Does verbal confidence actually predict answer correctness?
We often treat what a model says about its own certainty as evidence of accuracy. But does spoken confidence actually track whether the answer is right, or something else entirely?
Verbal confidence — the model's spoken answer to "how sure are you?" — is treated as the default reliability signal precisely because it is black-box: applicable even when internal probabilities are inaccessible, which is most deployed models. The tacit assumption is that it estimates the probability the answer is correct. This paper's two-stage abstention paradigm (answer + report confidence, then later commit-or-abstain) breaks that assumption. Across four non-reasoning models, prompt framings, and confidence formats, verbal confidence predicted the model's own commit/abstain decision substantially better than it predicted objective correctness.
Calibrated token log-probabilities show the opposite profile: their ability to predict abstention is tightly coupled to their ability to discriminate correct answers — the signature of a genuine truth-evidence signal. When the authors regress out the variance verbal confidence shares with log-probs, a residual decision-aligned component remains that is not reducible to truth-tracking. So the two signals are structurally different, not noisy versions of the same thing: log-probs carry an answer-evidence signal; verbal confidence carries an additional behavioral-control signal.
This matters because it undercuts using verbal confidence as a correctness proxy — which means it undercuts using self-reported "how sure are you?" numbers for calibration, deferral, or user trust. It grounds a broader pattern where models' spoken self-reports diverge from what their internal computation encodes: the model reports what it will do with the answer, not how true the answer is. This complicates the persuasion literature too — if models sound sure, that conviction indexes commitment, not evidence.
Inquiring lines that read this note 7
This note is a source for these research framings, grouped by the broader line of inquiry each explores. Scan the bold lines of inquiry; follow any specific question forward.
Do language models truly understand their encoded knowledge or merely retrieve patterns? Does polished AI output mislead users about actual accuracy and expertise? Why does self-revision increase model confidence despite degraded accuracy? How do different reward signals affect model reasoning and calibration differently? Why do users believe false claims despite fact-checking efforts? Why doesn't LLM knowledge translate to improved reasoning outputs?Related concepts in this collection 3
This note in its neighbourhood — explore the map, then jump to a related concept in the list below.
Click a node to walk · click center to open · click Open in graph to see this note in the full knowledge graph
-
Can models express uncertainty instead of just answering?
Most factuality work expands what models know rather than what they know they know. Can expressing calibrated uncertainty create a third path between confident errors and unhelpful abstention?
same answer/abstain framing; here the diagnosis is that verbal confidence is not aligned with intrinsic (truth) uncertainty
-
Does linguistic conviction explain why LLMs persuade more effectively?
Research investigates whether LLMs' persuasive advantage stems from expressing higher linguistic certainty than humans, and whether this confidence-loading effect operates independently of factual accuracy.
expressed conviction indexes commitment not correctness, complicating certainty-as-evidence
-
Do reasoning models actually use the hints they receive?
This explores whether language models acknowledge reasoning hints in their explanations when those hints causally influence their answers. Understanding this gap matters for evaluating whether chain-of-thought explanations can be trusted for safety monitoring.
another perception-action gap where spoken self-report diverges from what computation encodes
Related papers in this collection 8
Papers most semantically related to this note, ranked by cosine similarity in the embedding space.
- Reported Confidence in LLMs Tracks Commitment More Than Correctness
- Deep Research: A Systematic Survey
- What Large Language Models Know and What People Think They Know
- Linguistic Calibration of Long-Form Generations
- Post-Training Large Language Models via Reinforcement Learning from Self-Feedback
- Beyond Binary Rewards: Training LMs to Reason About Their Uncertainty
- Debating with More Persuasive LLMs Leads to More Truthful Answers
- Whether LLMs Can Navigate Beliefs and Facts Depends on How You Phrase It
Original note title
verbal confidence in LLMs tracks the commit/abstain decision more than answer correctness while log-probability confidence tracks truth