InfoChess: A Game of Adversarial Inference and a Laboratory for Quantifiable Information Control

Researcher Kieran A. Murphy has introduced InfoChess, a novel symmetric game designed as a pure laboratory for studying adversarial inference and information control in AI. Stripping away the traditional incentive of piece capture, the game elevates competitive information acquisition to the primary objective. Players use their pieces solely to alter visibility on the board, and their score is based on the accuracy of their probabilistic inference about the opponent's king location over time. This creates a controlled environment to disentangle the complex dynamics of multi-agent reasoning under partial observability, free from confounding material goals.

To explore strategic play, the researchers developed a hierarchy of heuristic agents with increasing levels of opponent modeling and trained a reinforcement learning (RL) agent that successfully outperformed these baselines. Crucially, they leverage the game's discrete structure to analyze gameplay through precise information-theoretic measures, including belief entropy and predictive log score. These metrics allow researchers to separately quantify epistemic uncertainty, calibration errors, and uncertainty induced by an adversary's deliberate movements. The project, which includes released code and a public interface, was accepted at the AAMAS 2026 workshop and is positioned as a foundational testbed for advancing multi-agent AI systems where information is the central currency of competition.