Tag
We present Neurograf v1, an experimental end-to-end pipeline for training Romanian-first language models from scratch, including a 635 GB corpus, an 80k byte-level BPE tokenizer, five decoder-only models (160M–6B params), instruction-tuned assistants, and embedding models. Neurograf-6B achieves 69.89% on the Romanian WWTBM benchmark, demonstrating the feasibility of low-resource European language model development within a unified pipeline.
The author argues that EmbeddingGemma 2 and embedding models in general are being underestimated, highlighting the open-source DigUp project — a Mac search app built on Google's EmbeddingGemma 2 that outperforms Spotlight with content, image, and audio search — and suggests embeddings could also improve RAG knowledge base search.
This paper defines a hierarchy of faithfulness criteria for knowledge base completion models and evaluates current embedding models, finding they are not logically faithful.
QontoFAQ introduces a new benchmark and metric for information retrieval, focusing on improving relevance measurement for answering product questions, with associated code and dataset released.
This paper presents a framework for funder name disambiguation in scientific publications using multi-functional embedding models trained with multi-task learning, achieving over 0.90 accuracy and outperforming general-purpose LLMs like GPT-5.2.
The article presents a method to migrate between embedding models without re-embedding the entire corpus by reranking a subset of documents, achieving similar retrieval quality, and introduces embedflow, a tool available on PyPI and GitHub for this purpose.
The author introduces embedflow, a method to migrate between embedding models without reindexing, saving significant time and resources. It's an open-source tool available on GitHub for easy integration into workflows.
This paper compares chunking and embedding strategies for RAG systems in Turkish, evaluating their impact on document question answering through a systematic study with multiple configurations.
Bonsai discovered a pervasive issue where embedding vectors are unnecessarily cast from float32 to float64, causing 'Float Bloat' that wastes storage and bandwidth, with an estimated global impact of over 20 Petabytes of overhead.
This blog post introduces how to train and finetune multi-vector embedding models using the Sentence Transformers library, showcasing its v6.0 update with a new MultiVectorEncoder type and demonstrating superior performance on medical retrieval tasks.
This article explains embedding models and their role in LLM inference, covering architectures, traffic profiles, and optimization techniques like quantization.
A discussion on the cost-effectiveness of embeddings versus LLMs, referencing a study called 'embedder's dilemma' that finds LLMs outperform embedding models at significantly higher cost.
Two AI models, LFM2.5-2.6B and LFM2.5-Embedding-350M, are now available for free on the OpenRouter platform.
Sentence Transformers v6.0 introduces MultiVectorEncoder for ColBERT-style late interaction retrieval, enhancing support for multi-vector embedding models with a familiar API.
The paper compares large language models and embedding models across 37 tasks, finding that while aggregate performance is similar, embedding models are far cheaper and faster, supporting a division of labor for cost-efficiency.
The authors propose Synthetic Query Probing, a simple yet effective approach to compare different embedding models by comparing similarity match scores across models instead of raw embedding spaces. The paper illustrates non-linear relationships between models like Ada and Titan, offering practical guidance for swapping embedding models and setting retrieval thresholds.
Introduces Synthetic Query Probing (SQP), a reference-free method for learning mappings between similarity score distributions across embedding models, enabling threshold portability in RAG systems. Experiments show systematic score distortions and isotonic regression as the best calibration approach.
Proposes RT, a test-time inference method that enhances semantic reasoning in text embedding models by running multiple forward passes, improving performance on reasoning tasks while maintaining general semantic understanding.
Analysis of 2.6 billion sketches from 236 countries reveals hidden cultural variation in how common concepts are visually represented, showing that visual imagery preserves rich semantic and cultural structure that language models compress.
This paper introduces a hierarchical framework for multimodal sexism detection in memes that models annotator disagreement using frozen Gemini Embedding 2 representations and a lightweight gated MLP, achieving 1st place on fine-grained sexism categorization at EXIST 2026.