REPO: CONTENT-AWARE ATTENTION REORDERING

Summary from the TLDR AI newsletter

RePo is a module that repositions tokens based on contextual relevance rather than fixed order, improving long-range attention and robustness to noisy or structured inputs. The method draws on Cognitive Load Theory to justify reorganizing input structure for more efficient model processing.