Think Through Uncertainty: Improving Long-Form Generation Factuality via Reasoning Calibration

Researchers Xin Liu and Lu Wang have introduced CURE (a framework for reasoning calibration), a novel method to tackle the persistent problem of hallucinations in long-form AI text generation. Unlike previous approaches that apply a single confidence score to an entire response or use post-hoc corrections, CURE teaches large language models (LLMs) to reason about uncertainty at the individual claim level. Its core innovation is the Claim-Aware Reasoning Protocol, which forces the model to structure its output into atomic, verifiable claims, each paired with an explicit confidence estimate. This granular approach is crucial because uncertainty varies significantly across different parts of a long answer.

The researchers then developed a multi-stage training pipeline. First, it aligns the model's stated confidence with the actual correctness of each claim. Then, it optimizes the overall response for factuality. This process results in a calibrated model that not only generates more accurate text but also knows which parts of its answer are most reliable. This calibrated confidence enables "selective prediction," meaning the model can choose to abstain from answering claims it deems too uncertain, a critical feature for trustworthy AI assistants.

Experimental results on four long-form factuality benchmarks, including Biography generation and FactBench, demonstrate significant gains. CURE consistently outperformed competitive supervised and reinforcement learning baselines, improving claim-level factual accuracy by up to 39.9%. Importantly, it maintained factual recall while achieving a 16.0% increase in AUROC (Area Under the Receiver Operating Characteristic curve), a key metric for calibration, proving the model's confidence scores are meaningful predictors of truth.