REZE: Representation Regularization for Domain-adaptive Text Embedding Pre-finetuning

TL;DR

REZE is a novel regularization framework for domain-adaptive text embedding pre-finetuning that controls representation shifts, improving robustness and performance across various benchmarks.

Contribution

It introduces a representation regularization method that explicitly manages representation shifts during embedding pre-finetuning, outperforming existing approaches.

Findings

REZE outperforms standard pre-finetuning methods in multiple benchmarks.

REZE maintains stable embeddings where other methods collapse.

Analysis shows REZE aligns embedding shifts with the original manifold.

Abstract

Recent text embedding models are often adapted to specialized domains via contrastive pre-finetuning (PFT) on a naive collection of scattered, heterogeneous tasks. However, this approach often introduces task-induced bias alongside domain knowledge, leading to uncontrolled representation shifts that distort the pretrained embedding geometry and cause substantial performance degradation. To address this issue, we propose REZE, a representation regularization framework that explicitly controls representation shift during embedding pre-finetuning. REZE operates on the relations of anchor-positive pairs and decomposes them in an eigenspace. It then measures task-wise dispersion along each eigencomponent to identify task-variant directions and applies adaptive soft-shrinkage to suppress task-induced noise while preserving task-invariant semantic structure, without inference-time overhead. Experiments across multiple embedding backbones and specialized benchmarks show that REZE outperforms standard pre-finetuning and isotropy-oriented post-hoc regularization in most settings, remaining stable where existing PFT variants collapse. Embedding space analyses further confirm that REZE induces controlled shifts aligned with the original embedding manifold, underscoring representation shift control as a key principle for robust embedding pre-finetuning under heterogeneous supervision.