Towards robust long-context understanding of large language model via active recap learning | ScienceToStartup | ScienceToStartup