A learning framework for neural algorithmic reasoning in which a network solving one problem instance may attend over representations of other instances in the training set, analogous to an open-book exam, rather than relying solely on the single input. A dataset encoder embeds auxiliary instances and an open-book processor fuses them into the reasoning state through cross-attention; the learned attention weights additionally expose interpretable relationships between distinct algorithmic tasks and recover much of the benefit of multi-task training at single-task cost.
Neural algorithmic reasoning is an emerging area of machine learning that focuses on building neural networks capable of solving complex algorithmic tasks. Recent advancements predominantly follow th…