Google Cloud launches Science One AI research framework

Google Cloud researchers have developed the Science One Framework, an experimental autonomous research prototype designed to solve the persistent issue of AI hallucinations in scientific writing. While existing autonomous agents like Sakana's AI-Scientist, AI Scientist v2, AutoResearchClaw, DeepScientist, and AI-Researcher can generate complete manuscripts, they often introduce errors such as non-existent citations or code-text mismatches. To address this, the new system implements a Chain-of-Evidence framework, which enforces that every claim is backed by a verifiable, recorded evidence chain.

The Science One Framework operates through three main components. First, a problem investigator uses the Semantic Scholar API to read up to 100 full-text PDFs per topic, grounding all citations. Next, a parallel discovery engine refines solutions across isolated branches. Finally, a paper writer and claim verifier cross-checks every assertion against actual workspace artifacts. To test this, researchers introduced the CoE Audit, an automated protocol that performs four integrity checks across 75 papers generated on five systems-optimization tasks—Prism, Cloudcast, EPLB, LLM-SQL, and transaction scheduling—from the Automated Design of Research Systems benchmark. While baseline systems hallucinated up to 21% of their references, the Science One Framework achieved zero phantom references and perfect score verification.

The system also demonstrated state-of-the-art performance on complex external benchmarks. On MLE-Bench, which spans five Kaggle competitions including medical imaging, fine-grained recognition, and 3D perception, the framework earned two Gold Medals—including a winning score on 3D Object Detection where baselines failed—and two Silver Medals. Additionally, in the Parameter-Golf live LLM-training competition, the framework successfully navigated strict hardware and file-size constraints to achieve a state-of-the-art score as of April 27, 2026, while baseline systems failed to produce valid submissions. It also achieved the best overall score on the Cloudcast and EPLB tasks within the Automated Design of Research Systems benchmark.

For AI practitioners and researchers, this development shifts the paradigm of autonomous scientific discovery from raw output generation to verifiable integrity. Instead of manually auditing AI-generated papers for phantom citations or unrepeatable experimental scores, practitioners can rely on built-in evidence chains that guarantee reproducibility. By treating verifiability as a core architectural constraint, the framework ensures that autonomous agents produce trustworthy, peer-level research that can be immediately integrated into real-world scientific pipelines.