Summary of What Do Learning Dynamics Reveal About Generalization in Llm Reasoning?, by Katie Kang et al.
What Do Learning Dynamics Reveal About Generalization in LLM Reasoning?
by Katie Kang, Amrith Setlur, Dibya Ghosh, Jacob Steinhardt, Claire Tomlin, Sergey Levine, Aviral Kumar
First submitted to arxiv on: 12 Nov 2024
Categories
- Main: Machine Learning (cs.LG)
- Secondary: None
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 The paper investigates the mechanisms behind the problem-solving abilities of large language models (LLMs) during finetuning. It aims to understand how learning dynamics shape downstream generalization on reasoning tasks. By analyzing memorization and performance, researchers find that a training metric called pre-memorization train accuracy can effectively characterize a model’s generalization behavior. This metric is able to predict test accuracy with high reliability across various models, datasets, and training configurations. On a per-example level, it also indicates whether individual predictions are robust to perturbations. By connecting learning behavior to generalization, this metric guides targeted improvements to training strategies. |
| Low | GrooveSquid.com (original content) | Low Difficulty Summary Large language models (LLMs) can do amazing things, like understand and answer questions. But we don’t really know how they learn to do these things. This paper tries to figure out what makes LLMs good at solving problems. They look at how the models learn on special kinds of tasks that require thinking and reasoning. The researchers find a way to measure how well the model is doing before it starts copying exactly from its training data. This measurement can predict how well the model will do later, and even tells us if individual answers are correct or not. By understanding how LLMs learn, we can make them better at solving problems without needing as much data. |
Keywords
» Artificial intelligence » Generalization
