Research map / Vision-language and multimodal / Multimodal LLMs
AnchorPrune: Relevance-Anchored Contextual Expansion for Visual Token Pruning
Kyuan Oh, Bumsoo Kim
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 importance-weighted contextual expansion instead of diversity-only selection in Stage 2 (ablation) | AnchorPrune · LLaVA-1.5-7B, 64 tokens | Rel.: 85.4 -> 91.0 | Improved |
| use importance-weighted contextual expansion instead of the best alternative coupling at 32 tokens (ablation) | AnchorPrune · LLaVA-1.5-7B, 32 tokens | Rel.: 81.1 -> 85.1 | Improved |
| use additive relevance-diversity coupling instead of diversity-only (ablation) | AnchorPrune Stage 2 · LLaVA-1.5-7B, 64 tokens | Rel.: 85.4 -> 87.3 | Improved |
| use negated instead of raw CLIP patch-text similarity in Stage 1 (ablation) | AnchorPrune · LLaVA-1.5-7B, 32 tokens | MME: 1245.4 -> 1394.7 | Improved |
| use negated instead of raw CLIP similarity in Stage 1 (POPE) (ablation) | AnchorPrune · LLaVA-1.5-7B, 32 tokens | POPE: 74.2 -> 85.3 | Improved |
| use negated instead of raw CLIP similarity in Stage 1 (TextVQA) (ablation) | AnchorPrune · LLaVA-1.5-7B, 32 tokens | TextVQA: 53.7 -> 54.2 | No clear effect |
| use token-wise max matching instead of mean aggregation of instruction embeddings (ablation) | AnchorPrune · Qwen2.5-VL-7B, 256 tokens | MME: 1647.4 -> 1657.5 | Improved |
| restore retained tokens to raster-scan order instead of raw selection order (ablation) | AnchorPrune · LLaVA-1.5-7B, 64 tokens | MME-OCR: 132.5 -> 140.0 | Improved |
| restore raster-scan order (OCRBench) (ablation) | AnchorPrune · LLaVA-1.5-7B, 64 tokens | OCRBench: 26.9 -> 28.9 | Improved |
| set patience P=3 instead of P=1 (ablation) | AnchorPrune · LLaVA-1.5-7B, 32 tokens | MME: 1384.8 -> 1394.7 | Improved |
| set patience P=3 instead of P=5 (ablation) | AnchorPrune · Qwen2.5-VL-7B, 64 tokens | MME: 1552.2 -> 1563.2 | Improved |
| vary novelty threshold tau from 0.1 to 0.4 (ablation) | AnchorPrune · LLaVA-1.5-7B, 32 tokens | MME: 1388.9 to 1394.7 | No clear effect |
Earlier methods this paper compares with
Named baselines from earlier accepted papers in this paper's results tables: better than 2 of 2 on most shared results.
- VisionZip — VisionZip: Longer is Better but Not Necessary in Vision Language Models (CVPR 2025) · better on 5 of 5 shared results
- DivPrune — DivPrune: Diversity-based Visual Token Pruning for Large Multimodal Models (CVPR 2025) · better on 4 of 4 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
- QuietPrune: Query-Guided Early Token Pruning for Vision-Language Models (CVPR 2026 · similarity 0.91)
- TransPrune: Token Transition Pruning for Efficient Large Vision-Language Model (CVPR 2026 · similarity 0.90)
- Rethinking Visual Token Reduction in LVLMs Under Cross-Modal Misalignment (AAAI 2026 · similarity 0.89)
- TopV: Compatible Token Pruning with Inference Time Optimization for Fast and Low-Memory Multimodal Vision Language Model (CVPR 2025 · similarity 0.89)
- Beyond Attention or Similarity: Maximizing Conditional Diversity for Token Pruning in MLLMs (NeurIPS 2025 · similarity 0.89)
- Don't Just Chase “Highlighted Tokens” in MLLMs: Revisiting Visual Holistic Context Retention (NeurIPS 2025 · similarity 0.89)
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.
