Teach a world model what actions change.

Causal
Action
Effect
Reweighting

Paper arXiv GitHub

Authors

Jianjie Fang^*¹ Xvyuan Liu^*¹ Ziyou Wang¹ Rongze Tang³ Zhaolu Wang¹ Zhuohang Li¹ Xin Zhang² Haisheng Su² Chen Gao¹^† Wei Wu² Xinlei Chen¹^† Yong Li¹^†

* Equal contribution. † Corresponding authors.

Affiliations

¹ Tsinghua University

² Manifold AI

³ University of Science and Technology of China

Date

2026

Abstract

World models are becoming core infrastructure for embodied intelligence, with action-conditioned video generation providing controllable predictions of how scenes evolve after agent interventions. Yet existing models are commonly trained with space–time-uniform mean squared error, allowing abundant background tokens to dominate the gradient while sparse interaction dynamics remain under-optimized; such uniform fitting rewards reconstructing appearance rather than learning how actions change the world. We introduce Causal Action Effect Reweighting (CAER), a general training paradigm that redistributes supervision toward the tokens whose predicted future is causally affected by the action. CAER contrasts the model's own predictions with and without action conditioning to localize these tokens online, then normalizes the resulting effect map into a weight that preserves the total gradient budget and changes only where it is spent. This online signal requires no external annotations or offline preprocessing, avoids additional data-processing time, and scales naturally with model and dataset size. Experiments across heterogeneous action-conditioned world-model tasks show that CAER converges to better solutions than uniform MSE training, with consistent improvements in the physical consistency, controllability, and visual quality of generated videos.

A small change to the objective.

A sharper world model.

01 — PROBLEM

What uniform MSE misses

Gradient is allocated by token count, so easy background pixels dilute sparse contact, object, and actor dynamics.

02 — EFFECT

Action-sensitive tokens

CAER contrasts the same noisy state under the real action and a learned null action to reveal where the future responds.

03 — BUDGET

Focus without inflation

Sample-wise unit-mean weights redistribute a fixed gradient budget; 10% action dropout keeps every token trainable.

04 — GENERALIZE

All settings

CAER is a general objective for action-conditioned world models. This paper validates it on four mainstream AC-WM settings with distinct control interfaces, where it improves over matched uniform MSE in each case.

A causal contrast inside
the training loop.

CAER framework: uniform MSE versus causal action effect reweighting. — **Figure 1.** Contrast → localize → reweight.click to enlarge

with action

v_\theta(z_{\tau_S};c_A)

observed action

null action

v_\theta(z_{\tau_S};c_{\emptyset})

action masked

effect map

S=\left\|v_\theta(z_{\tau_S};c_A)-v_\theta(z_{\tau_S};c_{\emptyset})\right\|_2

large difference = action-sensitive

$\mu^{(b)}(\rho)=1$

Uniform MSE sees pixels.
CAER sees consequences.

01 / 08LIBERO

scene

Reference frame for opening the middle drawer

Uniform MSE

CAER

02 / 08LIBERO

scene

Reference frame for pushing a plate to the stove

Uniform MSE

$Uniform MSE frame 7$

CAER

03 / 08RoboTwin

scene

Reference frame for ranking blocks by size

Uniform MSE

CAER

04 / 08RoboTwin

scene

Reference frame for placing a container on a plate

Uniform MSE

CAER

05 / 08Camera Control

scene

Reference frame for orbit-right camera control

Uniform MSE

CAER

06 / 08Camera Control

scene

Reference frame for outdoor camera trajectory

Uniform MSE

CAER

07 / 08PoseAnything

scene

Reference frame for rabbit pose-conditioned generation

Uniform MSE

CAER

08 / 08PoseAnything

scene

Reference frame for fish pose-conditioned generation

Uniform MSE

CAER

It learns the easy pixels first.
Then the hard interaction.

Uniform MSE cleans easy background first. CAER catches up when its effect map sharpens.

Self-improving loop

Better model → sharper map → harder tokens.

Training curves showing CAER overtaking uniform MSE on camera control and RoboTwin as training progresses. — **Figure 2.** CAER overtakes after the signal matures.click to enlarge

Experiment

Camera control · iWorld-Bench

All reported camera-control metrics.

Task	Objective	Aggregate Avg.	Generation quality				Trajectory following
Task	Objective	Aggregate Avg.	Image quality	Brightness consistency	Color temp. constraint	Sharpness retention	Motion smoothness	Trajectory accuracy
Camera Control	Uniform MSE	0.6412	0.6898	0.5513	0.5196	0.4752	0.9902	0.6211
	CAER	0.6614	0.6804	0.6693	0.6627	0.4150	0.9934	0.5474

Robot manipulation · WorldArena

Full EWMScore + 16 metrics, split into two readable tables.

Aggregate, visual, motion, and content metrics

Task	Objective	EWM score	Visual quality			Motion quality			Content consistency
Task	Objective	EWM score	Image quality	Aesthetic quality	JEPA similarity	Dynamic degree	Flow score	Motion smoothness	Subject consistency	Background consistency	Photometric consistency
LIBERO	Uniform MSE	57.66	0.3655	0.4950	0.5472	0.1383	0.0388	0.4853	0.6050	0.6280	0.6829
	CAER	61.79	0.3694	0.5076	0.5639	0.1623	0.0549	0.5020	0.7624	0.7915	0.8397
RoboTwin	Uniform MSE	62.35	0.5200	0.3895	0.8323	0.4667	0.2738	0.7930	0.8234	0.8967	0.1232
	CAER	63.13	0.5502	0.3879	0.8187	0.5268	0.3251	0.8231	0.8125	0.8957	0.0986

Physics, 3D accuracy, and controllability metrics

Task	Objective	Physics adherence		3D accuracy		Controllability
Task	Objective	Interaction quality	Trajectory accuracy	Depth accuracy	Perspectivity	Instruction following	Semantic alignment
LIBERO	Uniform MSE	0.5900	0.8111	0.9805	0.8080	0.5520	0.9207
	CAER	0.5885	0.8475	0.9793	0.8077	0.5615	0.9300
RoboTwin	Uniform MSE	0.6505	0.2781	0.8093	0.9071	0.6929	0.8955
	CAER	0.6620	0.2610	0.8212	0.9180	0.6860	0.8825

Pose-conditioned generation · VBench

All six VBench dimensions.

Task	Objective	Aggregate	Subject consistency	Background consistency	Motion smoothness	Dynamic degree	Aesthetic quality	Imaging quality
PoseAnything	Uniform MSE	0.7422	0.8460	0.9273	0.9686	0.74	0.4348	0.5366
	CAER	0.7746	0.8828	0.9202	0.9608	0.86	0.4269	0.5971

Hyperparameter sensitivity

RoboTwin + camera control. Purple = default / best.

Factor	Value	WorldArena ↑	iWorld-Bench ↑
Action dropout	5%	62.83	0.5447
	10%	63.13	0.6614
	15%	61.91	0.5797
	20%	61.40	0.5470
Fixed time $ \tau_S $	0.25	61.09	0.5652
	0.50	63.13	0.6614
	0.75	62.27	0.6167

Citation

@article{caer,
  title   = {CAER: Causal Action Effect Reweighting for World Model Training},
  author  = {Jianjie Fang and Xvyuan Liu and Ziyou Wang and Rongze Tang and Zhaolu Wang and Zhuohang Li and Xin Zhang and Haisheng Su and Chen Gao and Wei Wu and Xinlei Chen and Yong Li},
  year    = {2026},
  url     = {https://arxiv.org/abs/2608.30897}
}

CausalActionEffectReweighting

Authors

Affiliations

Date

Abstract

A small change to the objective.

What uniform MSE misses

Action-sensitive tokens

Focus without inflation

All settings

A causal contrast insidethe training loop.

Uniform MSE sees pixels.CAER sees consequences.

It learns the easy pixels first.Then the hard interaction.

Experiment

Camera control · iWorld-Bench

Robot manipulation · WorldArena

Pose-conditioned generation · VBench

Hyperparameter sensitivity

Citation

Causal
Action
Effect
Reweighting

A causal contrast inside
the training loop.

Uniform MSE sees pixels.
CAER sees consequences.

It learns the easy pixels first.
Then the hard interaction.