← News

ResearchSources: arXiv · 18 Nov 2026

Researchers probe what question-asking reveals about a language model's hidden state

An arXiv paper uses a student-teacher question-asking setup to detect whether a model's reasoning will succeed, but finds detection does not guarantee correction.

In plain words

A research paper studies what happens inside a large language model when it asks questions before answering. The authors found a measurable signal that predicts whether the model's final answer will be correct. However, using that signal to intervene often failed to fix wrong answers.

What was announced

The paper, posted on arXiv, examines test-time reasoning in large language models, where the same prompt and partial solution can yield different answers across samples. The authors use question-asking as an inference-time intervention: in a student-teacher setting, a student model asks questions of a teacher, and the researchers train a probe on the student's hidden state before and after a question is asked. That probe predicted the final correctness of the reasoning trajectory even before the teacher's answer arrived, which the authors read as evidence of a self-diagnosis signal during question generation rather than information transfer from the teacher.

The team then treats question-asking as a sequential decision problem, using the probe as a quality score and defining a gating policy that asks questions to maximize the likelihood of a correct outcome. The reported results show a gap between detection and recovery: the policy captures model correctness and uncertainty, yet interventions were about as likely to harm correct trajectories as to recover incorrect ones. The authors note the intervention's success depends heavily on the model's self-consistency, and they frame this diagnosis-correction gap as relevant to how language models refine themselves under uncertainty.

What it does not show

  • The source is only the paper's abstract, so no independent verification, full methodology details, or specific benchmark numbers are available from this evidence.

Sources

Source-support review completed 19 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 19 Nov 2026

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