Research map / Vision-language and multimodal / Robot action and navigation
VLANeXt: Recipes for Building Strong VLA Models
Xiao-Ming Wu, Bin Fan, Kang Liao, Jian-Jian Jiang, Runze Yang, Yihang Luo, Zhonghua Wu, Wei-Shi Zheng, Chen Change Loy
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 large policy module (16 tokens, 12 layers) over a separate head (ablation) | VLA with separate head · LIBERO-plus (spatial suite, Total) | success rate (%): 16.6 -> 34.0 | Improved |
| increase action chunk size from 1 to 8 (ablation) | VLA with large policy module · LIBERO-plus (spatial suite, Total) | success rate (%): 34.0 -> 43.4 | Improved |
| replace bin classification with regression (ablation) | VLA with chunk 8 · LIBERO-plus (spatial suite, Total) | success rate (%): 43.4 -> 48.4 | Improved |
| replace bin classification with VQ-VAE codebook classification (ablation) | VLA with chunk 8 · LIBERO-plus (spatial suite, Total) | success rate (%): 43.4 -> 36.5 | Hurt |
| use flow matching instead of bin classification (ablation) | VLA with chunk 8 · LIBERO-plus (spatial suite, Total) | success rate (%): 43.4 -> 45.0 | Improved |
| swap backbone from PaliGemma to Qwen3VL-4B (ablation) | VLA with flow matching · LIBERO-plus (spatial suite, Total) | success rate (%): 18.9 -> 64.8 | Improved |
| use soft instead of loose VLM-policy connection (ablation) | VLANeXt · LIBERO-plus (spatial suite, Total) | success rate (%): 53.7 -> 56.2 | Improved |
| add temporal observation history (ablation) | VLANeXt · LIBERO-plus (spatial suite, Total) | success rate (%): 56.2 -> 50.2 | Hurt |
| add wrist camera to third-person view (ablation) | VLANeXt · LIBERO-plus (spatial suite, Total) | success rate (%): 56.2 -> 80.5 | Improved |
| add proprioception conditioning in the VLM (ablation) | VLANeXt · LIBERO-plus (spatial suite, Total) | success rate (%): 80.5 -> 87.7 | Improved |
| add proprioception to the policy module only (ablation) | VLANeXt · LIBERO-plus (spatial suite, Total) | success rate (%): 80.5 -> 83.4 | Hurt |
| add auxiliary world modelling objective (ablation) | VLANeXt · LIBERO-plus (spatial suite, Total) | success rate (%): 87.7 -> 90.3 | Improved |
Later papers that compared with this method
1 later accepted paper reports VLANeXt as a baseline in their results tables; 0 of them report a better number on most of the results they share with it.
- Spatial Representation Distillation and Knowledge Routing for Vision-Language-Action Models (NeurIPS 2026 · better on 1 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
- VHELM: A Holistic Evaluation of Vision Language Models (NeurIPS 2024 · similarity 0.89)
- What Can RL Bring to VLA Generalization? An Empirical Study (NeurIPS 2025 · similarity 0.89)
- What matters when building vision-language models? (NeurIPS 2024 · similarity 0.89)
- EfficientVLA: Training-Free Acceleration and Compression for Vision-Language-Action Models (NeurIPS 2025 · similarity 0.88)
- Visual-Advantage On-Policy Distillation for Vision-Language Models (NeurIPS 2026 · similarity 0.88)
- 3D-VLA: A 3D Vision-Language-Action Generative World Model (ICML 2024 · similarity 0.88)
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.
