AI Paper Atlas

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

ICML 2026 · Vision-language and multimodal › Robot action and navigation · approach: Vision-language-action foundation models · 0 citations (OpenAlex)

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 changedWhereMeasuredEffect
measure nuisance mass ratio (NMR@10) of the policy's ISS saliencypi0.5 VLA on AGNOSTOS · AGNOSTOS success ratePearson correlation with success rate: r = -0.77Improved
use ISS instead of attention score for saliency under nuisance noisepi0.5 · robustness to nuisance noisesaliency cosine similarity: 0.959 -> 0.995Improved
use ISS instead of token normpi0.5 · robustness to nuisance noiseaction MSE: 0.011 -> 0.002Improved
token norm vs ISS on saliency cosine similaritypi0.5 · robustness to nuisance noisesaliency cosine similarity: 0.995 vs 0.999No clear effect
use ISS instead of attention scoresaliency fidelity · nuisance perturbation (first type)Pearson correlation with action MSE: 0.64 -> 0.78Improved
use ISS instead of token normsaliency fidelity · nuisance perturbation (first type)Pearson correlation with action MSE: 0.47 -> 0.78Improved
increase masking ratio to p=0.5 (ablation)ISS · average action MSEavg MSE (x1e-3): 3.7 -> 4.4 (N=100)Hurt
use N=100 interventions instead of 50 (ablation)ISS (p=0.3) · average action MSEavg MSE (x1e-3): 5.5 -> 3.7Improved
increase interventions to N=200 (ablation)ISS (p=0.5) · average action MSEavg MSE (x1e-3): 4.4 -> 6.6Hurt
use ISS rather than attention weightsexplanation cost · latencythroughput: 0.079s single inference vs 5.18s totalHurt
use wrist view versus front/overhead views for ISS attributionpi0.5 · Close Jar task, nuisance perturbationsnmr@10: wrist 0.09, 0.02, 0.06, 0.15 vs front 0.29, 0.24, 0.56, 0.39Improved

Closest papers by meaning

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.

Open scime draft

Get the next scime Dispatch and a monthly field update

Free. We send a confirmation link first; every email has a one-click unsubscribe. Privacy