Analyzing the Markov Property in Neural Algorithmic Reasoning

The author explores the misalignment between historical embeddings and the Markov nature of algorithmic reasoning tasks, proposing ForgetNet and G-ForgetNet to address this issue effectively.