RPM-guided search sets new SOTA on benchmarks, proving selection scales with compute

anirudhg9119 · x · 2026-09-01

RPM-guided search achieved breakthroughs on ML research tasks. Agentic RPM scored 94.1% on WinoGrande (vs 90.4% prior agentic SOTA), while Inference-only RPM hit 95.7% on SVAMP (vs 94.2% human SOTA). Results demonstrate that selection quality itself scales with compute.

Original post →

More from coding & agent

coding & agent channel →