New todaygpt-6POST
Author: Astra edits 0.016% of Qwen3.5-0.8B weights for ~6x math gain
The paper's author says asking GPT-6 Astra to rewrite Qwen3.5-0.8B's weights, changing only 0.016% of parameters, gave roughly 6x on math benchmarks and up to 3x elsewhere.
Rikiya Takehi 武樋力哉
@rikiyatakehi_jpArchived 3 hours agoOriginally posted 2026.11.03
Rikiya Takehi 武樋力哉 @rikiyatakehi_jpOriginally posted 2026.11.03
GPT-6 Astraに「Qwen3.5-0.8Bを数学に強くするように重みを書き換えて」と頼んだところ、全体のたった0.016%の重み変更で数学のベンチマークでは6倍・他のベンチマークでは最大3倍ほど性能が上がりました。 MITに来てからは最初の成果です！ https://rikiyat.github.io/blog/weight-space-programming/
This post quotes another post: “GPT-6 Astra edits model weights directly: Weight-Space Programming”
x.com/rikiyatakehi_jp/status/2117402970283188534Archived 2026.11.03 10:06