Weight-Space Programming: GPT-6 edits just 0.016% of Qwen3.5-0.8B weights, boosting math up to 6x
HaggaiMaron · x · 2026-11-04
A team had GPT-6 Astra directly edit Qwen3.5-0.8B's weights to improve math, dubbing the approach Weight-Space Programming.
Key results:
- Only 0.016% of weights changed — no gradients, no training
- Fixed the small model's reasoning loops
- Math improved up to 6x, instruction following up to 3x
The work extends the ICLR 2025 Weight Space Learning workshop's agenda of treating neural network weights as a new data modality, covering weight-space characterization, model analysis from weights, and model synthesis via pruning, merging, and generation.
More from Models
- OpenAI's Sol, Astra and Cyber models could drop as soon as Friday — haider1 · 2026-11-05
- Underdog alert: Xiaomi MiMo keeps competing with top models on cost-adjusted benchmarks — VictorSanh · 2026-11-05
- Pangram Image 1.0 detects 99.2% of AI images at 0.27% false positive rate — TuhinChakr · 2026-11-05
- Tencent reportedly rolls out AI model trained with limited cybersecurity data amid hacking fears — pstAsiatech · 2026-11-05
- Users can't tell frontier models from 50x cheaper ones, orchestration may win — RihardJarc · 2026-11-05
- Hydrogen-1.0, a science-tuned GLM-5.3 finetune, trends on Hugging Face — deep-principle · 2026-11-05