MIT shows one model can improve another without retraining
MIT researchers demonstrate that one model can directly modify another’s weights to fix a weakness without a new training phase. The approach paves the way for self-improving AI, but results remain highly variable.
MIT researchers have shown that one AI model can improve another by directly modifying its weights, without going through classical retraining. In the study reported on November 3, 2026, by The Rundown, GPT-6 Astra notably managed to correct a small math model by adjusting a tiny fraction of its parameters. This topic matters because it brings labs closer to AI capable of self-adjustment, with potential gains in performance, maintenance, and defect detection. For teams training models, it already looks like a shortcut that raises eyebrows.
Quand une IA réécrit les poids d’une autre
The idea is simple to state, harder to achieve: instead of sending a model into a new learning phase, another model inspects its weights—the numbers that determine its behavior—and then adjusts them directly. MIT researchers tested this method with OpenAI’s GPT-6 Astra on several small public models. On Alibaba’s Qwen, the tested AI scored 7.9% in math before correction, then 48.1% after modifying only 0.016% of its weights, according to the cited source. On Google Gemma 3, the effect was much more limited [to be verified].
This difference tells something important: the method seems mostly effective when it identifies a clear weakness, such as a tendency to repeat the same answers. In Qwen’s case, Astra removed this behavior, which raised the score. In Gemma 3’s case, no single flaw appeared as clearly, so the intervention changed little.
Un test utile, mais pas encore une baguette magique
The authors also report more exploratory results, such as asking the model to name a game learned only from its weights, or to generate a simple Breakout player without additional training. These trials remain modest but point to a broader idea: AI could eventually add or remove skills in another model or check for hidden problems. This is the part that interests labs, as it reduces dependence on costly and slow training cycles.
However, the study does not claim that a generalist model can already self-repair in production. Results depend on the starting point, defect type, and target model. In this context, talking about self-improvement is not an exaggeration, but it remains a lab promise, not a finished product.
Pourquoi ce travail change la conversation sur les agents
Major AI players have been working for months to build systems capable of conducting research, self-correcting, and progressing with almost no human in the loop. This MIT study feeds that scenario. If one model can modify another’s behavior with enough finesse, the next step is no longer just making chattier assistants but systems that iterate on themselves.
The consequences are twofold. On one hand, this could speed up the development of more reliable or specialized models at lower cost. On the other, it opens the door to harder-to-control systems, since a model able to rewrite other models also becomes a sensitive tool of technical governance. The day AI gives itself a polish, security teams will want to keep the manual handy.
Key points
- MIT published the study on November 3, 2026.
- GPT-6 Astra modified 0.016% of Qwen’s weights.
- Qwen’s math score rose from 7.9% to 48.1%.
- Gemma 3 reacted little to the same method.
- OpenAI also aims to measure everyday tasks with LifeVal.
By the numbers
- 0.016% — portion of weights modified on Qwen, according to MIT.
- 7.9% — Qwen’s initial math score, according to the study reported on November 3, 2026.
- 48.1% — score after correction on the same test.
- 4,000 people — size of the survey used by OpenAI for LifeVal.
- 16 tasks — scope of the LifeVal evaluation cited by The Rundown.