Maarten Sap - Home
Excerpt
Overarching Research Themes
Overarching Research Themes
Extracted by GPT-4, there may be inconsistencies.
Ethics in AI Development
My research group explores the complex dimensions of ethical AI and its implications for society. One of our pivotal works, HAICOSYSTEM: An Ecosystem for Sandboxing Safety Risks in Human-AI Interactions, introduces a framework designed to simulate potential safety risks associated with human-AI interactions. We also investigate the balance between usability and ethical considerations through our paper, AI-LieDar: Examine the Trade-off Between Utility and Truthfulness in LLM Agents, which highlights the dilemmas faced by LLMs in delivering truthful responses. Our efforts culminate in Particip-AI: A Democratic Surveying Framework that aims to engage a broader audience in discussions of future AI applications and their impacts.
Narrative Dynamics in AI
My research group explores the way narratives are constructed and perceived through AI technologies. We delve into emotional resonance in storytelling with our important study, HEART-felt Narratives: Tracing Empathy and Narrative Style in Personal Stories with LLMs, which examines how narrative styles can evoke empathy. Additionally, our work, Modeling Empathic Similarity in Personal Narratives, quantifies connections between audiences and narratives, enriching our understanding of storytelling mechanics. Another significant contribution, Quantifying the narrative flow of imagined versus autobiographical stories, provides insights into how different types of personal stories affect listener engagement.
Social Intelligence and Simulation
My research group explores how social intelligence can be embedded and evaluated in language models. We critically assess the limitations of simulating genuine interpersonal engagements in our study, Is This the Real Life? Is This Just Fantasy? The Misleading Success of Simulating Social Interactions With LLMs, which reveals the gaps between simulated responses and authentic human communication. To foster more effective models, we research the efficacy of context in understanding social dynamics through Clever Hans or Neural Theory of Mind? Stress Testing Social Reasoning in Large Language Models. Our ongoing efforts also led to SOTOPIA: Interactive Evaluation for Social Intelligence in Language Agents, which advances the measures of social reasoning capabilities among AI agents.
Addressing Toxic Language
My research group explores the pressing challenges in mitigating toxic language generation within AI systems. We highlight multilingual aspects of this issue with our notable study, PolygloToxicityPrompts: Multilingual Evaluation of Neural Toxic Degeneration in Large Language Models, which scrutinizes how toxic content manifests across languages. Another critical observation is presented in Counterspeakersâ Perspectives: Unveiling Barriers and AI Needs in the Fight against Online Hate, offering insights into the requirements for combating hate speech effectively. Additionally, our work, Leftover-Lunch: Advantage-based Offline Reinforcement Learning for Language Models, investigates reinforcement learning strategies to enhance ethical dialogue generation.