Summary of Do Large Language Models Perform the Way People Expect? Measuring the Human Generalization Function, by Keyon Vafa et al.

Do Large Language Models Perform the Way People Expect? Measuring the Human Generalization Function

by Keyon Vafa, Ashesh Rambachan, Sendhil Mullainathan

First submitted to arxiv on: 3 Jun 2024

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 A novel approach is proposed to evaluate large language models (LLMs), which have diverse uses and are challenging to assess. The paper introduces a framework that considers how humans make decisions about deploying LLMs based on their beliefs about where they will perform well. A dataset of 19,000 examples is collected from the MMLU and BIG-Bench benchmarks, showing that human generalization can be predicted using NLP methods. The study evaluates the alignment between LLMs and the human generalization function, revealing that more capable models (e.g., GPT-4) can perform worse in situations where the cost of mistakes is high due to misalignment.
Low	GrooveSquid.com (original content)	Low Difficulty Summary Large language models are really smart, but it’s hard to know how good they are. This paper tries to figure out why this is the case. It looks at how people decide when and where to use these models, based on what they think the model can do well. The study collected a big dataset of examples and found that people have consistent ways of thinking about which tasks an LLM will be good at. The results show that more powerful models can actually do worse in certain situations because they’re not aligned with how humans think.