Researchers introduce future confidence distillation to improve reliability in large language models by leveraging temporal aspects of confidence estimation. The method trains predictors on pre-solution hidden representations using teacher confidence estimates derived from post-solution correctness probes.
- Post-solution confidence is consistently better calibrated and more discriminative than pre-solution confidence.
- Linear probes trained on hidden representations recover substantially richer confidence-related information than models explicitly verbalize.
- Distilled predictors operate on pre-solution representations during inference, recovering much of the calibration improvement achieved by post-solution methods.
- The approach remains highly sample efficient and transfers across datasets within the same domain.
This technique enables significantly more reliable yet low-cost confidence estimation by anticipating confidence before answer generation is complete.