vec2vec-cross-backbone-convergence-3-of-15-vs-14-of-15

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

GAN-based vec2vec training is seed-sensitive: shared-backbone pairs converge at 80% top-1 for 14/15 random initializations, while cross-backbone pairs converge for only 3/15 within the same epoch budget.

Summary

Translating between embedding models that share the same architecture is reliable, but translating between models built on different architectures is mostly a coin flip that fails most of the time. In practice, any system relying on cross-model vec2vec translation needs to either budget for many training attempts, extend the training window, or treat cross-backbone translation as an unsupported capability.