Extensive experiments demonstrate that UP enhances exploration capacity and achieves superior reasoning accuracy across diverse RL algorithms (DAPO, GSPO, and GRPO), model architectures, and training modalities, validating UP as a truly universal plug-and-play enhancement for RL-based training.
A simple, budget-preserving pipeline is proposed that realigns the expert pool to the target domain before fine-tuning, and turns this redundant capacity into useful domain capacity and achieves lower training loss, with gains spanning all difficulty levels in downstream evaluation.
DataOrchestra, a framework that unifies different processing operations and orchestrates an example-specific pipeline for each example, is proposed and effective for math continued pretraining and outperforms stronger processing baselines, while reducing processing compute by skipping unnecessary downstream operations.
Zhen Huang, Yikun Wang, Shijie Xia et al.· 0 citations
Through a systematic study of four flagship models, a previously overlooked phenomenon is identified: under extensive context, models give up or provide uncertain incorrect answers long before exhausting the context window.
