大型语言模型的推理能力如何？在文本游戏环境中评估多步骤演绎推理

Deducing the culprit poses challenges for large language model (LLM) agents. A text-based multi-agent version of the classic board game Clue is implemented as a rule-based testbed for evaluating multi-step deductive reasoning, involving six agents sourced from GPT-4o-mini and Gemini-2.5-Flash. The study further investigates whether fine-tuning on structured logic puzzles translates to improved in-game reasoning and gameplay. Across 18 simulated games, agents achieved only four correct wins, indicating difficulties in maintaining consistent deductive reasoning throughout a full game. Additionally, it was found that fine-tuning does not reliably enhance performance and, in some instances, appears to increase the volume of reasoning without improving its precision.