🤖 AI Summary
This work addresses the representational bias and geometric distortion commonly induced by heterogeneous tasks during pre-finetuning of domain-adaptive text embeddings, which often degrade downstream performance. To mitigate this issue, the authors propose REZE, a novel framework that formalizes representation shift control as a core principle of pre-finetuning. REZE decomposes feature-space relationships between anchor-positive pairs to identify task-variant directions and applies adaptive soft shrinkage to constrain undesirable shifts. Notably, this approach incurs no inference overhead while effectively preserving task-invariant semantic structures. Extensive experiments across multiple embedding backbones and domain benchmarks demonstrate that REZE consistently outperforms standard pre-finetuning and post-hoc regularization methods. Embedding space analysis further reveals that the induced representation shifts align closely with—and remain stable relative to—the original data 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.