TL;DR
Quiet window (Dec 20, 2025–Jan 3, 2026). One material, retrieval‑relevant change landed: Milvus 2.5.24 (Dec 23) shipped stability fixes that touch hybrid (lexical + vector) search and sparse/BM25 stats handling. No other credible late‑interaction, learned‑sparse, filtered‑ANN, or ANN‑infrastructure releases/papers meeting the bar were found in this period.
Milvus 2.5.24: stability fixes for hybrid retrieval and sparse stats
- Key facts and current state of the topic
- Milvus is widely used as a first‑stage ANN engine in hybrid pipelines; correctness/latency often hinge on scalar filtering and lexical/sparse rescoring.
- Important context and background information
- Earlier 2.6.x releases added hybrid knobs (e.g., BM25 stats preload) and a Boost Ranker; 2.5.x remains deployed in many clusters where patch stability matters.
- Recent developments or changes
- On December 23, 2025, Milvus 2.5.24 released with security/stability updates and several fixes, including “improved startup fail‑fast,” hybrid‑search total count correction, JSON‑filter crash fixes, and “optimized loading performance of BM25 statistics for sparse vector search.” If you operate hybrid (vector + BM25/filters) on 2.5.x, plan an upgrade to 2.5.24 to reduce variance and avoid edge‑case failures. (github.com)
If you expected more: beyond this Milvus patch, our scan did not surface additional retrieval‑relevant releases or peer‑reviewed/preprint results between December 20, 2025 and January 3, 2026 that meet the inclusion criteria.