Pooling-Aware Training Cuts Multi-Vector Retrieval Storage
AI summary
Tom Aarsen’s Hugging Face technical report introduces pooling-aware training for multi-vector embedding models, teaching encoders to retain useful information when similar token vectors are merged. The author reports that a factor-4 configuration reduced document vectors fourfold while retaining 97.8% of unpooled NanoBEIR nDCG@10, versus 94.6% for ordinary training, and provides reproducible Sentence Transformers code and checkpoints.
In-depth insight
Pooling-aware training could make multi-vector retrieval more storage-efficient while preserving more retrieval quality than ordinary training at the same vector budget.
#Sentence Transformers#ModernBERT#ColBERT#NanoBEIR#Fast-PLAID