LLM edits small model weights directly with no backprop, fixing a stuck PPO policy by patching 3.5% of weights
bookwormengr · x · 2026-11-03
A striking demo shows a large model reading and directly editing a small model's weights without backpropagation: a PPO-trained car policy kept getting stuck on a ramp, and given the weights, the scene, and one failed trace, Astra wrote a patch in a single reply with no tool use — editing just 3.5% of weights to reach 100% success. The poster hails it as a possible backprop replacement for small models and asks the author about catastrophic forgetting. Single-case demo, unverified at scale, but a fascinating "model patches model" paradigm.
Related event: MIT: LLM Edits 0.016% of Weights to Boost Small Model Math Sixfold(40 posts)→
More from Research
- RoboArchitect turns frozen VLA policies into memory-equipped robot agents, lifting success rates — _Hao_Zhu · 2026-11-03
- ARC Prize 2026 closes with 18.5K entrants and 67K submissions, winners due December 4 — GregKamradt · 2026-11-03
- Should AI read cardiac MRI first? Lancet Digital Health study on ejection fraction — EricTopol · 2026-11-03
- "Emergent capability" claim mocked; Fleet hiring researchers to automate mech interp with frontier models — sebkrier · 2026-11-03
- Goodfire's Neural Geometry series: steering models along manifolds — burny_tech · 2026-11-03
- Prediction: an open model rated for 10M context within 3 months — teortaxesTex · 2026-11-03