Research map / Vision-language and multimodal / Multimodal LLMs
ToDRE: Effective Visual Token Pruning via Token Diversity and Task Relevance
Duo Li, Zuhao Yang, Xiaoqin Zhang, Ling Shao, Shijian Lu
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 relevance-driven Stage 2 on top of diversity Stage 1 (ablation) | ToDRE LLaVA-NeXT-7B · 4-benchmark avg, 10% retention | average relative performance (%): 95.8 -> 96.0 | Improved |
| add relevance-driven Stage 2 on top of diversity Stage 1 (ablation) | ToDRE LLaVA-NeXT-7B · 4-benchmark avg, 25% retention | average relative performance (%): 98.8 -> 98.9 | Improved |
| add relevance-driven Stage 2 on top of diversity Stage 1 (ablation) | ToDRE LLaVA-NeXT-7B · 4-benchmark total inference time, 10% retention | total time (min:sec): 31:18 -> 29:43 | Improved |
| apply Stage 2 alone to the unpruned model (ablation) | LLaVA-NeXT-7B · 4-benchmark avg | total time (min:sec): 77:04 -> 70:15 | Improved |
| select pivot by [CLS] attention rather than random (ablation) | ToDRE Stage 1 · 4-benchmark avg, 10% retention | average relative performance (%): 95.8 -> 96.0 | Improved |
| select pivot by [CLS] attention rather than center token (ablation) | ToDRE Stage 1 · 4-benchmark avg, 10% retention | average relative performance (%): 95.6 -> 96.0 | Improved |
| prune at layer 7L/8 instead of 5L/8 (ablation) | ToDRE Stage 2 · 4-benchmark avg | average relative performance (%): 96.2 -> 100.0 | Improved |
| prune at layer 7L/8 instead of 6L/8 (ablation) | ToDRE Stage 2 · 4-benchmark avg | average relative performance (%): 98.3 -> 100.0 | Improved |
| prune at layer 7L/8 instead of L/2 (ablation) | ToDRE Stage 2 · 4-benchmark avg | average relative performance (%): 87.9 -> 100.0 | Improved |
| set threshold tau=0.10 instead of 0.03 (ablation) | ToDRE Stage 2 · 4-benchmark avg | total time (min:sec): 79:22 -> 72:35 | Improved |
| vary threshold tau from 0.05 to 0.15 (ablation) | ToDRE Stage 2 · 4-benchmark avg | average relative performance (%): 100.0 / 100.1 / 100.0 | No clear effect |
| scale ToDRE to LLaVA-NeXT-13B vs DivPrune | ToDRE · 8 image benchmarks avg, 10% retention | average relative performance (%): 92.5 -> 93.6 | 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.
- 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
- TopV: Compatible Token Pruning with Inference Time Optimization for Fast and Low-Memory Multimodal Vision Language Model (CVPR 2025 · similarity 0.94)
- LearnPruner: Rethinking Attention-based Token Pruning in Vision Language Models (ICLR 2026 · similarity 0.94)
- SwiftVLM: Efficient Vision-Language Model Inference via Cross-Layer Token Bypass (NeurIPS 2026 · similarity 0.93)
- Balanced Token Pruning: Accelerating Vision Language Models Beyond Local Optimization (NeurIPS 2025 · similarity 0.93)
- QuietPrune: Query-Guided Early Token Pruning for Vision-Language Models (CVPR 2026 · similarity 0.93)
- Rethinking Visual Token Reduction in LVLMs Under Cross-Modal Misalignment (AAAI 2026 · similarity 0.93)
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.
