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

2026-04-20 · Source: cs.AI updates on arXiv.org · Field: Technology & Digital — Artificial Intelligence & Machine Learning, Robotics & Autonomous Systems, Gaming & Interactive Media · Depth: Expert, long

Summary

InfoChess is a novel symmetric adversarial game designed to isolate competitive information acquisition as its primary objective, unlike traditional games where information is instrumental to other goals like material gain. Players score by probabilistically inferring the opponent's king location throughout the game, with no piece capture. Pieces alter visibility, and all pieces move one square in any direction, but rooks and bishops cast extended lines of sight while pawns obstruct vision. The game introduces a hierarchy of heuristic agents with increasing opponent modeling, alongside a reinforcement learning (RL) agent that surpasses these baselines. Gameplay is analyzed using information-theoretic measures such as belief entropy, oracle cross entropy, and observer cross entropy, which help disentangle epistemic uncertainty, calibration mismatch, and uncertainty from adversarial movement. The authors provide code for the environment and agents, establishing InfoChess as a testbed for multi-agent inference under partial observability.

Key takeaway

For AI scientists and machine learning engineers developing multi-agent systems, InfoChess offers a controlled environment to study adversarial inference and strategic concealment. You should consider its design principles for isolating specific mechanisms in complex systems, particularly when disentangling information dynamics from other objectives. The provided code and public interface encourage experimentation with novel belief modeling and reinforcement learning strategies in partially observable settings.

Key insights

InfoChess isolates competitive information acquisition as the sole objective in a symmetric adversarial game.

Principles

• Information acquisition can be a primary game objective.
• Opponent modeling improves competitive performance.
• Information-theoretic metrics reveal strategic play regimes.

Method

InfoChess uses a fixed horizon of 25 turns per side, scoring players on the probability mass assigned to the true opponent king location at each turn, with initial king placement randomized across four back-row squares.

In practice

• Implement a hierarchy of heuristic agents for baseline comparison.
• Train RL agents using frozen state encoders.
• Utilize belief entropy to track epistemic uncertainty.

Topics

• Adversarial Inference
• Partial Observability
• Multi-agent Reinforcement Learning
• Information Theory
• Belief Modeling

Code references

• murphyka/infochess

Best for: AI Scientist, Machine Learning Engineer, Research Scientist

Related on AIssential

• When Data Lies: Finding Optimal Strategies for Penalty Kicks with Game Theory — Optimal decisions in strategic interactions cannot be inferred from raw historical averages alone.
• RTSGameBench: An RTS Benchmark for Strategic Reasoning by Vision-Language Models — RTSGameBench reveals leading VLMs struggle with multiagent coordination and scaled strategic reasoning in complex RTS environments.
• Kaggle Game Arena: Werewolf | Day 3 Highlights — AI models can engage in complex social deduction games, demonstrating strategic reasoning and adaptive deception.
• Interaction-Breaking Adversarial Learning Framework for Robust Multi-Agent Reinforcement Learning — Disrupting inter-agent mutual information via observation and action attacks improves MARL robustness.
• AIエージェントが最適化プログラミングコンテストで初優勝 — AI agents can now win complex, real-time optimization contests by scaling LLM inference and self-learning.
• Why Learn if You Can Infer: Active Inference for Robot Planning & Control — Active inference and object-centric priors enable rapid, human-level AI performance in complex game environments.

Open in AIssential →

Editorial summary, takeaway, and curation by AIssential. Original article published by cs.AI updates on arXiv.org.