● ResearchNov 3, 2026 · 04:408 sources
GPT-6 Astra Edits Neural Weights to Boost Qwen Model's Math Score to 48.1%
From marketbrief.now
AI;DR
Researchers at MIT CSAIL used GPT-6 Astra with Weight-Space Programming to repair the Qwen3.5-0.8B model. Astra edited 131,000 weights in two MLP layers, raising the MATH-500 score from 7.9% to 48.1% without traditional gradients or retraining. The experiments used a budget of 50 forward passes and cost approximately $1,000.
Read originalSign in to suggest
Sources
+ 7 more
- SUPPORT@rikiyatakehi“With a budget of 50 forward passes on practice problems, it located the cause and edited it out.”x.com
- SUPPORT@xennygrimmato_“Astra compared activations on looping vs. normal traces, found the direction that separates them, and edited 131k weights in two MLP layers to push against it.”x.com
- SUPPORT@timdarcet“We asked GPT-6 Astra to edit Qwen3.5-0.8B weights to improve math.”x.com
- SUPPORT@paul_cal“It can differentiate a neural net trained on pacman from one trained on pong, or space invaders, just by reading the weights.”x.com
- SUPPORT@teortaxestex“not as insane as it sounds from "Astra directly edited 131K weights", this is somewhat basic surgery humans can do too”x.com
- SUPPORT@willdepue“tldr very much signs of life but nothing super dramatic”x.com
- SUPPORT@tszzl“designed and not grown”x.com