GPT-6 Astra directly rewrote another model's weights in a research article published on 2 November 2026. Four researchers from MIT CSAIL had Astra read the raw weights of Alibaba's open small model Qwen3.5-0.8B, find why it kept looping when doing maths problems, and then change 0.016% of its weights. Under the same sampling settings, its MATH-500 score rose from 7.9% to 48.1%. This is a research demonstration backed by complete experiment logs, not yet a general training method.
A loop was the small model's main problem
Astra traced the problem to the multi-layer perceptrons in layers 15 and 19, removing one direction that kept the model going around in circles, changing 131,072 weight entries in total. Before the edit, 93.0% of MATH-500 answers hit the limit of 32,768 new tokens without producing a result; afterwards that share fell to 33.1%. GSM8K rose from 13.8% to 67.1%, and the instruction-following benchmark IFEval rose from 21.0% to 62.3%. The researchers stress that the gain came from removing an obstacle, not from stuffing mathematical knowledge into the weights that was not there before. The best run, with 50 forward passes, took 20.3 hours, 158 conversational turns and 80 Python executions; the cost sits in the search process, not in the edit itself.
Astra also tried other jobs with reading and writing weights
In the same article, Astra identified, from raw weights alone, which of four games each of four small networks had been trained to play; in a single reply with no tools, it changed 3.5% of the weights of a policy network so that a car that kept getting stuck before a ramp succeeded in all 1,024 tests; and game networks it wrote from scratch beat a PPO network trained for 500 million steps on Breakout. All of these results come from the researchers' own evaluations and have not yet been independently reproduced. A version of the edited model rebuilt from the public logs scored 51.0% on MATH-500, which points in the same direction, but it was a separate rerun and cannot be placed directly alongside the figures in the article.
It only works on models with a clear fault
A control experiment draws the boundary. Gemma 3 270M rarely loops, and Astra's repeated edits only moved its MATH-500 score from 6.8% to between 7.0% and 7.7%, which the researchers say does not establish an improvement. Nor could Astra tell from weights alone which of two Gemma models had been trained to be more sycophantic. The 48.1% figure is the best of four runs; with 10 forward passes the score was only 12.3%. The value of this work is in showing that a frontier model can already make targeted, reasoning-driven changes to weights themselves; replacing gradient training or automatically repairing arbitrary large models is still a long way off.
