Research map / Vision-language and multimodal / Multimodal LLMs
AIM: Adaptive Inference of Multi-Modal LLMs via Token Merging and Pruning
Yiwu Zhong, Zhuoming Liu, Yin Li, Liwei Wang
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 |
|---|---|---|---|
| merge tokens within frames (spatial) rather than across frames (ablation) | AIM token merging · VideoMME (25% retention) | accuracy: 55.8 -> 58.0 | Improved |
| merge tokens within frames (spatial) rather than across frames (ablation) | AIM token merging · VideoMME (3.1% retention) | accuracy: 47.4 -> 52.3 | Improved |
| do not prune text tokens during LLM pruning (ablation) | AIM · VideoMME | accuracy: 45.7 -> 58.2 | Improved |
| prune all visual tokens after layer 22 instead of layer 8 (ablation) | AIM scheduler · VideoMME | accuracy: 41.9 -> 58.1 | Improved |
| remove visual tokens after layer 15 instead of layer 22 (ablation) | AIM scheduler · VideoMME | accuracy: 58.1 -> 54.3 | Hurt |
| reduce token-merging retention from 100% to 25% (ablation) | AIM token merging · VideoMME | accuracy: 58.2 -> 58.0 | No clear effect |
| reduce token-merging retention from 25% to 1.6% (ablation) | AIM token merging · VideoMME | accuracy: 58.0 -> 50.9 | Hurt |
| start pruning at layer 14 instead of 21 (l2=22) (ablation) | AIM pruning scheduler · VideoMME | accuracy: 58.1 -> 58.2 | No clear effect |
| sample 192 frames instead of 32 with AIM (ablation) | AIM · MLVU | accuracy: 63.7 -> 69.3 | Improved |
| add AIM to LLaVA-OV with 128 frames (ablation) | LLaVA-OV · MLVU | accuracy: 67.7 -> 69.0 | Improved |
| apply AIM instead of FastV on Qwen2-VL base (ablation) | Qwen2-VL-7B · VideoMME | accuracy: 51.2 -> 52.8 | Improved |
| use AIM instead of LLaVA-Prumerge at 1.00 FLOPs on TextVQA | LLaVA-1.5-7B · TextVQA | accuracy: 53.8 -> 48.4 | Hurt |
Closest papers by meaning
- ATP-LLaVA: Adaptive Token Pruning for Large Vision Language Models (CVPR 2025 · similarity 0.92)
- One Layer's Trash is Another Layer's Treasure: Adaptive Layer-wise Visual Token Selection in LVLMs (CVPR 2026 · similarity 0.91)
- Improving Visual Token Reduction via Rectifying Distortions for Efficient Multimodal LLM Inference (ICML 2026 · similarity 0.91)
- Accelerating Multimodal Large Language Models via Dynamic Visual-Token Exit and the Empirical Findings (NeurIPS 2025 · similarity 0.91)
- Rethinking Visual Token Reduction in LVLMs Under Cross-Modal Misalignment (AAAI 2026 · similarity 0.91)
- Task-Related Token Compression in Multimodal Large Language Models from an Explainability Perspective (ICLR 2026 · similarity 0.90)
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.
