Research map / Vision-language and multimodal / Multimodal LLMs
LLaVA-PruMerge: Adaptive Token Reduction for Efficient Large Multimodal Models
Yuzhang Shang, Mu Cai, Bingxin Xu, Yong Jae Lee, Yan Yan
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 token supplement (key-similarity merging) to adaptive token selection (ablation) | LLaVA-1.5 Vicuna-7B + AITS · MME | score: 1221.6 -> 1350.3 | Improved |
| add token supplement to adaptive token selection (ablation) | LLaVA-1.5 Vicuna-7B + AITS · ScienceQA-IMG | accuracy: 66.5 -> 68.5 | Improved |
| add token supplement to adaptive token selection (ablation) | LLaVA-1.5 Vicuna-7B + AITS · TextVQA | accuracy: 54.8 -> 56.0 | Improved |
| add token supplement to adaptive token selection (ablation) | LLaVA-1.5 Vicuna-7B + AITS · POPE | accuracy: 75.7 -> 76.3 | Improved |
| use adaptive outlier-based selection instead of sequential sampling (ablation) | LLaVA-PruMerge (40 tokens) · TextVQA | accuracy: 42.72 -> 54.00 | Improved |
| use adaptive selection instead of spatial sampling (ablation) | LLaVA-PruMerge (40 tokens) · MME | score: 1180.23 -> 1250.07 | Improved |
| use adaptive selection instead of sequential sampling (ablation) | LLaVA-PruMerge (35 tokens) · POPE | accuracy: 11.7 -> 76.2 | Improved |
| use adaptive selection instead of spatial sampling (ablation) | LLaVA-PruMerge (16 tokens) · ScienceQA-IMG | accuracy: 66.29 -> 68.07 | Improved |
| add LoRA fine-tuning with reduced tokens (ablation) | LLaVA-PruMerge · MME | score: 1250.1 -> 1350.3 | Improved |
| add LoRA fine-tuning with reduced tokens (ablation) | LLaVA-PruMerge · TextVQA | accuracy: 54.0 -> 56.0 | Improved |
| add LoRA fine-tuning with reduced tokens (ablation) | LLaVA-PruMerge · POPE | accuracy: 76.2 -> 76.3 | No clear effect |
| add spatially uniform extra tokens (PruMerge+ over PruMerge) | LLaVA-1.5 Vicuna-7B · POPE | accuracy: 76.3 -> 84.0 | Improved |
Closest papers by meaning
- Recoverable Compression: A Multimodal Vision Token Recovery Mechanism Guided by Text Information (AAAI 2025 · similarity 0.92)
- DivPrune: Diversity-based Visual Token Pruning for Large Multimodal Models (CVPR 2025 · similarity 0.92)
- Matryoshka Multimodal Models (ICLR 2025 · similarity 0.92)
- Improving Visual Token Reduction via Rectifying Distortions for Efficient Multimodal LLM Inference (ICML 2026 · similarity 0.92)
- Accelerating Multimodal Large Language Models via Dynamic Visual-Token Exit and the Empirical Findings (NeurIPS 2025 · similarity 0.92)
- Rethinking Visual Token Reduction in LVLMs Under Cross-Modal Misalignment (AAAI 2026 · 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.
