Research map / Architectures and efficiency / Attention and transformers
HypRAG: Hyperbolic Dense Retrieval for Retrieval Augmented Generation
Hiren Madhu, Ngoc Bui, Ali Maatouk, Leandros Tassiulas, Smita Krishnaswamy, Menglin Yang, Sukanta Ganguly, Kiran Srinivasan, Rex Ying
Read the paperCheck your idea against 137,217 papers
AI-extracted. The measured effects below were read by AI from the paper's full text (no summary yet). Numbers and claims can be wrong or missing context, so check the paper before you cite or build on it.
Measured effects reported in the paper
| What was changed | Where | Measured | Effect |
|---|---|---|---|
| add OEM pooling instead of CLS pooling (ablation) | HyTE-FH · MTEB | mean over tasks: 49.33 -> 56.41 | Improved |
| add OEM pooling instead of Einstein midpoint (ablation) | HyTE-FH · MTEB | mean over tasks: 50.33 -> 56.41 | Improved |
| add OEM pooling instead of naive mean pooling (ablation) | HyTE-FH · MTEB | mean over tasks: 47.07 -> 56.41 | Improved |
| use fully hyperbolic model instead of Euclidean equivalent | EucBERT · MTEB | mean over tasks: 54.11 -> 56.41 | Improved |
| use fully hyperbolic model instead of Euclidean equivalent | EucBERT · RAGBench average | faithfulness: 0.596 -> 0.732 | Improved |
| use hybrid hyperbolic model initialized from EucBERT | EucBERT · RAGBench average | faithfulness: 0.596 -> 0.706 | Improved |
| use hyperbolic HyTE-FH instead of EucBERT with HNSW | EucBERT · MS MARCO ANN | Recall@50: 0.795 -> 0.859 | Improved |
| use hyperbolic HyTE-FH instead of EucBERT in ANN search | EucBERT · MS MARCO ANN | latency: 0.29 -> 0.09 ms | Improved |
| use hyperbolic HyTE-FH instead of EucBERT | EucBERT · MeSH hierarchy-sensitive retrieval | Recall@1: 0.810 -> 0.963 | Improved |
| use hyperbolic HyTE-FH instead of EucBERT | EucBERT · MeSH hierarchy-sensitive retrieval | SpecHR: 0.883 -> 0.970 | Improved |
| use HyTE-H(Qwen) instead of Qwen-Embedding | Qwen-Embedding · RAGBench average | faithfulness: 0.784 -> 0.803 | Improved |
| use hyperbolic fine-tuning on ModernBERT instead of Euclidean embedders on Emanual/ExpertQA | HyTE-H (bert) · RAGBench per-dataset | faithfulness | Mixed |
Closest papers by meaning
- Hierarchical Retrieval: The Geometry and a Pretrain-Finetune Recipe (NeurIPS 2025 · similarity 0.87)
- HELM: Hyperbolic Large Language Models via Mixture-of-Curvature Experts (NeurIPS 2025 · similarity 0.86)
- xRAG: Extreme Context Compression for Retrieval-augmented Generation with One Token (NeurIPS 2024 · similarity 0.86)
- RAPID: Long-Context Inference with Retrieval-Augmented Speculative Decoding (ICML 2025 · similarity 0.85)
- HiHPQ: Hierarchical Hyperbolic Product Quantization for Unsupervised Image Retrieval (AAAI 2024 · similarity 0.85)
- One Token Can Help! Learning Scalable and Pluggable Virtual Tokens for Retrieval-Augmented Large Language Models (AAAI 2025 · similarity 0.85)
Writing a paper in this area? scime draft is a free LaTeX editor with AI included, real-time co-authors and a check of your references and venue rules before you submit. It also imports OpenAI Prism projects.
