Research map / Architectures and efficiency / RNNs and normalization
Hierarchical Retrieval at Scale: Bridging Interpretability and Efficiency
Shubham Gupta, Zichao Li, Tianyi Chen, Cem Subakan, Siva Reddy, Perouz Taslakian, Valentina Zantedeschi
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 |
|---|---|---|---|
| use ReTreever instead of RAPTOR | tree-based retrieval · NQ | NDCG@10 (%): 40.28 -> 54.48 | Improved |
| use learned tree instead of Hier-KMeans | tree-based retrieval · NQ | NDCG@10 (%): 14.87 -> 54.48 | Improved |
| use cross-attention split instead of linear split (ablation) | ReTreever · NQ | NDCG@10 (%): 48.60 -> 54.58 | Improved |
| use MLP split instead of linear split (ablation) | ReTreever · VoxCeleb2 | NDCG@10 (%): 71.58 -> 39.21 | Hurt |
| add stochastic depth training for coarse levels (ablation) | ReTreever · NQ | NDCG@10 (%) fine level: 54.58 -> 54.48 | Hurt |
| add MRL training to encoder projection (ablation) | Encoder + Linear · NQ | NDCG@10 (%) fine level: 47.62 -> 46.60 | Hurt |
| increase tree depth from 10 to 12 (ablation) | ReTreever · VoxCeleb2 | NDCG@10 (%): 92.27 -> 89.60 | Hurt |
| increase tree depth from 6 to 12 (ablation) | ReTreever · NQ | NDCG@10 (%): 42.98 -> 52.30 | Improved |
| increase number of attention heads from 16 to 32 (ablation) | ReTreever · NQ | NDCG@10 (%): 54.58 -> 51.69 | Hurt |
| increase level embeddings from 1 to 8 (ablation) | ReTreever · ImageNet-1K | NDCG@10 (%): 78.23 -> 80.95 | Improved |
| use tree-based partitioning with HNSW inside leaves | ReTreever multi-index · NQ, HotpotQA, RepLiQA, TopiOCQA | NDCG@10 vs data access | Improved |
| use coarse tree-level representations | ReTreever-Stochastic · all datasets | inference cost / NDCG@10: 75% cost cut, 95% NDCG@10 retained | Improved |
Closest papers by meaning
- Hierarchical Retrieval: The Geometry and a Pretrain-Finetune Recipe (NeurIPS 2025 · similarity 0.88)
- RAPTOR: Recursive Abstractive Processing for Tree-Organized Retrieval (ICLR 2024 · similarity 0.87)
- SiReRAG: Indexing Similar and Related Information for Multihop Reasoning (ICLR 2025 · similarity 0.86)
- Adaptive Retrieval and Scalable Indexing for k-NN Search with Cross-Encoders (ICLR 2024 · similarity 0.85)
- MUVERA: Multi-Vector Retrieval via Fixed Dimensional Encoding (NeurIPS 2024 · similarity 0.85)
- Contrastive Retrieval Heads for Improved Attention-Based Reranking (NeurIPS 2026 · 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.
