Neuro-symbolic framework solves Zebra puzzles from natural language
Original title: From Natural Language to Exact Cover: A Neuro-Symbolic Approach to Zebra Puzzles
Combines LLM semantic parsing with symbolic exact cover solver for reliable logic puzzle solving.
The framework translates natural language Zebra puzzles into structured tool calls that construct exact cover instances. A symbolic solver computes the solution and provides feedback to iteratively refine invalid encodings using the uniqueness property. The hybrid approach consistently outperforms strong neural baselines including CoT prompting across multiple LLM families and scales.
- Bridges neural parsing and deterministic reasoning
- Exploits uniqueness of Zebra puzzles for feedback
- Evaluated at multiple parameter scales
- Guarantees correctness unlike pure neural methods
From the original article
1 min read at arxiv.orgChain-of-Thought (CoT) generation has substantially improved the performance of Large Language Models (LLMs) on complex reasoning tasks, including code generation, data analysis, and exam-style question answering. Despite these advances, purely neural LLMs continue to struggle with elementary logical reasoning problems and lack the determinism, soundness, and reliability characteristic of symbolic reasoning systems.
By Paulius Skaisgiris · arxiv.org
A short excerpt; the article and its images belong to the publisher. Summary, key facts and assessment above are AI-written from the article and can be wrong.