Training-seed variance in single-run comparisons of frozen-CLIP adaptation: evidence from cross-dataset facial expression recognition

Abstract Comparisons between adaptation routes for frozen vision-language models are usually run once per arm. We ask whether the resulting contrasts are stable across training seeds for cross-dataset facial expression recognition, with matched readout capacity and frozen evaluation sets. We retrain three source-domain arms - token attention pooling, CLIP-Adapter and a parameter-matched global head - under a seed-controlled recipe (ten seeds at ViT-B/16, fifteen at ViT-L/14) and re-evaluate them on three pre-specified axes. Five of the six contrasts a single run reports are not recovered. The ViT-B/16 margin over CLIP-Adapter on the crossed grid moves from +3.88 to -1.61 points; none of those five intervals excludes zero. One contrast survives: at ViT-L/14 the token route has a higher accuracy than CLIP-Adapter on the pre-specified corruption axis, by +3.50 points (95\% interval [+2.56, +4.43]), positive in all fifteen seeds and alone surviving family-wise correction. A second batch of fifteen seeds, fixed and hashed in advance, reproduces it at +3.77 points ([+2.99, +4.55]); pooled over thirty seeds, +3.63 points ([+3.06, +4.21]). Re-selecting the learning rate inside each seed moves one unstable contrast by 3.3 points, the size of its effect. Four of the six contrasts then remain unrecovered: the prospective batch also lifts the ViT-B/16 degradation margin above zero but its sign test does not confirm it, and the surviving contrast does not extend to other degradation families, a third encoder, or a degradation-specific increment on the JAFFE posed-expression corpus. Single-run comparisons in frozen-CLIP adaptation can report effects that do not replicate.