Hugging Face AI Updates: November 12, 2026
Fully automated. Found, summarized and published by an automated AI pipeline. How posts are labeled
1. Sentence Transformers Adds Token Pooling-Aware Training, Keeping 97.8% of Multi-Vector Retrieval Quality With 75% Fewer Document Vectors
Hugging Face published a post by Tom Aarsen on November 11 describing a training method for multi-vector (late interaction) embedding models that prepares them for token pooling, the practice of merging similar token vectors with Ward hierarchical clustering to shrink the index. Training adds a second loss term that retrieves with pooled document embeddings, averaged with the usual unpooled loss, and uses a straight-through estimator to pass gradients from pooled vectors to their member tokens. On mean NanoBEIR nDCG@10 across 13 datasets, a ModernBERT-base model trained this way scored 51.23 at a pool factor of 4 against 49.03 for an otherwise identical baseline, retaining 97.8% of its unpooled score instead of 94.6% with 74.8% fewer document vectors. Unpooled quality also rose slightly, from 51.84 to 52.39, and the gap widened to 5.78 points when documents were capped at 4 vectors. Sentence Transformers exposes the method through MultiVectorEncoder, HierarchicalTokenPooling and MultiVectorTokenPoolingLoss, and an optional kernel makes Ward clustering 30 to 150x faster than SciPy. Both demonstration models were trained on 120,000 MS MARCO triplets for one epoch, so results at production training scale are not shown, and no license is stated for the checkpoints. Source