Updated todayNewin the edition since 3 Nov 20269 sources

MIT CSAIL weight-space programming

The gist

Latest: MIT "Weight-Space Programming" uses GPT-6 Astra to edit rival model weights

The technique enables a frontier model to diagnose and repair another model's learned behavior by operating directly on weights without any retraining. Alex Krentsel noted that a system able to rewrite weights can in principle modify itself, connecting the approach to questions about recursive self-improvement.

When stories like this move, I write about them in an email each morning.

One free email each morning, around 8 AM US Eastern.

The full picture

MIT CSAIL researchers Rikiya Takehi, Jayden Teoh, Alex Zhang, and Omar Khattab introduced "weight-space programming," using GPT-6 Astra to inspect and directly edit another model's weights without gradient-based training. Applied to Qwen3.5-0.8B, Astra diagnosed a reasoning loop caused by two layers generating a repetitive internal signal and fixed it by editing 131,072 weights (0.016% of total), raising MATH-500 accuracy from 7.9% to 48.1% and improving instruction following up to 3x. In the paper's experiments, GPT-6 Astra correctly identified all four Atari game policies (Breakout, Pong, Pac-Man, Freeway) from raw weights alone with no tools provided. Astra also fixed a stuck PPO car-ramp policy by writing a weight patch in a single reply with no tool-use, editing 3.5% of weights to achieve 100% task success. The technique also covers writing small neural networks from scratch in weight space. A follow-up on Gemma 3 270M IT showed only marginal improvement, MATH-500 rising from 6.8% to 7.0%-7.7%, with the team noting the Qwen result did not automatically carry over. The work has not been peer reviewed.

How it developed

3 November 2026

MIT CSAIL researchers published results showing GPT-6 Astra raised Qwen3.5-0.8B's MATH-500 accuracy from 7.9% to 48.1% by directly editing its weights, subtracting a repetitive signal from two layers that caused it to loop, touching only 131,072 parameters.

Alex Krentsel independently reproduced the core result and surfaced one additional finding. A follow-up on Gemma 3 270M IT found MATH-500 rose from 6.8% to between 7.0% and 7.7%, and the researchers noted the Qwen outcome did not automatically carry over.

First citedRohan Paul on X daily.dev kingy.ai latent.space

Sources

Rohan Paul on X