Research map / Vision-language and multimodal / Multimodal LLMs
Why and When Visual Token Pruning Fails? A Study on Relevant Visual Information Shift in MLLMs Decoding
Jiwan Kim, Kibum Kim, Wonjoong Kim, Byung-Kwan Lee, Chanyoung Park
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 DSTP (detect RVIS, swap in reserved tokens) to FastV | Qwen3-VL-4B · visual reasoning avg, 33.3% tokens | relative accuracy: 55.68% -> 82.90% | Improved |
| add DSTP to DivPrune | InternVL3.5-8B · visual reasoning avg, 33.3% tokens | relative accuracy: 63.55% -> 90.09% | Improved |
| add DSTP to VisionZip | Qwen3-VL-4B · visual reasoning avg, 22.2% tokens | relative accuracy: 46.40% -> 68.74% | Improved |
| add DSTP to FastV | Qwen3-VL-4B · ScienceQA (33.3%) | accuracy: 87.98 -> 91.84 | Improved |
| add DSTP to VisionZip | Qwen3-VL-4B · VQAT (33.3%) | accuracy: 77.10 -> 77.20 | No clear effect |
| RISD fixed-threshold detection instead of random trigger (ablation) | DSTP · MathVerse | accuracy: 37.69 -> 52.54 | Improved |
| fixed threshold instead of running-average threshold (ablation) | DSTP · MathVerse | accuracy: 51.51 -> 52.54 | Improved |
| CPTS union swap instead of Hard selection (ablation) | DSTP · MathVerse | accuracy: 45.05 -> 52.54 | Improved |
| CPTS union swap instead of Merge (ablation) | DSTP · MMMU-Pro | accuracy: 28.38 -> 30.52 | Improved |
| CPTS instead of full visual tokens (ablation) | DSTP · MathVerse | accuracy: 53.04 -> 52.54 | No clear effect |
| add DSTP to FastV | Qwen3-VL-4B · MathVerse | latency per example: 119.35s -> 102.55s | Improved |
| add DSTP to FastV | InternVL3.5-8B · MathVerse | latency per example: 9.14s -> 9.99s | Hurt |
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.
- DivPrune — DivPrune: Diversity-based Visual Token Pruning for Large Multimodal Models (CVPR 2025) · better on 3 of 3 shared results
- VisionZip — VisionZip: Longer is Better but Not Necessary in Vision Language Models (CVPR 2025) · better on 3 of 3 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
- SwiftVLM: Efficient Vision-Language Model Inference via Cross-Layer Token Bypass (NeurIPS 2026 · similarity 0.91)
- What Kind of Visual Tokens Do We Need? Training-Free Visual Token Pruning for Multi-Modal Large Language Models from the Perspective of Graph (AAAI 2025 · similarity 0.91)
- When Token Pruning is Worse than Random: Understanding Visual Token Information in VLLMs (CVPR 2026 · similarity 0.91)
- Rethinking Visual Token Reduction in LVLMs Under Cross-Modal Misalignment (AAAI 2026 · similarity 0.91)
- IF-Prune: Information-Flow Guided Token Pruning for Efficient Vision-Language Models (CVPR 2026 · similarity 0.91)
- Beyond Attention or Similarity: Maximizing Conditional Diversity for Token Pruning in MLLMs (NeurIPS 2025 · similarity 0.91)
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.
