Exploring the Analogical Reasoning Ability of Large Language Models: With a Focus on the Copycat Problem

Haebom

Nov 3, 20232y ago

Dalle3 - 유추문제를 푸는 인공지능

The development of artificial intelligence and machine learning is advancing at a remarkable pace, and among them, Large Language Models (LLMs) are getting particular attention. In this post, I focus on the analogical reasoning capabilities of LLMs, especially their ability to tackle problems like 'Copycat.' 'Copycat' is a simple task devised by Douglas Hofstadter in the 1980s to capture key aspects of human analogical reasoning.

Understanding the Copycat problem

A Copycat problem typically looks like this: "If the string 'abc' changes to 'abd', how should the string 'pqr' change?" These kinds of problems test human reasoning while also offering insights into how machines could approach analogy tasks.

https://melaniemitchell.me/ExplorationsContent/analogy-problems.html

LLMs and Copycat

LLMs can solve these analogy problems to some degree. In particular, Melanie Mitchell's research has shown that GPT-3 can do quite well with Copycat tasks. Still, in this post, I explore how the Llama-7B-chat model tackles Copycat. While Llama-7B-chat may not be as strong as GPT-3.5, it's small and convenient enough for experimentation.

Experiments with Llama-7B-chat

I tried solving a variety of Copycat problems using the Llama-7B-chat model. For example, I presented problems like these:

The answer to "0 1 2 to 2 1 0, 1 2 3 to 3 2 1, 4 5 6 to " is "6 5 4".

The answer to "0 1 2 to 0 1 3, 1 2 3 to 1 2 4, 4 5 6 to " is "4 5 7".

The answer to "0 1 to 0 1 1, 1 2 to 1 2 2, 4 5 to " is "4 5 5".

These problems gave me a clearer picture of how Llama-7B-chat approaches analogy tasks.

Analyzing the mechanism of analogical reasoning

To analyze the internal workings of the Llama-7B-chat model, I used tools like 'logit lens' and 'zero ablation.' This gave me insights into how the model arrives at specific answers and what components are most important.

In this post, I explored the analogical abilities of LLMs and examined how the Llama-7B-chat model tackles Copycat problems. This analysis helped deepen my understanding of how LLMs work and reason by analogy. Going forward, I plan to use a wider range of models and problems to explore LLM analogy skills even further. Personally, I'd love to try it out on 70B as well... but sadly, I haven’t been able to get a 70B grant yet. ㅠㅜ

Hofstadter - Analogy as the Core of Cognition.pdf344.21KB

Circuit Breaking: Removing Model Behaviors with Targeted Ablation

Language models often exhibit behaviors that improve performance on a pre-training objective but harm performance on downstream tasks. We propose a novel approach to removing undesirable behaviors...

arxiv.org

The ConceptARC Benchmark: Evaluating Understanding and...

The abilities to form and abstract concepts is key to human intelligence, but such abilities remain lacking in state-of-the-art AI systems. There has been substantial research on conceptual...

arxiv.org

grgv.xyz

Subscribe to 'haebom'

📚 Welcome to Haebom's archives.
---
I post articles related to IT 💻, economy 💰, and humanities 🎭.
If you are curious about my thoughts, perspectives or interests, please subscribe.
haebom@kakao.com