Qwen3.5-0.8B: 0.016% Weight Edit Delivers 6x Math Gain
Rikiya Takehi reports that GPT-6 Astra performed Weight-Space Programming on Qwen3.5-0.8B, directly rewriting 0.016% of weights to raise math performance sixfold and instruction following threefold with no gradients or training required.
SourceAnalysis
GPT-6 Astra directly rewrote 0.016% of Qwen3.5-0.8B weights to fix reasoning loops, lifting math scores up to 6x and instruction following 3x without gradients or training. The demonstration, posted by Rikiya Takehi on 3 November 2026, frames the technique as Weight-Space Programming and points to mechanistic interpretability work that treats network parameters as readable, editable programs rather than opaque artifacts produced only through gradient descent.
傅盛
@FuSheng_0306Chairman and CEO of Cheetah Mobile, Chairman of OrionStar