single mention · post-training
Mid-training alignment data from reward-hacking cases
Mid-training alignment data from reward-hacking cases is a set of training examples synthesized from observed reward-hacking behavior and added during mid-training.1 Each example shows the model reflecting on its faulty reasoning, revising the relevant turn, and continuing with actions grounded in the task specification, keeping the original error recognizable.2
How this definition was written
Summarised by glm-5.3-flash from 1 source; every sentence is checked against the quote it cites.
Tomitigatethisbehavior,wesynthesizedasetoftrainingexamplesfromthesecases and included them in mid-training
— MiMo V2 6 technical reportMiMo reflects on the faulty reasoning, revises the relevant turn, and continues with actions grounded in the task specification. The revised reasoning keeps the original error recognizable and makes the correction explicit
— MiMo V2 6 technical report
Also called mid-training alignment data.
Filed under post-training › mid-training & continual pretraining