Research map / Architectures and efficiency / RNNs and normalization
You Need Multiple Exiting: Dynamic Early Exiting for Accelerating Unified Vision Language Model
Shengkun Tang, Yaqing Wang, Zhenglun Kong, Tianchi Zhang, Yao Li, Caiwen Ding, Yanzhi Wang, Yi Liang, Dongkuan Xu
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 modality decomposition of the encoder (separate image/text early exits) (ablation) | MuE · SNLI-VE test | accuracy: 88.5 vs 75.0 without decomposition | Improved |
| add modality decomposition of the encoder (ablation) | MuE · MS COCO Karpathy test | CIDEr: 137.0 vs 117.5 without decomposition | Improved |
| add layer-wise task loss on every decoder layer (ablation) | MuE · MS COCO Karpathy test | CIDEr: 137.0 vs 117.9 without task loss | Improved |
| add layer-wise task loss on every decoder layer (ablation) | MuE · SNLI-VE test | accuracy: 87.0 vs 88.5 without task loss | No clear effect |
| multiple early exiting via layer-wise similarity across encoder and decoder | OFA Base · SNLI-VE test | expected time reduction: -50% time at 88.5 accuracy vs OFABase 89.2 | Improved |
| multiple early exiting via layer-wise similarity across encoder and decoder | OFA Base · MS COCO Karpathy test | CIDEr: 137 at -40.2% time vs DeeCap 129 at -38% | Improved |
| DeeBERT-style decoder-only early exiting | OFA Base · SNLI-VE test | accuracy: 78.8 vs 89.2 full model | Hurt |
| training with smaller static model OFATiny instead of early exiting | OFA · SNLI-VE test | accuracy: 85.2 vs 89.2 full model | Hurt |
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.
- DeeCap — DeeCap: Dynamic Early Exiting for Efficient Image Captioning (CVPR 2022) · 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
- Frozen Transformers in Language Models Are Effective Visual Encoder Layers (ICLR 2024 · similarity 0.86)
- iLLaVA: An Image is Worth Fewer Than 1/3 Input Tokens in Large Multimodal Models (ICLR 2026 · similarity 0.86)
- Investigating Redundancy in Multimodal Large Language Models with Multiple Vision Encoders (ICLR 2026 · similarity 0.86)
- Beyond Encoder Accumulation: Measuring Encoder Roles in Multi-Encoder VLMs (NeurIPS 2026 · similarity 0.86)
- Large Vision-Language Models Get Lost in Attention (ICML 2026 · similarity 0.86)
- HiRED: Attention-Guided Token Dropping for Efficient Inference of High-Resolution Vision-Language Models (AAAI 2025 · similarity 0.86)
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.
