Generalize attribution behavior beyond RhythmFormer

Investigate whether the attribution behavior observed for RhythmFormer also appears in other neural-network architectures.

Background

The study evaluates four attribution methods exclusively on RhythmFormer models trained for remote photoplethysmography. Although the experiments compare conditions, datasets, and recording challenges, they do not test whether the observed relationships among skin coverage, SaCo, and model performance persist for other model architectures. Consequently, the external validity of the attribution findings beyond RhythmFormer remains unresolved.

References

Every statement here is about RhythmFormer, and whether the same attribution behaviour appears in another architecture is untested.

— Cross-Dataset Transfer and Reliability of Explainable Artificial Intelligence for RhythmFormer Remote Photoplethysmography  (2609.03663 - Chen et al., 3 Sep 2026) in Section 5.6, “Reproducibility, limitations, and future work”

Whether the faithfulness result generalizes to other domains and forecasters is open; a replication on a second domain is the most direct test and a natural next step.

— How Faithful Is Attribution for Sales Forecasting? A Counterfactual Study  (2609.04797 - Kechyn, 4 Sep 2026) in Section 6, Limitations, paragraph “Single dataset and metric”

It remains unclear whether the ranking of saliency methods generalizes to stronger models.

— Faithful Faithfulness Evaluations: Challenges & Pitfalls Learned from a Breast MRI Case Study  (2609.25978 - Poolpol et al., 22 Sep 2026) in Discussion

Finally, our evaluation is limited to VideoMamba classifiers with a class-token readout and does not establish whether the same construction applies to other readout mechanisms or selective state-space architectures.

— READ: Reading the Selective Scan for Faithful Video Attribution  (2610.13776 - Trędowicz et al., 9 Oct 2026) in Limitations, Section 5, p. 9