Meta's Research Preference Models paper teaches AI agents 'research taste'
Ibrahimdidamson · x · 2026-09-04
Meta's new paper on Research Preference Models (RPMs) tackles a key problem for automated R&D: how to instil "research taste" in agents running experiments.
Why it matters:
- Scaling of Nano-GPT speed-runs (Elie Bakouch) showed agents get stuck in local optima, burning compute on hyperparameter tuning instead of novel ideas.
- PostTrainBench found most models focus on algorithmic tuning rather than curating high-quality datasets — agents hate working with data as much as researchers.
The RPM approach:
- Treat each experiment as a node in a tree;
- Score nodes by downstream eval performance;
- Mutate nodes to generate candidate experiments;
- Use an LLM-based preference model to steer exploration toward genuinely novel directions and out of local optima.
More from coding & agent
- Free Design Skill for Lovable brings Linear-grade design systems to AI-built apps — damienghader · 2026-09-04
- Dev keeps forgetting he installed a Vegeta skill, gets delightful Dragon Ball replies — tekbog · 2026-09-04
- Claude Code vs Codex: one-shot migration takes 37 min with zero rework vs 27 min but broken twice — notherealironman · 2026-09-04
- Open source coding agent opencode passes 200k GitHub stars — anomalyco · 2026-09-04
- Post-mortem: why in-memory agent state graphs failed in 24/7 production, and the disk-backed fix — Remarkable_Plant7820 · 2026-09-04
- Ian Nuttall burns through all Ahrefs MCP credits in one week, falls back to computer use — iannuttall · 2026-09-04