Quantify the relative temporal variance of token contrasts and individual view scores

Measure the variance of the PACE token contrast across training steps relative to the variance of the individual teacher-view log-probabilities, in order to determine the extent to which pairwise contrast reduces sensitivity to training drift.

Background

PACE argues that pairwise hint contrasts cancel offsets shared by the positive and negative teacher views. The paper reports that the single-view vision gap changes substantially during training, whereas the batch means of the full-vision and null-vision hint contrasts remain comparatively stable. However, the authors note that these batch-level means do not establish whether token-level or view-specific variability is reduced.

The unresolved task is to measure the variance of the combined token contrast across training steps and compare it with the variance of the individual view log-probabilities. This would assess the robustness of the claimed drift-mitigation property beyond shared batch-level offsets.

References

These are batch means, so they do not rule out token- or view-specific shifts; measuring the variance of $c_{i,t}$ across steps relative to that of the individual view log-probabilities remains future work.

— Enhancing Multimodal RLVR with Pairwise Token Contrast and Adaptive Trajectory Confidence  (2610.21671 - Zhang et al., 15 Oct 2026) in Appendix, Section “Shared-Offset Cancellation in Pairwise Token Contrast,” paragraph “Empirical verification”