Explain the exception of tidal networks and the depth-dependent cost of learned weighting

Determine why learned Q/K weighting improves forecasting on tidal multivariate groups but harms in-context groups, and why the harm becomes costly only through repeated application across network depth, including what the learned weighting responds to and whether depth-dependent mechanisms such as rank collapse account for the behavior.

Background

The paper decomposes Chronos-2 group attention into a V/O pathway that performs pooled cross-variate representation and a Q/K pathway that determines variate-specific weights. Uniform pooling generally helps both multivariate and in-context forecasting, whereas learned weighting helps multivariate groups but often harms in-context groups. The authors further find that the harm is not attributable to a single later block: it emerges through repeated composition across the twelve attention blocks, while the first block is the most effective location for an inference-time repair.

The authors explicitly leave unresolved both the reason tidal networks behave differently and the mechanism by which depth amplifies the cost of learned weighting. They note that the interventions localize the effect to a pathway and block but do not reveal what the block computes, and identify rank collapse and other depth-dependent explanations as possibilities that the experiments cannot adjudicate.