Seeing Through the Illusion: Test-Time Scaling for World Action Models
Seeing Through the Illusion: Test-Time Scaling for World Action Models
This paper introduces WISE (World Action Inference Scaling through Evaluation), a pre-execution selector that evaluates sampled rollouts from world action models on action validity, temporal regularity, inverse-dynamics consistency, and task progress. By executing the highest-scoring candidate while keeping the model frozen, WISE raises closed-loop success from 6.0% to 24.0% on RoboLab with Cosmos~3 Edge and from 75.1% to 78.3% mean success on LIBERO with DreamZero, exceeding the best existing selector in aggregate by 4.0 and 0.9 points respectively.
WISE improves closed-loop success rates by selecting the highest-scoring candidate from four sampled rollouts based on action validity, temporal regularity, inverse-dynamics consistency, and task progress, without modifying the underlying frozen world action model.
Source: arXiv Robotics — research abstracts · Read original article ↗
Loading article text…
What the source reports
Publisher-reported claims, with original evidence. These results have not been independently verified by RoboSignal.
Reported numbers
aggregate improvement over best selector on RoboLab
4
View original evidence
roposed action and the imagined motion, and instruction-conditioned task progress, and executes the highest-scoring candidate while keeping the WAM frozen. With four candidates per decision, WISE raises closed-loop success from 6.0\% to 24.0\% on RoboLab with Cosmos~3 Edge and from 75.1\% to 78.3\% mean success on LIBERO with DreamZero, and exceeds the best existing selector in aggregate by 4.0 and 0.9 points. Compon
Open source S3aggregate improvement over best selector on LIBERO
0.9
View original evidence
roposed action and the imagined motion, and instruction-conditioned task progress, and executes the highest-scoring candidate while keeping the WAM frozen. With four candidates per decision, WISE raises closed-loop success from 6.0\% to 24.0\% on RoboLab with Cosmos~3 Edge and from 75.1\% to 78.3\% mean success on LIBERO with DreamZero, and exceeds the best existing selector in aggregate by 4.0 and 0.9 points. Compon
Open source S3
- paper: not_reported
Source excerpts and review record
Automatically extracted; no manual editorial approval recorded.
roposed action and the imagined motion, and instruction-conditioned task progress, and executes the highest-scoring candidate while keeping the WAM frozen. With four candidates per decision, WISE raises closed-loop success from 6.0\% to 24.0\% on RoboLab with Cosmos~3 Edge and from 75.1\% to 78.3\% mean success on LIBERO with DreamZero, and exceeds the best existing selector in aggregate by 4.0 and 0.9 points. Compon
Open source S3
Source：arXiv Robotics — research abstracts · arxiv.org