RRC: relevance-guided calibration squeezes more from dense retrieval embedding compression

_reachsumit · x · 2026-10-30

A new paper proposes RRC (Relevance-Guided Residual Calibration) to improve spectral compression of dense retrieval embeddings:

• Existing spectral compression (PCA, Whitening, SpecTemp) projects embeddings onto a smaller principal subspace to cut storage and scoring costs, but corpus-derived weighting rules don't account for each direction's contribution to relevance.
• RRC is a closed-form method: it estimates margin-based weights from relevance judgments along each retained direction, decomposes log weights into a fitted power-law trend plus a residual, and folds the residual into the compressor's shared transformation — no iterative training or extra inference stage.
• Across 3 encoders and 9 datasets (conventional and reasoning-intensive retrieval), SpecTemp + RRC achieves the highest mean effectiveness on BEIR and BRIGHT among tested compressors.

Original post →

More from Research

• AWS's EasyProver solves 669/672 Putnam problems, plus new harder benchmark Aegis — amazonwebservices · 2026-10-30
• Paper shows algebraic emergence in neural circuits: composed motifs yield group structure absent in primitives — neurovium · 2026-10-30
• "More Is Different" in neural circuits: compositionality as a lens for interpretability — neurovium · 2026-10-30
• Researchers credit tabpack for making large MLP grid experiments feasible — antoine_chaffin · 2026-10-30
• Lychee-Reranker: decoupled FBNL reranker up to 4B params, strong on BEIR/MIRACL — tomaarsen · 2026-10-30
• Sentence Transformers adds tunable compression: R2 rerankers keep 92% nDCG@10 at 128x — tomaarsen · 2026-10-30

Research channel →