Reproduce “AS-ES Learning: Towards Efficient CoT Learning in Small Models” on OpenResearch
Turn your favorite coding agents into research agents that review literature, develop hypotheses, and run experiments.
Universal app for Apple silicon and Intel. Requires macOS 11 or later.
About this paper
Nuwa Xi, Yuhan Chen, Sendong Zhao, Haochun Wang, Bing Qin, Ting Liu
AS-ES Learning introduces a data-efficient paradigm for enabling Chain-of-Thought (CoT) reasoning in small models by deconstructing CoT into Abstractive and Extractive Segments. This method improves performance on tasks like Math Word Problems and PET summarization without requiring additional data augmentation, with a single small model effectively handling both segment types.
Key insights
- A single small model, when trained with the AS-ES paradigm, can effectively perform both extractive and abstractive reasoning steps, challenging the assumption that small models require specialized modules for different CoT phases.
- The optimal segmentation strategy is highly dependent on the task's inherent CoT structure and model size; for instance, interleaving worked well for sequential Math Word Problems, while entropy-oriented segmentation was better for parallel PET summarization.
- AS-ES learning works by achieving a significantly lower loss boundary during training compared to direct seq2seq approaches, indicating that breaking CoT into components allows the model to learn each step more effectively and reduce overall error.
Summary by alphaXiv.
Tip: change arxiv.org to autoarxiv.org in any arXiv abstract link to open the paper here.