vec2vec-qwen-pair-outperforms-ot-rank-249-vs-425

IN premise — summaries/2026/08/24/jha-2025-vec2vec-s3-mean-cosine-similarity-average-cosine-similarity-between-bar.md

Created 2026-08-24T17:10:57+00:00

On the (Qwen, GTE) cross-backbone pair—where Qwen is 14× larger than the next model and 37× larger than GTE—vec2vec achieves mean rank 2.49 versus 425.07 for the best OT baseline, after 20 days of A100 training.

Summary

The vector-space comparison method (vec2vec) finds the correct cross-model match almost immediately on a Qwen-to-GTE pairing, while the best optimal-transport alternative is off by roughly 170x in ranking, even after 20 days of GPU training. This means that when comparing models of very different sizes and architectures, the transport-based approach is effectively unusable, and the system should rely on the vector-space method for cross-backbone alignment.