Weight-Space Programming Improves Qwen3.5-0.8B via GPT-6 Astra
November 3, 2026
Using the ML Intern harness, GPT-6 Astra edited only 0.016% of Qwen3.5-0.8B weights to improve math reasoning by 6x and instruction following by 3x without traditional gradient training.
HOW THIS AFFECTS YOU
●
builderYou may soon be able to optimize small models via direct weight editing rather than expensive fine-tuning.
●
researcherYou can explore weight-space programming as a gradient-free method for model optimization.