Research map / Vision-language and multimodal / Robot action and navigation
MergeVLA: Cross-Skill Model Merging Toward a Generalist Vision-Language-Action Agent
Yuxia Fu, Zhizhen Zhang, Yuqi Zhang, Zijian Wang, Zi Huang, Yadan 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 |
|---|---|---|---|
| add task-specific LoRA mask to TA merge (ablation) | OpenVLA TA merge (all) · LIBERO avg | success rate: 0.0 -> 62.4 | Improved |
| add task mask to VLA-Adapter merge (ablation) | VLA-Adapter TA · LIBERO avg | success rate: 0.0 -> 0.0 | No clear effect |
| leave final action expert block unmerged (expert head) (ablation) | VLA-Adapter TA + mask · LIBERO avg | success rate: 0.0 -> 23.1 | Improved |
| remove self-attention and use sigmoid gate in action expert (MergeVLA architecture) (ablation) | VLA-Adapter TA + mask except head · LIBERO avg | success rate: 23.1 -> 89.7 | Improved |
| MergeVLA architecture single-task finetuning vs VLA-Adapter | MergeVLA · LIBERO avg | success rate: 98.5 -> 96.7 | Hurt |
| MergeVLA architecture robustness vs VLA-Adapter (single-task) | MergeVLA · LIBERO-Plus avg | success rate: 59.0 -> 72.4 | Improved |
| route with value subspace only instead of key (ablation) | test-time task router · LIBERO avg | success rate: 53.6 -> 89.7 | Improved |
| route with value only instead of K&V (ablation) | test-time task router · LIBERO avg | success rate: 65.1 -> 89.7 | Improved |
| use TIES instead of TA for merging | MergeVLA · LIBERO avg | success rate: 89.7 -> 90.2 | Improved |
| use TIES instead of TA for merging | MergeVLA head L-1..L · RoboTwin Setting A | avg success: 78.7 -> 88.7 | Improved |
| extend expert head from last 2 to last 3 blocks with TIES (ablation) | MergeVLA · RoboTwin Setting B | avg success: 59.3 -> 70.7 | Improved |
| extend expert head from last 2 to last 3 blocks with TA (ablation) | MergeVLA · RoboTwin Setting B | avg success: 48.7 -> 49.3 | No clear effect |
Closest papers by meaning
- NestedVLA: Learning to Consolidate and Generate Skills for Vision-Language-Action Model (NeurIPS 2026 · similarity 0.93)
- ElegantVLA: Learning When to Think for Efficient Vision-Language-Action Models (NeurIPS 2026 · similarity 0.92)
- PriorVLA: Prior-Preserving Adaptation for Vision-Language-Action Models (NeurIPS 2026 · similarity 0.91)
- VLA-REPLICA: A Low-Cost, Reproducible Benchmark for Real-World Evaluation of Vision-Language-Action Models (NeurIPS 2026 · similarity 0.91)
- Beyond Data Scaling: Representation-Centric Pre-training for Vision-Language-Action Models (NeurIPS 2026 · similarity 0.90)
- VLA-Cache: Efficient Vision-Language-Action Manipulation via Adaptive Token Caching (NeurIPS 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.
