Related video: Anthropic mathematician claims Claude helped prove Riemann Hypothesis in six weeks, @TheAIGRID
Trending on Hacker NewsLanguage models can rewrite their own weights without retraining
arxiv.org · · 40-second read
Top comments from YouTube.
November 20, 2026 · Emerging AI
Researchers have demonstrated that a language model can rewrite its own weights autonomously to fix errors and improve performance, without gradients or retraining. A 27-billion-parameter , given access to its weights, graded answers, and the ability to run itself, diagnosed its own failures and edited its weights to become more accurate—improving by 6.6 points on unseen problems.
The model tested its edits on practice problems and deployed only successful changes. Over eight consecutive sessions, accuracy continued rising with each iteration. The researchers also showed the model could repair a damaged copy of itself to over 80% original performance on image tasks it had never encountered.
The ability appears to scale with model size: in one family, smaller models made no improvements or degraded themselves, while only the largest model successfully improved and repaired itself. The researchers frame this as a step toward recursive self-improvement, where models could continuously upgrade themselves during rather than requiring separate training phases.
This is an AI-written summary. The original reporting belongs to arxiv.org. How we write these · Spot a mistake?
Get the AI brief every morning.
The five AI stories that matter each morning: startups, money, jobs, tech and what's coming next. Free, and one click to unsubscribe. See our Privacy Policy.