New framework CURE teaches AI models to identify uncertainty at the claim level, reducing confident false statements in long-form text generation

Large language models frequently hallucinate in extended text generation, with existing methods failing to teach models which parts are unreliable

Previous approaches using reinforcement learning with correctness rewards only provide single scalar confidence scores for entire responses, insufficient for claim-by-claim variation

CURE framework improves factuality by implementing claim-level reasoning about uncertainty through a Claim-Aware Reasoning Protocol

The approach builds on recent advances in LLM reasoning and calibration to help models express appropriate confidence in individual claims rather than stating incorrect information with unwarranted certainty