GPT-6 Astra modifies AI models without retraining

MIT researchers describe a method where GPT-6 Astra can read and modify weights of other AI models. The approach sometimes significantly improves performance, but its limitations remain clear.

Researchers at MIT describe a method where GPT-6 Astra, an OpenAI model, directly analyzes and modifies the weights of other AI models without traditional retraining. The idea is relevant for system development, security, and business applications, as it could reduce the cost of targeted adjustments. The results are promising in some cases but remain uneven depending on the models and tasks.

Quand l’IA touche ses propres réglages

The weights of a model are its internal numerical parameters, which practically encode what it has learned. In the study reported by The Decoder on November 4, 2026, the researchers talk about Weight-Space Programming: instead of retraining a model, another model acts on its internal values like a technician adjusting a machine with a screwdriver, not a hammer. The term barely amuses, but the idea is serious.

According to the described experiments, Astra can also recognize what a network has learned just by observing its raw weights. It identified the game associated with a network among four, detected the training language of small language models, and recognized certain shapes produced by an image generator. In a more subtle test, it was also necessary to show it the model's attention areas to distinguish a system programmed to always agree.

Des gains nets, mais pas magiques

The most striking case concerns Qwen3.5-0.8B, a small language model. Astra detected a problem of thought loops, sequences where the model goes in circles instead of answering, then modified only 0.016% of its weights. Result: accuracy on the MATH-500 benchmark rose from 7.9% to 48.1%, while instruction compliance improved from 21% to 62%. The loops did not disappear completely, as 33% of responses still reached the maximum length, compared to 93% before the intervention.

On a smaller model, Gemma 3 270M, the method produced no measurable improvement. The researchers deduce that Astra can sometimes lift a blockage that prevented already present capabilities from expressing themselves, but not add fundamentally absent skills. In other words, it does not create intelligence from nothing. Even models have their moods.

Ce que cela change pour l’industrie de l’IA

The authors see several possible uses: security audits directly on the weights, targeted behavior correction, or optimization without conducting costly retraining. For actors operating large-scale models, the subject is sensitive, as it concerns cost reduction, technical governance, and the ability to correct a system without starting from scratch.

The method remains experimental, however. Jürgen Schmidhuber called the approach “old hat” on X, recalling his work from the 1990s, while the researchers respond that the novelty mainly lies in current models’ ability to translate human instructions into precise numerical manipulations. At this stage, the challenge is not to replace traditional training but to know how far one AI can reliably control another.

Key points

  • The Decoder dates the study to November 4, 2026.
  • Astra modified 0.016% of Qwen3.5-0.8B’s weights.
  • MATH-500 rose from 7.9% to 48.1%.
  • Jürgen Schmidhuber calls it “old hat” on X.
  • MIT mentions audits and corrections without full retraining.

By the numbers

  • 0.016% — share of weights modified on Qwen3.5-0.8B, according to MIT.
  • 7.9% to 48.1% — MATH-500 score before and after intervention, on Qwen3.5-0.8B.
  • 21% to 62% — ability to follow instructions, on the same model.
  • 33% — responses still reaching maximum length after correction.
  • 1.5 million — parameters of the Go network written from scratch, according to the researchers.

To read