embeddings

This paper introduces the World Embedding Benchmark, with 8,000 controlled simulation cases spanning fluid/solid mechanics, dynamics, and optics, to evaluate how video representations encode physical information. It finds pretrained multimodal embeddings struggle with physical alignment, probes recover physics from frozen embeddings, contrastive training trades off alignment vs. recoverability, and retrieved reference videos improve physical fidelity of MiniMax-H3-generated videos.

The author demonstrates how frontier models like Opus 5.5 can be used for historical research, using embedding-based semantic search over the GLOBALISE Dutch East India Company archive to surface a previously unnoticed 1615 Dutch eyewitness account of hunting dodos, alongside Carter Church's use of GPT-6 Astra to crack a Napoleonic-era cipher.

The paper proposes Doc2LoRA, a hypernetwork that maps scientific papers to LoRA adapters so that every embedding point — including mixtures of papers — is an interrogable LLM, enabling both interpretable generation of new ideas and competitive search performance on APS papers.

This paper explores graph neural network encoders for projecting heterogeneously sampled signal sets from sensor networks and radars into fixed-size, sampling-invariant embedding spaces, evaluated via waveform discrimination on synthetic complex-valued RF signals.

Cohere released Embed 5, its most powerful embedding family yet, with pro and fast variants sharing an embedding space, multimodal support, 100+ language coverage, 128k context, and matryoshka embeddings across multiple dimensions.

Researchers have developed a new RAG approach called PageIndex that bypasses traditional dependencies like vector databases and embeddings, potentially disrupting the RAG industry.

This paper proposes Clinical Guardrail Probes (CG-Probes) to measure risks in patient-facing AI assistants for oncology by recovering guardrail directions from query embeddings using linear probing methods.

REALMS is a conversational AI system for real-time exact audience sizing using LLMs and embeddings, enabling marketers to query high-dimensional profile data with natural language and receive precise counts quickly in production.

The article demonstrates that using embedding models for emoji search is faster, cheaper, and comparable in quality to Jev, with benchmarks and a browser-based demo.

This article critiques Retrieval-Augmented Generation (RAG) systems, demonstrating through experiments that embeddings fail to capture contextual details like contradictions, leading to hallucinations, and emphasizes the essential role of strong underlying models for accurate AI responses.

The paper proposes a zero-inference prospective term for personal memory retrieval that boosts memory items linked to future commitments without query-time computation. It shows improved recall on a synthetic task set and positions this as part of a layered architecture for proactive AI assistants.

The article introduces Contrastive-LM (CLM), a new language model trained with contrastive learning for connecting states and actions, offering significant improvements in latency and performance on computer-use, gaming, and tool-calling tasks.

The paper introduces Ovis-Embedding, a state-of-the-art omni-modal embedding model that uses a shared backbone to encode text, image, video, and audio in a common representation space, achieving top performance on benchmarks like MMEB-v3 and MVEB.

The tweet notes that Sentence Transformers models for search have gained popularity on Hugging Face, and expresses hope for a resurgence of encoders with zero-shot capabilities.

MongoDB Atlas introduces auto-embedding, enabling databases to generate and manage embeddings internally for semantic search, eliminating the need for external services and improving data synchronization.

Memorable introduces a method to optimize memory in AI agents using embeddings instead of tokens, enabling procedural memory that persists across multiple runs.

Manticore Search introduces automatic document chunking for vector search, improving recall for long documents by splitting them into chunks and embedding each chunk.

This paper develops a vector logic for formal semantics, characterizing when compressed vector representations allow exact linear or affine readouts of truth conditions, with experiments on GloVe and word2vec embeddings.

Antfly has rewritten their search-and-inference database from Go to pure Zig for improved performance and zero dependencies. The article details their technical decisions, first principles, and implementations of vector search algorithms.

The paper introduces ePID, a scalable pipeline using supervised embeddings to compute partial information decomposition for symptom networks, enabling the separation of redundant and synergistic information in mental health data.