AI Paper Atlas

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

ICML 2026 · Architectures and efficiency › RNNs and normalization · approach: Learned embeddings for retrieval

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 changedWhereMeasuredEffect
use ReTreever instead of RAPTORtree-based retrieval · NQNDCG@10 (%): 40.28 -> 54.48Improved
use learned tree instead of Hier-KMeanstree-based retrieval · NQNDCG@10 (%): 14.87 -> 54.48Improved
use cross-attention split instead of linear split (ablation)ReTreever · NQNDCG@10 (%): 48.60 -> 54.58Improved
use MLP split instead of linear split (ablation)ReTreever · VoxCeleb2NDCG@10 (%): 71.58 -> 39.21Hurt
add stochastic depth training for coarse levels (ablation)ReTreever · NQNDCG@10 (%) fine level: 54.58 -> 54.48Hurt
add MRL training to encoder projection (ablation)Encoder + Linear · NQNDCG@10 (%) fine level: 47.62 -> 46.60Hurt
increase tree depth from 10 to 12 (ablation)ReTreever · VoxCeleb2NDCG@10 (%): 92.27 -> 89.60Hurt
increase tree depth from 6 to 12 (ablation)ReTreever · NQNDCG@10 (%): 42.98 -> 52.30Improved
increase number of attention heads from 16 to 32 (ablation)ReTreever · NQNDCG@10 (%): 54.58 -> 51.69Hurt
increase level embeddings from 1 to 8 (ablation)ReTreever · ImageNet-1KNDCG@10 (%): 78.23 -> 80.95Improved
use tree-based partitioning with HNSW inside leavesReTreever multi-index · NQ, HotpotQA, RepLiQA, TopiOCQANDCG@10 vs data accessImproved
use coarse tree-level representationsReTreever-Stochastic · all datasetsinference cost / NDCG@10: 75% cost cut, 95% NDCG@10 retainedImproved

Closest papers by meaning

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.

Open scime draft

Get the next scime Dispatch and a monthly field update

Free. We send a confirmation link first; every email has a one-click unsubscribe. Privacy