In-Context Multiple Instance Learning

arXiv·medium signal

Multiple Instance Learning (MIL) handles supervision at the level of bags of instances and is widely used in pathology and satellite imagery. This paper reframes MIL to operate in-context — leveraging a model's context window to do bag-level reasoning without task-specific retraining. A lightweight path to weak-supervision problems for practitioners who want to avoid building dedicated MIL pipelines. Published 2026-06-04 (cs.LG/cs.CV).