I learned that a technique called Weight‑Space Programming lets a next‑gen LLM edit the weights of another model directly, without gradients or traditional training.
I used GPT‑6 Astra to edit only 0.016 % of the total parameters of Qwen3.5‑0.8B.
The edit fixed reasoning loops and boosted the model’s math capabilities by up to 6× compared to the unmodified baseline.
Instruction‑following performance also improved by up to 3× after the weight edit.
The team chose this tiny modification fraction to demonstrate that massive gains can be realized with minimal perturbation, supporting their claim of “native” weight manipulation capability.
🤯