Research map / Vision-language and multimodal / Robot action and navigation
Revisiting Parameter Redundancy in Vision-Language-Action Models: Insights from VLM-to-VLA Adaptation
Fengnian Zhang, Tao Huang, Siyu Xu, Zhong Jin, Chang Xu
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 |
|---|---|---|---|
| prune highest-diff instead of lowest-diff Llama2 FFN channels (20%, no recovery) (ablation) | OpenVLA · LIBERO-Spatial | success rate (%): 1.5 -> 76.3 | Improved |
| add LoRA fine-tuning after lowest-diff pruning (ablation) | OpenVLA · LIBERO-Spatial | success rate (%): 1.5 -> 86.5 | Improved |
| prune highest-diff instead of lowest-diff DINOv2 FFN channels (ablation) | OpenVLA · LIBERO-Spatial | success rate (%): 0.0 -> 82.0 | Improved |
| prune lowest-diff instead of highest-diff DINOv2 attention heads (ablation) | OpenVLA · LIBERO-Spatial | success rate (%): 1.6 -> 76.7 | Improved |
| prune highest-diff instead of lowest-diff Llama2 attention heads (ablation) | OpenVLA · LIBERO-Spatial | success rate (%): 0.0 -> 84.3 | Improved |
| prune highest-diff instead of lowest-diff Llama2 FFN channels (ablation) | OpenVLA · LIBERO-Spatial | success rate (%): 2.7 -> 72.0 | Improved |
| prune highest-diff instead of lowest-diff Gemma attention heads (ablation) | pi0.5 · LIBERO-Spatial | success rate (%): 0.0 -> 85.0 | Improved |
| prune highest-diff instead of lowest-diff Gemma FFN channels (ablation) | pi0.5 · LIBERO-Spatial | success rate (%): 5.0 -> 95.0 | Improved |
| choose highest-diff vs lowest-diff pruning for the projector FFN (30%) (ablation) | OpenVLA · LIBERO-Spatial | success rate (%): 0.0 -> 54.0 | Mixed |
| remove SigLIP FFN entirely (ratio 1.0) (ablation) | OpenVLA · LIBERO-Spatial | success rate (%): 84.7 -> 70.0 | Hurt |
| use the multi-module joint pruning scheme instead of LLM-Pruner at matched 12.4GB | OpenVLA · LIBERO (Long, Moderate config) | success rate (%): 1.0 -> 62.3 | Improved |
| apply Moderate joint pruning (7.3GB to 5.6GB) | pi0.5 · LIBERO (avg of 4 suites) | success rate (%): 96.9 -> 89.0 | Hurt |
Closest papers by meaning
- EfficientVLA: Training-Free Acceleration and Compression for Vision-Language-Action Models (NeurIPS 2025 · similarity 0.93)
- VLA-REPLICA: A Low-Cost, Reproducible Benchmark for Real-World Evaluation of Vision-Language-Action Models (NeurIPS 2026 · similarity 0.92)
- PriorVLA: Prior-Preserving Adaptation for Vision-Language-Action Models (NeurIPS 2026 · similarity 0.91)
- ElegantVLA: Learning When to Think for Efficient Vision-Language-Action Models (NeurIPS 2026 · similarity 0.91)
- Beyond Data Scaling: Representation-Centric Pre-training for Vision-Language-Action Models (NeurIPS 2026 · similarity 0.91)
- Knowledge Insulating Vision-Language-Action Models: Train Fast, Run Fast, Generalize Better (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.
