Research map / Vision-language and multimodal / Robot action and navigation
Embodied Interpretability: Linking Causal Understanding to Generalization in Vision-Language-Action Models
Hanxin Zhang, Mingshuo Xu, Abdulqader Dhafer, Shigang Yue, Hongbiao Dong, Zhou Daniel Hao
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 |
|---|---|---|---|
| measure nuisance mass ratio (NMR@10) of the policy's ISS saliency | pi0.5 VLA on AGNOSTOS · AGNOSTOS success rate | Pearson correlation with success rate: r = -0.77 | Improved |
| use ISS instead of attention score for saliency under nuisance noise | pi0.5 · robustness to nuisance noise | saliency cosine similarity: 0.959 -> 0.995 | Improved |
| use ISS instead of token norm | pi0.5 · robustness to nuisance noise | action MSE: 0.011 -> 0.002 | Improved |
| token norm vs ISS on saliency cosine similarity | pi0.5 · robustness to nuisance noise | saliency cosine similarity: 0.995 vs 0.999 | No clear effect |
| use ISS instead of attention score | saliency fidelity · nuisance perturbation (first type) | Pearson correlation with action MSE: 0.64 -> 0.78 | Improved |
| use ISS instead of token norm | saliency fidelity · nuisance perturbation (first type) | Pearson correlation with action MSE: 0.47 -> 0.78 | Improved |
| increase masking ratio to p=0.5 (ablation) | ISS · average action MSE | avg MSE (x1e-3): 3.7 -> 4.4 (N=100) | Hurt |
| use N=100 interventions instead of 50 (ablation) | ISS (p=0.3) · average action MSE | avg MSE (x1e-3): 5.5 -> 3.7 | Improved |
| increase interventions to N=200 (ablation) | ISS (p=0.5) · average action MSE | avg MSE (x1e-3): 4.4 -> 6.6 | Hurt |
| use ISS rather than attention weights | explanation cost · latency | throughput: 0.079s single inference vs 5.18s total | Hurt |
| use wrist view versus front/overhead views for ISS attribution | pi0.5 · Close Jar task, nuisance perturbations | nmr@10: wrist 0.09, 0.02, 0.06, 0.15 vs front 0.29, 0.24, 0.56, 0.39 | Improved |
Closest papers by meaning
- CofactVLA: Deconfounding Vision-Language-Action Models via Counterfactual Intervention (NeurIPS 2026 · similarity 0.90)
- Not All Features Are Created Equal: A Mechanistic Study of Vision-Language-Action Models (NeurIPS 2026 · similarity 0.89)
- Causal Tracing of Object Representations in Large Vision Language Models: Mechanistic Interpretability and Hallucination Mitigation (AAAI 2026 · similarity 0.88)
- ATI-VLA: Action-Centric Predictive Vision–Language–Action Models via Actionable Alignment Then Adaptive Injection (NeurIPS 2026 · similarity 0.88)
- Event-Grounded Sparse Autoencoders for Vision-Language-Action Policies (NeurIPS 2026 · similarity 0.88)
- Perception Without Engagement: Dissecting the Causal Discovery Deficit in LMMs (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.
