Summary of Aggregation Artifacts in Subjective Tasks Collapse Large Language Models’ Posteriors, by Georgios Chochlakis et al.
Aggregation Artifacts in Subjective Tasks Collapse Large Language Models’ Posteriors
by Georgios Chochlakis, Alexandros Potamianos, Kristina Lerman, Shrikanth Narayanan
First submitted to arxiv on: 17 Oct 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 The paper explores the limitations of In-Context Learning (ICL) in Large Language Models (LLMs), specifically in complex subjective domains like emotion and morality. Researchers find that ICL predominantly relies on retrieving task priors rather than learning to perform tasks, which is attributed to aggregation in corresponding datasets creating annotation artifacts. The study evaluates posterior bias towards certain annotators using quantitative measures of LLM priors and finds that aggregation is a confounding factor. While addressing this issue improves results, it does not fully explain the gap between ICL and the state-of-the-art. The paper also highlights the potential for minority annotators to better align with LLMs and amplify their perspectives. |
| Low | GrooveSquid.com (original content) | Low Difficulty Summary This research investigates how Large Language Models (LLMs) learn from tasks. It seems that these models are mostly relying on what they learned before, rather than learning new things. This is especially true when it comes to complex topics like emotions and morality. The study thinks this might be because the way we combine different opinions (aggregation) creates problems in the data. By looking at individual annotators’ labels, researchers find that some people’s perspectives are actually better represented by LLMs than others. |
