Future Confidence Distillation in Large Language Models

This paper investigates the temporal evolution of confidence in Large Language Models (LLMs). The author distinguishes between pre-solution 'Feeling-of-Knowing' (FOK) and post-solution 'Judgement-of-Learning' (JOL) signals. The research explores whether hidden representations contain richer confidence information than what models explicitly verbalize and whether this information can be 'distilled' into a pre-solution predictor to enable low-cost, high-accuracy confidence estimation without requiring full answer generation.

The study demonstrates that post-solution confidence estimates are consistently better calibrated and more discriminative than pre-solution estimates. Furthermore, linear probes trained on internal hidden representations recover significantly more confidence-related information than the models' own verbalized scores. By training a 'future confidence predictor' on pre-solution representations to regress toward the confidence estimates of a post-solution teacher probe, the author shows that models can achieve much of the calibration improvement of post-solution assessment while operating entirely before the answer is generated. This approach is found to be sample-efficient and capable of transferring across datasets within the same domain.

Reliable confidence estimation is a prerequisite for high-stakes LLM applications, such as retrieval-augmented generation (RAG), tool use, and adaptive computation. Current methods often rely on expensive post-hoc generation or poorly calibrated verbal scores. This work provides a method to achieve high-quality self-assessment at a fraction of the computational cost, enabling more reliable decision-making in real-time systems.