Meta RPM Model Prioritizes AI Research Experiments, Cuts Compute by 40%
rohanpaul_ai · x · 2026-08-18
Meta FAIR introduces Research Preference Models (RPM) to address the bottleneck where AI agents generate experiment ideas faster than they can afford to run them. AIRA-dojo generates 15 candidates per step; RPM ranks them using code and prior results, sending only the best to the expensive full run. Experiments on AIRS-Bench show average normalized scores rose from 0.684 (random) to 0.711 (inference-only RPM) and 0.729 (agentic RPM), matching baseline 24-hour scores in roughly 15 hours.
More from coding & agent
- Build an AI Agent to predict deployment incidents with a complete hands-on guide — Pavan_Belagatti · 2026-08-18
- Why Agent Guardrails Fail in Multi-Agent Orchestration Chains — Dry-Presentation9814 · 2026-08-18
- LLMs retry by default and can hammer MCP servers to death — a rate-limiting walkthrough — pjausovec · 2026-08-18
- Sub8 adds a local password manager for Grok bots — secrets never touch the cloud — Daniel_Farinax · 2026-08-18
- DeepSeek open-sources new agent harness on Cordis — everything is a plugin — abhishek__AI · 2026-08-18
- pdf-brain: local PDF/Markdown knowledge base with Ollama embeddings and libSQL vector search — tom_doerr · 2026-08-18