NBA benchmark

New method boosts LLM long context handling with attention-state memory

Researchers have developed a new method called attention-state memory to improve how large language models handle long context inputs. This training-free approach externalizes the prefix into a memory of precomputed att…