IBM paper expands BM25 with cross-encoder terms to close the retrieval gap
_reachsumit · x · 2026-08-04
IBM’s CE-QE paper tackles the vocabulary gap in lexical retrieval by expanding BM25 queries with terms that a cross-encoder identifies as semantically decisive.
- Instead of relying on BM25’s own top results, it seeds expansion from semantically retrieved passages, reducing self-reinforcing drift.
- Unlike generative query expansion methods such as HyDE or Query2doc, every added term is copied verbatim from retrieved text, so the method cannot hallucinate corpus vocabulary.
- The paper reports strong gains on BEIR, including NQ Recall@100 rising from 0.32 to 0.47, and says the SESF variant beats cross-encoder score fusion by 2.5% in recall.
Related event: IBM's Cross-Encoder Query Expansion Boosts Hybrid Search Recall(3 posts)→
More from Research
- NJU Introduces AVE-Compass: A Benchmark for Audio-Video Editing — NJU-LINK · 2026-08-06
- Deep Dive into Kimi K3 Architecture: 2.8T Parameters and LatentMoE Details — AxSaucedo · 2026-08-06
- NBER Paper: 19% of Workers Retroactively Edit Profiles, AI Skills Surge — steverathje2 · 2026-08-06
- Nature Publishes Expanded Codebook of Human Transcription Factor DNA-Binding Specificity — anshulkundaje · 2026-08-06
- Hundred-Page Language Models Book: build LLMs from scratch with PyTorch — burkov · 2026-08-06
- Dark Hundred-Page Language Models Book released, hands-on LLM guide — burkov · 2026-08-06