中央研究院 資訊科學研究所

Even large language models (LLM) nowadays are able to answer a wide variety of questions, and Wei et al. (2022) also proposed a novel chain-of-thought method to enhance the ability of interpretability for LLM by multi-step reasoning, however, these systems are not good at answering complex, open-ended questions that are ambiguous, and require a broad, multifaceted perspective to give complete answers. In our recent work we explore a new way to enhance LLM’s ability to “multifaceted thinking” by taking inspiration from how humans tackle complex problems. We proposed a thinking framework that has the model ask a few critical questions and figure out the answer before deriving a conclusion. We can further extract the whole “thought process” to form a hierarchical tree called tree-of-thought. Tree-of-thought not only serves as a transparent view into the LLM’s reasoning process but also provides a way for humans to take part in the thought process. This leads to an opportunity to combine artificial intelligence (AI) with human intelligence (HI) to foster human-AI collaboration that can tackle the nature of multifaceted thought of complex open-ended questions. This excellent ongoing research project in academia is sponsored by Google and is one of the very few R&D cooperation project in Asia.

Wen-Hsiang Lu (盧文祥) is a professor in the department of Computer Science and Information Engineering, National Cheng Kung University. He received his M.S. and Ph.D. degrees from the department of Computer Science and Information Engineering, National Chiao Tung University, in 1990 and 2003. In 2004, he received Dragon Golden Paper Award sponsored by the Acer Foundation. His research interests include natural language processing, speech processing, machine translation, information retrieval, web/text mining, medical informatics. His research results have been published in top-tier journals and conferences, such as ACM TOIS, ACM TALIP, ACM SIGIR, ACL, etc.