SENPAI Auto-Research Agent Writes Kernels, Cuts Training Time 22% (26h to 22h)

morgymcg · x · 2026-09-22

The author's team used W&B's open-source autonomous research agent SENPAI to run an automated ML research loop: the agent researched and wrote two optimizations on its own — a SliceAttention kernel (for Transolver) and a BatchedMuon implementation — cutting a full training run by 22% (26h → 22h, 4 hours saved) with no significant change in error or quality.

SENPAI is a semi-autonomous research harness: an Advisor proposes and reviews experiments, and Student agents each implement one assigned experiment, iterating on observability data. The repo ships a full LLM-training-optimization guide, paper, and tooling to reproduce the workflow.

Original post →

More from coding & agent

coding & agent channel →