Language models can rewrite their own weightsImage: arxiv.org
Developing

Language models can rewrite their own weights

Heat 84 · #1 in AI · Rising5 reports · 5 sources · First reported Nov 16 · Updated Nov 20

The story

Researchers report that a 27-billion-parameter open language model can rewrite its own weights by its own reasoning and become better, with no gradients and no training run. Given its weights, its graded answers and the ability to run itself, the model diagnoses its own failure and edits the weights behind it; its own edit makes it 6.6 points more accurate on 490 problems it never saw, rising from 57.2% to 63.8%. The same approach repairs a damaged copy of the model, lifting it from 7% to 70% accuracy on six image benchmarks it never saw, without ever reading the undamaged weights. The ability emerges only with scale: across four sizes of one model family, the 2-billion-parameter model makes no edit and the 4- and 9-billion-parameter models break themselves, while only the 27-billion-parameter model works. The work builds on earlier weight-space programming research, in which a frontier model edited the weights of a much smaller model.
Latest development
builds on the weight-space programming line in which a frontier model edited a smaller model.

How it unfolded

  1. builds on the weight-space programming line in which a frontier model edited a smaller model.
  2. welcome.ai, Substack and VibeLeaderboard summaries corroborate
  3. paper submitted Nov 16, 2026, revised Nov 18 (v2)
  4. the ability emerges with scale (2B makes no edit, 4B/9B break themselves, only 27B works)
  5. it also repairs a damaged copy of itself, lifting it on six image benchmarks it never saw from 7% to 70% accuracy
  6. its own edit makes it 6.6 points more accurate on 490 held-out problems it never saw (57.2% to 63.8%)
  7. Nov 21
    Jiajun Fan and colleagues (UIUC, CMU, Rutgers, Princeton, Stanford) report a 27B open model (Qwen3.5) that rewrites its own weights by its own reasoning, with no gradients and no training

The angle

A 27B model that fixes its own weights and gets 6.6 points better - what self-rewriting is, and what it is not (no recursive self-improvement, sessions restart from base weights)Suggested format: Short video · for AI watchers, developers, tech-curious general audience

Coverage (5)

Hacker NewsHacker News
arXivarXiv
WelcomeWelcome
SubstackSubstack
VibeleaderboardVibeleaderboard

Heat

Now 84 · Peak 84 · 9 recorded checks since Nov 20
Full ranking →