Embodied video world models predict scene changes caused by robot actions, yet generated object responses may remain inconsistent with task intent and contact dynamics. Critical responses in embodied scenes often concentrate in local regions and specific time intervals. Inspired by chain-of-thought reasoning, we introduce ReactWorld, an embodied video generation model guided by structured interaction anticipation. In the generation backbone's feature space, an interaction response field predictor anticipates how the interaction will unfold, explicitly describing participant roles, future motion response intensity, and local motion changes. A response feedback adapter forms increments to object features from the difference between the object branch's outputs with predicted and zeroed local dynamics inputs. It also converts predicted displacements into positional biases for temporal attention. These two mechanisms guide object feature updates and information retrieval across frames, respectively. Through interaction anticipation and generative feedback, ReactWorld achieves an average score of 77.51 on RBench, outperforming PhysisForcing by 3.29 points with improvements in all nine categories. On EZS-Bench, it achieves a Domain score of 86.98 and an average score of 81.87.
Figures & tables
Figure 2: Overview of ReactWorld. Top: response prediction and generative feedback within each denoising forward pass. Bottom: (a) predicting roles, future responses, and interval dynamics to construct interaction context and local dynamic conditions; (b) extracting object-response increments relative to a zero-dynamics reference and updating features by role; and (c) converting gated, clipped displacements into positional biases within a fixed candidate window. Biases are added to query–key logits before softmax and feature aggregation, followed by bounded residual feedback. Scene and response visualizations illustrate the mechanism.
Figure 3: Structured response predictions and displacement biases. (a) The future response field referenced to the first frame, together with object roles, motion support, and displacement at frames 0, 56, and 76, overlaid on the corresponding RGB frames. (b) The same query and a shared 7×7 candidate window for the interval from frame 56 to 60, showing attention without displacement bias, the displacement bias, and attention with bias. Red and green circles indicate the query and displacement endpoint, respectively.
Method
Avg. ↑
Task category ↑
Robot morphology ↑
Manipulation
Spatial
Multi- entity
Long- horizon
Reasoning
Single arm
Dual arm
Quadruped
Humanoid
Wan2.2-A14B ∗ [ 22 ]
69.80
70.90
66.00
92.10
68.10
55.00
68.80
67.80
67.00
72.80
Cosmos3-Nano [ 17 ]
66.33
69.55
78.00
73.40
56.98
63.25
61.65
66.95
60.80
66.35
ABot-PhysWorld [ 9 ]
57.68
64.40
59.50
67.00
42.93
38.00
61.25
63.60
59.55
62.85
PhysisForcing [ 31 ]
74.21
75.65
80.00
94.20
74.73
78.13
67.75
67.60
60.85
69.00
Video-FT
71.94
73.50
71.50
91.90
70.80
63.80
68.20
68.40
66.90
72.50
Table 1: RBench category scores. ∗ denotes results from the official Qwen leaderboard.
Figure 4: Embodied task generation across methods. Columns show Wan2.2, Cosmos3-Nano, ABot-PhysWorld, PhysisForcing, and ReactWorld. Rows show stacking cups, dispensing jam, bagging apples, and opening an oven.
Method
Domain ↑
Quality ↑
Avg. ↑
Wan2.2-A14B [ 22 ]
77.42
76.89
77.16
Cosmos3-Nano [ 17 ]
84.32
75.60
79.96
ABot-PhysWorld [ 9 ]
83.66
76.94
80.30
PhysisForcing [ 31 ]
84.49
76.58
80.54
Video-FT
83.45
76.64
80.05
ReactWorld
86.98
76.75
81.87
Table 2: EZS-Bench196 results.
Method
Task progress ↑
Contact plausibility ↑
Visual quality ↑
Video-FT
3.74 ( p=0.063 )
3.24 ( p<0.001 )
3.92 ( p=0.211 )
PhysisForcing
3.70 ( p=0.006 )
3.78 ( p=0.140 )
4.16 ( p=0.749 )
ReactWorld
4.08
4.06
4.12
Table 3: Human evaluation. We report mean ratings on a five-point scale. Parentheses give Holm-adjusted p -values from paired t -tests against ReactWorld.
Variant
RBench
EZS-Bench
PSS ↑
TAC ↑
Domain ↑
Quality ↑
Video-FT
4.01
3.81
83.45
76.64
Heads-only
4.13
3.90
84.11
76.66
Heads+Adapter
4.24
4.12
84.73
76.69
Full w/o bias
4.37
4.15
85.38
76.71
Full (ReactWorld)
4.62
4.20
86.98
76.75
Table 4: Ablation of response supervision and generative feedback.
Variant
RBench
EZS-Bench
PSS ↑
TAC ↑
Domain ↑
Quality ↑
Full
4.62
4.20
86.98
76.75
Whole-entity context
4.31
4.14
84.85
76.71
No zero reference
4.43
4.12
85.12
76.68
Amplitude-matched
4.51
4.16
85.95
76.72
Table 5: Ablation of local responses and the zero-dynamics reference.
Figure 5: Directional interventions on feedback displacement. (a) Normal feedback. (b,c) Leftward and rightward offsets applied to the feedback displacement input shift the robot arm’s motion in the corresponding directions.
Figure 6: Displacement bias strength. Filled markers denote the default setting. A multiplier of zero disables positional bias while retaining object feature updates and content-based temporal attention.
Figure 7: Insertion depth. (a) Input image and PCA visualizations of backbone features at the same forward call and time position. (b) EZS-Bench Domain scores for different insertion blocks.