Astra reads and edits other models' weights without full retraining
Original title: GPT-6 Astra sees the matrix and reprograms other AI models directly through their weights
Researchers say changing 0.016 percent of Qwen3.5-0.8B weights raised MATH-500 accuracy from 7.9 to 48.1 percent.
Researchers describe Weight-Space Programming, in which a language model called Astra inspects raw network weights and, in some tests, writes or edits them using tools and test feedback. Astra identified which of four games a network was trained for and generated untrained game-playing weights that they say beat a standard baseline in Breakout. On Qwen3.5-0.8B, 50 iterations reduced thinking loops and raised scores; Gemma 3 270M showed no measurable gain. Experiment logs are on GitHub and an edited Qwen model is on Hugging Face.
Unconfirmed: Performance gains and the claim that Astra only unblocked existing Qwen skills are the researchers' own results.
- Breakout weights averaged 517 points versus 25 for the trained baseline.
- A 9x9 Go network of about 1.5 million parameters beat a model trained on roughly 20 million games.
- Instruction-following on the edited Qwen model rose from 21 to 62 percent.
- A third of responses still hit the length limit, down from 93 percent.
- Jürgen Schmidhuber called the idea old; authors say language-to-weight translation is new.
- AI models
- AstraQwen3.5-0.8BGemma 3 270M
From the original article
3 min read at the-decoder.com
MIT researchers show that OpenAI's GPT-6 Astra can read and deliberately rewrite the weights of other AI models without traditional training. In one experiment, Astra boosted a Qwen model's math accuracy from 7. 9 to 48.
By Maximilian Schreiner · the-decoder.com
A short excerpt; the article and its images belong to the publisher. Summary, key facts and assessment above are AI-written from the article and can be wrong.