Exploring the inference ability of Large Language Models: Focusing on the Copycat problem
Haebom
Dalle3 - Artificial intelligence that solves analogy problems
The development of artificial intelligence and machine learning is progressing at a dazzling pace, and among them, Large Language Models (LLM) are particularly noteworthy. In this post, we will focus on the inference ability of LLM, and in particular, the ability to solve the problem called 'Copycat'. 'Copycat' is a simple problem conceived by Douglas Hofstadter in the 1980s, and it was intended to capture the core aspects of human inference reasoning ability.
Understanding the Copycat Problem
Copycat problems take the form: "If the string 'abc' is changed to 'abd', what happens to the string 'pqr'?" These problems test the reasoning ability of humans, while also providing insight into how machines can solve such analogy problems.
LLM has the ability to solve these analogy problems to some extent. In particular, GPT-3 was found to be able to solve the 'Copycat' problem quite well according to Melanie Mitchell's research. However, in this post, we will look at the process of solving the Copycat problem using a model called Llama-7B-chat. Llama-7B-chat is less powerful than GPT-3.5, but it is a small and convenient model for experimentation.
Experiments using Llama-7B-chat
We have tried to solve various Copycat problems using the Llama-7B-chat model. For example, we have presented the following problems:
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 helped us gain a better understanding of how the Llama-7B-chat model solves inference problems.
Mechanism analysis of analogical ability
To analyze the internal mechanism of the Llama-7B-chat model, we used the 'logit lens' and 'zero ablation' techniques. This allowed us to gain insight into how the model arrives at specific answers and which parts play important roles.
In this post, we will look at the inference ability of LLM, and in particular, analyze the process of solving the Copycat problem using the Llama-7B-chat model. Through this analysis, we were able to increase our understanding of how LLM works and its inference ability. In the future, we will use more diverse models and problems to explore LLM's inference ability in more depth. Personally, I want to try it in 70B too... but I haven't received a 70b grant. ㅠㅜ
Hofstadter - Analogy as the Core of Cognition.pdf352KB
📚 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. Would you like to be notified when new articles are posted? 🔔 Yes, that means subscribe. haebom@kakao.com
