Research World Models: predict which AI research experiments are worth the compute
yuyinzhou_cs · x · 2026-11-11
Researchers introduce Research World Models (RWMs): research agents propose more ideas than GPUs can run, so a fixed LLM reads the research environment, a proposed code/recipe change, and past experiment records to predict which changes will pay off before spending compute.
Key results (built on 2,653 real experiment records, 9 research environments, 171,862 H100 GPU-hours):
- 13/13 models with 320 past records in context beat the best model without them
- Transfer: using only OLMo3/Nanochat/Marin records (zero Qwen3 records), selection regret on Qwen3-153M experiments drops 78%
- Decisions: starting search from past records yields +15.8% higher final best gain over 24 experiments in multi-round autoresearch
Model weights stay fixed — experience is supplied purely in context, turning researchers' intuition into reusable, transferable knowledge. Paper and website available.
Related event: Research World Models: LLMs Predict Which Experiments Are Worth Running(2 posts)→
More from coding & agent
- Salesforce's Headless Toolkit maps a seven-stage lifecycle for agent-driven development — msrivastav13 · 2026-11-11
- 7 production AI agents replaced manual claims work in 4 months: a healthcare automation breakdown — alex_verem · 2026-11-11
- 100 parallel Harbor rollouts on Tensorlake hit 47 rollouts/min at just 37¢ per batch — Wild_Independent_518 · 2026-11-11
- Google donates Agent Substrate, gVisor, OpenRL to CNCF; post reveals who runs agent swarms in prod — michaelTM_ai · 2026-11-11
- Inkbox AI raises $5.3M to give agents their own email, phone and WhatsApp — EXM7777 · 2026-11-11
- Everyone wants AI agents, but nobody owns the "is it actually done?" check — AIFlow_ML · 2026-11-11