School of Engineering Department of Computer Science and Engineering 102 Generative AI Supervisor: CHEN Qifeng / CSE Student: HUNG Ming Kin / DSCT RUAN Junhong / COSC TANG Kewei / COMP Course: UROP 1100, Spring UROP 1100, Spring UROP 1100, Spring In this UROP project, our group analyse on different generative models, including Pixel2Pixel, Generative Adversarial Networks (GANs),variational autoencoder (VAE) and recent diffusion models. In this report, our group is going to firstly introduce the principles of these generative models. Then, our group is going to examine and evaluate the results when we tried to implement these models. Generative AI Supervisor: CHEN Qifeng / CSE Student: KONG Lingcheng / COSC Course: UROP 1000, Summer This report focus on my UROP1000 progress during this summer 2024. The first part will be the knowledge I learnt as the foundation of further research. This report will also include what I learnt about how to do research by entering the group directed by professor Qifeng Chen and Chenyang Lei. In the last part I will present some tasks I have done. Generative AI Supervisor: CHEN Qifeng / CSE Student: LEUNG Hok To / DSCT Course: UROP 1100, Fall This project studies autonomous large language models (LLMs) and their applications in reasoning tasks. It begins with a survey on transformers, autonomous LLM, and its limitations. Then, it proceeds to explore the different prompt engineering techniques to enhance the reasoning capability of LLM via in-context learning. Finally, an experiment is conducted to inquire into the performance of those techniques in arithmetical operations evaluation tasks and logical thinking, where the final outputs of the models are accessed and intermediate steps and the model’s explainability of its answer are analyzed. The obtained results demonstrate the effectiveness of a blended framework of multiple agent debate (MAD) and Chain of Thought (CoT) in solving reasoning tasks, in comparison to other configurations. The report underscores the potential of incorporating MAD as a context learning technique aside from the prevalent CoT and Selfrefinement (SR) approach, contributing towards a better prompt engineering design paradigm.
RkJQdWJsaXNoZXIy NDk5Njg=