Weight-space programming: AI edits another model’s weights
MIT CSAIL researchers report GPT-6 Astra reading and editing neural-network weights directly. In their Qwen3.5-0.8B experiment, changing 0.016% of weights reduced reasoning loops and raised MATH-500 from 7.9% to 48.1%, without gradient-based retraining. These are author-reported experiments, not a released ChatGPT feature.
Source checked Nov 4, 2026
Listen to this articleAbout 2 min
Preparing audio…
AI narration. Your voice and speed choices stay in this browser.
EDITORIAL TAKEAWAY
A useful research direction for repairing specific model failures. Reducing repetitive output can reveal existing ability; it does not establish a general way to teach missing reasoning skills.
Before you try it
- Published November 2 by Rikiya Takehi, Jayden Teoh, Alex Zhang and Omar Khattab. We reviewed the post and linked model card; we have not reproduced the experiments.
- The Qwen result used Python weight editing and 50 batched forward passes for feedback. It was not a zero-feedback repair, and no gradient training does not mean no compute.
- A separate Gemma 3 270M experiment showed no clear math improvement. Identifying sycophancy from weights alone also failed; attention probes were needed.
A small next step
Our suggested way to explore the change.
- Read the experiment setup beside each headline result.
- For a model you maintain, measure repeated-output failures separately from incorrect reasoning.
- If exploring weight edits, compare a copied checkpoint with the original on held-out tasks and check for regressions.
Go to the source
Takehi et al. · Weight-Space Programming research post ↗Qwen · Qwen3.5-0.8B model card (model context) ↗Availability can change. Check the maker's current requirements before making a decision.