New paper explores weight-space programming for native neural network manipulation

On November 3, 2026, a post about a new paper appeared on Reddit's LocalLLaMA community. The paper reports that frontier LLMs are surprisingly adept at natively reading and writing neural network weights, performing well beyond what one would expect from humans or conventional algorithms. Its authors suggest the finding could matter for both optimization and interpretability research.

According to the post's summary, GPT-6 Astra can tell what a given network was trained to do just by looking at its raw weights, and it can also write small game-playing networks from scratch without any environment feedback. The networks it produces outperform a baseline trained with PPO for 100 million steps. As of now, these claims come solely from the community post's retelling, and no further details from the paper itself are available.

What happens next is unclear: the post gives no information about whether the paper has been peer-reviewed, whether its code or weights will be released, or how the benchmarks were designed. Readers interested in the claims currently have only the community summary to go on.