[KongchangAI]
ModelLeWM

LeWorldModel

A world model for model-based reinforcement learning that learns environment dynamics and supports planning; it is the analytical subject of research on local action geometry fidelity.

Timeline (last 90 days)

Nov 3

A study titled "Does LeWorldModel Preserve Local Action Geometry?" was accepted as a poster by the NeurIPS 2026 NeurReps Workshop.

Unverified50%
Nov 3

The study tests LeWorldModel (LeWM) on two environments, Reacher and OGBench Cube.

Unverified50%
Nov 3

The study compares the action-Jacobian Gram matrix extracted from the learned world model with the action-Jacobian Gram matrix obtained from counterfactual queries to the simulator.

Unverified50%
Nov 3

At the smallest perturbation scale, the dominant action direction of LeWorldModel deviates from the simulator's true direction: about 36 degrees on Reacher and about 42 degrees on Cube.

Unverified50%
Nov 3

After the perturbation is magnified tenfold, the action geometry discrepancy of LeWorldModel shrinks significantly: by about 70% on Reacher and about 47% on Cube.

Unverified50%
Nov 3

The action geometry measured in LeWorldModel's latent representation varies far more across perturbation scales than the geometry measured in physical state coordinates.

Unverified50%
Nov 3

The local geometry given by LeWorldModel is less accurate than the simulator-induced geometry when predicting the response magnitude for unseen action directions.

Unverified50%
Nov 3

The authors explicitly state that it has not yet been established whether these local action geometry deviations directly cause planning failures, which is the question to be answered next.

Unverified50%
Nov 3

The authors are interested in incorporating local action geometry into CEM's sampling covariance to improve exploration and control performance, but whether standard CEM's covariance adaptation mechanism already implicitly compensates for geometric deviations remains undetermined.

Unverified50%

All Facts (9)

Unverified

The authors explicitly state that it has not yet been established whether these local action geometry deviations directly cause planning failures, which is the question to be answered next.

50%
Unverified

The study tests LeWorldModel (LeWM) on two environments, Reacher and OGBench Cube.

50%
Unverified

The study compares the action-Jacobian Gram matrix extracted from the learned world model with the action-Jacobian Gram matrix obtained from counterfactual queries to the simulator.

50%
Unverified

At the smallest perturbation scale, the dominant action direction of LeWorldModel deviates from the simulator's true direction: about 36 degrees on Reacher and about 42 degrees on Cube.

50%
Unverified

After the perturbation is magnified tenfold, the action geometry discrepancy of LeWorldModel shrinks significantly: by about 70% on Reacher and about 47% on Cube.

50%
Unverified

The action geometry measured in LeWorldModel's latent representation varies far more across perturbation scales than the geometry measured in physical state coordinates.

50%
Unverified

The local geometry given by LeWorldModel is less accurate than the simulator-induced geometry when predicting the response magnitude for unseen action directions.

50%
Unverified

The authors are interested in incorporating local action geometry into CEM's sampling covariance to improve exploration and control performance, but whether standard CEM's covariance adaptation mechanism already implicitly compensates for geometric deviations remains undetermined.

50%
Unverified

A study titled "Does LeWorldModel Preserve Local Action Geometry?" was accepted as a poster by the NeurIPS 2026 NeurReps Workshop.

50%

Source Articles