Summary of Active-dormant Attention Heads: Mechanistically Demystifying Extreme-token Phenomena in Llms, by Tianyu Guo et al.
Active-Dormant Attention Heads: Mechanistically Demystifying Extreme-Token Phenomena in LLMs
by Tianyu Guo, Druv Pai, Yu Bai, Jiantao Jiao, Michael I. Jordan, Song Mei
First submitted to arxiv on: 17 Oct 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 This research paper investigates the perplexing behavior of large language models (LLMs) based on transformers. The authors identify three distinct phenomena: attention sinks, value-state drains, and residual-state peaks, collectively referred to as extreme-token phenomena. These occurrences are marked by specific tokens receiving an abnormally high level of attention, displaying significantly lower value states, and exhibiting larger residual-state norms compared to other tokens. The paper explores the implications of these extreme tokens on LLM inference, quantization, and interpretability. |
| Low | GrooveSquid.com (original content) | Low Difficulty Summary This study looks at a strange issue with large language models. Some special words get way more attention than others, which makes it hard for computers to understand what the model is really saying. This problem affects how well the model works and makes it tricky to figure out what’s going on inside the model. The researchers are trying to understand why this happens and how we can fix it. |
Keywords
» Artificial intelligence » Attention » Inference » Quantization » Token
