Research map / Vision-language and multimodal / Human and embodied modeling
LightSTAR: Efficient Visual Document Retrieval via Lightweight Selection with Vision-Adaptive Refinement
Tongkun Guan, Haocheng Wang, Wei Shen, Xiaokang Yang
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 scale-adaptive top-k late interaction in place of max-pooling (ablation) | LLM-free Visual Selection · ViDoRe average | Recall@100: 96.8 -> 97.8 | Improved |
| add content-grounded query embedding (function-word filtering) (ablation) | LLM-free Visual Selection · ViDoRe average | Recall@100: 97.0 -> 97.8 | Improved |
| add adaptive region-wise feature fusion with layout-aware features (ablation) | Vision-adaptive Semantic Refinement · ViDoRe average | NDCG@5: 88.3 -> 89.1 | Improved |
| add hardness-aware objective (ablation) | Vision-adaptive Semantic Refinement · ViDoRe average | NDCG@5: 87.9 -> 89.1 | Improved |
| add refinement stage to selection-only pipeline (ablation) | LightSTAR · ViDoRe average | NDCG@5: 80.8 -> 89.1 | Improved |
| replace full-corpus refinement with selection then refinement (ablation) | LightSTAR · ViDoRe average | NDCG@5: 89.3 -> 89.1 | No clear effect |
| replace full-corpus refinement with selection then refinement (ablation) | LightSTAR · ViDoRe latency (5000 pages) | latency (s): 465.1 -> 123.9 | Improved |
| use unified LightSTAR selection instead of SigLIP first stage before ColQwen2.5 | cascade pipeline · ViDoRe average | NDCG@5: 81.3 -> 89.1 | Improved |
| use unified LightSTAR selection instead of CLIP first stage before ColQwen2.5 | cascade pipeline · ViDoRe average | NDCG@5: 73.3 -> 89.1 | Improved |
| use LightSTAR instead of ColQwen2.5 | retriever · 7,000-page corpus | end-to-end latency: 4.1x faster (656.7/161.1) | Improved |
| use LightSTAR instead of VisRAG-Ret | retriever · 7,000-page corpus | end-to-end latency: 10.4x faster (1678.9/161.1) | Improved |
| use LightSTAR instead of ColPali | retriever · 7,000-page corpus | end-to-end latency: 2.3x faster (362.6/161.1) | Improved |
Earlier methods this paper compares with
Named baselines from earlier accepted papers in this paper's results tables: better than 1 of 1 on most shared results.
- ColPali — ColPali: Efficient Document Retrieval with Vision Language Models (ICLR 2025) · better on 2 of 2 shared results
Matched by method name in results tables read by AI. Authors' own tables tend to favour the newer method, and a comparison can be missed when a method is renamed.
Closest papers by meaning
- ColPali: Efficient Document Retrieval with Vision Language Models (ICLR 2025 · similarity 0.91)
- Document Haystacks: Vision-Language Reasoning Over Piles of 1000+ Documents (CVPR 2025 · similarity 0.87)
- RobustVisRAG: Causality-Aware Vision-Based Retrieval-Augmented Generation under Visual Degradations (CVPR 2026 · similarity 0.87)
- Visual Abstraction: A Plug-and-Play Approach for Text-Visual Retrieval (ICML 2025 · similarity 0.87)
- Towards Natural Language-Based Document Image Retrieval: New Dataset and Benchmark (CVPR 2025 · similarity 0.87)
- MSTAR: Box-free Multi-query Scene Text Retrieval with Attention Recycling (NeurIPS 2025 · similarity 0.87)
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.
