Xiaomi's MiMo-V2.6 report details scaling reinforcement learning for agents

Xiaomi's MiMo team published a technical report on MiMo-V2.6, a family of models that handle text, images and other media, featured on Hugging Face Daily Papers. The focus is on scaling reinforcement learning, the training stage where models practise tasks and are rewarded for success. They trained with very large batches at context lengths up to 1 million tokens, across coding, general, visual and cybersecurity tasks run inside different agent setups. To reward long agent tasks accurately, they graded groups of attempts together, which the authors say also pushed the model towards shorter, more token-efficient answers. They froze part of the model during training for stability and added several layers of defence against reward hacking, where a model games the scoring. The team says it is open-sourcing records of how training progressed, its training environments and its reinforcement learning framework. This is a research report; it does not by itself make a new model available to use.