Astra can read and write weights to improve Qwen

Researchers at MIT CSAIL report that GPT-6 Astra can improve a neural network by reading and editing its weights directly, the stored numbers that determine what the network computes, instead of retraining it with an optimizer. Their main case is a small Qwen model, Qwen3.5-0.8B, that often failed math problems because it got stuck repeating the same text until it ran out of room. Astra traced that looping to a single direction in the model's internal activity, written by the feed-forward (MLP) blocks of layers 15 and 19, and subtracted that direction from 64 output columns in each of those layers. The edit changed 131,072 of the model's roughly 0.8 billion weights, about 0.016% (MIT CSAIL blog post; OpenAI introduced Astra in September).

How well the edit worked depended on how much Astra was allowed to watch the Qwen model run. The post counts this in forward passes, where one pass means the Qwen model answers up to 512 practice problems taken from the GSM8K and MATH training sets, not the test sets. With a budget of 50 passes, the edited model scored 48.1% on the MATH-500 test against 7.9% for the original, 67.1% against 13.8% on GSM8K grade-school math, and 62.3% against 21.0% on IFEval, an instruction-following test. The share of MATH-500 answers that ran to the 32,768-token limit fell from 93.0% to 33.1%. With 10 or 25 passes, that share barely moved (87.2% and 86.4%) and MATH-500 reached only 12.3% and 14.1%.

The authors read the result as uncovering ability the model already had, not adding new ability. On Gemma 3 270M, a model that rarely loops, Astra produced no meaningful gain (MATH-500 moved from 6.8% to between 7.0% and 7.7%), and the authors conclude that editing weights "could remove an obstacle that hid an ability, but it could not add an ability the model lacked." The post also reports smaller demonstrations, such as naming which of four arcade games a policy network plays from its weights alone. It calls the results "exciting but still limited," it is a blog post with published logs and code rather than a reviewed paper, and it does not say how the researchers accessed Astra.