MATE: Solving Contextual Markov Decision Processes with Memory of Accumulated Transition Embeddings

In three linesMATE is a memory architecture for solving Contextual Markov Decision Processes (CMDPs). It replaces the intractable posterior belief with sum-aggregated memory, avoiding growing computational costs of Transformers and gradient issues of RNNs. Evaluations demonstrate computational advantages while achieving performance comparable to standard sequence-model baselines.

Summary generated by Claude — human-verified