Pretty useful paper, the framework they propose inside is user-friendly and designed to simplify the process of developing and experimenting with new RLM architecture. Also it supports different granularities of reasoning steps, ranging from individual tokens to full sentences or structured segments. And enables diverse training schemes, like OBS which is pretty neat overall