This is a page that curates AI-related papers published worldwide. All content here is summarized using Google Gemini and operated on a non-profit basis. Copyright for each paper belongs to the authors and their institutions; please make sure to credit the source when sharing.
LM-Searcher: Cross-domain Neural Architecture Search with LLMs via Unified Numerical Encoding
Created by
Haebom
Author
Yuxuan Hu, Jihao Liu, Ke Wang, Jinliang Zhen, Weikang Shi, Manyuan Zhang, Qi Dou, Rui Liu, Aojun Zhou, Hongsheng Li
Outline
This paper proposes LM-Searcher, a neural architecture search (NAS) framework applicable across diverse tasks, leveraging large-scale language models (LLMs). While existing LLM-based NAS approaches have limitations due to their heavy reliance on prompt engineering and domain-specific tuning, LM-Searcher performs neural network architecture optimization across diverse domains without domain-specific adaptation. To achieve this, we utilize NCode, a universal numeric string representation for neural network architectures, to enable cross-domain architecture encoding and exploration. Furthermore, we reframe the NAS problem as a ranking task and train the LLM using directed tuning samples derived from a novel pruning-based subspace sampling strategy to select high-performing architectures from a candidate pool. A refined dataset containing diverse architecture-performance pairs promotes robust and transferable learning. Extensive experiments demonstrate that LM-Searcher achieves competitive performance on both within-domain (e.g., CNNs for image classification) and across-domain (e.g., LoRA architectures for segmentation and generation), presenting a new paradigm for flexible and generalizable LLM-based architecture search. The dataset and model will be made public at https://github.com/Ashone3/LM-Searcher .
We present a novel LLM-based NAS framework that enables neural network architecture optimization across diverse domains without domain-specific adaptation.
◦
Cross-domain architecture encoding and exploration possible using NCode, a universal numeric string representation.
◦
Efficient architecture exploration through pruning-based subspace sampling strategies.
◦
Achieve competitive performance in both in-domain and out-of-domain operations.
◦
Code and data made public for reproducibility
•
Limitations:
◦
Further analysis is needed on the generalization performance and limitations of the proposed NCode representation.
◦
Further research is needed to further increase its versatility for various tasks.
◦
There is room for optimization and improvement of pruning-based subspace sampling strategies.
◦
Due to the high dependence on the performance of LLM, there is a possibility that the limitations of LLM may affect the performance of LM-Searcher.
