LLMs Generate Improved Neural Networks with Source-Guided Adaptation

Large language models (LLMs) are increasingly being explored for their ability to generate code and even neural network architectures. However, unrestricted generation often leads to invalid or suboptimal designs. This research focuses on a more constrained and effective approach: using LLMs to improve a weaker target neural network model by leveraging guidance from a stronger source model within the same architectural family, drawing from a neural-network database. The proposed protocol involves generating candidate modifications with source-conditioned inputs, alongside non-source controls and a no-LLM ablation. The key finding is that LLMs can significantly enhance the target model's performance. For instance, on CIFAR-10, the source-guided approach achieved a substantial accuracy advantage over non-source candidates, improving a weak target model from 0.1254 to 0.5049. Similar gains were observed on SVHN AlexNet. Crucially, the study disentangles transfer from adaptation, showing that the LLM does not merely copy the source model's recipe but actively adapts it to the target model's context. This indicates a more sophisticated form of knowledge transfer, where LLMs can intelligently modify and optimize network designs based on learned patterns and guidance.