Summary of Improving Mathematical Reasoning Capabilities Of Small Language Models Via Feedback-driven Distillation, by Xunyu Zhu et al.
Improving Mathematical Reasoning Capabilities of Small Language Models via Feedback-Driven Distillation
by Xunyu Zhu, Jian Li, Can Ma, Weiping Wang
First submitted to arxiv on: 22 Nov 2024
Categories
- Main: Computation and Language (cs.CL)
- Secondary: Artificial Intelligence (cs.AI)
GrooveSquid.com Paper Summaries
GrooveSquid.com’s goal is to make artificial intelligence research accessible by summarizing AI papers in simpler terms. Each summary below covers the same AI paper, written at different levels of difficulty. The medium difficulty and low difficulty versions are original summaries written by GrooveSquid.com, while the high difficulty version is the paper’s original abstract. Feel free to learn from the version that suits you best!
| Summary difficulty | Written by | Summary |
|---|---|---|
| High | Paper authors | High Difficulty Summary Read the original abstract here |
| Medium | GrooveSquid.com (original content) | Medium Difficulty Summary Large Language Models (LLMs) have shown remarkable reasoning capabilities, achieving state-of-the-art performance in various tasks. However, their significant computational and memory demands, resulting from billions of parameters, impede deployment on resource-constrained devices. Knowledge distillation is a promising solution, where LLMs transfer reasoning capabilities to Small Language Models (SLMs) with fewer parameters, enabling broader deployment. Our proposed Feedback-Driven Distillation (FDD) framework enhances SLMs’ mathematical reasoning capabilities by constructing a distillation dataset through prompting LLMs to pair mathematical problems with corresponding rationales. We categorize problems into easy and hard categories based on SLM performance, generating more complex variations for easy problems and synthesizing new questions of similar complexity for hard problems. A multi-round distillation paradigm is proposed to iteratively enrich the distillation datasets, improving SLMs’ mathematical reasoning abilities. Experimental results demonstrate that our method achieves state-of-the-art (SOTA) mathematical reasoning performance. |
| Low | GrooveSquid.com (original content) | Low Difficulty Summary This paper talks about how big language models can help smaller ones learn better. The big models are really good at solving math problems, but they need a lot of computer power and memory to work. This makes it hard for them to be used on devices that don’t have as much power. To fix this, the researchers came up with a way to make the small models smarter by letting the big models teach them. They did this by making the big models create lots of math problems with explanations, and then used those problems to train the small models. The results show that this method makes the small models much better at solving math problems. |
Keywords
» Artificial intelligence » Distillation » Knowledge distillation » Prompting
