Skip to main content
Latest Update
2026-11-02 12:17 UTC

MIT CSAIL: GPT-6 Astra rewrites weights to boost Qwen math

MIT CSAIL: GPT-6 Astra rewrites weights to boost Qwen math

According to The Rundown AI citing MIT CSAIL’s Rikiya Takehi, GPT-6 Astra edited 0.016% of Qwen3.5-0.8B’s weights to raise a math score to 48.1% from 7.9%, a method termed Weight-Space Programming.

Source

Analysis

An MIT CSAIL team has shown that GPT-6 Astra can improve certain AI models by directly rewriting their weights.

According to The Rundown AI on November 2, 2026, a single edit raised the math score of the Qwen3.5-0.8B model from 7.9% to 48.1%.

The method, called Weight-Space Programming, altered only 0.016% of the weights, fixed reasoning loops, boosted math performance up to six times and instruction following up to three times, and required no gradients or training, see the blog post by Rikiya Takehi.

The Rundown AI

@TheRundownAI

Updating the world’s largest AI newsletter keeping 2,000,000+ daily readers ahead of the curve. Get the latest AI news and how to apply it in 5 minutes.