Modality Linear Representation-Steering (MoReS)

A trained, reversible linear map that steers a modality's tokens (e.g. visual tokens in an MLLM) into and back out of a low-rank subspace at every layer, rather than adding a single fixed direction — the intervention itself, not just its effect, is constrained to a learned k-dimensional subspace.