MIT Researchers Show GPT-6 Astra Editing Qwen3.5 Weights Without Training
MIT：GPT-6 Astra能直接修改Qwen权重，数学能力暴涨6倍
- Who
- MIT researchers
- What
- demonstrated using GPT-6 Astra to inspect and modify neural network weights through weight-space programming
- Quote
MIT刚刚做了个实验，瞬间引爆大模型圈—— 0.8B的「源神」Qwen3.5交给Astra，仅改动其中 0.016% 的权重，模型性能可以直接飞升。
What happened
MIT researchers used GPT-6 Astra to inspect and edit a Qwen3.5-0.8B model, changing 0.016% of its weights without training or gradient descent. Its MATH-500 score rose from 7.9% to 48.1%, while instruction-following performance also improved. The article describes additional experiments in which Astra identified network functions, modified a car game policy, and generated weights for a MinAtar network.
Why it matters
The experiments trace how Astra inferred a small model’s looping failure from its weights, then changed selected parameters without gradient training. The article also reports tests of weight inspection and network generation across several tasks.
Read next
See the full storyMIT Researchers Use GPT-6 Astra to Improve Qwen3.5 by Editing Its WeightsThe story so far
MIT researchers used GPT-6 Astra to inspect and edit a Qwen3.5-0.8B model, changing 0.016% of its weights without training or gradient descent. The model’s MATH-500 score rose from 7.9% to 48.1%, and its instruction-following performance also improved. In other experiments, Astra identified neural network functions, modified a car-game policy, and generated weights for a MinAtar network.
Keep up with AI, one email a day
Model releases, price changes and compute developments, with original sources. Free, confirmed by email, and easy to unsubscribe. Read a past issue · 2026-11-03