AI That Does Its Own Research: Recursive Posts SOTA GPU and Training Results
Recursive released early results June 11, 2026 from an automated AI research system that runs many research threads over long horizons, combines promising branches, and validates results against reward hacks before treating improvements as real progress.
Three Benchmark Wins
| Benchmark | Recursive | Previous Best | Metric |
|---|---|---|---|
| NanoChat Autoresearch | 0.9109 BPB | 0.9372 BPB | Validation bits per byte |
| NanoGPT Speedrun | 77.5s | 79.7s | Time to 3.28 val loss |
| SOL-ExecBench | 0.754 SOL | 0.699 SOL | Mean across 235 kernels |
What the System Found
On NanoChat, the biggest gains came from a richer short-context memory mechanism — hashed bigram and trigram embedding tables mixed into the attention value path through learned gates. Solutions combined architecture changes, auxiliary losses, optimizer behavior, weight decay schedules, and compiler settings rather than a single trick.
On GPU kernels, the system invented new optimizations, recast known ideas under tighter constraints, and composed improvements across modeling and systems layers.
Context
Richard Socher co-founded Recursive. The work builds on Karpathy’s autoresearch repos and the broader movement toward “AI scientist” systems that automate the research loop. Artifacts are being open-sourced for community verification.