# EGA: Adapting Frozen Encoders for Vector Search with Bounded Out-of-Distribution Degradation

Source: [EGA: Adapting Frozen Encoders for Vector Search with Bounded Out-of-Distribution Degradation](<https://www.sota2.com/research/paper/69fd6e385497b3b670c89534>)

Collection index: [llms.txt](<https://www.sota2.com/research/llms.txt>)

Data updated: 2026-05-12T05:00:04.826Z

Authors: Dongfang Zhao

Published: 2026-05-07T04:56:37.000Z

## Abstract

Vector search systems built on frozen vision encoders face queries from unseen classes at deployment, yet existing adapter training collapses under this shift: high-capacity adapters with global contrastive losses silently reassign unseen-class samples to wrong seen-class clusters, dropping worst-case Label Precision by over 40 points below the frozen baseline in our tests. We propose Euclidean Geodesic Alignment (EGA), a residual adapter that couples three principles: zero initialization, local triplet loss, and hypersphere projection. These collectively induce a self-limiting dynamic: triplets that already satisfy a small margin stop producing gradients, so the adapter automatically stops updating where the local geometry is already correct. Our experiments show that at convergence $96.5\\%$ of triplets are gradient-free, leaving unseen-class regions largely untouched while still enabling full-capacity refinement of seen classes. Across five diverse out-of-distribution (OOD) benchmarks, EGA achieves the highest worst-case Label Precision on the four primary splits and a consistent improvement on the fifth. The design also transfers to stronger backbones in addition to CLIP, and we provide an analytical justification linking gradient sparsity to bounded OOD perturbation.

## Recorded benchmark results

A research result applies to its stated dataset, split, metric and evaluation configuration. Compare like-for-like settings, preserve the metric's unit and direction, and distinguish paper-reported results from independent reproduction. The best result recorded here is not a claim of exhaustive, current global coverage.

- [Out-of-Distribution Retrieval on FGVC-Aircraft OOD (test)](<https://www.sota2.com/research/sota/out-of-distribution-retrieval-on-fgvc-aircraft-ood-test.md>): Out-of-Distribution Retrieval; FGVC-Aircraft OOD (test); AR@1: 90.5.

- [Out-of-Distribution Retrieval on Oxford-IIIT Pet 80/20 class-disjoint OOD protocol](<https://www.sota2.com/research/sota/out-of-distribution-retrieval-on-oxford-iiit-pet-80-20-class-disjoint-ood-protocol.md>): Out-of-Distribution Retrieval; Oxford-IIIT Pet 80/20 class-disjoint OOD protocol; LP@1: 0.9646.

- [In-distribution Retrieval on CIFAR-100](<https://www.sota2.com/research/sota/in-distribution-retrieval-on-cifar-100.md>): In-distribution Retrieval; CIFAR-100; LP@1: 70.5.

- [In-distribution Retrieval on ImageNet 1K](<https://www.sota2.com/research/sota/in-distribution-retrieval-on-imagenet-1k.md>): In-distribution Retrieval; ImageNet-1K; LP@1: 40.7.

- [Out-of-Distribution Retrieval on Food-101 (test)](<https://www.sota2.com/research/sota/out-of-distribution-retrieval-on-food-101-test.md>): Out-of-Distribution Retrieval; Food-101 OOD (test); AR@1: 87.9.

- [Out-of-Distribution Retrieval on ImageNet-1K OOD (test)](<https://www.sota2.com/research/sota/out-of-distribution-retrieval-on-imagenet-1k-ood-test.md>): Out-of-Distribution Retrieval; ImageNet-1K OOD (test); AR@1: 86.8.

- [Out-of-Distribution Retrieval on CIFAR-10 OOD (test)](<https://www.sota2.com/research/sota/out-of-distribution-retrieval-on-cifar-10-ood-test.md>): Out-of-Distribution Retrieval; CIFAR-10 OOD (test); AR@1: 83.3.

## Original sources and implementation

[Original paper](<https://arxiv.org/abs/2605.05674>)

[PDF](<https://arxiv.org/pdf/2605.05674.pdf>)