Summary of Number Cookbook: Number Understanding Of Language Models and How to Improve It, by Haotong Yang et al.
Number Cookbook: Number Understanding of Language Models and How to Improve It
by Haotong Yang, Yi Hu, Shijia Kang, Zhouchen Lin, Muhan Zhang
First submitted to arxiv on: 6 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 This paper investigates the numerical understanding and processing ability (NUPA) of large language models (LLMs). Current LLMs are capable of solving complex reasoning tasks but struggle with basic numerical understanding, such as determining whether 9.11 is greater than 9.9. To address this gap, the authors introduce a comprehensive benchmark covering four numerical representations and 17 distinct numerical tasks in four major categories. The benchmark is derived from primary and secondary education curricula and encompasses everyday numerical understanding scenarios. Through evaluation, the authors find that current LLMs frequently fail on many of these tasks. They then explore techniques to enhance NUPA, including tokenizers, PEs, and number formats, and evaluate their effectiveness using the proposed testbed. The results show that naive finetuning can improve NUPA for many but not all tasks, and surprisingly, techniques designed to enhance NUPA are ineffective for finetuning pretrained models. The authors also explore the impact of chain-of-thought techniques on NUPA. |
| Low | GrooveSquid.com (original content) | Low Difficulty Summary This paper looks at how well large language models do with numbers. These models can solve lots of complex problems but often get basic math wrong, like saying 9.11 is greater than 9.9. To help them improve, the authors created a test to see what kind of math they’re good at and what they struggle with. They found that most large language models aren’t very good at simple math problems. Then, they tried different ways to make the models better at numbers and saw if it worked. |
