MICE-P: new retrieval pipeline matches ColBERT throughput while keeping cross-encoder effectiveness
tomaarsen · x · 2026-10-27
- MICE architecture: lower layers encode queries and documents separately so document representations are pre-computable; upper layers let query tokens cross-attend to them, keeping richer interactions than ColBERT's MaxSim and near cross-encoder effectiveness.
- MICE-P pipeline: the first end-to-end MICE retrieval pipeline — Fast-PLAID indexes 128d compressed document vectors, SPLADEv3 recalls candidates, which are decompressed and scored by MICE.
- Results: on DBPedia latency vs decontaminated BEIR13 nDCG@10, MICE-P-61m/130m achieve late-interaction-level throughput while matching cross-encoder quality, beating LateOn + PLAID — a new efficiency-effectiveness trade-off class.
- Cost: bigger indexes — 4-bit residual compression (vs LateOn's 2-bit) plus a separate SPLADE index: 100.1 GB vs 51.8 GB on Climate-FEVER's 5M docs.
- Availability: Apache 2.0 checkpoints (61M/130M, Ettin/ModernBERT-based, 4 interaction layers) loadable via Sentence Transformers CrossEncoder; indexed pipeline uses the authors' open-source implementation.
More from Models
- Opus 5.5 keeps getting blocked as [cyber] when writing SASS code for GPUs — DimitrisPapail · 2026-10-27
- Prague lab BottleCapAI releases ThinkingCap models cutting thinking tokens up to 69% with negligible accuracy loss — JFPuget · 2026-10-27
- Recursive SI trained a model from scratch in 4 months using auto research — altryne · 2026-10-27
- New AI fingerprint in support tickets: "please do not upgrade our plan" sign-offs — glcst · 2026-10-27
- 36 blind AI judges pick a solo-session agent setup over Superpowers 36-0 — Kai_ThoughtArchitect · 2026-10-27
- New benchmark shows how often AI models hallucinate nonexistent skin conditions — Individual-War3274 · 2026-10-27