Token pooling-aware training proposed for multi-vector embeddings
The Hugging Face post title describes a training method; no method, models or results are in the available text.
Stories per day · 30 days
5 RSS
Relatedretrieval6gemma4language-models3multimodal3quantization3rag3alphaearth2contrastive-learning2dense-retrieval2differential-privacy2earth-observation2information-retrieval2mteb2nemotron2nvidia2open-weights2privacy2representation-learning2
70 stories, newest first · one card per story
Stories tagged #embeddings
Geometric RAG access control isolates policies but not embeddings
A curious user can recover original embeddings from a few hundred pairs and forge policy-compatible queries.
Vendi Anomaly Score flags outliers via dataset diversity
The score needs no hyperparameter tuning, scales linearly, and led 148 embedding benchmarks plus ImageNet tests.
Paper introduces e-means, a faster mini-batch k-means variant
Authors report 10-45x GPU speedups and a no-sweep rule for decay rate and batch size.
Saidzoda fine-tune lifts Belebele recall for four Central Asian languages
Belebele recall@1 rises 14.4 to 26.0 points, to 75.2%–77.7%, while English and Russian recall fall.
Sparse autoencoders decompose Earth embeddings into sparse features
On AlphaEarth, trained SAEs leave eight times less variance unexplained than a frozen random decoder.
Dense retrievers favor passages naming high-scoring answer objects
Across eight models, higher token-prediction scores predict higher retrieval ranks even when the relation is wrong.
onóna uses twin tied embeddings to cut validation loss
At matched token budgets it lowered loss in 11 of 12 runs and raised average accuracy in 10 of 12.
None of state-of-the-art molecular embedding models surpass fingerprints
Analysis shows between 225 million and 4 billion molecules in the Enamine REAL database have non-unique representations.
Box to embed Nvidia Nemotron in content retrieval
Enterprise Advance customers get it early next year; Nvidia expects faster queries and 90% fewer tokens.
Trainable [DUMP] prefix tokens cut interference in retrieval embeddings
The method leaves architecture, pooling, and contrastive loss unchanged and reports gains on retrieval and long-context benchmarks.
Near-zero embeddings stall gated model training
Untied gated checkpoints contain near-zero embeddings that amplify gradients 30+ orders of magnitude and trigger FP32 clipping, freezing rows in 20% of Telugu pretraining steps.
SD-ViT trains every ViT block as a usable face-recognition exit
Early-exit probes keep over 96% of full-depth accuracy at half the compute without rebuilding the gallery.
Survey outlines unification of omnimodal embeddings
Position paper traces convergence of embedding research across modalities toward joint embeddings for compositionality.
Embedding search finds 92% of trial evidence versus 45% for labels
On 7,354 randomized trials, a two-stage embedding-plus-language-model filter recovers far more systematic-review evidence, with cost rising by search depth.
AlphaEarth embeddings evaluated for crop classification across US and India
64-dimensional embeddings separate similar crops better than NDVI or harmonics, but transfer worse from the US to India.
EmbeddingGemma 2 unifies text, image, audio and video embeddings
The open model claims a new Pareto frontier while supporting quantization and selective encoder loading.
Small coding agents gain little from code graphs past one hop
On SWE-bench Verified, structured tools added 17–21 points at 1.7–4B but only 4–5 points at 32B.
Yalda Embedding released as a 595.8M Persian retrieval model
Adapted from Qwen3-Embedding-0.6B on more than one million pairs, it scores 0.5488 macro-average on eight FaMTEB tasks.
Simplex embeddings cut the multimodal modality gap by 97–99%
Softmax adapters and Total Variation similarity match or beat unconstrained retrieval on five benchmarks.
OlmoEarth v1.3 State-of-the-Art Embeddings Released
AllenAI releases OlmoEarth v1.3 with state-of-the-art embeddings covering the whole planet.
OPTIC compresses Gemma 4 E4B per-layer embeddings up to 1,002×
Authors report 97.8% of uncompressed performance at about 100× PLE compression, versus 66.3% for product quantization.
NVIDIA announces Nemotron 3.5 Embed, Rerank, and Parse models
The post body was unavailable, so sizes, benchmarks, licenses, and release status are not stated.
Perplexity releases pplx-embed-v2-late 0.6B and 9B models
Both models, distilled from an 18B teacher, share an embedding space so the 9B can index and the 0.6B can query.