School of Engineering Department of Computer Science and Engineering 139 Mental State Reasoning for Large Language Models Supervisor: SONG Yangqiu / CSE Student: HE Jiaxuan / DSCT Course: UROP 1100, Summer Right now, AI tools for online shopping platforms like Amazon mainly look at what you’ve browsed or bought before to guess what you might like and suggest products. While this works fairly well for people who shop there often, it struggles to bring in new customers or introduce new items — a common problem known as out-of-domain generalization. To overcome this, we use eCeLLM, a series of AI models customized for ecommerce that are built by fine-tuning general-purpose language models through instruction training. These models get better at understanding search commands because they analyze them through a framework called Belief-Desire-Intention (BDI). In this article, we focus on how the BDI process helps create a solid database that guides eCeLLM’s decision-making. Mental State Reasoning for Large Language Models Supervisor: SONG Yangqiu / CSE Student: JIA Yusheng / COMP Course: UROP 1100, Summer This summer vacation, I studied the relevant content of large language models. By reading literature, I have learned the most fundamental ToM (Theory of Mind) theory, as well as how large language models reason based on words and sentences and finally form their own language logic. At the same time, I have gained a relatively clear understanding of ToM’s ability deficiencies. Through practice, I have learned how to initiate the training of large language models, detect the training process, adjust parameters, and perform other detailed operations, thereby achieving better inference results. In this report, I will summarise the knowledge I gained during my vacation, as well as some practical experiences. Mental State Reasoning for Large Language Models Supervisor: SONG Yangqiu / CSE Student: LIANG Hao / COMP Course: UROP 1100, Fall UROP 2100, Spring UROP 3100, Summer This report summarizes my experience and work in studying the Theory of Mind (ToM) in large language models. In Part 1, I discussed nine recent papers that I read this semester, which focus on ToM performance and evaluation in LLMs, as well as CoT in ToM. I identify common themes such as LLMs’ current limitations on ToM tasks and emerging strategies to improve performance. In Part 2, I show my annotation work on a Product Relation Prediction dataset from my PhD mentor, Chunkit Chan. For each product, I selected relevant Beliefs, Desires, and Intentions from given choices and wrote Intention_choice_rationale and BDI_reasoning_process. I list some sample annotation examples to help me illustrate how I reasoned about customer BDI states. I also discuss challenges in mapping product information to mental states and reflect on what I learned. Overall, this project enhanced my research skills in reading papers, reasoning about ToM, and performing data annotation.
RkJQdWJsaXNoZXIy NDk5Njg=