Research map / Vision-language and multimodal / Robot action and navigation
Discrete Diffusion VLA: Bringing Discrete Diffusion to Action Decoding in Vision-Language-Action Policies
Zhixuan Liang, Yizhuo Li, Tianshuo Yang, Chengyue Wu, Sitong Mao, Liuao Pei, Tian Nian, Shunbo Zhou, Xiaokang Yang, Jiangmiao Pang, Yao Mu, Ping Luo
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 adaptive max-confidence decoding instead of one-shot parallel decoding (ablation) | Discrete Diffusion VLA · LIBERO-Goal | success rate (%): 95.6% -> 96.8% | Improved |
| use random decoding order instead of confidence ordering (ablation) | Discrete Diffusion VLA · LIBERO-Goal | success rate (%): 96.8% -> 95.8% | Hurt |
| max confidence rather than confidence gap as the ranking metric (ablation) | Discrete Diffusion VLA · LIBERO-Goal | success rate (%): 96.6% -> 96.8% | No clear effect |
| add linear temperature decay from 1.0 to 0.0 instead of hard argmax (ablation) | Discrete Diffusion VLA · LIBERO-Goal | success rate (%): 96.2% -> 96.8% | Improved |
| increase denoising steps T (ablation) | Discrete Diffusion VLA · LIBERO-Goal | success rate and throughput | Mixed |
| use discrete diffusion action head instead of AR on a Qwen2.5-VL backbone (ablation) | Qwen2.5-VL, no robot pretraining · LIBERO 4 suites | avg success rate (%): 81.2 -> 95.1 | Improved |
| use discrete diffusion instead of continuous (GR00T-style) diffusion head on Qwen2.5-VL (ablation) | Qwen2.5-VL, no robot pretraining · LIBERO 4 suites | avg success rate (%): 94.9 -> 95.1 | No clear effect |
| replace discrete diffusion by FAST head on Qwen2.5-VL (ablation) | Qwen2.5-VL, no robot pretraining · LIBERO 4 suites | avg success rate (%): 95.1 -> 91.1 | Hurt |
| discrete diffusion head instead of separate continuous diffusion head for language robustness | OpenVLA-OFT · LIBERO-Goal OOD language augmentation | performance degradation: 2.4% -> 0.8% | Improved |
| discrete diffusion head instead of continuous diffusion for vision robustness | OpenVLA-OFT · LIBERO-Goal OOD vision augmentation | performance degradation: 29.0% -> 20.4% | Improved |
| discrete diffusion VLA versus parallel decoding on language augmentation | OpenVLA-OFT (Discrete) · LIBERO-Goal OOD language augmentation | performance degradation: 8.0% -> 0.8% | Improved |
| 12-step discrete diffusion decoding instead of autoregressive decoding | OpenVLA · LIBERO-Goal inference | latency (ms): 136.2 -> 68.8 | Improved |
Later papers that compared with this method
1 later accepted paper reports Discrete Diffusion VLA as a baseline in their results tables; 1 of them report a better number on most of the results they share with it.
- Fast-dVLA: Accelerating Discrete Diffusion VLA to Real-Time Performance (ECCV 2026 · 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
- Fast-dVLA: Accelerating Discrete Diffusion VLA to Real-Time Performance (ECCV 2026 · similarity 0.95)
- ElegantVLA: Learning When to Think for Efficient Vision-Language-Action Models (NeurIPS 2026 · similarity 0.92)
- VLA-Cache: Efficient Vision-Language-Action Manipulation via Adaptive Token Caching (NeurIPS 2025 · similarity 0.91)
- EfficientVLA: Training-Free Acceleration and Compression for Vision-Language-Action Models (NeurIPS 2025 · similarity 0.91)
- Beyond Data Scaling: Representation-Centric Pre-training for Vision-Language-Action Models (NeurIPS 2026 · similarity 0.91)
- NS-VLA: Towards Neuro-Symbolic Vision-Language-Action Models (NeurIPS 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.
