Title: Addressing Zebra Puzzles Through Constraint-Driven Multi-Agent Architectures

Original: arXiv:2407.03956v3 Announce Type: replace-cross Abstract: Previous studies have sought to boost the logic puzzle-solving capabilities of Large Language Models (LLMs) by employing methods like chain-of-thought prompting or integrating symbolic representations. However, these frameworks often fall short when tackling intricate logical challenges, such as Zebra puzzles, primarily because converting natural language clues into formal logical statements is inherently complex. To address this, we present ZPS, a multi-agent system that combines LLMs with a standard theorem prover. This system manages the demanding task of puzzle resolution by decomposing the problem into smaller, more tractable components, producing SMT (Satisfiability Modulo Theories) code for resolution via a theorem prover, and leveraging iterative feedback among agents to refine their outputs. Additionally, we developed an automated grader for grid puzzles to verify solution accuracy, demonstrating its reliability through a user study. Our methodology yielded performance gains across all three LLMs evaluated, with GPT-4 achieving a 166% increase in the count of completely correct solutions.

Source: arXiv Generated at: 2026-06-04 00:00:00 UTC