报告主题：Improving LLM Reasoning with Argumentation

报告人：Ramon Ruiz-Dolz

报告时间：2025年11月14日 上午 10:00

报告地点：北京大学王选计算机研究所106报告厅

Abstract: Since the release of the latest versions of popular LLMs such as o3, Gemini 2.0 Flash Thinking, or DeepSeekV3-R1, the race for Artificial General Intelligence (AGI) has put its focus on the “reasoning” capabilities of these models. Advances in the “reasoning” capabilities of LLMs are measured and quantified with the use of benchmarks including a wide set of tasks and challenges that address reasoning from different perspectives (e.g., commonsense, mathematical, or multi-hop reasoning). However, none of these tasks involve argumentation, the natural form of reasoning in natural language.

In this presentation, Ramon will briefly cover the nature of some of the most popular natural language reasoning benchmarks, highlighting their limitations and the need of developing more complex tasks and challenges that capture the nuances of human natural language reasoning. Then, he will present some os his recently published work in which these limitations are partially addressed. Finally, the presentation will conclude with current trends and promising intersections between the areas of computational argumentation and natural language processing aimed at improving the reasoning capabilities of LLMs.

Bio: Ramon Ruiz-Dolz (València, 1996) is a Lecturer in Computing at the University of Dundee (United Kingdom). Before that, he has been a visiting lecturer at the Dundee International Institute of Central South University (Changsha, China), a postdoctoral researcher at ARG-tech, and a visiting researcher at ELLIS and at the National Institute of Informatics (Tokyo, Japan) and the Universitat Politècnica de València (Spain). Ramon obtained his PhD in Computer Science from the Universitat Politècnica de València with a Cum Laude distinction and awarded with the Extraordinary Prize.

Ramon’s main research topics are Computational Argumentation and Natural Language Processing. His current research focuses on the analysis of the natural language reasoning capabilities of LLMs from an argumentative perspective, and the integration of concepts from argumentation theory into NLP algorithms to improve reasoning and create tools for developing critical thinking skills.