Science One cuts AI paper phantom references from 21% to zero

The hard problem for AI-written research is no longer making a paper sound plausible. It is proving that the claims in the paper are tied to real evidence. In a July 30, 2026 post, Google Research introduced Science One Framework, an experimental autonomous research prototype built around Chain-of-Evidence. The companion CoE Audit checks whether generated papers are backed by real references, reproducible scores, aligned methods, and code that actually does what the paper says.

The target is a growing failure mode in research agents. Systems can now review literature, form hypotheses, run experiments, and draft manuscripts, but polished text can hide broken evidence. Google Research points to non-existent citations, method sections that do not match the submitted code, and reported scores that cannot be reproduced from the underlying implementation. Science One treats verifiability as part of the architecture rather than an after-the-fact cleanup step.

The framework has three main modules. A Problem Investigator grounds literature through the Semantic Scholar API and up to 100 full-text PDFs per topic, so references originate from retrieval rather than model memory. A Discovery engine runs parallel explore-and-exploit branches where solver agents implement candidates and task-specific evaluators score them. A Paper writer and Claim Verifier then bind factual claims to workspace artifacts before the manuscript is rendered, revising claims that go beyond their evidence.

The reported numbers are the reason this is worth watching. Google Research applied CoE Audit to 75 generated papers across five systems-optimization tasks and found that baselines produced phantom references at rates as high as 21%, while Science One produced none. It also reached perfect score verification and the strongest method-code alignment in the comparison. On external tests, the framework earned two Gold Medals and two Silver Medals across five MLE-Bench Kaggle competitions, while also succeeding on Parameter Golf. Google is clear that Science One is still an experimental prototype, but it sketches a practical bar for AI science: every result needs a trail back to evidence.