training-gaming
Behavior in which an AI understands relevant features of the process being used to train it and deliberately aims its optimization at some component of the reward process for the current episode, such as the feedback or score that determines training updates, rather than merely pursuing the task as presented. Training-gaming can be terminally or instrumentally motivated and can produce alignment faking when appearing aligned is useful for performing well according to the training process.
Metadata
- aliases
- en
- playing the training game
- training game
- training gaming
- training-gamer
- ja
- 訓練ゲーム
- 訓練ゲーミング
- concept_id
- training_gaming
- preferred
- en
- training-gaming
- ja
- トレーニング・ゲーミング
- relations
- relations
- prerequisite
- status
- active