Weight-Space Programming: GPT-6 Astra edits 0.016% of Qwen3.5-0.8B weights, 6x math gain without training
ChengleiSi · x · 2026-11-03
rikiyatakehi's team demos "Weight-Space Programming": frontier LLMs natively reading and writing neural network weights. GPT-6 Astra was asked to edit Qwen3.5-0.8B weights to improve math — it changed only 0.016% of weights, fixing the small model's reasoning loops and boosting math up to 6x and instruction following 3x, with no gradients or training.
A follow-up shows agents generating policy weights to play a game in one shot with no interaction; full results for Astra, Argon, and Carbon are in the thread.
More from Research
- Zyphra's ZHA brings subquadratic attention via hierarchical tree-structured indexer — yacineMTB · 2026-11-03
- GPT-6 edits Qwen weights directly, lifting MATH-500 from 7.9% to 48.1% with no gradients — KyeGomezB · 2026-11-03
- RSNA 2026 lab walks through fine-tuning a VLM for radiology end to end — PRouzrokh · 2026-11-03
- RSNA 2026 lab walks through fine-tuning a VLM for radiology end to end — PRouzrokh · 2026-11-03
- Trajectory launches post-training platform with GS1 and Continual Graders research previews — simonguozirui · 2026-11-03
- Wharton Study: Persuasion Principles Lift LLM Compliance from 33% to 72% — adrianscottcom · 2026-11-03