What's this about?
XConf retrieves graded past episodes to revise confidence; authors report matching/beating ten-sample consistency on 23 of 24 AUROC comparisons at one-tenth generation cost (49 HF upvotes).
What exactly is confidence calibration?
An 80% confidence should correspond to roughly 80% correctness over similar cases. Historical feedback can help only when outcomes are valid and future tasks are comparable.
Why does this matter?
Author-reported preprint; no discussion sampled.
- Significance: past outcomes are stronger than unsupported self-confidence.
- Significance: changing task distributions can break calibration.
- Significance: retrieval and feedback quality need checking.