Research map / Architectures and efficiency / Pruning and efficient tuning
AD-BTS: Adaptive Dual-Branch Token Sparsification via Spatial Information Density
Xinpei Gao, Xin Luo, Ming Liu, Chunjiang Wang, S Kevin Zhou
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 |
|---|---|---|---|
| route inputs adaptively between RSB and SFB (delta=0.4) instead of RSB only (ablation) | AD-BTS · ChartQA (30% tokens) | Relaxed accuracy: 73.12 -> 83.68 | Improved |
| route inputs adaptively instead of SFB only (ablation) | AD-BTS · MME (30% tokens) | score: 2233.70 -> 2351.76 | Improved |
| dual-branch routing at delta=0.4 versus single-branch variants (ablation) | AD-BTS · Nine image benchmarks, 30% tokens | normalized avg score (%): 97.11% / 98.01% -> 98.72% | Improved |
| add LoRA adapters (r=256) in SFB versus r=0 (ablation) | AD-BTS · ChartQA (30% tokens) | Relaxed accuracy: 80.52 -> 83.68 | Improved |
| increase LoRA rank from 256 to 512 (ablation) | AD-BTS · Nine image benchmarks, 30% tokens | normalized avg score (%): 98.72% -> 96.82% | Hurt |
| lower LoRA rank to 64 (ablation) | AD-BTS · Nine image benchmarks, 30% tokens | normalized avg score (%): 98.72% -> 97.50% | Hurt |
| add context fusion on top of SFB with LoRA (ablation) | AD-BTS · ChartQA (30% tokens) | Relaxed accuracy: 79.96 -> 83.68 | Improved |
| use dual-branch routing instead of always-RSB with persistent LoRA at matched capacity (ablation) | AD-BTS · ChartQA (30% tokens) | Relaxed accuracy: 78.64 -> 83.68 | Improved |
| AD-BTS versus VisionSelector at extreme compression | Qwen2.5-VL-7B (10% tokens) · ChartQA | Relaxed accuracy: 61.6 -> 68.24 | Improved |
| AD-BTS versus uncompressed model on video | Qwen2.5-VL-7B · Video prefill time | prefill time (ms): 1413.34 -> 746.25 | Improved |
| RSB-only routing on redundancy-dominated MME (ablation) | AD-BTS · MME (30% tokens) | score: 2351.76 -> 2355.26 | Improved |
Closest papers by meaning
- One Layer's Trash is Another Layer's Treasure: Adaptive Layer-wise Visual Token Selection in LVLMs (CVPR 2026 · similarity 0.88)
- SparseVLM: Visual Token Sparsification for Efficient Vision-Language Model Inference (ICML 2025 · similarity 0.88)
- TwinPrune: Density-Aware Two-Phase Token Pruning for Vision-Language Models (NeurIPS 2026 · similarity 0.88)
- EvoComp: Learning Visual Token Compression for Multimodal Large Language Models via Semantic-Guided Evolutionary Labeling (CVPR 2026 · similarity 0.87)
- An Empirical Study of Attention and Diversity for Adaptive Visual Token Pruning in Large Vision-Language Models (ICLR 2026 · similarity 0.87)
- Balanced Token Pruning: Accelerating Vision Language Models Beyond Local Optimization (NeurIPS 2025 · similarity 0.87)
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.
