Causal attribution of the wedge channel’s cross-phase routing effect

Investigate whether the wedge channel moves attention mass onto cross-phase token pairs characterized by small dot products and large wedge magnitudes, thereby establishing the causal mechanism underlying the observed performance gains.

Background

The paper proposes that the wedge-product component of Time–Frequency Geometric Cross-Attention directly captures cross-phase relationships whose representations are near-orthogonal and therefore poorly distinguished by dot-product attention. Although the experiments show performance improvements and the paper reports Temporal Orthogonal Division-of-Labor patterns in bimanual demonstrations, these results do not establish that the wedge channel actually routes attention toward the hypothesized cross-phase token pairs.

The authors identify a concrete unresolved analysis: examine query/key routing to determine whether attention mass is redirected toward pairs with small dot products but large wedge magnitudes, and combine this analysis with a dot-only ablation under out-of-distribution conditions. Resolving this issue would turn the proposed mechanism from a motivating hypothesis into causal evidence.