// AI news · 2026-11-11 · Hugging Face Blog
Pooling-Aware Training for Multi-Vector Embeddings
Hugging Face introduces pooling-aware training for multi-vector embedding models to preserve retrieval performance during vector compression.
Published 2026-11-11 08:00 (Beijing time)Source Hugging Face BlogAuto-summarised, fact-checked against source
01Summary
Using HierarchicalTokenPooling with Ward clustering, the method reduces document vectors by up to four times while maintaining or improving nDCG@10 scores. Training adapts to compressed representations, achieving 97.8% retention vs. 94.6% in standard models under same storage budget.
Summarised from the original article by an open-source model, then checked automatically: every number and English name must appear in the source. The original article is authoritative.
02Key points
- Improves retrieval quality during vector pooling with 97.8% nDCG@10 retention.
- Reduces document vectors by 4x while preserving semantic fidelity.
- Uses Ward clustering to group similar token embeddings hierarchically.
- Supports adaptive threshold calibration based on corpus characteristics.
03Source
Token Pooling-Aware Training for Multi-Vector Embedding Models
Hugging Face Blog · published 2026-11-11 08:00 Beijing time
05Related news
- Same source · Hugging Face Releases Fine-Tuning Guide for TRELLIS.22026-11-12 · Hugging Face Blog
- Same source · Link Hugging Face to Google Cloud Account2026-11-10 · Hugging Face Blog
- Same topic · Google launches A2UI 1.0 for generative UI protocols2026-11-11 · Google AI Changelog
- Same topic · Gigabyte Launches AI TOP AERO Mini PC2026-11-11 · IT之家
- Same day · Google releases Gemini Robotics 2.5 with 96% success rate in production tasks2026-11-11 · Korea AI Times
From AIMap AI news: collected from 1,100+ official and media sources. Summaries only restate the original article; follow the link for the full story. All stories of 2026-11-11 · Archive by date · All AI news · Today