ADAPT Framework: Online Reweighting Improves LLM Generalization
Researchers have introduced ADAPT, a novel online data curation framework for training large language models (LLMs), challenging traditional offline methods. Published on arXiv, the study argues that existing offline data selection and mixing techniques are brittle, engineering-heavy, and often harm generalization by reducing data diversity through hard filtering. In contrast, ADAPT treats data curation as an online reweighting problem, dynamically adjusting sample importance via loss weighting during training without altering the total number of samples. The framework utilizes adaptive per-sample learning rates guided by similarity-based quality signals, effectively acting as an implicit curriculum learner. This approach allows the model to progressively shift focus from coarse-grained patterns to fine-grained semantic distinctions as training evolves. Experimental results across instruction tuning and large-scale pretraining demonstrate that ADAPT consistently outperforms both offline selection methods and prior online techniques. Notably, it achieves stronger cross-benchmark generalization under equal computational costs (FLOPs), offering a more efficient and robust solution for optimizing LLM performance through dynamic data management.
Wire timeline
ADAPT Framework: Online Reweighting Improves LLM Generalization
Researchers have introduced ADAPT, a novel online data curation framework for training large language models (LLMs), challenging traditional offline methods. Published on arXiv, the study argues that existing offline data selection and mixing techniques are brittle, engineering-heavy, and often harm generalization by reducing data diversity through hard filtering. In contrast, ADAPT treats data curation as an online reweighting problem, dynamically adjusting sample importance via loss weighting during training without altering the total number of samples. The framework utilizes adaptive per-sample learning rates guided by similarity-based quality signals, effectively acting as an implicit curriculum learner. This approach allows the model to progressively shift focus from coarse-grained patterns to fine-grained semantic distinctions as training evolves. Experimental results across instruction tuning and large-scale pretraining demonstrate that ADAPT consistently outperforms both offline selection methods and prior online techniques. Notably, it achieves stronger cross-benchmark generalization under equal computational costs (FLOPs), offering a more efficient and robust solution for optimizing LLM performance through dynamic data management.
cs.AI updates on arXiv.org