The Pulse
GPT-6 Astra Edits Qwen Weights, Raising Math Score to 48.1%
MIT CSAIL researchers say GPT-6 Astra edited 0.016% of Qwen3.5-0.8B’s weights without gradient training. On MATH-500, the edited model scored 48.1%, up from 7.9%, in the team’s evaluation.
AI.info Team · Written with AI · fact-checked ·
On November 2, four MIT CSAIL researchers described using GPT-6 Astra to rewrite part of another model’s weights directly, without gradient descent or a training loop. Their target was Qwen3.5-0.8B, whose tendency to repeat text until it hit its output limit was dragging down its math results. After Astra changed 131,072 of the model’s roughly 800 million weights, its MATH-500 score rose from 7.9% to 48.1% in the researchers’ evaluation.
Astra targeted Qwen’s reasoning loops
The researchers call the method “Weight-Space Programming”: a language model examines a neural network’s numerical parameters and edits them to meet a natural-language goal. For the Qwen experiment, Astra worked in a Python session with the model’s weights loaded and was allowed up to 50 forward passes on practice problems from GSM8K and MATH. It received no gradient information and did not train the model, but it could inspect outputs from those practice runs.
Astra traced the looping behavior to a direction associated with the model’s multilayer perceptron blocks in layers 15 and 19. It subtracted that direction from 64 output columns in each layer, changing 0.016% of the weights. The researchers say the change reduced the share of MATH-500 answers that ran to the 32,768-token limit from 93% to 33.1%.
The gains extended beyond MATH-500
With the same 50-pass edit, Qwen3.5-0.8B’s GSM8K score rose from 13.8% to 67.1%, while its IFEval instruction-following score increased from 21.0% to 62.3%. Those figures are results from the authors’ evaluation, which used separate test benchmarks Astra did not see while making the edit. The team also tried smaller forward-pass budgets: at 10 and 25 passes, MATH-500 scores reached 12.3% and 14.1%, respectively, far below the 48.1% result at 50.
The post places the Qwen result among eight experiments testing whether Astra could identify what networks do from their weights, fix a failing controller, or write small networks from scratch. In one control task, an edited policy cleared an obstacle in all 1,024 evaluation episodes after its original policy failed in 256. In a separate experiment, Astra wrote a small Breakout-playing network that scored 517.4, compared with 25.3 for the researchers’ PPO network trained for 500 million steps.
“Very early results,” with a clear failure case
The researchers’ own caution is explicit: Our current results do show that capabilities are exciting but still limited, specifically due to inefficiencies in context length and harnesses. Rikiya Takehi, Jayden Teoh, Alex Zhang, and Omar Khattab, MIT CSAIL researchers. Their experiment on Gemma 3 270M shows another limit: Astra could not meaningfully raise its math score, which moved from 6.8% to at most 7.7% across the tested edits. The authors argue that removing a specific obstacle is different from giving a model a skill it does not already have.
The researchers published experiment logs, prompts, code, evaluations, and weight files, and a rebuilt Qwen checkpoint is available on Hugging Face. The checkpoint page reports a rerun scoring 51.0% on MATH-500, with 31.6% of answers hitting the token limit. The post’s headline result remains the researchers’ 48.1% score; the rerun offers a second measured result for the released edit.