Consistency Is Not Correction: Separating Confidence, Abstention, and Recovery in LLM Agents
4.60T1 sourcearXiv cs.MA
Source record
Published by arXiv cs.MA (T1 source). The original is at https://arxiv.org/abs/2611.21651.
Pipeline notes
The summary and note below are generated by the signal pipeline — they are Beyond Desk’s reading, not quotations from the source.
SummaryThe paper separates confidence-based answer ranking, abstention, and correction into distinct properties, derives a cap on improvements from correcting wrong proposals, and audits two model scales. It finds some controller comparisons lacked enough erroneous cases to detect a working gate and proposes two pre-experiment checks, alongside sensitivity to ties and output limits.
Why it mattersUseful for designing agent evaluations: it shows when apparent gains are statistically impossible and identifies concrete checks to run before spending inference budget.
Cited by
No citations on record.