Large Language Models as a Universal Searcher for Open-World Automated Machine Learning
Trained on search trajectories from one vision task, the LLM searcher transferred to other tasks and modalities without task-specific fine-tuning, the authors report.
Their NCode encoding turns model architectures, training hyperparameters and data ratios into discrete sequences that the searcher ranks jointly. The paper reports state-of-the-art results across image classification, image segmentation, speaker verification and text-to-image benchmarks. A billion-parameter Diffusion Transformer experiment required 896 GPU-days and achieved a GenEval score of 0.91. Analysis of reinforcement-learning-trained reasoning traces showed a shift from exploring diverse candidates early to exploiting promising anchors later.
ArXiv · AI/CL/LG's note
Their NCode encoding turns model architectures, training hyperparameters and data ratios into discrete sequences that the searcher ranks jointly. The paper reports state-of-the-art results across image classification, image segmentation, speaker verification and text-to-image benchmarks. A billion-parameter Diffusion Transformer experiment required 896 GPU-days and achieved a GenEval score of 0.91. Analysis of reinforcement-learning-trained reasoning traces showed a shift from exploring diverse candidates early to exploiting promising anchors later.
ArXiv · AI/CL/LG's note
score 5