Research map / Large language models / Instruction and fine-tuning
Rewarding Doubt: A Reinforcement Learning Approach to Calibrated Confidence Expression of Large Language Models
David Bani-Harouni, Chantal Pellegrini, Paul Stangel, Ege Özsoy, Kamilia Zaripova, Nassir Navab, Matthias Keicher
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 |
|---|---|---|---|
| fine-tune verbalized confidence with log-scoring-rule PPO reward | Llama-3-8B-Instruct · TriviaQA | ECE: 0.3459 -> 0.0226 | Improved |
| fine-tune verbalized confidence with log-scoring-rule PPO reward | Llama-3-8B-Instruct · TriviaQA | AUROC: 0.5858 -> 0.8592 | Improved |
| fine-tune confidence with RL reward | Llama-3-8B-Instruct · TriviaQA | accuracy: 0.6310 -> 0.6309 | No clear effect |
| use RL instead of supervised trained probe | Rewarding Doubt · TriviaQA | ECE: 0.0189 -> 0.0226 ECE; AUROC 0.8173 -> 0.8592 | Mixed |
| apply Rewarding Doubt to multi-answer training | Llama-3-8B-Instruct · QAMPARI | ECE: 0.5319 -> 0.0816 | Improved |
| apply Rewarding Doubt to other base models (ablation) | Qwen-2.5-3B, Qwen-2.5-7B, Gemma-2-9B, LLaMA-3.1-8B · TriviaQA | ECE: Qwen-2.5-3B 0.5330 -> 0.1483 | Improved |
| train on TriviaQA, evaluate out of domain | Rewarding Doubt · MedQA | ECE: 0.4480 -> 0.1145 | Improved |
| train on TriviaQA, evaluate out of domain | Rewarding Doubt · CommonsenseQA | ECE: 0.2820 -> 0.2930 | No clear effect |
| train on single-answer, evaluate on multi-answer QAMPARI (ablation) | Rewarding Doubt · QAMPARI multi fact | ECE: 0.5319 -> 0.1777 | Improved |
| train specifically on multi-fact data rather than single fact (ablation) | Rewarding Doubt · QAMPARI multi fact | ECE: 0.1777 -> 0.1061 | Improved |
Later papers that compared with this method
1 later accepted paper reports RewardingDoubt as a baseline in their results tables; 1 of them report a better number on most of the results they share with it.
- From Sampling to Cognition: Modeling Internal Cognitive Confidence in Language Models for Robust Uncertainty Calibration (AAAI 2026 · better on 4 of 4 shared results)
Matched by method name in results tables read by AI. Authors' own tables tend to favour the newer method, and a comparison can be missed when a method is renamed.
Closest papers by meaning
- Calibrating LLMs with Semantic-level Reward (NeurIPS 2026 · similarity 0.91)
- Taming Overconfidence in LLMs: Reward Calibration in RLHF (ICLR 2025 · similarity 0.91)
- Beyond Binary Rewards: Training LMs to Reason About Their Uncertainty (ICLR 2026 · similarity 0.91)
- Linking Perception, Confidence and Accuracy in MLLMs (CVPR 2026 · similarity 0.90)
- Decoupling Reasoning and Confidence: Resurrecting Calibration in Reinforcement Learning from Verifiable Rewards (ICML 2026 · similarity 0.90)
- QA-Calibration of Language Model Confidence Scores (ICLR 2025 · similarity 0.89)
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.
