embedding

Ternlight is a 7 MB embedding model that runs in the browser via WASM, enabling fast semantic search without any API calls or servers. It ships as an npm package and can be used in three lines of code.

BaryGraph introduces a novel knowledge graph where every relationship is a first-class embedded document (BaryEdge) rather than an edge between nodes, enabling recursive abstraction triads that surface structural bridges between distant concepts. The preprint includes benchmarks showing structural metrics correlate with human similarity judgments better than cosine similarity alone.

The paper introduces Ember, a lightweight optimizer for embedding and LM-head matrices that exploits gradient geometry to improve efficiency and performance across supervised finetuning, RL, and pretraining, while using far less optimizer state than Adam.

Gbrain's dream cycle is a 24/7 automated loop that ingests daily data, enriches entities, consolidates conversations, merges duplicates, fixes citations, and ranks pages to maintain a fresh and accurate knowledge base.

This paper introduces and empirically evaluates methods for measuring semantic similarity between knowledge graphs using KG embeddings, proposing EmbPairSim and AvgEmbSim scoring functions that outperform baselines like Sentence-BERT on WikiText-2 and CC-News datasets.

YouTube founder Chad Hurley recounts how the ability to embed videos for free, inspired by PayPal's payment button strategy, drove YouTube's explosive growth. The article highlights the 'come for the tool, stay for the network' strategy advocated by a16z partner Chris Dixon.

DREAM trains dense retrieval embeddings by using autoregressive language model attention to supervise query-document similarity, eliminating the need for labeled data. It consistently outperforms baselines on BEIR and RTEB benchmarks across model scales.

HAKARI-Bench is a lightweight benchmark for comparing retrieval methods across multiple configurations and languages, enabling efficient model selection and performance analysis. It reproduces full benchmarks like MTEB at high correlation while being faster to run.

EvoEmbedding is a dynamic embedding model that maintains a continuously updated latent memory to generate adaptive representations for long-context retrieval, outperforming larger specialist models and improving agentic workflows.

Liquid AI introduces LFM2.5-Embedding-350M and LFM2.5-ColBERT-350M, two multilingual retrieval models optimized for fast and accurate search across 11 languages, with latency as low as 1.5ms.

This paper introduces CADE, a framework for time-series question answering that maps each timestep directly into the LLM embedding space and uses a one-directional supervised contrastive loss to align time-series representations with frozen text anchors, outperforming existing baselines on the Time-MQA benchmark.

A Chinese science tweet that intuitively explains the core chain of LLMs (Large Language Models): from token, embedding, position encoding, attention, FFN to residual stream and next-token prediction, helping readers without a math background understand AI papers.

This paper applies the likelihood ratio framework for forensic authorship attribution to Japanese texts, fusing stylometric features with embedding-based systems to improve discrimination and calibration.

This blog post explains the connection between JEPA (Joint Embedding Predictive Architecture) models and Canonical Correlation Analysis (CCA), a statistical method from 1936, arguing that CCA is the conceptual precursor to JEPA and that the idea of maximizing correlation in embedding space dates back to Hotelling.

An open-source project called turbovec has reached 10K stars on GitHub. It is a Rust-based vector index with Python bindings that uses Google Research's TurboQuant algorithm to compress embeddings to near the theoretical Shannon limit, enabling fully local RAG with 10 million documents fitting in 4 GB RAM and searching faster than FAISS.

This paper investigates how lexical overlap, rather than semantic content, influences LLM representations across layers and architectures, and demonstrates that this lexical effect persists even in models trained for semantic similarity, leading to degraded performance on downstream tasks.

This paper proposes set-distance rewards for reinforcement learning in chest X-ray report generation, using embedding-based set-to-set distances between generated and reference reports. Post-training with these rewards via GRPO consistently outperforms supervised fine-tuning and exact-match rewards, and enables efficient test-time scaling.

ScaleMAP is a new nonlinear dimensionality reduction method that preserves local density and neighborhood structure by rescaling embedding distances based on original-space local radii, achieving better density preservation than DensMAP while maintaining UMAP-level neighborhood preservation.

This paper proposes a method for cross-table retrieval and alignment of heterogeneous numerical tabular datasets using statistical descriptors and sentence embeddings, enabling similarity matching and interpretable variable-level correspondence without shared column names.

This paper proposes a model-based approach to assess massively multilingual parallel data by decomposing it into parallelism assessment and reference-free quality estimation, finding that no single universal metric works across all language directions.