MOOSE-Star: Efficient LLM Training for Science скачать в хорошем качестве

MOOSE-Star: Efficient LLM Training for Science

In this AI Research Roundup episode, Alex discusses the paper: 'MOOSE-Star: Unlocking Tractable Training for Scientific Discovery by Breaking the Complexity Barrier' MOOSE-Star is a new framework designed to solve the mathematical intractability of training LLMs for hypothesis generation. It transforms the discovery process into a Hierarchical Markov Decision Process, reducing search complexity from exponential to logarithmic levels. The approach decouples the task into inspiration retrieval and hypothesis composition to significantly lower sample complexity. By using recursive clustering to organize knowledge, the model can navigate massive databases of millions of papers efficiently. This enables end-to-end training for scientific reasoning that was previously considered mathematically ill-posed. Paper URL: https://arxiv.org/abs/2603.03756 #AI #MachineLearning #DeepLearning #LLMs #ScientificDiscovery #HierarchicalSearch #ReasoningModels Resources: GitHub: https://github.com/ZonglinY/MOOSE-Star Hugging Face model: https://huggingface.co/ZonglinY/MOOSE... Hugging Face model 2: https://huggingface.co/ZonglinY/MOOSE...