My paper titled “Beyond Hamming: Query-Aware Decoding of Binary Cosine Sketches” is accepted to ICML 2026. The paper is about angular distance estimator for vector DB like RAG.

This study identifies the near-background region as the key bottleneck for retrieval performance in 1-bit binary-sketch-based cosine estimation, and proposes QA-Cos, a new decoder that improves estimation by exploiting only query-side information while keeping the stored sketches unchanged. We further provide a theoretical formulation via query-aware maximum likelihood estimation and show through retrieval experiments that QA-Cos improves both candidate selection quality and retrieval efficiency.