IMPACT Attention Is the Interaction Map forScalable Interaction-Aware World Model Training

Rongze Tang^1,2,*, Jianjie Fang^3,*, Zhaolu Wang³, Ziyou Wang³, Xvyuan Liu³, Haisheng Su⁴, Xin Zhang⁴, Wei Wu⁴, Chen Gao^2,3,†, Yong Li³, Zhibo Chen^1,2,†

¹USTC ²Zhongguancun Academy ³Tsinghua University ⁴Manifold AI

* Equal contribution † Corresponding authors

Paper arXiv Code

Abstract

World models have made remarkable progress in action-conditioned future prediction for embodied agents, yet still struggle to model physically plausible interactions. Existing approaches address this limitation by constraining generation with external representations encoding motion, geometry, or semantics, whose construction depends on auxiliary estimators or manual annotations.

We identify a supervision-allocation mismatch under the globally averaged MSE denoising objective: prevalent static content dominates optimization, while sparse dynamic-object regions are disproportionately under-supervised. IMPACT uses manipulated-object cross-attention as an internal spatiotemporal prior, calibrates sampled candidates with detached local prediction errors, and targets denoising supervision with the resulting interaction map.

Contributions

Attention Is the Interaction Map.

01 — PROBLEM

Uniform supervision spends gradient on what is abundant.

Standard MSE spreads its signal across the frame. IMPACT reallocates supervision toward sparse object motion and contact.

Ground TruthStandard MSEIMPACT

Robot armStacking blocks

Human handWhite box

Method

Calibrate in the forward pass.
Target in the backward pass.

Object-token groundingIdentify “blue bowl” as the manipulated object.

Results

Consistent across backbones and control signals.
Effective across embodiments.

WorldArena62.53EWMScore · Cosmos + IMPACT

WorldArena+3.81EWMScore · Wan 2.2-AC

EgoDex110.94FVD ↓

EgoDex0.772Hand IoU ↑

Robot-arm manipulation

WorldArena

Model	EWMScore	Visual	Motion	Content	Physics	3D	Control
Wan 2.6	61.86	61.62	68.31	60.36	42.31	75.88	60.85
CtrlWorld	59.70	55.33	50.28	63.99	54.89	86.30	54.65
WoW	54.88	52.98	50.02	64.52	38.11	74.77	49.89
Wan 2.2-AC	58.65	56.57	48.23	60.55	53.41	86.16	54.40
Wan 2.2-AC + IMPACT	62.46	60.60	53.28	60.29	55.87	92.56	60.00
Cosmos-Predict 2.5 (action)	55.91	57.87	38.89	68.73	42.23	82.53	49.51
Cosmos-Predict 2.5 (action) + IMPACT	62.53	56.32	68.22	57.90	41.82	77.99	71.19

Human-hand manipulation

EgoDex

Model	FVD ↓	FID ↓	CLIP-Hand ↑	Hand IoU ↑
VACE	358.42	50.65	0.895	0.493
Wan 2.2-AC	366.12	44.71	0.921	0.693
Wan 2.2-AC + IMPACT	110.94	5.79	0.952	0.772

Component ablation

WorldArena

Configuration	EWMScore	Visual	Motion	Physics
Wan 2.2-AC	58.65	56.57	48.23	53.41
+ IWS	61.54	60.44	49.16	55.16
+ IWS + ADS	62.46	60.60	53.28	55.87

Gallery

Contact stays coherent.
Objects follow the action.

Robot arm

WorldArena

Human hand

EgoDex

Citation

@article{tang2026impact,
  title  = {IMPACT: Attention Is the Interaction Map for Scalable Interaction-Aware World Model Training},
  author = {Tang, Rongze and Fang, Jianjie and Wang, Zhaolu and Wang, Ziyou and Liu, Xvyuan and Su, Haisheng and Zhang, Xin and Wu, Wei and Gao, Chen and Li, Yong and Chen, Zhibo},
  journal = {arXiv preprint arXiv:2609.00161},
  year    = {2026},
  url     = {https://arxiv.org/abs/2609.00161}
}

IMPACT Attention Is the Interaction Map forScalable Interaction-Aware World Model Training

Attention Is the Interaction Map.

Uniform supervision spends gradient on what is abundant.

An internal prior replaces dense preprocessing.

ADS sharpens the prior; IWS delivers the gradient.

Calibrate in the forward pass.
Target in the backward pass.

Consistent across backbones and control signals.
Effective across embodiments.

Robot-arm manipulation

Human-hand manipulation

Component ablation

Contact stays coherent.
Objects follow the action.

WorldArena

EgoDex

IMPACT Attention Is the Interaction Map forScalable Interaction-Aware World Model Training

Attention Is the Interaction Map.

Uniform supervision spends gradient on what is abundant.

An internal prior replaces dense preprocessing.

ADS sharpens the prior; IWS delivers the gradient.

Calibrate in the forward pass.Target in the backward pass.

Consistent across backbones and control signals.Effective across embodiments.

Robot-arm manipulation

Human-hand manipulation

Component ablation

Contact stays coherent.Objects follow the action.

WorldArena

EgoDex

Calibrate in the forward pass.
Target in the backward pass.

Consistent across backbones and control signals.
Effective across embodiments.

Contact stays coherent.
Objects follow the action.