CKNNA representational alignment

methods / Representation Alignment / CKNNA representational alignment

Techniqueintermediate

Centered Kernel Nearest-Neighbor Alignment (CKNNA): measures similarity between two representations by combining a CKA-style kernel-alignment score with a mutual k-nearest-neighbor restriction, used to track how similar a multimodal model's internal visual-token representations are to a frozen vision foundation model's features across depth.

Used in (1 observation)

MLLM visual tokens diverge from frozen VFM features; alignment loss recovers accuracymeasured

structure: Representational-similarity trajectory across depth · models: LLaVA-1.5-7B, LLaVA-1.5-13B · paper: Visual Representation Alignment for Multimodal Large Language Models