The Signs Were Always There: Training-Free Detection and Control in Raw Transformer Dimensions
The Bag of Dims (BoD) framework treats standard transformer hidden states as a training-free feature basis. It allows for unsupervised detection of semantic concepts by analyzing sign patterns across dimensions, and for steering generation using a training-free 'write target.' This enables concept suppression and induction without any model retraining, proving that reading and controlling concepts are separate skills on the same substrate.