single source · post-training
Multi-harness reinforcement learning
Multi-harness reinforcement learning treats harness diversity as a training dimension, training the agent across several harnesses so it generalizes to unseen interaction mechanisms.1 Instead of production harnesses, whose safeguards and coupled modules make credit assignment unreliable, it trains on minimal, decoupled mini-harnesses that can be freely recombined.2
How this definition was written
Summarised by glm-5.3-flash from 1 source; every sentence is checked against the quote it cites.
We therefore treat harness diversity as an additional training dimension alongside diversity in tasks and environments
— MiMo V2 6 technical reportThese extras fall outside the task-completion reward signal, making credit assignment unreliable
— MiMo V2 6 technical report
Also called multi-harness training.
Filed under post-training › agentic post-training