In an embedding caching system, embeddings generated from previous problems are re-used to improve performance on future problems. A data structure stores problems and their corresponding embeddings. When computing future embeddings, this data structure can be queried to determine whether an embedding has already been computed for a problem with the same structure. If it has, the embedding can be retrieved from the data structure, saving the time and computational expense of generating a new embedding. In one variation, the query is not based on exact matches. If a new problem is similar in structure to previous problems, those embeddings may be used to accelerate the generating of an embedding for the new problem, even if they cannot be used directly to embed the new problem.