MIT's weight-space programming: GPT edits 0.016% of Qwen's weights, 6x math boost
AI寒武纪 · wechat · 2026-11-03
MIT CSAIL PhD student Rikiya Takehi's "weight-space programming" experiments go viral: GPT-6 Astra reads raw neural network weights and identifies RL game policies, hand-writes a patch fixing a PPO cart (3.5% of weights, 100% success), writes a Breakout network from scratch outscoring a PPO model trained 500M steps by 20x, and fixes Qwen3.5-0.8B's reasoning death-loop by editing just 0.016% of weights — boosting math up to 6x and instruction-following 3x, all without gradient descent or training.
More from Research
- spaCy creator deconstructs the "LLM writes weights" paper: it detects loops, not mental simulation — honnibal · 2026-11-03
- MIT: AI edits 0.016% of another model's weights, lifting MATH-500 from 7.9% to 48.1% — rohanpaul_ai · 2026-11-03
- Schmidhuber: AI weight editing is 'old hat' from his 1990s fast weight programmers — SchmidhuberAI · 2026-11-03
- UCL's TAL lets LLM agents learn from tutorial videos with no parameter updates, +113% on VidLearn — UniversityCollegeLondon · 2026-11-03
- SIMIT: vision-language models self-improve at test time by imagining their own practice data — dtudk · 2026-11-03
- Face quality scores can backfire: higher quality often signals a morphing attack — FraunhoferIGD · 2026-11-03