Ralf Widenhorn

Justin C. Dunlap, Anne-Simone Parent, Ralf Widenhorn

We present a systematic framework of indices designed to characterize Large Language Model (LLM) responses when challenged with rebuttals during a chat. Assessing how LLMs respond to user dissent is crucial for understanding their reliability and behavior patterns, yet the complexity of human-LLM interactions makes systematic evaluation challenging. Our approach employs a fictitious-response rebuttal method that quantifies LLM behavior when presented with multiple-choice questions followed by deliberate challenges to their fictitious previous response. The indices are specifically designed to detect and measure what could be characterized as sycophantic behavior (excessive agreement with user challenges) or stubborn responses (rigid adherence to the fictitious response in the chat history) from LLMs. These metrics allow investigation of the relationships between sycophancy, stubbornness, and the model's actual mastery of the subject matter. We demonstrate the utility of these indices using two physics problems as test scenarios with various OpenAI models. The framework is intentionally generalizable to any multiple-choice format question, including on topics without universally accepted correct answers. Our results reveal measurable differences across OpenAI model generations, with trends indicating that newer models and those employing greater "Reasoning Effort" exhibit reduced sycophantic behavior. The FR pairing method combined with our proposed indices provides a practical, adaptable toolkit for systematically comparing LLM dialogue behaviors across different models and contexts.

Gerd Kortemeyer, Marina Babayeva, Giulia Polverini et al.

We investigate the multilingual and multimodal performance of a large language model-based artificial intelligence (AI) system, GPT-4o, using a diverse set of physics concept inventories spanning multiple languages and subject categories. The inventories, sourced from the PhysPort website, cover classical physics topics such as mechanics, electromagnetism, optics, and thermodynamics, as well as relativity, quantum mechanics, astronomy, mathematics, and laboratory skills. Unlike previous text-only studies, we uploaded the inventories as images to reflect what a student would see on paper, thereby assessing the system's multimodal functionality. Our results indicate variation in performance across subjects, with laboratory skills standing out as the weakest. We also observe differences across languages, with English and European languages showing the strongest performance. Notably, the relative difficulty of an inventory item is largely independent of the language of the survey. When comparing AI results to existing literature on student performance, we find that the AI system outperforms average post-instruction undergraduate students in all subject categories except laboratory skills. Furthermore, the AI performs worse on items requiring visual interpretation of images than on those that are purely text-based. While our exploratory findings show GPT-4o's potential usefulness in physics education, they highlight the critical need for instructors to foster students' ability to critically evaluate AI outputs, adapt curricula thoughtfully in response to AI advancements, and address equity concerns associated with AI integration.