LLM AutoML searcher jointly tunes architecture, hyperparameters, and data
Original title: Large Language Models as a Universal Searcher for Open-World Automated Machine Learning
Trained on one vision task, the searcher reports a 0.91 GenEval score after 896 GPU-days on a billion-parameter Diffusion Transformer.
The paper proposes an LLM-based AutoML framework that jointly optimizes model architecture, training hyperparameters, and data ratio without task-specific fine-tuning. NCode encodes design choices as discrete sequences so search becomes a ranking task. Trained only on trajectories from a single-domain vision task, the searcher is reported to generalize to image classification, segmentation, speaker verification, and text-to-image benchmarks. Reinforcement learning is used to add chain-of-thought reasoning so the search strategy can be interpreted.
Unconfirmed: Cross-task and state-of-the-art claims are the authors' reported results, not independently verified here.
- Joint optimization of a billion-parameter Diffusion Transformer used 896 GPU-days.
- The authors report a GenEval score of 0.91.
- The searcher was trained only on a limited single-domain vision task.
- Reported strategy shifts from diversity-driven exploration to late-stage anchor exploitation.
From the original article
1 min read at arxiv.orgDeveloping effective machine learning models requires complex manual tuning, while current Automated Machine Learning (AutoML) frameworks remain domain-specific and isolate model architecture, training hyperparameter, and data ratio optimization.
By Yuxuan Hu · arxiv.org
A short excerpt; the article and its images belong to the publisher. Summary, key facts and assessment above are AI-written from the article and can be wrong.