Summary of Attention Speaks Volumes: Localizing and Mitigating Bias in Language Models, by Rishabh Adiga et al.
Attention Speaks Volumes: Localizing and Mitigating Bias in Language Models
by Rishabh Adiga, Besmira Nushi, Varun Chandrasekaran
First submitted to arxiv on: 29 Oct 2024
Categories
- Main: Computation and Language (cs.CL)
- Secondary: Artificial Intelligence (cs.AI); Machine Learning (cs.LG)
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 how bias emerges in large language models (LLMs) when given ambiguous comparative prompts. It proposes a new metric to quantify LLMs’ preferences and introduces ATLAS, a technique to localize and reduce bias by analyzing attention scores. The method is tested on four datasets using different LLM architectures, demonstrating that bias is concentrated in later layers and can be effectively mitigated without compromising performance. |
| Low | GrooveSquid.com (original content) | Low Difficulty Summary The paper investigates how large language models (LLMs) become biased when given ambiguous prompts. It develops a new way to measure the model’s preferences and creates ATLAS, a method to understand where the bias comes from and fix it. The approach is tested on several datasets using different types of LLMs, showing that bias tends to happen in later parts of the model and can be reduced without hurting performance. |
Keywords
* Artificial intelligence * Attention
