GPT-6 edits 0.016% of Qwen3.5-0.8B weights, boosting math scores up to 6x with no training
SonglinYang4 · x · 2026-11-03
A team demonstrated "Weight-Space Programming": GPT-6 Astra directly read and edited Qwen3.5-0.8B weights to improve math — no gradients, no training.
- Only 0.016% of weights were changed, fixing the small model's reasoning loops
- MATH-500 jumped from 7.9% to 48.1% (50 forward passes; 49.45% in non-thinking mode), GSM8K from 13.8% to 67.1% — up to 6x improvement
- IFEval rose from 21.0% to 62.3% (3x), and hit-token-limit rate fell from 93% to 33.1%
The authors argue frontier LLMs can now natively read and write neural network weights. Details in their blog post.
More from Models
- Claude's charming support menu: vent, think-it-thru or rest mode — repligate · 2026-11-03
- Miles Brundage: Claude Code has gotten pushy with suggested follow-up projects — Miles_Brundage · 2026-11-03
- Hand-assigning every weight: blogger builds a working transformer with zero training — voooooogel · 2026-11-03
- ComfyUI User Details How Flux2Klein and Qwen 2.1 Fail at Character Height Differences — Ishelle91 · 2026-11-03
- Community rebuilds edited Qwen3.5-0.8B from logs, scores 51.0% on MATH-500 — _akhaliq · 2026-11-03
- Claude's alignment answers aren't literal defaults but a posture of supplying the right answer — 1a3orn · 2026-11-03