IC-390The parallelotope volume of modality embeddings from LanguageBind, VAST, and Valor on MSR-VTT is strongly correlated with their downstream R@1 retrieval performance

Giordano Cicchetti, Eleonora Grassucci, Luigi Sigillo, Danilo Comminiello

SourceGramian Multimodal Representation Learning and Alignment

The authors extract video, audio, and text embeddings from LanguageBind, VAST, and Valor (both zero-shot and fine-tuned), compute the GRAM parallelotope volume for each model's three modality vectors, and correlate this geometric quantity with each model's R@1 on MSR-VTT. Over 1000 test samples, the Pearson correlation between 1−GRAM and R@1 is ρ = 0.923. The claim is that the geometric alignment of a model's modality embeddings, as captured by the volume of the parallelotope they span, is a reliable indicator of that model's downstream retrieval capability.

Evidence
correlational
Key metric
Pearson correlation coefficient ρ = 0.923 between 1−GRAM and R@1, computed over 6 model configurations (LB zs, VAST zs, GRAM zs, LB ft, VAST ft, Valor ft) on MSR-VTT, averaged over 1000 test samples
Caveat
The correlation is computed over only 6 data points, 3 of which are the authors' own GRAM model; the paper does not report the correlation restricted to released models alone. The GRAM volume is the authors' own proposed metric, so the finding is partly circular: it validates the metric using the models it was designed to evaluate.
Model
VAST, LanguageBind, Valor
Datasets
MSR-VTT [eval]
Extraction
automatic-extraction