Researchers tested 40 questions drawn from newly accepted papers. GPT-6-Astra correctly judged 98 percent as unstudied yet stated that status plainly in only 8 percent of ordinary answers. All six frontier systems showed reporting gaps of 30-90 percentage points. The study attributes the asymmetry to models recognizing studied questions at 88-100 percent when supplied the answering papers.
Higher capability correlated with more accurate internal judgments yet larger reporting gaps across 22 models. The authors built a fixed answer combining each model's judgment with separate checks of cited papers. This raised Astra's clear reporting of open status from 8 percent to 95 percent and produced gains in all six systems tested.
The method predicts reviewers' novelty grades by combining several models' written ratings with a rule learned from graded examples. Performance reached the range reported by Think-Probe-Respond. Unlike TPR the approach requires no access to model weights or internal activations and therefore works with closed-weight frontier models.
The supplied evidence covers only the 40-question sample drawn from newly accepted papers and the six frontier systems plus 22-model capability comparison. No broader claim about overall model behavior on other question types or domains is supported by the excerpts.