MIT Researchers Show GPT-6 Astra Can Read and Rewrite Other AI Models' Weights

A neural network's weights are large tables of numbers, and humans cannot read them. Models are normally adjusted by retraining, a step-by-step optimisation that is expensive. In the MIT experiments Astra worked instead like a mechanic turning individual dials. Depending on the test, it either read the lists of numbers directly or used programming tools and feedback from test runs.

Reading What a Model Has Learned

In the first set of tests the researchers gave Astra the weights of trained networks and asked what they did. According to The Decoder, it correctly named which of four computer games a network had been trained to play and identified the basic shapes an image generator produced, although it confused filled shapes with outlines. It also matched small language models to the language they were trained on, without seeing any training data and without running the models.

The method had limits. Given two fine-tuned language models, one of them trained always to agree with users, Astra could not tell them apart from the weights alone. It identified the sycophantic model in all three pairs only after it was allowed to see what the models paid attention to while they processed text.

Writing Networks Without Training

Astra also generated weights from scratch for networks that play three simple games, using only the game rules and source code. The researchers say these beat networks trained with a standard method over 100 million rounds. In Breakout, Astra's network averaged 517 points against 25 for the trained baseline, but in the other two games it fell behind the longest-trained versions. For Go on a 9x9 board, Astra wrote a network of about 1.5 million parameters. It had 50 rounds of revisions, each based on up to 128 practice games, and the result beat a similar-sized model trained on roughly 20 million games.

The most practical test used Alibaba's small Qwen3.5-0.8B language model. Astra was asked to improve its maths and was evaluated on problems it had not seen. It found a known weakness that Qwen's own documentation warns about: thinking loops, in which the model repeats the same text instead of reaching an answer. Astra changed 0.016 per cent of the weights. Accuracy on the MATH-500 benchmark rose from 7.9 per cent to 48.1 per cent, and instruction-following improved from 21 per cent to 62 per cent. The loops were reduced but not eliminated: a third of responses still hit the maximum length, down from 93 per cent.

Astra produced no measurable gain on Google's smaller Gemma 3 270M model, which rarely fell into such loops. The researchers think Astra removed a blockage that was suppressing abilities Qwen already had, and that rewriting individual weights could not add skills a model fundamentally lacked.

Why It Matters

If the approach scales, it could offer cheaper ways to audit, debug and adjust models than full retraining. The researchers point to security audits carried out directly on model weights and to targeted behaviour changes. For companies that deploy open-weight models, editing weights to remove a known failure without retraining would be a real saving. The same ability, applied by one model to another, also raises oversight questions for regulators and safety teams.

Not everyone thinks the idea is new. AI researcher Jürgen Schmidhuber wrote on X that the approach was "old hat" and pointed to his own work from the 1990s. Khattab, one of the researchers, replied that the team cites those papers, which he described as "far ahead of its time". He said the new element is how well today's language models can translate between human language and a network's numbers.

The researchers acknowledge that the capabilities are still limited. They say a benchmark to measure the method systematically is their next step. They have published the experiment logs on GitHub and the edited Qwen model on Hugging Face.