GPT-6 Astra learns to edit neural network weights directly

MIT CSAIL researchers demonstrated that the GPT-6 Astra model can natively read and write neural network weights without gradient descent (the Weight-SSpace Programming approach), and that precise, gradient-free edits to fractions of a percent of weights eliminate defects like reasoning loops in Qwen3.5-0.8B, raising MATH-500 scores from 7.9% to 48.1%.

Compiled by Sergey KostenchukPublished 2026-11-02Updated 2026-11-02

2026-11-02 Research

🤖 GPT-6 Astra learns to edit neural network weights directly

MIT CSAIL researchers demonstrated: the frontier model GPT-6 Astra natively reads and writes neural network weights without gradient descent, training cycles, or environmental feedback — an approach called Weight-Space Programming. Based solely on raw weights (17,253 numbers), the model identified which game was trained in an MLP, and the language of a 7.2M-parameter transformer.

🌍 If a model optimizes the weights themselves rather than prompts and surrounding code, the training paradigm changes: gradient optimizers (Adam, Muon) may yield to 'reasoning about weights,' and precise edits are cheaper than full fine-tuning.

👤 On Qwen3.5-0.8B, editing 0.016% of weights removed the reasoning loop and raised MATH-500 from 7.9% to 48.1%, and IFEval by 3x; in physical control, a one-time edit of 3.5% of weights achieved success from 0% to 100%. 100% identification accuracy is currently claimed without control runs.

Source 1: https://rikiyat.github.io/blog/weight-space-programming/

Sources