Summary of The Geometry Of Categorical and Hierarchical Concepts in Large Language Models, by Kiho Park et al.
The Geometry of Categorical and Hierarchical Concepts in Large Language Models
by Kiho Park, Yo Joong Choe, Yibo Jiang, Victor Veitch
First submitted to arxiv on: 3 Jun 2024
Categories
- Main: Computation and Language (cs.CL)
- Secondary: Artificial Intelligence (cs.AI); Machine Learning (cs.LG); Machine Learning (stat.ML)
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 extension of the linear representation hypothesis is proposed to formalize the encoding of features, such as categorical concepts like “is_animal”, as vectors in the representation space of large language models (LLMs). Building upon previous work on binary concepts with natural contrasts, this study demonstrates how to represent polytopes in the representation space for categorical concepts. Theoretical results are validated using Gemma and LLaMA-3 LLMs, estimating representations for over 900 hierarchically related concepts from WordNet. |
| Low | GrooveSquid.com (original content) | Low Difficulty Summary Large language models (LLMs) have a secret: they use linear directions to represent semantic concepts. This idea is called the “linear representation hypothesis”. Researchers have already figured out how to make this work for simple ideas like “male” and “female”, but what about more complex concepts? In this study, scientists show how to extend this idea to represent features like “is_animal” as special kinds of vectors in the LLM’s space. This lets them turn categorical concepts into polytopes (like geometric shapes) in that space. They also prove a connection between how these concepts are organized and their geometry. To test this, they used two big language models to represent over 900 related ideas from WordNet. |
Keywords
» Artificial intelligence » Llama
