COPU: Conformal Prediction for Uncertainty Quantification in Natural Language Generation

• DOI: 10.48550/arxiv.2502.12601
• Publication Date: 2025-02-18
• Authors: Shaoyu wang; Yaming Jiang; Yuxin Tang; Lu Cheng; Hanjie Chen
• Keywords: FOS: Computer and information sciences; Computer Science - Computation and Language; Computation and Language (cs.CL)

Uncertainty Quantification (UQ) for Natural Language Generation (NLG) is crucial assessing the performance of Large Models (LLMs), as it reveals confidence in predictions, identifies failure modes, and gauges output reliability. Conformal Prediction (CP), a model-agnostic method that generates prediction sets with specified error rate, has been adopted UQ classification tasks, where size set indicates model's uncertainty. However, when adapting CP to NLG, sampling-based generating candidate outputs cannot guarantee inclusion ground truth, limiting its applicability across wide range rates. To address this, we propose \ourmethod, explicitly adds truth uses logit scores measure nonconformity. Our experiments six LLMs on four NLG tasks show \ourmethod outperforms baseline methods calibrating rates empirical cover rates, offering accurate user-specified