Google Research builds an AI system designed to keep its own science honest

Researchers Rui Meng and Tomas Pfister published Science One on Thursday as a response to a specific, growing problem: autonomous AI research systems are getting better at producing polished looking papers while getting worse, proportionally, at producing trustworthy ones, with hallucinated citations, mismatches between described methods and actual code, and non reproducible results all common failure modes. Science One’s Chain of Evidence approach requires that every claim carry a complete and correct evidence chain, enforced through three components: a Problem Investigator that builds citation graphs from the Semantic Scholar API and reads up to 100 full text PDFs to ground references rather than relying on model memory, a Discovery Engine that runs parallel research branches with isolated evaluation cycles, and a Paper Writer paired with a dedicated Claim Verifier that reconciles every factual statement against its declared source. A four part audit protocol then re-runs code to verify scores, checks for specification violations, verifies references, and analyzes whether the described method actually matches the implementation. In testing, the system produced zero phantom references against baselines that hallucinated citations up to 21 percent of the time, achieved perfect score verification, and matched or exceeded human expert performance on five benchmark research tasks, including two gold and two silver medal finishes across Kaggle competitions in MLE-Bench and a state of the art result on the Parameter-Golf benchmark. The team said the goal is a system where, as they put it, every reference in a generated paper traces back to a grounded citation rather than the model’s memory.