Google's Science One Hits MLE-Bench Gold via Evidence Chains

On July 30, 2026, Google Research announced the Science One Framework, an experimental prototype built to perform verifiable autonomous scientific discovery. The system tackles AI hallucinations by requiring native evidence linking at the exact moment a claim is generated.

Rather than relying on post-generation citations, the framework introduces an architectural constraint called Chain-of-Evidence (CoE). Every scientific claim is inextricably linked to a traceable path of source data, prior literature, or experimental results. If the model cannot ground the output immediately, the claim fails the constraint and is not produced.

To enforce this constraint across long-running tasks, Google shipped CoE Audit. This automated protocol evaluates AI-generated scientific papers by tracing and validating the underlying evidence chains. If you evaluate and test AI agents, this shifts the focus from checking the final answer to auditing the continuous proof path.

Benchmark Results

The system was tested against the MLE-Bench (Machine Learning Engineering Benchmark), a suite of complex Kaggle competitions designed to break automated agents. The framework secured two Gold Medals and two Silver Medals across five competitions, handling tasks in medical imaging, fine-grained recognition, and 3D perception.

Competition Type	Baseline System Result	Science One Result
3D Object Detection	Complete Failure	Winning Score
MLE-Bench Suite	Standard Accuracy	2 Gold, 2 Silver Medals

During a live LLM-training competition called Parameter-Golf, which imposes strict hardware and file-size limits, Science One logged a state-of-the-art score as of April 27, 2026. Researchers noted it discovered genuine algorithmic techniques rather than relying on standard hyperparameter sweeps.

Practical Implications

Led by Google Cloud researchers Rui Meng and Tomas Pfister, the project signals a structural shift in how autonomous systems are designed. Multi-agent systems often prioritize solver quality, which measures the ability to find an answer. Science One prioritizes verifiability, proving why an answer is correct.

For teams building domain-specific autonomous tools, the takeaway is architectural. Treating grounding as a post-generation step leaves a gap where hallucinations thrive. Building verifiability into the core generation constraint ensures the research output scales safely without requiring manual fact-checking of every intermediate claim.