AI Paper Atlas

Research map / Vision-language and multimodal / Human and embodied modeling

HieRD: Hierarchical Relational Distillation for Vision-Language Embedding Models

Vinh Le, Nguyen Hong Dang, Tu Vu, Linh Ngo Van, Duc Anh Nguyen, Trung Le

ICML 2026 · Vision-language and multimodal › Human and embodied modeling · approach: Scene graph generation

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
distill with HieRD instead of standard fine-tuning (FastVLM-0.5B)FastVLM-0.5B student · MMEB CLS averagePrecision@1: 64.93 -> 67.11Improved
distill with HieRD instead of standard fine-tuning (LLaVA-OneVision-0.5B)LLaVA-OneVision-0.5B student · MMEB CLS averagePrecision@1: 65.83 -> 70.08Improved
distill with HieRD instead of standard fine-tuning (LLaVA-OneVision-0.5B)LLaVA-OneVision-0.5B student · DocVQAPrecision@1: 38.43 -> 49.23Improved
distill with HieRD on FastVLM-0.5B VQAFastVLM-0.5B student · MMEB VQA benchmarksPrecision@1: avg 51.56 -> 52.76Mixed
add intra-modal structural alignment loss (ablation)HieRD · SUN397Precision@1: 66.2 -> 67.2Improved
add cross-modal structural alignment loss (ablation)HieRD · ImageNet-1K and VOC2007Precision@1: ImageNet 54.7 -> 56.0; VOC2007 78.9 -> 80.6Improved
add hidden representation alignment loss (ablation)HieRD · ImageNet-1KPrecision@1: 55.7 -> 56.0Improved
use both intra-modal and cross-modal losses versus removing both (ablation)HieRD · ImageNet-1KPrecision@1: 54.2 -> 56.0Improved
use attention-weighted group aggregation instead of mean pooling (ablation)HieRD · SUN397Precision@1: 63.7 -> 67.2Improved
use attention-weighted group aggregation instead of mean pooling (ablation)HieRD · ImageNet-1KPrecision@1: 53.7 -> 56.0Improved
change DBSCAN min_samples from 8 to 4 (ablation)HieRD · MMEB CLS overallPrecision@1: 67.16 -> 66.12Hurt
change DBSCAN min_samples from 8 to 12 (ablation)HieRD · MMEB CLS overallPrecision@1: 67.16 -> 66.58Hurt

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