MIT researchers use GPT-6 Astra to edit Qwen model weights

MIT CSAIL researchers used OpenAI's GPT-6 Astra to directly edit the weights of a small Qwen model, according to VentureBeat. The changes raised Qwen3.5-0.8B's MATH-500 score from 7.9% to 48.1% while modifying about 0.016% of its weights.

The researchers call the method weight space programming. Astra inspected the model through a Python environment and proposed weight changes without using the gradient based process found in conventional training.

The patch changed 131,072 weights in layers 15 and 19. It removed a numerical direction associated with repetition, reducing cases where the model repeated text until reaching its output limit and allowing it to complete more answers.

The strongest attempt used 50 evaluation requests, 158 conversational turns and 80 Python runs over 20.3 hours, including 4.7 hours lost to a service outage. The experiment showed a targeted model repair, but did not establish that direct weight editing is faster, cheaper or more effective than conventional training.