Beyond Binary Rewards: Training LMs to Reason About Their Uncertainty

• URL: http://arxiv.org/abs/2507.16806v1
• Date: Tue, 22 Jul 2025 17:56:01 GMT
• Title: Beyond Binary Rewards: Training LMs to Reason About Their Uncertainty
• Authors: Mehul Damani, Isha Puri, Stewart Slocum, Idan Shenfeld, Leshem Choshen, Yoon Kim, Jacob Andreas,
• Abstract summary: This paper describes RLCR, an approach to training reasoning models that jointly improves accuracy and confidence estimation.<n>We show that across diverse datasets, RLCR substantially improves calibration with no loss in accuracy.<n>We also demonstrate that verbalized confidence can be leveraged at test time to improve accuracy and calibration.
• Score: 59.97939500426759
• License: http://creativecommons.org/licenses/by/4.0/
• Abstract: When language models (LMs) are trained via reinforcement learning (RL) to generate natural language "reasoning chains", their performance improves on a variety of difficult question answering tasks. Today, almost all successful applications of RL for reasoning use binary reward functions that evaluate the correctness of LM outputs. Because such reward functions do not penalize guessing or low-confidence outputs, they often have the unintended side-effect of degrading calibration and increasing the rate at which LMs generate incorrect responses (or "hallucinate") in other problem domains. This paper describes RLCR (Reinforcement Learning with Calibration Rewards), an approach to training reasoning models that jointly improves accuracy and calibrated confidence estimation. During RLCR, LMs generate both predictions and numerical confidence estimates after reasoning. They are trained to optimize a reward function that augments a binary correctness score with a Brier score -- a scoring rule for confidence estimates that incentivizes calibrated prediction. We first prove that this reward function (or any analogous reward function that uses a bounded, proper scoring rule) yields models whose predictions are both accurate and well-calibrated. We next show that across diverse datasets, RLCR substantially improves calibration with no loss in accuracy, on both in-domain and out-of-domain evaluations -- outperforming both ordinary RL training and classifiers trained to assign post-hoc confidence scores. While ordinary RL hurts calibration, RLCR improves it. Finally, we demonstrate that verbalized confidence can be leveraged at test time to improve accuracy and calibration via confidence-weighted scaling methods. Our results show that explicitly optimizing for calibration can produce more generally reliable reasoning models.

Related papers

• Post-Training Large Language Models via Reinforcement Learning from Self-Feedback [3.73824942136665]
  Large Language Models (LLMs) often produce plausible but poorly-calibrated answers.<n>We present Reinforcement Learning from Self-Feedback (RLSF), a post-training stage that uses the model's own confidence as an intrinsic reward.
  arXiv  Detail & Related papers  (2025-07-29T15:46:26Z)
• SGIC: A Self-Guided Iterative Calibration Framework for RAG [45.17496149653415]
  Large language models (LLMs) capitalize on their robust in-context reasoning.<n>We present a new framework that employs uncertainty scores as a tool.<n>We also introduce an innovative approach for constructing an iterative self-calibration training set.
  arXiv  Detail & Related papers  (2025-06-19T09:45:13Z)
• The Hallucination Dilemma: Factuality-Aware Reinforcement Learning for Large Reasoning Models [63.98194996746229]
  Large language models (LLMs) have significantly advanced in reasoning tasks through reinforcement learning (RL) optimization.<n>However, reasoning-oriented RL fine-tuning significantly increases the prevalence of hallucinations.<n>We propose Factuality-aware Step-wise Policy Optimization (FSPO), an innovative RL fine-tuning algorithm incorporating explicit factuality verification.
  arXiv  Detail & Related papers  (2025-05-30T14:23:32Z)
• Thinking Out Loud: Do Reasoning Models Know When They're Right? [19.776645881640178]
  Large reasoning models (LRMs) have recently demonstrated impressive capabilities in complex reasoning tasks.<n>We investigate how LRMs interact with other model behaviors by analyzing verbalized confidence.<n>We find that reasoning models may possess a diminished awareness of their own knowledge boundaries.
  arXiv  Detail & Related papers  (2025-04-09T03:58:19Z)
• Beyond Accuracy: The Role of Calibration in Self-Improving Large Language Models [15.638622371475853]
  Large Language Models (LLMs) have demonstrated remarkable self-improvement capabilities.<n>We investigate the impact on confidence estimation by investigating the impact on confidence estimation.
  arXiv  Detail & Related papers  (2025-04-03T04:39:54Z)
• Calibrating Language Models with Adaptive Temperature Scaling [58.056023173579625]
  We introduce Adaptive Temperature Scaling (ATS), a post-hoc calibration method that predicts a temperature scaling parameter for each token prediction. ATS improves calibration by over 10-50% across three downstream natural language evaluation benchmarks compared to prior calibration methods.
  arXiv  Detail & Related papers  (2024-09-29T22:54:31Z)
• Training Language Models to Self-Correct via Reinforcement Learning [98.35197671595343]
  Self-correction has been found to be largely ineffective in modern large language models (LLMs) We develop a multi-turn online reinforcement learning approach, SCoRe, that significantly improves an LLM's self-correction ability using entirely self-generated data. We find that SCoRe achieves state-of-the-art self-correction performance, improving the base models' self-correction by 15.6% and 9.1% respectively on MATH and HumanEval.
  arXiv  Detail & Related papers  (2024-09-19T17:16:21Z)
• Calibrating Long-form Generations from Large Language Models [34.72041258464477]
  Large Language Models' (LLMs) confidence scores should align with the actual likelihood of its responses being correct. Current confidence elicitation methods and calibration metrics rely on a binary true/false assessment of response correctness. We introduce a unified calibration framework, in which both the correctness of the LLMs' responses and their associated confidence levels are treated as distributions across a range of scores.
  arXiv  Detail & Related papers  (2024-02-09T17:00:32Z)
• Just Ask for Calibration: Strategies for Eliciting Calibrated Confidence Scores from Language Models Fine-Tuned with Human Feedback [91.22679548111127]
  A trustworthy real-world prediction system should produce well-calibrated confidence scores. We show that verbalized confidences emitted as output tokens are typically better-calibrated than the model's conditional probabilities.
  arXiv  Detail & Related papers  (2023-05-24T10:12:33Z)
• A Close Look into the Calibration of Pre-trained Language Models [56.998539510508515]
  Pre-trained language models (PLMs) may fail in giving reliable estimates of their predictive uncertainty. We study the dynamic change in PLMs' calibration performance in training. We extend two recently proposed learnable methods that directly collect data to train models to have reasonable confidence estimations.
  arXiv  Detail & Related papers  (2022-10-31T21:31:07Z)

This list is automatically generated from the titles and abstracts of the papers in this site.

This site does not guarantee the quality of this site (including all information) and is not responsible for any consequences.