RuozhiBench: Evaluating LLMs with Logical Fallacies and Misleading Premises
Zenan Zhai, Hao Li, Xudong Han, Zhenxuan Zhang, Yixuan Zhang, Timothy Baldwin, Haonan Li
TL;DR
RuozhiBench introduces a bilingual benchmark to probe LLMs' abilities to identify and reason about deceptive premises and logical fallacies. It comprises RuozhiBench-Gen (data collection, filtering, translation, and annotation) and RuozhiBench-MC (a multiple-choice evaluation) to assess generation and decision-making under misleading prompts. Across 17 models, results show larger models generally outperform smaller ones, yet top performers (e.g., Claude-3-haiku) reach only about 62% vs human performance exceeding 90%, highlighting persistent gaps in logical reasoning under deception. The work also analyzes evaluation protocols, annotates fallacy types, and demonstrates the benefits and limitations of different evaluation formats, offering guidance for developing more robust LLMs and future research directions in logical reasoning and deception detection.
Abstract
Recent advances in large language models (LLMs) have shown that they can answer questions requiring complex reasoning. However, their ability to identify and respond to text containing logical fallacies or deliberately misleading premises remains less studied. To address this gap, we introduce RuozhiBench, a bilingual dataset comprising 677 carefully curated questions that contain various forms of deceptive reasoning, meticulously crafted through extensive human effort and expert review. In a comprehensive evaluation of 17 LLMs from 5 Series over RuozhiBench using both open-ended and two-choice formats, we conduct extensive analyses on evaluation protocols and result patterns. Despite their high scores on conventional benchmarks, these models showed limited ability to detect and reason correctly about logical fallacies, with even the best-performing model, Claude-3-haiku, achieving only 62% accuracy compared to the human of more than 90%.