When Does Dense Retrieval Need Asymmetric Geometry?

Introduction

Dense retrieval systems encode queries and documents into vectors and then rank candidates with a similarity function. One important design choice is whether both sides should use the same projection matrix. A shared projection is compact and structurally constrained. Separate query and document projections are more expressive, but their additional parameters can increase estimation variance when training data are limited.

The paper When Does Dense Retrieval Need Asymmetric Geometry? studies this choice through a bias-variance lens. Rather than assuming that greater flexibility is automatically beneficial, it asks when the extra expressiveness of dual projections is worth the statistical cost.

Key findings

• The two designs represent different operator classes. Shared projections induce positive-semidefinite operators, which constrain the relationships that can be expressed between queries and documents. Dual projections can realize arbitrary low-rank operators and therefore cover a broader family of cross-space relationships.
• Flexibility has a statistical price. A dual projection can reduce the approximation error caused by the shared structure, but it also introduces more degrees of freedom. The paper derives a local Gaussian boundary in which dual projections are preferable precisely when the squared directional signal exceeds the estimation cost of those additional parameters.
• The preferred geometry depends on sample size. In the reported rank-versus-sample-size experiments, shared projections win in most low-data settings. As the amount of training data increases, dual projections become consistently preferable across the examined grids. This does not make dual geometry universally superior; it highlights the need to match model capacity to available evidence.
• Geometric mismatch favors asymmetry. In controlled experiments, the mean Dual-minus-Shared NDCG@10 advantage more than doubles as query rotation increases from 0 to 90 degrees. When query and document spaces are less aligned, forcing both sides through the same transformation can create a larger approximation bias.
• CARS turns the theory into a selection procedure. The Cross-fitted Asymmetry Risk Selector estimates reproducible directional signal from training pairs and uses it to choose between shared and dual geometry. The paper reports a 49–96% reduction in held-out regret relative to fixed-geometry baselines, along with 90.1% mean geometry-selection accuracy.

Why it matters

The main contribution is a shift in how projection design can be approached. Instead of treating shared versus dual projections as a purely architectural preference, practitioners can view it as a model-selection problem governed by signal, data volume, and geometric mismatch.

For small or noisy datasets, the lower variance of a shared projection may produce more reliable retrieval. For larger datasets, or for systems in which query and document representations exhibit clearly different directions, dual projections may use their extra capacity to reduce bias. The result also cautions against making the decision solely from training-set ranking scores. Validation risk, sample size, low-rank constraints, and the reproducibility of asymmetric signals should be considered together.

CARS provides a practical route toward that decision, although its robustness across retrieval tasks, encoders, and training procedures will require further evaluation. The broader lesson is conditional rather than absolute: asymmetric geometry is worthwhile when its reproducible signal is larger than the estimation burden introduced by its extra freedom.

Source: Hugging Face Daily Papers