN> for uncertain) directly to their output. This embeds self-assessment into the very vocabulary of the model.

Applications in High-Stakes Environments

Verbalized confidence is not just a theoretical curiosity; it is a functional tool for routing and safety. In robotics, Kai Cheng et al. used verbalized scores from vision-language models to guide exploration, allowing robots to stop early when they are confident they have found an object. In multi-agent systems, Xinyi Zeng showed that internal confidence signals can be used to maintain system accuracy even when the majority of agents in a network are providing malicious or erroneous information.

Ultimately, the goal of this research direction is to move away from "blind" generation. By teaching models to verbalize their doubt, we create systems that can abstain from answering Jiefeng Chen et al. or signal when a human expert needs to step in Leila Tavakoli et al..

Key Foundational Work

Prudent Silence or Foolish Babble? Examining Large Language Models' Responses to the Unknown (2023) — contributed 1 concept elements to this direction.
How (not) to ensemble LVLMs for VQA (2023) — contributed 1 concept elements to this direction.
When to Read Documents or QA History: On Unified and Selective Open-domain QA (2023) — contributed 1 concept elements to this direction.
History & citationVersion 1 · 2026-10-04

Current · Permanent link

View frozen version 1

Cite this direction

Latest version (a living page):

@misc{lacuna_verbalized_confidence_estimation_08d4cf00,
  title = {verbalized confidence estimation},
  author = {{Lacuna}},
  howpublished = {Lacuna research direction},
  year = {2026},
  url = {https://lacuna.tiptreesystems.com/direction/verbalized-confidence-estimation/rdr_08d4cf004dd04264a4d6fb562d9dd036},
  note = {Latest version of a living research direction; cite a version for a fixed text. Accessed 2026-11-21},
}

This exact version (version 1, frozen text):

@misc{lacuna_verbalized_confidence_estimation_08d4cf00_v1,
  title = {verbalized confidence estimation},
  author = {{Lacuna}},
  howpublished = {Lacuna research direction},
  year = {2026},
  url = {https://lacuna.tiptreesystems.com/direction/verbalized-confidence-estimation/rdr_08d4cf004dd04264a4d6fb562d9dd036/v/1},
  note = {Version 1, frozen 2026-10-04. Accessed 2026-11-21},
}
About

These directions overlap with this one through shared papers.

See something wrong? Report a mistake