GPT-6 Edits Qwen3.5-0.8B Weights Directly: 0.016% Change Boosts Math 6x

coallaoh · x · 2026-11-04

rikiyatakehi's team demonstrates "Weight-Space Programming": frontier LLMs natively reading and writing neural network weights to improve other models—no gradients or training required.

Experiment: GPT-6 Astra directly edited Qwen3.5-0.8B's weights to improve math
By changing only 0.016% of the weights, it fixed the 0.8B model's reasoning loops, improving math up to 6x and instruction following 3x
@oodgnas breaks down the qwen-math experiment: Astra uses training examples (inputs, labels) plus activations (e.g., model.layers.{l}.mlp.downproj.weight) to statistically diagnose and edit suspicious neurons—a new kind of optimization paradigm

Original post →

GPT-6 Edits Qwen3.5-0.8B Weights Directly: 0.016% Change Boosts Math 6x

More from Models