Recursive Open-Sources Automated Research Artifacts After SOTA Benchmark Run

Recursive announced June 11, 2026 that it is open-sourcing artifacts from its automated AI research system after achieving state-of-the-art results on three benchmarks measuring training efficiency and GPU kernel optimization.

Key Metrics

• NanoChat validation BPB: 0.9109 (previous community best 0.9372)
• NanoGPT Speedrun: 77.5 seconds (previous 79.7s)
• NVIDIA SOL-ExecBench mean score: 0.754 (previous 0.699 on 235 kernels)

Automated Research Loop

The system automates the full research cycle: propose an idea, implement it, run an experiment, validate the result, and use learnings to choose the next experiment. It is designed to scale using open-ended algorithm principles and recursively self-improving AI.

Independent Verification

Open-sourcing artifacts enables the ML community to independently verify benchmark claims — a positive signal amid growing interest in automated research systems from companies including OpenAI (Erdős problem), Sakana AI (AI Scientist), and Recursive.

Richard Socher, Recursive co-founder and former Salesforce chief scientist, leads the effort. The company positions the release as early results, not a solved automated research problem.