MIT Paper: Random Weight Perturbations Can Improve Large Models, Rivaling PPO
serrjoa · x · 2026-11-04
A new paper by Yulu Gan and Phillip Isola (MIT), 'Neural Thickets: Diverse Task Experts Are Dense Around Pretrained Weights,' argues pretraining yields a distribution over parameter vectors rather than a single point. In small models, task-specific experts occupy a negligible volume, requiring gradient descent; in large well-pretrained models, task-improving specialists densely populate the neighborhood of pretrained weights—even random perturbations help, and this strengthens with scale. The authors propose a fully parallel post-training method: sample N random perturbations, keep top K, ensemble via majority vote—competitive with PPO, GRPO, and ES on contemporary large models. Commenters verified the code behind the surprising claim.
More from Research
- Karpathy Boosts Open Hacker News Dataset: 10M+ Posts Usable as a Personal Knowledge Base — huangyun_122 · 2026-11-04
- PixelCraft: A Multi-Agent Framework Targets Multimodal AI's Chart-Reading Failures — burkov · 2026-11-04
- UBC's Kev-Bio VLM guides ultrasound landmark detection via clinical questions — UBC-V · 2026-11-04
- PixelBench: testing multimodal models pixel-by-pixel instead of with QA — bdsqlsz · 2026-11-04
- François Chollet: modeling brain wiring as an engineering problem constrained by genome size and time — CatAstro_Piyush · 2026-11-04
- OffsecCyber dataset combines HackerOne reports and CVEs to train pentesting LLMs — AccBalanced · 2026-11-04