Counterfactual consistency training makes action-conditioned world models' predictions respond to actions, reducing zero-action drift and improving average visual planning success from 70.1% to 73.1%.
Title resolution pending
1 Pith paper cite this work. Polarity classification is still indexing.
1
Pith paper citing it
fields
cs.CV 1years
2026 1verdicts
CONDITIONAL 1representative citing papers
citing papers explorer
-
Overcoming Statistical Bias in Action-Controllable World Models
Counterfactual consistency training makes action-conditioned world models' predictions respond to actions, reducing zero-action drift and improving average visual planning success from 70.1% to 73.1%.