Recursive’s Automated AI Research System Hits SOTA on Training Benchmarks

Recursive, the AI research company co-founded by Richard Socher, published early results on June 11, 2026 from an automated system that proposes, implements, tests, and validates ML improvements without human intervention in the loop.

Benchmark Results

The system achieved state-of-the-art on three benchmarks:

• NanoChat Autoresearch: 0.9109 validation BPB vs 0.9372 community best (autoresearch@home), representing a 1.3× speedup to reach equivalent loss
• NanoGPT Speedrun: 77.5s vs 79.7s to reach 3.28 validation loss, improving on a baseline refined by the community over 2+ years
• NVIDIA SOL-ExecBench: Mean SOL score 0.754 vs 0.699 across 235 GPU kernels — 18% reduction in gap to hardware limits

Karpathy Ecosystem Connection

The benchmarks tie directly to Andrej Karpathy’s autoresearch ecosystem. NanoChat autoresearch is a popular starting point for automated research experiments; autoresearch@home extends it into a collaborative setting where dozens of humans and hundreds of agents collectively improve performance.

Recursive’s system outperformed the entire autoresearch@home community starting from the same seed solution.

Open Source

Recursive is open-sourcing artifacts from the runs so independent researchers can inspect and reproduce results. The company frames the release as “first steps” toward fully automated AI research.

Company Background

Recursive was co-founded by Richard Socher, former Salesforce chief scientist and pioneer in neural language understanding. The company focuses on recursively self-improving AI systems that compound discoveries across modeling, optimization, and systems layers.