AI
When Does Dense Retrieval Need Asymmetric Geometry? A Bias-Variance Theory of Shared and Dual Projections
Publication: Hugging Face Daily Papers
Author: Not available in collected metadata
Source published: 25/09/2026, 20:00:00 UTC
Briefing created: 30/09/2026, 05:07:15 UTC
Based on a feed excerpt; full article content has not been acquired. Source evidence retrieved 30/09/2026, 03:18:26 UTC.
Briefings
A theoretical framework explains when dual query-document projections outperform shared ones in dense retrieval, using a bias-variance theory for low-rank bilinear scoring. Experiments show Dual minus Shared NDCG@10 advantage doubles as query rotation increases, while Shared dominates at small sample sizes. The Cross-fitted Asymmetry Risk Selector (CARS) reduces regret by 49-96% compared to fixed-geometry baselines.
Key points
- Dual projections realize arbitrary low-rank operators while shared projections induce positive-semidefinite operators, creating an exact approximation gap in scoring.
- The Cross-fitted Asymmetry Risk Selector (CARS) estimates reproducible directional signal from training pairs to select between shared and dual projections.
Context and limitations
Results are limited to multiple datasets and embedding models; specific performance metrics like NDCG@10 and regret reduction percentages are reported but not broken down by dataset type or model architecture.
Generated by Bittrees-hosted qwen35-2b sha256:aaf42c8b7c3c on 30/09/2026, 05:07:15 UTC. Based on publisher-provided feed evidence; may not cover the full original article. Check the source for details.
View archived briefing on IPFS
bafybeia6se3cjozcfn2ifyrqzfbqworyemujjn5t66cdxjqgc3stg3hefa