Research map / Vision-language and multimodal / Multimodal LLMs
TOP-RL: Task-Optimized Progressive Token Pruning with Reinforcement Learning for Vision Language Models
Hengyi Wang, Weiying Xie, Hui Jiang, Yaotao Wei, Kai Jiang, Mingxiang Cao, Chenhe Hao, Leyuan Fang
Read the paperCheck your idea against 137,217 papers
AI-extracted summary. Written by AI from the paper's full text. Numbers and claims can be wrong or missing context, so check the paper before you cite or build on it.
Proposes TOP-RL, a task-optimized progressive visual token pruning method for vision-language models that uses reinforcement learning to decide pruning across tasks and layer depths.
Problem
LVLM inference is expensive because of the large number of visual tokens; existing compression methods prune at the input or in early layers and ignore variation across tasks and depths.
Approach
Progressive, task-aware visual token pruning across model layers, with the pruning policy optimized by reinforcement learning rather than fixed input-stage or early-layer rules.
Measured effects reported in the paper
| What was changed | Where | Measured | Effect |
|---|---|---|---|
| use TOP-RL task-optimized progressive pruning instead of FastV | LLaVA-1.5-7B, 192 tokens · 9-benchmark average | avg accuracy retention: 87.7% -> 97.6% | Improved |
| use TOP-RL instead of VisionZip at 64 tokens | LLaVA-1.5-7B · 9-benchmark average | avg accuracy retention: 87.9% -> 91.0% | Improved |
| use TOP-RL on Video-LLaVA instead of VisionZip | Video-LLaVA · Video QA average | avg accuracy retention: 93.4% -> 93.9% | Improved |
| use TOP-RL instead of VisionZip for efficiency | LLaVA-NeXT-7B · Efficiency (320 tokens) | speedup: 2.11x -> 1.91x | Hurt |
| use TOP-RL instead of VisionZip for accuracy at same budget | LLaVA-NeXT-7B · Efficiency (320 tokens) | F1 score: 86.4 -> 87.2 | Improved |
| transfer pruning agent across cognitive levels (perception to reasoning) (ablation) | TOP-RL agent · SQA | accuracy: 69.6 -> 46.4 | Hurt |
| transfer pruning agent within the same cognitive domain (POPE agent to TextVQA) (ablation) | TOP-RL agent · TextVQA | accuracy: 58.2 -> 58.9 | No clear effect |
| apply progressive multi-stage pruning (FastV-P) instead of one-shot FastV (ablation) | FastV · TextVQA, SQA | accuracy | Improved |
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 8 of 10 shared results
- DivPrune — DivPrune: Diversity-based Visual Token Pruning for Large Multimodal Models (CVPR 2025) · better on 2 of 2 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 0.93)
- Balanced Token Pruning: Accelerating Vision Language Models Beyond Local Optimization (NeurIPS 2025 · similarity 0.92)
- ATP-LLaVA: Adaptive Token Pruning for Large Vision Language Models (CVPR 2025 · similarity 0.91)
- ZOO-Prune: Training-Free Token Pruning via Zeroth-Order Gradient Estimation in Vision-Language Models (CVPR 2026 · similarity 0.91)
- SwiftVLM: Efficient Vision-Language Model Inference via Cross-Layer Token Bypass (NeurIPS 2026 · similarity 0.91)
- LearnPruner: Rethinking Attention-based Token Pruning in Vision Language Models (ICLR 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.
