Mitosis Bench: an open pathology benchmark and RL environment for finding mitotic figures in tumor images

Mitosis Bench is a benchmark and reinforcement-learning environment where an AI agent must locate every mitotic figure in 768x768 tumor images (100 tasks drawn from the MIDOG++ dataset, human and dog tumors). The author reports the best model reaching 32.4% pass@1. The environment (midog-mitosis) was merged into Prime Intellect's open residency-environments repository on 26 October 2026 (Apache-2.0 repository); announced by the author on 27 October 2026.