MiMo-V2.6-Flash · MiMo-V2.6-Pro
Xiaomia practical approach to advancing foundation models through large-scale agentic reinforcement learning.
general family · filed under post-training
Agentic reinforcement learning applies RL to multi-step agent tasks, where rollouts span multiple turns and dialogue contexts in an interactive environment.1,2 Because outcome rewards provide only coarse supervision for these long trajectories, it adopts an agent-centric execution model with fine-grained learning signals such as loss masking and advantage shaping.3
Summarised by glm-5.3-flash from 1 source; every sentence is checked against the quote it cites.
Agentic rollouts span multiple turns and dialogue contexts— MiMo V2 6 technical report
scaling reinforcement learning (RL) on complex agentic tasks— MiMo V2 6 technical report
outcome rewards provide only coarsesupervision. Wethereforeadoptanagent-centricexecutionmodelandorganizetrajectory data as a hierarchy. Within this hierarchy, a configurablePenalty Module applies loss masking 26 and advantage shaping— MiMo V2 6 technical report
Also called agentic RL, Large-scale agentic reinforcement learning.
8 quotes from 8 sources, one card per model, each opening with its strongest.
a practical approach to advancing foundation models through large-scale agentic reinforcement learning.
trained with reinforcement learning inside our agent harness
Post-training from this checkpoint (SFT, mixed RL, multi-domain on-policy distillation, and agentic RL) produces LFM3-1.7B.
tasks from web browsing to office work are run inside real agent harnesses, preparing the model for deployment.
tasks from web browsing to office work are run inside real agent harnesses, preparing the model for deployment
Post-training incorporates SFT, large-scale agentic RL, and Multi-Teacher On-Policy Distillation (MOPD)
Post-training incorporates SFT, large-scale agentic RL, and Multi-Teacher On-Policy Distillation (MOPD)
large-scale agentic RL