THE CRUNCH
Computer scientists at MIT's Computer Science and Artificial Intelligence Laboratory (CSAIL) report that frontier AI models can directly rewrite the weights of another model, changing its behaviour without any training. In a demonstration, the team asked OpenAI's GPT-6 Astra to edit the weights of Qwen3.5-0.8B to make it better at maths. Astra changed just 0.016 percent of the model's weights, cutting its reasoning loops and improving maths performance by as much as sixfold on the MATH-500 benchmark, while roughly tripling its instruction-following score on IFEval.
Crucially, there was no training via gradient descent, the standard optimisation method behind most model updates. Astra simply revised the numerical values that make up Qwen's weights. "We find early signs that frontier LLMs are surprisingly good at natively reading, writing, and reasoning about neural network weights," said Rikiya Takehi, a PhD student at MIT CSAIL, in a social media post. The researchers call the approach Weight-Space Programming, and describe it as models treating raw weights like readable, editable code: diagnosing failures and patching them directly.
The technique goes further than tweaking. Asked to write, from scratch, the weights of a policy network that plays the game Breakout as well as possible, Astra did so zero-shot, with no training at all. Takehi suggests the method could improve models while avoiding catastrophic forgetting, the loss of existing knowledge that can occur when a model is revised to include new knowledge. The authors also expect it to open doors to better model interpretability, reduced sycophancy and, eventually, recursive self-improvement.
The work remains early. The team's next step is building a benchmark and harness for Weight-Space Programming on LLMs, which they hope will turn what "currently seems like black magic into a rigorous, reproducible engineering discipline."