Llama2 self-improvement using memory-of-thought

• Main Author: Dong, Yuxiu
• Other Authors: Lihui Chen
• Format: Thesis-Master by Coursework
• Language: English
• Published: Nanyang Technological University 2024
• Subjects: Engineering; Memory-of-thought; Llama2; Large language model; Natural language processing
• Online Access: https://hdl.handle.net/10356/179097
• Institution: Nanyang Technological University

Memory-of-Thought (MoT) is a newly proposed mechanism for LLMs to self-improve without annotated datasets and expensive resource consumption for fine-tuning models. This project evaluates the effectiveness of MoT on a pre-trained Large Language Model Llama2 and compares the performance of improved Llama2 with ChatGPT3.5 API on 10 benchmark data sets. The experiments demonstrate that MoT is able to improve the comprehensive reasoning capabilities of Llama2 on those downstream applications. Based on the experimental study, we also discover that MoT might yield a more substantial improvement in scenarios where the CoT process is employed. Meanwhile, further analysis indicates a greater improvement percentage of MoT on Llama2 than that on ChatGPT. Additionally, the experiments on long conversations prove that MoT can improve the performance of Llama2 in long open-domain conversations, resulting in better consistency, engagingness, and response selection.