Answers are grounded in this paper and cite their sources.
High accuracy on independent and identically distributed (IID) test data does not reveal which temporal cue a sequence model uses. A directional process can leave a cue in individual frames, so out-of-distribution (OOD) failure alone does not establish temporal-order use. We propose a six-gate audit that attributes failure to a nuisance and distinguishes frame-local, order-invariant multi-frame, and order-encoded cues. A companion controlled benchmark provides the factor-level probes and interventions the audit requires. Applied to this benchmark, the audit revealed that a variant designed to encode temporal order was instead solvable from individual frames. In the strict order-encoded variant (OE-Strict), direction is readable only from interior frame order. With nuisance direction aligned with the label with probability 0.97 during training, a gated recurrent unit (GRU) trained by empirical risk minimization (ERM) selects the shortcut in thirty of thirty seeds (0.970 IID and 0.030 OOD accuracy). Channel-selective interventions attribute the collapse to nuisance frame ordering. The audit distinguishes which process supports a prediction and whether using that process requires temporal order.
No takes yet. Share an insight, caveat, or question.
Cho et al. (2026) studied this question.
Synapse has enriched 10 related papers. Consider them for comparative context: