When an AI says 'I'm fairly sure,' can those words be trusted even when its underlying confidence numbers disagree?
Can linguistic uncertainty expression be calibrated independently from numerical confidence?
This explores whether the words a model uses to signal doubt ("probably", "I'm fairly sure") can be made trustworthy as their own skill, separate from a numeric confidence score such as a token probability or a stated percentage.
This explores whether the words a model uses to signal doubt can be made trustworthy separately from a numeric confidence score. The corpus suggests yes, and that they have to be, because by default the two channels don't agree. One study found that what models *say* about their confidence predicts whether they will commit to an answer or abstain far better than it predicts whether the answer is right. Log-probabilities show the opposite pattern and track truth directly (Does verbal confidence actually predict answer correctness?). Verbal confidence behaves like a separate behavior that can drift from the model's internal numbers.
That separate behavior can be trained directly. Linguistic calibration rewards a model for long-form text whose hedges let a reader make good downstream decisions. It produces a Llama-2-7B that expresses calibrated confidence at accuracy comparable to factuality baselines, and it generalizes across science, biomedicine, and biography (Can models express calibrated confidence in long-form text?). The calibration target here is what the reader can do with the words, not whether a number matches an accuracy rate. Language can therefore be tuned on its own terms, and it can express doubt across a whole paragraph, not just a single scalar.
The independence has a limit, though. The goal is described as *faithful* uncertainty: expressed doubt should match the model's actual internal uncertainty, because hallucination comes partly from models not knowing where their knowledge ends (Can models express uncertainty instead of just answering?). Words need something real to be faithful to. The corpus offers several ways to sharpen that internal signal. Semantic entropy measures disagreement over meanings rather than tokens, which catches confabulations that token-level scores miss (Can we detect when language models confabulate?). A model's own past outcomes can ground its confidence at a tenth of the cost of self-consistency sampling (Can past performance predict when a model will be right?). Confidence can also serve as a reward, which reverses the calibration damage RLHF does (Can model confidence work as a reward signal for reasoning?). So the numbers and the words are separate dials, but a good numeric signal is what makes the verbal dial worth turning.
The verbal channel matters most because it is the one people read. Users in every language studied trust confident-sounding outputs even when they're wrong, and they follow confidence cues rather than accuracy (Do users worldwide trust confident AI outputs even when wrong?). A model with well-calibrated internal probabilities but overconfident phrasing still misleads. A neighboring finding points the same way: the channel a model answers through changes what you see. Survey answers that looked pathologically skewed became realistic once the model wrote text and that text was mapped to a scale afterward (Why do LLMs give unrealistic survey responses?).
The corpus has no direct test of the strictest version of your question: training the wording while leaving numeric confidence untouched, then comparing the two. The closest evidence is the decision-based training above, which suggests calibrated language is achievable but doesn't show it comes free of the underlying signal.
Sources 8 notes
Across multiple models and settings, what LLMs say about their confidence predicts whether they will commit or abstain far better than it predicts objective correctness. Log-probabilities show the opposite pattern, tracking truth directly.
Training with confidence statements and user-decision rewards produces Llama-2-7B that achieves calibrated long-form generation at comparable accuracy to factuality baselines. The approach generalizes across domains including science, biomedicine, and biography.
Models hallucinate because they lack awareness of their own knowledge boundaries, not just knowledge itself. Expressing uncertainty calibrated to intrinsic uncertainty—faithful uncertainty—offers a metacognitive solution beyond the answer-or-abstain tradeoff.
Clustering sampled answers by bidirectional entailment and computing entropy over semantic clusters catches confabulations invisible at token level. This self-referential approach works across tasks without task-specific training data.
XConf matches ten-sample self-consistency at a tenth of the cost by retrieving the model's past episodes with similar confidence levels and reading their historical success rates. Ablations show the signal depends entirely on stored outcomes, not on the retrieval prompt itself.
Show all 8 sources
RLSF uses answer-span confidence to rank reasoning traces, creating synthetic preferences that strengthen step-by-step reasoning while reversing RLHF's calibration degradation—without requiring human labels or external verifiers.
Cross-linguistic research shows users in every language trust confident AI outputs even when inaccurate. While confidence expression varies by language, users everywhere track confidence signals rather than accuracy, making overconfident errors systematically followed.
Semantic Similarity Rating—prompting for text then mapping to scales via embeddings—achieves 90% of human test-retest reliability with realistic distributions. Pathological skew and over-positivity disappear when output channels change, proving these are measurement artifacts, not intrinsic failures.
Papers this line draws on 8
The research behind the notes this line reads — ranked by how closely each paper relates.
- Reported Confidence in LLMs Tracks Commitment More Than Correctness
- Post-Training Large Language Models via Reinforcement Learning from Self-Feedback
- Linguistic Calibration of Long-Form Generations
- Confidence Comes from Experience: Experiential Confidence Estimation from Reasoning to Agents
- Humans overrely on overconfident language models, across languages
- Beyond Accuracy: The Role of Calibration in Self-Improving Large Language Models
- Beyond Binary Rewards: Training LMs to Reason About Their Uncertainty
- Reinforcement Learning with Metacognitive Feedback Elicits Faithful Uncertainty Expression in LLMs