Semantic Calibration Prevails Where Token Confidence Fails: Benchmarking Long-Form Scientific QA
This work introduces the first large-scale benchmark for UQ calibration in long-form, reasoning-demanding scientific QA, evaluating four UQ methods on 685,000 responses across up to 20 LLMs and seven datasets, supported by an extensible open-source framework whose shared-generation design enables reproducible cross-met...
Philip Müller, Nicholas Popovic, Michael Färber et al.
· 4 citations
· ⚡1