Stanford paper: 27B model rewrites its own weights via reasoning, no training needed
rohanpaul_ai · x · 2026-11-21
A new paper from Stanford and other top labs, "Language Models Can Rewrite Their Own Weights" (arXiv:2611.21885), shows for the first time a language model improving itself by editing its own weights through reasoning alone — no gradients, no training runs.
- An open 27B model is given its own weights, graded practice answers, and the ability to run itself; it diagnoses failures and rewrites the weights behind them, effectively acting as its own optimizer.
- Its self-edits boost accuracy by 6.6 points on unseen problems, and across 8 consecutive sessions accuracy keeps rising as it tests each edit and keeps only what works.
- Self-repair: a damaged, nearly non-functional copy of the model is restored to over 80% of original accuracy without retraining or ever reading the original weights.
- The ability emerges with scale: smaller models in the same family either make no edits or break themselves; only the largest improves and repairs itself.
Related event: Paper: LLMs Can Rewrite Their Own Weights Without Training(2 posts)→
More from Models
- Google ships 8 updates in 2 weeks: Gemini 4 Neon/Hydrogen/Carbon, Genie 4, Robotics 2.5 — GoogleAI · 2026-11-21
- OpenAI expands Daybreak security model access for defenders and opens GPT-6 Astra Blue alpha — OpenAIDevs · 2026-11-21
- Stop picking speech-to-speech models on price per minute — TCO is what matters — graceisford · 2026-11-21
- Persimmon v0.1 Pro and Ultra Now Live in Playground With API Access — gharik · 2026-11-21
- Unsloth appears set to release 4 new models, hidden in HF collection — coder543 · 2026-11-21
- Hume builds safety framework from scratch for human-like user simulator models — LukeW · 2026-11-21