Research map / Vision-language and multimodal / Robot action and navigation
Towards More Efficient Decoding for Autoregressive Vision-language-action Models
Wenxuan Song, Jiayi Chen, Pengxiang Ding, Yuxin Huang, Han Zhao, Yinchuan Li, Yingcong Chen, Donglin Wang, Haoang Li
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 adaptive early-exit decoding (ablation) | CEED-VLA on OpenVLA · LIBERO-Long | speedup (x): 2.5 -> 4.1 | Improved |
| add adaptive early-exit decoding (ablation) | CEED-VLA on OpenVLA · LIBERO-Long | success rate (%): 61.2 -> 62.3 | No clear effect |
| add consistency training (ablation) | CEED-VLA on OpenVLA · LIBERO-Long | speedup (x): 1.6 -> 2.5 | Improved |
| add consistency training to Jacobi plus early-exit (ablation) | CEED-VLA on OpenVLA · LIBERO-Long | success rate (%): 54.2 -> 62.3 | Improved |
| add Jacobi decoding over AR (ablation) | CEED-VLA on OpenVLA · LIBERO-Long | speedup (x): 1.0 -> 1.6 | Improved |
| add adaptive early-exit to Jacobi decoding (ablation) | CEED-VLA · decoding speed (average) | tokens/s: 57.57 -> 79.24 | Improved |
| use consistency training model with intermediate output at j=12 and 8 (no consistency training) (ablation) | LLaVA-VLA · CALVIN ABC->D | average length: 1.99 and 1.44 vs above 3.34 | Hurt |
| increase AR loss weight to 10x (ablation) | CEED-VLA LLaVA-VLA · CALVIN ABC->D | speedup (x): 2.00x -> 1.79x | Mixed |
| raise confidence threshold sigma from 0.5 to 0.7 (ablation) | CEED-VLA LLaVA-VLA · CALVIN ABC->D | speedup (x): 2.0x -> 1.3x | Mixed |
| lower confidence threshold sigma from 0.5 to 0.4 (ablation) | CEED-VLA · Real-world | success rate (%): 77.5 -> 72.5 | Hurt |
| increase training trajectories beyond 120k (ablation) | CEED-VLA LLaVA-VLA · CALVIN ABC->D | speedup | No clear effect |
| replace AR with CEED-VLA decoding | pi0-FAST · LIBERO | decoding speed (tokens/s): 107.5 -> 420.6 | Improved |
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.
- SparseVLM — SparseVLM: Visual Token Sparsification for Efficient Vision-Language Model Inference (ICML 2025) · better on 1 of 1 shared result
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
- FASTer: Toward Powerful and Efficient Autoregressive Vision–Language–Action Models with Learnable Action Tokenizer and Block-wise Decoding (ICLR 2026 · similarity 0.94)
- ElegantVLA: Learning When to Think for Efficient Vision-Language-Action Models (NeurIPS 2026 · similarity 0.91)
- ActionUNet: Improving Robustness of VLA Models with Efficient Multi-scale Fine-tuning (NeurIPS 2026 · similarity 0.90)
- VLA-Cache: Efficient Vision-Language-Action Manipulation via Adaptive Token Caching (NeurIPS 2025 · similarity 0.89)
- NS-VLA: Towards Neuro-Symbolic Vision-Language-Action Models (NeurIPS 2026 · similarity 0.89)
- EfficientVLA: Training-Free Acceleration and Compression for Vision-Language-Action Models (NeurIPS 2025 · similarity 0.89)
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.
