Research map / Vision-language and multimodal / Robot action and navigation
Fast-dVLA: Accelerating Discrete Diffusion VLA to Real-Time Performance
Wenxuan Song, Jiayi Chen, Shuai Chen, Jingbo Wang, Pengxiang Ding, Han Zhao, qin yikai, Xinhu Zheng, Yan Wang, 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 |
|---|---|---|---|
| use block size as multiple of action dimensionality instead of random (ablation) | Fast-dVLA (Dream-VLA) · LIBERO-Long | success rate: 73.3% -> 74.7% | Improved |
| use block size as multiple of action dimensionality instead of random (ablation) | Fast-dVLA (Dream-VLA) · LIBERO-Long | speedup: 3.95x -> 4.01x | Improved |
| lower confidence threshold tau_conf to 0.5 (ablation) | Fast-dVLA · LIBERO | success rate and speed: 2.8x acceleration, 2% drop | Mixed |
| distill asymmetrically from finetuned weight instead of training from scratch (ablation) | Fast-dVLA (Dream-VLA) · LIBERO | training steps: 2,000 steps, 1/10 of scratch | Improved |
| raise tau_add and tau_act from 0.4/0.4 to 0.6/0.8 (ablation) | Fast-dVLA (UD-VLA) · CALVIN ABCD->D | avg length / speed: 4.42 -> 4.57 length, 253.1 -> 160.7 tokens/s | Mixed |
| use dual-state thresholds tau_add 0.5, tau_act 0.7 instead of equal 0.5/0.5 (ablation) | Fast-dVLA (UD-VLA) · CALVIN ABCD->D | avg length: 4.44 -> 4.54 | Improved |
| use more conservative radical decoding log4 instead of log2 (ablation) | Fast-dVLA (UD-VLA) · CALVIN ABCD->D | speed: 186.67 -> 144.71 tokens/s | Hurt |
| use log4 instead of log2 radical decoding (ablation) | Fast-dVLA (UD-VLA) · CALVIN ABCD->D | avg length: 4.54 -> 4.58 | No clear effect |
| apply Fast-dLLM KV reuse to Dream-VLA | Dream-VLA · LIBERO | success rate: 0.856 -> 0.828 | Hurt |
| apply Fast-dVLA to Dream-VLA | Dream-VLA · LIBERO | speed: 98.8 -> 313.1 tokens/s | Improved |
| apply Fast-dVLA to DD-VLA | Discrete Diffusion VLA · LIBERO | speed: 152.1 -> 402.7 tokens/s | Improved |
| apply Fast-dVLA to UD-VLA | UD-VLA · CALVIN ABCD->D | avg length: 4.64 -> 4.54 | Hurt |
Earlier methods this paper compares with
Named baselines from earlier accepted papers in this paper's results tables: better than 3 of 3 on most shared results.
- Dream-VLA — DreamVLA: A Vision-Language-Action Model Dreamed with Comprehensive World Knowledge (NeurIPS 2025) · better on 4 of 4 shared results
- Discrete Diffusion VLA — Discrete Diffusion VLA: Bringing Discrete Diffusion to Action Decoding in Vision-Language-Action Policies (ICML 2026) · better on 2 of 2 shared results
- UP-VLA — UP-VLA: A Unified Understanding and Prediction Model for Embodied Agent (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
- ElegantVLA: Learning When to Think for Efficient Vision-Language-Action Models (NeurIPS 2026 · similarity 0.92)
- D-VLA: A High-Concurrency Distributed Asynchronous RL Framework for Vision-Language-Action Models (NeurIPS 2026 · similarity 0.91)
- EfficientVLA: Training-Free Acceleration and Compression for Vision-Language-Action Models (NeurIPS 2025 · similarity 0.91)
- VLA-Cache: Efficient Vision-Language-Action Manipulation via Adaptive Token Caching (NeurIPS 2025 · similarity 0.91)
- VLA-REPLICA: A Low-Cost, Reproducible Benchmark for Real-World Evaluation of Vision-Language-Action Models (NeurIPS 2026 · similarity 0.90)
- DiffusionVLA: Scaling Robot Foundation Models via Unified Diffusion and Autoregression (ICML 2025 · 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.
