RE-TRIANGLE: Geometric alignment boosts multimodal retrieval by +8.7%

A new reproducibility study published on arXiv (2605.27436) by Ghosh, Bandyopadhyay, Bindra, and Qiao validates the geometric alignment approach of the TRIANGLE framework for multimodal retrieval. Traditional pairwise alignment methods (e.g., text↔video, text↔audio) create a geometric blind spot: they neglect mutual consistency between peripheral modalities like video and audio. TRIANGLE addresses this by minimizing the area of modality triplets on a hypersphere, forcing holistic alignment across all modalities. The study confirms that this geometric objective outperforms pairwise baselines in zero-shot settings, with Recall@1 gains of up to +8.7 points—though benefits are domain-dependent.

However, the researchers could not reproduce the original paper’s learning-from-scratch results. Analysis on a synthetic toy dataset revealed that jointly optimizing the geometric alignment loss with Data-Text Matching (DTM) loss leads to instability. Additionally, cosine regularization was found to primarily stabilize text-to-video retrieval, and fine-tuning with domain supervision amplified geometric alignment benefits but at the cost of reduced cross-dataset generalization. These findings support the efficacy of geometric alignment for multimodal retrieval while highlighting critical optimization sensitivities. Code is available at the provided URL.