Title: Prioritizing Certainty: A New Approach to Streamlining LLM Uncertainty Assessment

Abstract:

The trustworthy integration of large language models (LLMs) hinges on the ability to generate precise uncertainty estimates. Current techniques largely follow an "answer-first" approach, calculating confidence levels only after a response has been fully generated. This method evaluates the accuracy of a single output, which restricts its practical utility. In contrast, we investigate a "confidence-first" paradigm in which the model declares its confidence prior to providing an answer. This score is interpreted as the likelihood of answering correctly based on the model’s existing policy. To implement this, we introduce CoCA (Co-optimized Confidence and Answers), a reinforcement learning framework built on GRPO. CoCA achieves simultaneous optimization of answer accuracy and confidence calibration through segmented credit assignment. By allocating distinct rewards and group-relative advantages to the confidence and answer components, the framework ensures stable joint training and mitigates the risk of reward hacking. Our evaluations across benchmarks for mathematics, coding, and factual question-answering demonstrate that CoCA enhances both calibration and uncertainty discrimination without compromising answer quality, thus expanding the potential for diverse downstream applications.

Source: arXiv Generated at: 2026-06-04 00:00:00 UTC