MLReplicate: Benchmarking Autonomous Research Systems for Machine Learning Reproducibility

Automated review systems are unreliable gatekeepers: Only 10 of 37 valid manuscripts passed automated conference-style review, yet human experts found that 59% of these accepted papers contained fabricated or unsupported claims, revealing a significant gap between automated and human evaluation standards.

All autonomous research systems produce reproducibility failures: Across six state-of-the-art systems tested, human reviewers consistently identified methodological flaws, hallucinated experimental results, and reproducibility issues, indicating that current autonomous research systems are not yet ready for independent scientific work.

Cost and computational resources don't guarantee quality: The benchmark found that "neither token budget nor computational cost predicts output quality: the cheapest system outperforms the most resource-intensive system," suggesting that throwing more computational power at autonomous research doesn't solve fundamental reliability problems.

Human oversight remains essential: The dual-protocol evaluation combining automated review with expert human assessment proved necessary to catch critical failures, with human reviewers identifying problems that automated systems missed entirely.

Evaluation frameworks must keep pace with AI capabilities: The paper highlights that while autonomous research systems have advanced rapidly, "robust and realistic evaluation frameworks have failed to keep pace," emphasizing the need for better benchmarks before deploying such systems in real scientific contexts.