Recent advances in dense retrieval expose fundamental issues in how embedding spaces behave under domain shift, particularly in zero-shot scenarios.
This thesis examines representation-level principles that strengthen generalization by focusing on two...
Recent advances in dense retrieval expose fundamental issues in how embedding spaces behave under domain shift, particularly in zero-shot scenarios.
This thesis examines representation-level principles that strengthen generalization by focusing on two directions: balanced isotropy and query-focused referentiability.
First, we propose Hybrid Isotropy Learning (HIL), which stabilizes multi-vector retrieval models by combining isotropic and anisotropic properties of token representations.
Through an interaction-based isotropy metric (InterIso) and a hybrid late-interaction architecture, HIL improves zero-shot retrieval performance.
Second, we introduce Learning Referentiable Representation (LRR), a framework that ensures passages remain reliably referenced by their gold queries even when domains shift.
LRR enhances representation locality and query-awareness through two targeted metrics, Self-P and Self-Q, preventing degeneration into non-referentiable states.
Both studies offer a unified perspective on controlling representational geometry for dense retrieval. On experiments with BEIR and MS MARCO benchmarks, these approaches consistently improve zero-shot performance, demonstrating the importance of geometric regularization in retrieval models.