Sources of surrogate-ranking failure

Determine the respective contributions of higher-order interactions, finite calibration noise, and objective mismatch to the failure of the surrogate allocation model to predict rollout FID rankings.

Background

The paper evaluates unary and pairwise surrogate objectives for selecting horizon changes. On an ImageNet NFE-16 experiment, the pairwise surrogate achieves lower surrogate energy but worse rollout FID than the unary method, showing that surrogate optimization does not reliably predict endpoint quality. The paper lists higher-order effects, calibration noise, and mismatch between the surrogate and rollout objective as possible explanations, without determining their contributions.

References

Higher-order effects, finite calibration noise and objective mismatch are possible explanations, whose contributions are unresolved.

— Rethinking Prediction Horizons: Decoupling Model Queries from State Updates in Few-Step MeanFlow  (2610.20071 - Wang et al., 14 Oct 2026) in Appendix, subsection “Earlier surrogate and reference diagnostics”