Research map / Architectures and efficiency / RNNs and normalization
SafeSAE-VLA: Interpreting OpenVLA Progress Dynamics with Sparse Feature Analysis
Socrates Osorio, Joy Yang
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 SAE top-20 features instead of full 16384-d SAE (ablation) | SAE progress probe · LIBERO progress quartiles | AUROC: 0.918 -> 0.894 | Hurt |
| SAE features versus motion-only controls (ablation) | SAE progress probe · LIBERO progress quartiles | AUROC: 0.572-0.711 -> 0.918 | Improved |
| dense raw-activation LR instead of SAE LR (ablation) | progress probe · pooled encoding | AUROC: 0.947 -> 0.975 | Improved |
| choose top-ranked 20 SAE features instead of random 20 (ablation) | SAE LR · pooled encoding | AUROC: 0.783 -> 0.926 | Improved |
| layer 20 versus layers 16 and 24 (ablation) | SAE monitor · LIBERO progress | AUROC: 0.884 / 0.896 / 0.871 | No clear effect |
| tertile or median split instead of quartile split (ablation) | full SAE · LIBERO progress | AUROC: 0.948 tertile, 0.929 median | No clear effect |
| top-20 SAE instead of full dictionary under task-family shift (ablation) | SAE readout · held-out tasks | AUROC: 0.537 -> 0.658 | Improved |
| set top-20 features to high-progress class means (ablation) | OpenVLA · raw-activation progress readout | readout shift: +0.763 vs +0.078 random | Improved |
| set 13 active features to class mean (ablation) | OpenVLA · LIBERO goal-task-1 | any-violation incidence: 1.00 -> 0.50 | Improved |
| progress-tuned top-20 used to predict success (ablation) | Top-20 SAE · LIBERO success audit | AUROC: 0.968 (full SAE) -> 0.471 (top-20) | Hurt |
Closest papers by meaning
- Event-Grounded Sparse Autoencoders for Vision-Language-Action Policies (NeurIPS 2026 · similarity 0.88)
- SAFE: Multitask Failure Detection for Vision-Language-Action Models (NeurIPS 2025 · similarity 0.88)
- ActionUNet: Improving Robustness of VLA Models with Efficient Multi-scale Fine-tuning (NeurIPS 2026 · similarity 0.88)
- VLA-Corrector: Lightweight Detect-and-Correct Inference for Adaptive Action Horizon (NeurIPS 2026 · similarity 0.88)
- ElegantVLA: Learning When to Think for Efficient Vision-Language-Action Models (NeurIPS 2026 · similarity 0.88)
- Beyond Data Scaling: Representation-Centric Pre-training for Vision-Language-Action Models (NeurIPS 2026 · similarity 0.87)
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.
