We introduce a text‑based, multi‑agent adaptation of the classic board game Clue as a testbed for assessing multi‑step deductive reasoning in large language models (LLMs). Players must infer hidden cards from a sequence of observations, keep their inferences consistent across turns, and revise beliefs when new constraints appear, exposing current models’ limits in evidence integration and constraint maintenance.
Six LLM‑driven agents are instantiated: three using GPT‑4o‑mini and three using Gemini‑2.5‑Flash. Baseline performance is measured over repeated games, with each turn’s reasoning logged.
We then augment the agents with a tool: a structured possibility matrix that converts implicit game state from the reasoning logs into an explicit representation of remaining possibilities. The matrix updates each turn, encoding extended memory and deductive constraints, offloading these tasks from the agents.
Comparisons between tool‑augmented agents and the baseline reveal that the possibility matrix substantially improves consistency and constraint reasoning, especially in long‑horizon scenarios, leading to higher task success rates.
Review