Researchers say LLMs can be good at natively reading and writing neural network weights; GPT-6 Astra edited Qwen3.5-0.8B's weights to make it better at math (Rikiya Takehi)

Researchers have found that Large Language Models (LLMs) demonstrate proficiency in natively reading and writing neural network weights. For instance, GPT-6 Astra successfully edited the weights of Qwen3.5-0.8B, resulting in improved mathematical capabilities for the latter. This suggests a novel approach to neural network development and optimization, potentially allowing LLMs to directly manipulate and enhance other models' functionalities.

This report details the first instance of an LLM, GPT-6 Astra, directly editing another model's weights to improve performance, unlike previous methods that relied on fine-tuning or architectural changes.