NV-Retriever: Improving text embedding models with effective hard-negative mining

By Gabriel de Souza Pereira Moreira, Radek Osmulski, Mengyao Xu, Ronay Ak, Benedikt Schifferer, Even Oldridge. 2024

Subfields: Information Retrieval & Search, Representation & Self-Supervised Learning

Citations in the corpus

• Training Sparse Mixture Of Experts Text Embedding Models (2025, unknown): “Recent work has shown that using stronger teacher models for mining yields higher quality finetuning datasets (Moreira et al., 2024; Yu et al., 2024).”
• CoRNStack: High-Quality Contrastive Data for Better Code Retrieval and Reranking (2024, unknown): “To eliminate false negatives, we follow Moreira et al. (2024) to remove any c − j which is sufficiently close to t i .”
• NEAR: Neighborhood-Aware Residuals for Memory Retrieval (2026, unknown): “We evaluate nine frozen embedding models (Zhang et al., 2025; Vera et al., 2025; Zhao et al., 2026a; Nussbaum et al., 2025; Moreira et al., 2024; Babakhin et al., 2025).”
• RePlay: Retrieval-Based Voice Playback for Multi-Turn spoken dialogue (2026, unknown): “C ( i ) contains i + and hard negatives mined by embedding similarity, excluding near-paraphrases of the target [26]. y ij =+1 if j = i + and − 1 otherwise, τ is a learned inverse temperature and b a learned bias, and λ weights the alignment term.”
• MLAIRE: Multilingual Language-Aware Information Retrieval Evaluation Protocal (2026, unknown): “…ranging from widely adopted encoder-only families such as multilingual-e5 [32], bge-m3 [33], gte [34], snowflake-arctic [35], nomic-embed [36], embeddinggemma [37] and jina [38, 39], to recent LLM-based embedding models including Qwen3-Embedding [40], llama-nemotron [41], and pplx-embed [42].”
• Fantastic (small) Retrievers and How to Train Them: mxbai-edge-colbert-v0 Tech Report (2025, unknown): “We follow NV-Embedv2 [26] in our mining process, where we used Qwen3-Embedding-8B to mine hard negatives and set the threshold to 0.95.”
• SitEmb-v1.5: Improved Context-Aware Dense Retrieval for Semantic Association and Long Story Comprehension (2025, unknown): “…in embedding models, either by designing efficient bidirectional models (Chen et al., 2024; Sturua et al., 2024; Nussbaum et al., 2024), or by repurposing powerful unidirectional pre-trained LLMs as embedding generators (Li et al., 2023; Wang et al., 2024a; Moreira et al., 2024; Kim et al., 2024).”
• Redundancy, Isotropy, and Intrinsic Dimensionality of Prompt-based Text Embeddings (2025, unknown): “Notably, one common modification is the use of bidirectional attention (BehnamGhader et al., 2024; Lee et al., 2024a; de Souza P. Moreira et al., 2024).”
• mFollowIR: a Multilingual Benchmark for Instruction Following in Retrieval (2025, unknown): “Some of the earliest work on the topic of training with instructions prepended a dataset prefix that defined relevance [5, 47] which has continued with more recent models [34, 46].”
• LLMs Are Also Effective Embedding Models: An In-depth Overview (2024, unknown): “First, we explore how LLMs can be directly prompted to produce embeddings, with a focus on prompt design and embedding derivation strategies that require no parameter updates.”
• Nemotron ColEmbed V2: Top-Performing Late Interaction embedding models for Visual Document Retrieval (2026, unknown): “Such global representation has been shown to significantly enhance retrieval accuracy for LLM-based embedding models [39, 2, 40].”
• CodeXEmbed: A Generalist Embedding Model Family for Multiligual and Multi-task Code Retrieval (2024, unknown): “For general retrieval models, we evaluated E5-Base (Wang et al., 2022), GTE-Base (Li et al., 2023b), BGE-Base (Xiao et al., 2023), Con-triever (Izacard et al., 2023), E5-Mistral (Wang et al., 2023), BGE-M3 (Chen et al., 2024), NV-Embed-V2 (Moreira et al., 2024), SFR-V2 (Meng* et al., 2024) and OpenA”
• DenseOn with the LateOn: Fully Open Dense and Late-Interaction Models for Multilingual, Long-Context, and Code Search (2026, unknown): “For each sample, we retain 10 negatives after filtering with NV-Retriever at a threshold of 0.95, from which we sample 7 random negatives at each step.”