Publication
arXiv
Stage
Preprint
What we read
Summary of the paper
Authors
Paulius Skaisgiris, Thomas Pammer, Veronika Semmelrock, Mykyta Ielanskyi, Maximilian Heisinger, Erich Kobler
Universities and research institutions
Institute of Logic and Computation, TU Wien, Austria, Institute for Application-Oriented Knowledge Processing, JKU Linz, Austria, Department of Artificial Intelligence and Cybersecurity, University of Klagenfurt, Austria, Institute for Machine Learning, JKU Linz, Austria, Institute for Symbolic Artificial Intelligence, JKU Linz, Austria, Institute for Machine Learning, LIT AI Lab, Department of Virtual Morphology, Clinical Research Institute Medical AI, JKU Linz, Austria

Institution metadata: Original paper ↗

What they did and found

A system that blends language understanding with a math-like checker turned Zebra Puzzle clues into a set of rules and used a solver to verify results. It outperformed several plain neural models on medium and large Zebra puzzles, though performance varied with setup.

Why it matters

This shows hybrids can be more accurate for hard problems, but they need careful setup and ongoing feedback to stay reliable as problems grow.

NeuralExactCover blends language skills with a formal solver. The solver’s results guide the language model to refine its tool calls, aiming for a correct, unique solution without exhausting the model’s capacity.

Practically, this means you can handle tougher problems by sharing the work between an AI that understands text and a solver that enforces exact rules, but you must manage the setup and feedback properly to keep results trustworthy.

What remains uncertain

Preprint; results are based on Zebra Logic benchmarks and ZebraExtreme, not peer-reviewed; performance varies with puzzle size and model/configuration.

Read the paper PDF ↗

Original sources · 1
  1. From Natural Language to Exact Cover: A Neuro-Symbolic Approach to Zebra Puzzles ↗arXiv · 2026-11-02

Check the original paper for its authors, methods, version and access terms.