Adnan KhanAdnanProduct & AI
San Francisco Bay Area
Get in touch

Source note3 min read

Future Confidence Distillation in Large Language Models

Source

arXiv

arxiv.org/abs/2607.07626v1

Read the source

Source

Plain-Language Brief

The paper investigates how the confidence of large language models evolves during the answering process. It compares pre-solution and post-solution confidence estimates across different models and finds that post-solution confidence is better calibrated and more discriminative. The authors introduce a new method called 'future confidence distillation' which trains predictors using teacher confidence estimates produced by post-solution correctness probes, requiring only pre-solution representations for inference.

What the Document Says

  • The paper discusses the importance of reliable confidence estimation in deploying large language models (LLMs).
  • Existing approaches treat confidence as a property of completed responses and overlook its evolution during the answering process.
  • The authors investigate confidence from a temporal perspective by comparing pre-solution Feeling-of-Knowing (FOK) and post-solution Judgement-of-Learning (JOL) confidence estimates.
  • They find that post-solution confidence is better calibrated and more discriminative than pre-solution confidence.
  • Linear probes trained on hidden representations recover richer confidence-related information than models explicitly verbalizing.
  • The authors introduce 'future confidence distillation', a method to train predictors using teacher confidence estimates produced by post-solution correctness probes.
  • These distilled predictors require only pre-solution representations for inference and transfer across datasets within the same domain.
  • The findings suggest that confidence-related information evolves throughout the answering process and can be anticipated before answer generation is complete.

Key Concepts Mentioned

  • large language models
  • confidence estimation
  • Feeling-of-Knowing (FOK)
  • Judgement-of-Learning (JOL)
  • future confidence distillation

Claims Made by the Document

  • Reliable confidence estimation is essential for deploying large language models.
  • Post-solution confidence is better calibrated and more discriminative than pre-solution confidence.
  • Linear probes trained on hidden representations recover richer confidence-related information.
  • Future confidence distillation can train predictors using teacher confidence estimates produced by post-solution correctness probes.

Evidence Used

  • The paper discusses the importance of reliable confidence estimation...
  • They find that post-solution confidence is better calibrated and more discriminative than pre-solution confidence...
  • Linear probes trained on hidden representations recover richer confidence-related information than models explicitly verbalizing...
  • The authors introduce 'future confidence distillation'...

Methods / Approach

UNKNOWN

Limitations Stated by the Authors

UNKNOWN

Limitations Noted by the Agent

Agent interpretation — not from the source.

UNKNOWN

Open Questions

UNKNOWN

Notes for Future Q&A

May include interpretation. Labeled where it does.

UNKNOWN

Raw Abstract / Excerpt

Reliable confidence estimation is essential for deploying large language models (LLMs) in confidence-aware systems, where downstream decisions such as retrieval, tool use, and adaptive computation depend on accurately estimating answer reliability. Existing approaches, however, largely treat confidence as a property of completed responses, overlooking how confidence-related information evolves throughout the answering process. In this work, we investigate confidence from a temporal perspective by comparing pre-solution Feeling-of-Knowing (FOK) and post-solution Judgement-of-Learning (JOL) confidence estimates across frontier and open-source LLMs. We show that post-solution confidence is consistently better calibrated and more discriminative than pre-solution confidence, while linear probes trained on hidden representations recover substantially richer confidence-related information than models explicitly verbalise. Building on this observation, we introduce future confidence distillation, which trains predictors operating on pre-solution hidden representations using teacher confidence estimates produced by post-solution correctness probes. Despite requiring only pre-solution representations for inference, distilled predictors recover much of the calibration improvement achieved by post-solution confidence, remain highly sample efficient, and transfer across datasets within the same domain. Together, our findings demonstrate that confidence-related information evolves throughout the answering process and can be anticipated before answer generation is complete, enabling significantly more reliable yet low-cost confidence estimation.