ClassicLogic: A Knowledge-Driven Benchmark of Classic Puzzle Games for Evaluating Compositional Generalization
Modern artificial intelligence excels at pattern matching but often struggles with the systematic, multi-step reasoning required to solve complex problems. While humans can easily combine known rules to tackle new, unfamiliar challenges—a skill known as compositional generalization—AI models frequently fail when they must plan and execute logical strategies. ClassicLogic is a new benchmark suite designed to test and improve these reasoning capabilities by using four classic logic puzzles: Sudoku, KenKen, Kakuro, and Futoshiki.
A Hierarchical Approach to Reasoning
The core innovation of ClassicLogic is its hierarchical knowledge base. Unlike other benchmarks that simply measure whether an agent gets the right answer, ClassicLogic defines exactly how a puzzle should be solved. It breaks down complex strategies into smaller, foundational rules. For example, in Sudoku, an advanced move like an "X-Wing" is defined as a specific combination of simpler pattern-recognition steps and basic constraint-propagation rules. This structure allows researchers to see exactly where an AI model fails: whether it cannot learn the basic rules, struggles to combine them, or is unable to plan multi-step strategies.
Three Dimensions of Generalization
ClassicLogic evaluates an agent’s reasoning across three distinct levels:
Entity Composition: The ability to perceive and interpret visual inputs, such as identifying handwritten digits in a puzzle grid.
Relational Composition: The ability to follow specific rules, such as integrating arithmetic constraints in KenKen or inequality constraints in Futoshiki.
Procedural Composition: The ability to chain simple, atomic rules together to form complex, multi-step strategies that solve the puzzle.
Strategy-Driven Puzzle Generation
To ensure the benchmark is both rigorous and scalable, the authors developed a two-stage generation process. First, they create a library of puzzle "templates" that are mathematically validated to require specific strategies to solve. This ensures that every puzzle has a unique solution and a known difficulty level. Second, the system uses these templates to generate a wide variety of playable puzzles in real-time. By applying symmetry-preserving transformations—such as rotating the grid or swapping values—the benchmark can produce many distinct puzzles that all test the same underlying logical skills.
A Tool for Neuro-Symbolic AI
ClassicLogic is designed to support the development of neuro-symbolic AI, a field that aims to combine the perceptual strengths of neural networks with the logical rigor of symbolic systems. By providing a transparent environment where the reasoning process is as important as the final result, the benchmark helps researchers diagnose the limitations of their models. It serves as a diagnostic tool to guide the creation of AI systems that are not just accurate, but also capable of the systematic, interpretable reasoning characteristic of human intelligence.
Comments (0)
to join the discussion
No comments yet
Be the first to share your thoughts!