Anirudh Goyal Details Agentic RPM Research Selection

Anirudh Goyal, a Senior Staff Research Scientist at Google DeepMind, posted several replies explaining RPM variants. He describes how models receive a small sandbox budget to run pilot experiments before committing to full runs. The approach generates multiple candidate solutions and uses tournaments, with prior search tree context, to select which receive expensive execution. He notes that selection quality improves with added history, reasoning steps, and larger candidate pools. RPM-guided search produced new results on AIRS-Bench tasks. Inference-only and agentic versions reached baseline performance levels faster while using a smaller share of the execution budget.