Orthogonalized Multi-Direction Superposition Steering (GEMS)

Combines several simultaneously-injected steering directions by first Gram-Schmidt-orthogonalizing them against the current residual-stream update, then recombining as a norm-constrained weighted sum (so the perturbation's magnitude matches the local residual-stream update rather than growing across layers), restricted to a targeted attention-output pathway — addressing a measured collapse mode where naive additive multi-direction steering (a) accumulates norm out of the training distribution and (b) suffers mutual interference proportional to how non-orthogonal the injected directions are.