MIT: GPT-6 edits 0.016% of Qwen weights, boosts MATH-500 from 7.9% to 48.1%
Puzzleheaded-King584 · reddit · 2026-11-03
MIT CSAIL researchers (Rikiya Takehi, Jayden Teo, Alex Zhang, Omar Khattab) published "Sparks of Weight-Space Programming":
- Asked to improve Qwen3.5-0.8B's math without gradients or training, GPT-6 Astra edited only 0.016% of its weights, fixing the reasoning loops even Qwen's model card warns about
- MATH-500 jumped from 7.9% to 48.1% (6x), and instruction following (IFEval) improved 3x out of the box
- Just by reading raw weights, GPT-6 can tell what a network was trained to do, and can write small game-playing networks from scratch that outplay networks PPO-trained for 100M steps
They call this Weight-Space Programming — frontier LLMs natively reading and writing neural network weights, potentially a big deal for optimization and interpretability.
More from Models
- The '7x accuracy' weight-editing claim was mostly just turning off thinking mode — ChengleiSi · 2026-11-03
- Dev compares astra and sol to Opus 5.5: rigid on open-ended tasks — haider1 · 2026-11-03
- Open-source models now beat every 'frontier' model from just 11 months ago — Hesamation · 2026-11-03
- Sebastian Raschka Maps October's 'Open October': 10 Open LLM Architectures in One Month — rasbt · 2026-11-03
- First hands-on: Meta's Muse Wonder 1.0 vs DeepSeek V4.1 Flash in Codex CLI — Richidm · 2026-11-03
- "Claude Code has been terrible today" — developer smells a new Opus launch coming — hermanschutte · 2026-11-03