← News

ResearchSources: arXiv · 16 Nov 2026

Researchers propose validity screen for LLM confidence signals before benchmark metrics

An arXiv preprint adapts a clinical-assessment practice: check whether a confidence signal actually tracks correctness before interpreting it, then applies the screen to 20 frontier models.

In plain words

Language models can report how confident they are in each answer, and systems rely on those signals to decide when to defer or escalate. The paper argues that evaluations should first verify a signal varies with correctness, borrowing the idea of validity screening from clinical personality testing. Applying its screen, most tested models pass, but verdicts were often unstable across data splits.

What was announced

The authors specify a portable screen computed from the correctness-by-confidence table. It measures how often a model is highly confident when wrong, how often it lacks confidence when right, and the contrast between the two, then labels a signal Valid only when it measurably tracks correctness, distinguishing blanket confidence, blanket withdrawal and inversion. No new statistics are introduced.

On 20 frontier models answering 452 two-alternative items, one model was rated Invalid because DeepSeek-R1 withdrew 95% of its correct answers, four were Indeterminate and 15 were Valid. Two Indeterminate models kept confidence on 96% of their errors but made too few errors to certify the pattern. With verbalized 0-100 confidence across MMLU and other datasets, verdicts depended on the cut point: at 50 most models were Invalid because answers piled at the ceiling, while at 90-95 most passed. Only 6 of 20 verdicts reproduced reliably on half the items, and the authors state the verdicts are descriptive because protocol amendments were settled on the same data, pending a held-out test.

What it does not show

  • The findings come from the paper's abstract only; the full article was not reviewed.
  • Verdicts are descriptive because protocol amendments were settled on the same data, pending a held-out test.
  • Only 6 of 20 verdicts reproduced reliably on half the items, so classifications may not be stable.

Sources

Source-support review completed 17 Nov 2026. This story was drafted with AI assistance and checked against retained source passages. Automated checking is not independent verification.

Coverage: arXiv: abstract.

Updates and corrections

No corrections so far.

First published 17 Nov 2026

New to the subject? AI from the Ground Up starts from the beginning.