FutureWorld: A Plug-and-Play Causal Next-State Predictor for Interactive Video World Exploration

Guangyuan Li; Kelu Yao; Shuolin Xu; Xinlei Yu; Wei Xing; Lei Zhao; Peng-Tao Jiang; Xiaogang Xu

cs.CVOct 29, 2026

FutureWorld: A Plug-and-Play Causal Next-State Predictor for Interactive Video World Exploration

Authors: Guangyuan Li, Kelu Yao, Shuolin Xu, Xinlei Yu, Wei Xing, Lei Zhao, Peng-Tao Jiang, Xiaogang Xu,

Organizations: Zhejiang University · National University of Singapore · Zhejiang Lab · Bournemouth University ·

Abstract

Interactive video world models aim to follow user actions while maintaining scene consistency over extended interactions. History-based methods reuse generated latents to preserve scene content, but three limitations of this reuse contribute to error recirculation. 1) The latent generated at each interaction is retained as history, potentially containing blurred details, geometric distortions, or semantic drift. 2) Although retrieved historical content matches the current scene, it can still contain errors that are carried into the next chunk. 3) New outputs then become history for later interactions, potentially reinforcing the same errors through repeated reuse. To limit this cycle, we aim to predict the future scene state from the current generated latent and the next action before generating the next interaction's video chunk. This future-state prior provides guidance for the next chunk, with the aim of reducing the errors carried forward from history. To this end, we propose a plug-and-play Causal Next-Chunk Predictor (C-NCP) with cross-model adaptability and prediction-guided modulation. We integrate C-NCP into four base generators and evaluate the augmented models across three benchmarks for interactive long-video generation and action-conditioned world modeling. Extensive experiments demonstrate reduced error accumulation and improved long-horizon consistency across diverse generation settings. The project page is at https://GuangYuanKK.github.io/futureworld.

Figures & tables

Figure 1: Improving long-horizon consistency through explicit next-state prediction. (a) Reusing generated history can carry existing errors into subsequent generations, allowing scene distortions and semantic drift to accumulate across interactions (red dashed boxes). (b) Our plug-in predictor estimates the next latent state from the current generated latent and the next action. This future-state prior guides the autoregressive video DiT through feature modulation, helping preserve scene structure and semantic consistency as interactions proceed (green dashed boxes).

Figure 2: Overview of FutureWorld. Panel a presents the Causal Next-Chunk Predictor, which combines VAE-specific adapters, action conditioning, and a causal spatiotemporal backbone to predict the complete next latent chunk. Panel b illustrates next-chunk prediction training, where the predictor learns to forecast subsequent latent chunks from input chunks and control conditions under supervision from corresponding ground-truth chunks. Panel c shows autoregressive inference, where the predicted future state guides a frozen autoregressive video DiT through modulation to generate the next latent chunk.

Method	T.F. $\uparrow$	M.S. $\uparrow$	S.C. $\uparrow$	B.C. $\uparrow$	A.Q. $\uparrow$	I.Q. $\uparrow$	Avg. $\uparrow$
YUME 1.5	0.9127	0.9151	0.6928	0.9036	0.4128	0.7132	0.7584
FantasyWorld	0.9149	0.9128	0.6975	0.9020	0.4149	0.7104	0.7588
Infinite-World	0.9112	0.9170	0.6902	0.9067	0.4116	0.7160	0.7588
SANA-WM	0.9246	0.9764	0.7994	0.9034	0.5050	0.7401	0.8081
SANA-WM + Ours	0.9366	0.9867	0.8437	0.9202	0.5192	0.7532	0.8266
Matrix-game-3.0	0.9241	0.9748	0.7946	0.9042	0.5025	0.7740	0.8124

Table 1: Quantitative comparison on RealEstate10K-test using VBench. All videos are 60 seconds long. Higher is better for every metric. Bold rows denote base generators augmented with our predictor. T.F.: Temporal Flickering; M.S.: Motion Smoothness; S.C.: Subject Consistency; B.C.: Background Consistency; A.Q.: Aesthetic Quality; I.Q.: Imaging Quality; Avg.: Average.

Figure 3: Qualitative comparisons over 60-second interactive rollouts with four base generators. Each pair shows the original generator above and its predictor-augmented counterpart below, with six frames ordered from left to right. Action controls appear in the lower-left corners, and cyan dashed boxes highlight subjects and scene regions for comparison. The examples show improved preservation of subject appearance and scene layout with our predictor.

Figure 4: The upper panel compares the generator’s main latent before and after modulation, with the predicted future state shown in blue. The lower panel compares main latents obtained with historical retrieval and predictor-guided modulation.

Variant	Causal	Cond.	Temp.	Concat	Mod.	T.F. $\uparrow$	M.S. $\uparrow$	S.C. $\uparrow$	B.C. $\uparrow$	A.Q. $\uparrow$	I.Q. $\uparrow$	Avg. $\uparrow$
Baseline	$\times$	$\times$	$\times$	$\times$	$\times$	0.9241	0.9748	0.7946	0.9042	0.5025	0.7740	0.8124
Bidirectional Predictor	$\times$	$\checkmark$	$\checkmark$	$\times$	$\checkmark$	0.9307	0.9823	0.8189	0.9134	0.5103	0.7815	0.8229
w/o Condition	$\checkmark$	$\times$	$\checkmark$	$\times$	$\checkmark$	0.9295	0.9789	0.8144	0.9113	0.5103	0.7809	0.8209
w/o Temporal Loss	$\checkmark$	$\checkmark$	$\times$	$\times$	$\checkmark$	0.9283	0.9789	0.8312	0.9180	0.5139	0.7850	0.8259
Direct Concat	$\checkmark$	$\checkmark$	$\checkmark$	$\checkmark$	$\times$	0.9332	0.9850	0.8300	0.9180	0.5110	0.7848	0.8270
Full Ours	$\checkmark$	$\checkmark$	$\checkmark$	$\times$	$\checkmark$	0.9361	0.9884	0.8387	0.9210	0.5167	0.7877	0.8314

Table 2: Component ablation of the predictor with Matrix-game-3.0 as the base generator on RealEstate10K-test using VBench. All videos are 60 seconds long. Higher is better for every metric. The full method is highlighted in bold. Causal: causal prediction; Cond.: control conditioning; Temp.: temporal loss; Concat: direct concatenation; Mod.: modulation.

Metric	Baseline	w/ Predictor
Total (s)	26.707	29.174
Avg. (s/interaction)	6.677	7.294
Extra (s/interaction)	–	0.617
Overhead	–	9.24%

Table 3: Inference-time comparison of Matrix-game-3.0 with and without the predictor when generating four 720P video chunks on an A800 GPU. Avg. denotes the time per interaction.

Appendix figures & tables5 assets

Supplementary material from the paper’s appendix.

Appendix

Method	T.F. $\uparrow$	M.S. $\uparrow$	S.C. $\uparrow$	B.C. $\uparrow$	A.Q. $\uparrow$	I.Q. $\uparrow$	Avg. $\uparrow$
YUME 1.5	0.9002	0.9321	0.7115	0.8921	0.4270	0.6877	0.7584
FantasyWorld	0.9024	0.9298	0.7162	0.8905	0.4291	0.6849	0.7588
Infinite-World	0.8987	0.9340	0.7089	0.8952	0.4258	0.6905	0.7588
SANA-WM	0.9413	0.9713	0.7364	0.9119	0.4522	0.7237	0.7895
SANA-WM + Ours	0.9534	0.9850	0.7804	0.9288	0.4743	0.7371	0.8098
Matrix-game-3.0	0.9360	0.9708	0.7600	0.9104	0.4535	0.7555	0.7977

Appendix

Table A1: Quantitative comparison on Sekai-val using VBench. All videos are 60 seconds long. Higher is better for every metric. Bold rows denote base generators augmented with our predictor. T.F.: Temporal Flickering; M.S.: Motion Smoothness; S.C.: Subject Consistency; B.C.: Background Consistency; A.Q.: Aesthetic Quality; I.Q.: Imaging Quality; Avg.: Average.

Method	Qual. $\uparrow$	Set. $\uparrow$	Inter. $\uparrow$	Cons. $\uparrow$	Phys. $\uparrow$	Avg. $\uparrow$
YUME 1.5	77.6	72.4	72.0	80.1	65.2	73.5
FantasyWorld	72.4	71.3	72.1	86.4	66.8	73.8
Infinite-World	77.0	69.3	75.9	80.0	62.1	72.9
DreamX-World	77.5	80.8	78.4	74.9	63.3	75.0
SANA-WM	79.3	76.1	82.1	80.7	61.9	76.0
SANA-WM + Ours	80.3	77.1	83.7	85.2	63.1	77.9

Appendix

Table A2: Quantitative comparison on the official WBench test set using 60-second videos. Higher is better for every metric. Bold rows denote base generators augmented with our predictor. Qual.: Quality; Set.: Setting; Inter.: Interaction; Cons.: Consistency; Phys.: Physical; Avg.: Average.

Variant	Stage I Input	Stage II Input	T.F. $\uparrow$	M.S. $\uparrow$	S.C. $\uparrow$	B.C. $\uparrow$	A.Q. $\uparrow$	I.Q. $\uparrow$	Avg. $\uparrow$
One-stage Pred	Pred	–	0.9290	0.9785	0.8110	0.9110	0.5050	0.7831	0.8196
One-stage GT	GT	–	0.9315	0.9825	0.8205	0.9145	0.5090	0.7840	0.8237
One-stage Pred+GT	Pred+GT	–	0.9330	0.9842	0.8268	0.9168	0.5118	0.7846	0.8262
Two-stage GT $\rightarrow$ Pred	GT	Pred	0.9345	0.9860	0.8320	0.9185	0.5135	0.7860	0.8284
Full Ours	Pred+GT	Pred	0.9361	0.9884	0.8387	0.9210	0.5167	0.7877	0.8314

Appendix

Table A3: Training-strategy ablation with Matrix-game-3.0 on RealEstate10K-test using VBench. All videos are 60 seconds long. Stage inputs specify inter-chunk feedback; Pred+GT denotes element-wise addition. Higher is better for every metric, and the full method is highlighted in bold. T.F.: Temporal Flickering; M.S.: Motion Smoothness; S.C.: Subject Consistency; B.C.: Background Consistency; A.Q.: Aesthetic Quality; I.Q.: Imaging Quality; Avg.: Average.

Figure 5: Additional qualitative comparisons over 60-second interactive rollouts in desert and coastal scenes. Each pair shows the base generator above and its predictor-augmented counterpart below, with six snapshots ordered from left to right. Action controls appear in the lower-left corners, and cyan dashed boxes highlight the camel in the desert example.

Figure 6: User-study preferences across four base generators on scene consistency, action adherence, temporal consistency, visual quality, and overall preference. Blue and red bars show the preference percentages for the base model and its predictor-augmented counterpart, respectively. The dashed line marks equal preference at 50%.

Interactive World Models

Interactive world models predict and generate how environments change in response to actions, supporting simulation, planning, and control.

See all 122 papers

Next Forcing: Causal World Modeling with Multi-Chunk Prediction
Jun 9, 2026Gangwei Xu, Qihang Zhang, Jiaming Zhou +4Causal World Models Autoregressive Video Generation
Robbyant · HUST · HKUST (GZ) +1
A Plausible Future Is Not a Causal Future: Can Video Models Generate How the World Unfolds?
Oct 11, 2026Jialiang Li, Zhengning Xu, Zongyang Ma +1Causal Video Generation Video Generation
DeepDelta Lab, School of Artificial Intelligence, Shanghai Jiao Tong University · Xi’an Jiaotong-Liverpool University
Mix Forcing: Causal Pre-Training via Mixture Attention for Interactive World Models
Oct 24, 2026Zelin Gao, Hao Ouyang, Jiapeng Zhu +6Video Generation Causal Video Generation
The Hong Kong University of Science and Technology · Robbyant

Jun 9, 2026J/K move · Enter open · S save

Next Forcing: Causal World Modeling with Multi-Chunk Prediction

Gangwei Xu, Qihang Zhang, Jiaming Zhou, Xing Zhu+3

Robbyant · HUST · HKUST (GZ)+1

Causal World Models Autoregressive Video Generation Video World Models Causal Video Generation

Open paper AskPDF arXiv

Autoregressive video generation has emerged as a powerful paradigm for World Action Models (WAMs). However, existing approaches suffer from slow training convergence and limited converged accuracy, particularly at high frame rates, as the training supervision is confined to the current chunk without explicit signals about future dynamics; they also suffer from slow inference due to iterative video denoising. In this paper, we present Next Forcing, a multi-chunk prediction (MCP) framework for causal world modeling that enables faster training, higher accuracy, and accelerated inference. Inspired by multi-token prediction in large language models, Next Forcing introduces an MCP training objective that augments the main model with lightweight auxiliary MCP modules to simultaneously denoise video chunks at multiple future temporal horizons (next

^1

, next

^2

, next

^3

chunks). These MCP modules form a causal chain across prediction depths, where intermediate features fused from multiple layers of the main model are leveraged to predict future dynamics, allowing near-future predictions to inform farther-future ones and providing dense multi-scale temporal supervision back to the main model. During training, the MCP modules significantly accelerate convergence and improve converged accuracy, especially at high frame rates: at 50 fps, Next Forcing achieves a 93.1% relative improvement over LingBot-VA at 5k training steps and 2.3x faster convergence, and establishes new state-of-the-art results on the RoboTwin benchmark (94.1/93.5% on Clean/Random). At inference, the MCP modules can be retained to predict the next video chunk in parallel with the current one, achieving 2x inference acceleration. Next Forcing also demonstrates significant improvements on PhyWorld, a benchmark evaluating adherence to physical laws in video generation, and over 50% FVD reduction on general video pretraining.

Abstract

Interactive World Models

Explore similar work

Next Forcing: Causal World Modeling with Multi-Chunk Prediction

A Plausible Future Is Not a Causal Future: Can Video Models Generate How the World Unfolds?

Mix Forcing: Causal Pre-Training via Mixture Attention for Interactive World Models