Give a model the reference list of an unpublished paper and ask what the paper says. The best ones get it 15 percent of the time.
A benchmark called Reconstruction, posted to arXiv on 17 August by Qingqing Mao and colleagues, tests something narrower and harder than the usual reasoning evaluations: can a model work out a paper's central idea from nothing but its pre-publication bibliography? The seed paper is withheld along with everything published at the same time or later, references are stripped of titles and given anonymous IDs, and a separate language model judges whether the hypothesis a model proposes matches the real one. Across 643 papers in six scientific fields, seven frontier models matched the withheld idea between roughly 3 and 15 percent of the time. The authors then wired the same models into a pipeline that has them review each other's hypotheses and run a Swiss-system tournament over competing slots; that lifted the match rate to roughly 23 to 42 percent, about 2.4 times the best single model.
Related stories
Google's open models have been downloaded a billion times, and outsiders have built 100,000 versions of them2026-08-22Three-quarters of Americans do not want a data centre near them. A year ago they were evenly split.
2026-08-22The model on its own scored 30 percent. Wrapped in Nvidia's scaffolding, the same model cleared everything.
2026-08-22Machine learning read the shape of sick brain cells and picked out nine already-approved drugs that calmed them down
2026-08-21DeepSeek's cheap workhorse can now see — and on agent tasks that need eyes it says it is close to Anthropic's best
2026-08-21