Overview
- MIT CSAIL researchers described a technique called weight-space programming where an LLM analyzes and writes another model’s neural network weights inside a controlled Python runtime.
- In their key experiment the team had OpenAI’s GPT-6 Astra locate a repeated MLP write direction in Qwen3.5-0.8B and change 131,072 weights, about 0.016% of the model’s parameters.
- Those localized edits reduced a looping signal that caused answers to hit token limits and raised benchmark scores sharply, with MATH-500 jumping from 7.9% to 48.1% and GSM8K from 13.8% to 67.1%.
- A control test on a smaller Gemma 3 model produced minimal gains, which the authors say shows the method clears obstacles to a model’s existing skills rather than creating new capabilities out of nothing.
- The authors suggest uses for interpretability and targeted optimization but note the need for full weight access, independent replication, and more work to assess robustness and safety implications.