Title:	Understanding Chain-of-Thought reasoning in Open Source Large Language Models
DNr:	Berzelius-2024-38
Project Type:	LiU Berzelius
Principal Investigator:	Amir Hossein Akhavan Rahnama <amiakh@kth.se>
Affiliation:	Kungliga Tekniska högskolan
Duration:	2024-01-24 – 2024-08-01
Classification:	10201
Keywords:

Abstract

Chain-of-thought (CoT) prompting can significantly improve the reasoning abilities of large language models (LLMs) for more complex prompts. In Chain-of-Thought reasoning, LLMs generate intermediate responses that help to improve the accuracy of the responses. Even though these types of prompts are becoming more popular, our understanding of them is limited. In our project, we aim to analyze the performance of LLMs for COT commands using a combination of hidden states and attention weights for open-source LLM models such as Llama-2.