GPT-6 Astra Edited A Small AI's Weights Directly And Made It 6x Better At Maths
0:00 / 0:53
News
GPT-6 Astra Edited A Small AI's Weights Directly And Made It 6x Better At Maths
calendar_today Date:
schedule Duration: 0:53
database
Summary Report
GPT-6 Astra edited 0.016% of Qwen3.5-0.8B's weights to fix reasoning loops, lifting MATH-500 from 7.9% to 48.1%.
- 01. Edited 0.016% of weights with no gradients or training
- 02. MATH-500 7.9% to 48.1%; IFEval about 3x; not peer reviewed
MIT CSAIL researchers asked GPT-6 Astra to improve Qwen3.5-0.8B at maths without training. Astra read the raw weights, found a flaw that caused reasoning loops, and edited just 0.016% of them directly. MATH-500 rose from 7.9% to 48.1% and instruction following improved about 3x. The team calls it weight-space programming. The result is early and not peer reviewed.
Meta Data
Model:
