A new arXiv paper introduces a Rater Ising-Potts model, an extension of the Ising model designed for multinomial rating data. The approach builds pairwise agreement indicators and category labels into the model, drawing its weights from large language model embeddings. The authors position the work as a link between network psychometrics and AI methods.
A new arXiv paper introduces a method that grounds chain-of-thought reasoning in retrieved evidence to improve universal multimodal embeddings, which aim to represent text, images and other modalities in one shared space. The authors argue that reasoning steps should be tied to retrieval so that only relevant information shapes the final embedding. The work targets a single model that can handle a range of cross-modal retrieval tasks.
A new arXiv preprint describes a method for corporate and financial-disclosure analysis that avoids training and cross-model alignment steps. The authors argue that dense text embeddings and large language models struggle with context limits, hallucination risk, compute cost, and inconsistent vector spaces across independently trained models. The work is demonstrated on SEC filings, with a revised version now posted.
A new arXiv paper proposes treating the geometry of document embeddings as a measurable signal of how scientific concepts reorganize over time. The authors add a counterfactual ablation method that estimates how much any single concept contributes to the overall structure of a body of knowledge. The work aims to give scientometrics a quantitative tool for spotting shifts that resemble scientific revolutions.
Researchers propose a scalable method for partial information decomposition (PID) that uses supervised embeddings to analyze mental-health symptom networks. Standard pairwise edge weights cannot capture redundant information shared about a third symptom or synergistic information that only emerges from combinations. The approach aims to make PID practical for larger symptom networks.
A new arXiv preprint proposes embedding temporal knowledge graphs in biquaternionic space, arguing that relying on a single geometric space limits how well models capture varied relational patterns. The method pairs that representation with complex-valued attention to score facts whose validity changes over time. The work targets the link-prediction task of inferring missing facts in evolving knowledge graphs.
A new arXiv preprint proposes MoME, a technique that combines conditional memory — token-indexed embedding tables that give a model cheap parametric lookups — with sparse capacity ideas inspired by Mixture-of-Experts. The method aims to make those lookups context-aware by routing them selectively, as part of broader efforts to scale language models more efficiently.
A new arXiv preprint examines how speaker recognition networks organise the latent representations they learn from speech utterances. The work targets the limited transparency of these models by analysing whether speaker embeddings are arranged hierarchically. The authors position the study as a step toward explaining the internal mechanisms behind speaker identification systems.
Researchers propose PACE, a training method for multimodal embedding models that shifts the contrastive objective from an angular (cosine-based) formulation toward a norm-based one over the course of training. The approach is presented as an alternative to standard cosine contrastive objectives, which the authors say offer stable but potentially limited training dynamics. The work is posted as an arXiv preprint in the cs.AI and cs.LG categories.
A revised arXiv paper addresses a known weakness in dense retrieval: compressing a passage into a single embedding tends to weight early text more heavily, so retrieval quality drops when the relevant span sits later in the passage. The authors propose attention calibration as a way to reduce this positional bias, building on earlier inference-time approaches. It is a research contribution rather than a released product.
A new arXiv paper examines when inexpensive hashing-based embeddings can match pre-trained protein language models like ESM-2 on biological sequence classification. The authors argue that PLMs are costly to embed and fine-tune because they need GPUs, and offer a theoretically grounded hashing approach as a lighter substitute. The work is a revised preprint and is not yet peer-reviewed.
A new arXiv paper introduces SALT, a technique for strengthening how individual tokens are represented within multilingual sentence encoders. These encoders are optimized to align whole sentences across many languages, supporting applications like translation mining and zero-shot learning for low-resource languages. The paper targets the weaker token-level alignment that results from this sentence-focused training.
A cs.CL paper on arXiv examines how dense retrieval models build question representations incrementally by encoding 2,144 held-out headlines from the Thu Vien Phap Luat Vietnamese legal library one word at a time. The authors track these embedding trajectories using Nemotron-3-Embed (8B/1B) and Qwen3-Embedding (8B/0.6B) models, offering insight into how vector representations evolve as words arrive. The v2 listing replaces the earlier version of the paper.
Researchers have published an extended version of DGCPath, a self-supervised framework that learns numerical representations of travel paths from vehicle trajectory data. The approach combines generative and contrastive objectives while explicitly modeling the distribution of trajectory data. It targets applications in intelligent transportation systems, where analyzing routes at scale remains a core challenge.
A new arXiv study investigates why low-bit quantization delivers high recall on some embedding types while failing sharply on others, a difference that average distortion and global rank correlation cannot explain. The authors analyze quantized vector search at the level of individual comparisons to identify when compressed indexes still reproduce full-precision retrieval decisions.
A newly posted arXiv paper introduces multi-functional embedding models that link messy funder name strings in publication records to consistent identities. The work addresses the problem that inconsistent funder naming makes it hard to study at scale how research money is distributed across disciplines, institutions, and regions. Cleaner funder data could improve large-scale analyses of research funding patterns.
Researchers have introduced SEA-LION-Embedding, a set of text embedding models built for Southeast Asian languages and released with open, documented training resources. The work addresses a persistent gap in the field, where leading embedding models cannot be independently reproduced because their training corpora remain private. The release aims to support reliable performance on downstream tasks across the region's many languages.
A new arXiv paper proposes estimating urban livability by applying foundation model embeddings to high-resolution geospatial data. The approach targets areas where socioeconomic indicators are hard to measure, aiming to inform policy interventions and resource allocation. It explores whether pretrained model representations can capture neighborhood-level quality-of-life signals in data-scarce regions.
A new arXiv machine learning paper examines whether odor-quality descriptors, typically encoded as high-dimensional profiles, can be arranged in a simple two-dimensional space. The researchers test a hyperbolic embedding in the Poincaré disk to see if it yields an interpretable geometric organization of smell-related terms. The work reflects growing interest in applying non-Euclidean geometry to sensory and linguistic data.
A new arXiv paper tackles the mismatch between high-dimensional text embeddings and the limited input capacity of near-term quantum circuits. The authors experiment with compressing semantic representations so they can feed into hybrid quantum-classical classifiers for NLP tasks. The work offers an empirical look at how much representation reduction is viable for quantum machine learning applications.
NeoMME is a newly introduced encoder model designed to natively process multiple input modalities while supporting a broad range of languages. Its architecture emphasizes efficiency, aiming to deliver competitive performance at lower computational cost for workloads such as embeddings and retrieval. The release is detailed on the Hugging Face blog, making the model accessible to developers building multimodal, multilingual applications.
WHY IT MATTERS ↘For teams running embedding and retrieval pipelines, a single efficient encoder covering many languages and modalities could reduce inference costs and simplify architectures that otherwise chain separate per-modality or per-language models. It also signals growing competition in a segment long dominated by English-centric encoders, potentially lowering the barrier to shipping multilingual, multimodal search applications.
A Hugging Face blog post walks through training and finetuning multi-vector embedding models using the Sentence Transformers library. It covers the practical workflow for building models that represent text as multiple vectors rather than a single embedding. The write-up is aimed at developers who want to apply these techniques to their own retrieval or search tasks.
WHY IT MATTERS ↘Multi-vector retrieval models typically deliver meaningfully better recall than single-embedding approaches, but their higher storage and latency costs have kept adoption limited to teams with in-house IR expertise. A practical, library-level guide lowers that barrier, which pushes more teams toward late-interaction retrieval and raises the pressure on vector database and search vendors to handle multi-vector indexes economically.