Research map / Large language models / Language, safety and interpretability
How do LLMs Compute Verbal Confidence?
Dharshan Kumaran, Arthur Conmy, Federico Barbero, Simon Osindero, Viorica Patraucean, Petar Veličković
Read the paperCheck your idea against 137,217 papers
AI-extracted. The measured effects below were read by AI from the paper's full text (no summary yet). Numbers and claims can be wrong or missing context, so check the paper before you cite or build on it.
Measured effects reported in the paper
| What was changed | Where | Measured | Effect |
|---|---|---|---|
| steer activations at PANL with high/low confidence vectors | Gemma 3 27B · TriviaQA | verbal confidence | Improved |
| steer at answer-colon (AC) position | Gemma 3 27B · TriviaQA | verbal confidence | No clear effect |
| patch clean PANL activation after answer corruption (ablation) | Gemma 3 27B · TriviaQA high-confidence | confidence recovery | Improved |
| patch clean PANL+1 activation (control) (ablation) | Gemma 3 27B · TriviaQA high-confidence | confidence recovery | No clear effect |
| mean-ablate PANL or CC residual (ablation) | Gemma 3 27B · TriviaQA | verbal confidence disruption | Hurt |
| cross-confidence activation swap at PANL (ablation) | Gemma 3 27B · TriviaQA | confidence change | Improved |
| block attention from CC to question and answer tokens (ablation) | Gemma 3 27B · TriviaQA categorical | first-token change rate: ~10% change rate | No clear effect |
| block attention from CC to PANL (minimal numeric prompt) (ablation) | Gemma 3 27B · TriviaQA numeric 0-9 | first-token change rate: ~21% (layers 30-36) | Hurt |
| block attention from CC to PANL (categorical prompt) (ablation) | Gemma 3 27B · TriviaQA categorical | first-token change rate | No clear effect |
| block attention from PANL to answer tokens (ablation) | Gemma 3 27B · TriviaQA numeric 0-9 | first-token change rate: ~20% (layers 22-28) | Hurt |
| add PANL activations to log-probability predictors of verbal confidence | Gemma 3 27B · TriviaQA | variance explained: 10.0% -> 38.0% | Improved |
| mean-ablate PANL in reasoning model (ablation) | Magistral Small 24B · TriviaQA | first-token change rate | No clear effect |
Closest papers by meaning
- Can LLMs Express Their Uncertainty? An Empirical Evaluation of Confidence Elicitation in LLMs (ICLR 2024 · similarity 0.91)
- LLMs Show No Signs Of Individuated Metacognition (NeurIPS 2026 · similarity 0.90)
- Learning to Route LLMs with Confidence Tokens (ICML 2025 · similarity 0.89)
- SteerConf: Steering LLMs for Confidence Elicitation (NeurIPS 2025 · similarity 0.89)
- On the Robustness of Verbal Confidence of LLMs in Adversarial Attacks (NeurIPS 2025 · similarity 0.88)
- Evaluating language models as risk scores (NeurIPS 2024 · similarity 0.88)
Writing a paper in this area? scime draft is a free LaTeX editor with AI included, real-time co-authors and a check of your references and venue rules before you submit. It also imports OpenAI Prism projects.
