Neural Algorithmic Reasoning Must Explain When Neuralization Adds Value

Neural algorithmic reasoning (NAR) has emerged as a promising direction for integrating mathematical and algorithmic reasoning into end-to-end trainable neural pipelines, aiming to bring the structure and reliability of algorithms into practical pipelines, such as real-world mathematical problem solving. This position paper argues that, despite growing interest, NAR remains a broad and ambiguously defined collection of ideas. This vagueness obscures its relationship to neighboring areas, such as differentiable programming and machine learning for combinatorial optimization, and risks slowing progress. We argue that NAR has distinctive leverage, but realizing this leverage requires clearer answers to two central questions: what should count as NAR, and when is neuralizing an algorithmic procedure genuinely justified? To this end, we propose a working definition of NAR, clarify where it should and should not be applied, and identify theoretical goals for understanding correctness, generalization, and benefits of neuralized algorithms. We further propose empirical protocols that emphasize cross-paradigm comparisons and track practical value beyond synthetic algorithmic tasks. Together, these objectives provide a roadmap for developing NAR into a coherent research area with broader scientific and practical impact.