Recently, video world models have advanced interactive generation, with large models offering higher quality and smaller models enabling faster inference. Existing speculative methods combine their strengths through fast draft generation and selective target regeneration guided by rewards on decoded frames. However, verification covers only the external scorer's evaluation dimensions and adds decoding and scoring costs. Moreover, these scores provide no correction, requiring costly target regeneration for rejected drafts. To address these limitations, we propose Speculative World Model (SpecWM), a unified, training-free inference framework for streaming world models. It verifies draft chunks using the target's latent velocity field. Verification also yields a reusable correction; two thresholds select Accept, Refine, or Reject. Specifically, one-step refinement reuses it to improve perceptual quality, while three-step refinement strengthens camera following and trajectory consistency. These findings motivate transition-aware refinement. On WBench Navigation, SpecWM improves visual quality at 1.63× speedup over the target and raises NavScore from the draft's 75.99 to 85.63, achieving the best quality-speed-control trade-off. Additionally, our LongSpecBench shows sustained quality and control over long, multi-turn rollouts, while experiments on Hy-Nomas validate SpecWM across backbones and denoising budgets.
Figures & tables
Figure 2: SpecWM overview. Top: decoded-frame reward verification and SpecWM’s latent verification with transition-aware refinement. Main: a fixed final-stage target probe supplies a score and a reusable clean prediction. Two thresholds select Accept, Refine, or Reject. One-step Refine reuses the prediction; longer corrections follow the 3→2→1 schedule near control transitions. Fig. 3 details overlapped execution and re-proposal. Right: selected quality and control scores, and speedups of Draft-only, Target-only, and SpecWM on WBench Navigation and LongSpecBench (Tables 1 and 3 ), allowing direct comparison across both evaluation settings.
Figure 3: Overlapped speculative execution. Top: Draft-only and Target-only rollouts. Bottom: SpecWM drafts the next chunk while verifying the current one. Accept preserves the next proposal; committing a refined or regenerated chunk invalidates it. Dashed proposals are discarded and re-proposed after rollback, using the updated history that includes the committed output.
Efficiency
Routing
Visual ↑
Consistency ↑
Camera ↑
Method
Speedup ↑
A/F/R/T (%)
HPSv3
Imaging
Aesth.
Plaus.
Backgr.
Spatial
NavScore
Nav. Acc.
Nav. Cons.
Single-model references
Target-only (14B)
1.00 ×
—
74.57
66.82
66.38
61.69
92.46
79.12
85.49
83.76
87.21
Draft-only (1.3B)
4.35 ×
—
72.49
64.67
66.66
60.07
91.67
69.29
75.99
73.07
78.91
Uniform policies
Accept-all
2.50 ×
100/0/0/0
72.71
64.67
66.16
60.07
91.67
69.29
75.94
73.13
78.74
Table 1: Results on the full 158-case WBench Navigation split with the LingBot-World 2.0 1.3B draft and 14B target. Quality and control scores are multiplied by 100. Parentheses specify refinement length; 3→2→1 is the full transition-aware schedule. Speedup is relative to Target-only. A/F/R/T gives chunk percentages for Accept, Refine, Reject, and target-only fallback. All SpecWM variants retain fallback and overlapped execution. Bold and underline mark the best and second-best values in each performance column.
Figure 4: WBench quality-speed-control comparison. Top: Imaging versus NavScore; dotted arrows indicate increasing refinement length within each policy family. Bottom: generation-loop speedup over Target-only. Longer fixed refinements improve navigation but reduce speed. In comparison, SpecWM achieves the highest Imaging and NavScore at 1.63 × speedup (Table 1 ).
Figure 5: Qualitative comparison under changing controls. Target-only, Draft-only, SDVG-style, and SpecWM share the inputs and control sequence in each case. Control inputs are shown below the frames. Highlighted regions mark control failures and geometry collapse. SpecWM follows the changing controls while preserving scene structure across both sequences.
Variant
Speedup ↑
HPSv3 ↑
Imag. ↑
Aesth. ↑
Spatial ↑
Nav. ↑
SpecWM ( 3→2→1 )
1.63 ×
77.66
67.19
68.19
80.83
85.63
Refinement schedule
SpecWM (1-step)
1.84 ×
76.73
65.91
67.19
70.09
76.39
SpecWM ( 3→1 )
1.67 ×
76.39
66.02
67.21
76.90
81.49
Routing
w/o Refine (A/R)
1.21 ×
75.01
66.10
67.08
75.35
82.18
Table 2: SpecWM ablations on the same 158-case WBench Navigation split. Scores are multiplied by 100; Nav. denotes NavScore. All variants use the same final-stage probe. Random routing matches route shares; synchronous execution disables draft-target overlap while retaining the full policy. Bold / underline marks the best/second-best value per column.
Efficiency
Visual ↑
Consistency ↑
Control ↑
Composite ↑
Method
Speedup ↑
Imaging
Aesth.
Long-T.
Revisit
Dir.
Resp.
Cam.
WMQ
QES
Target-only (14B)
1.00 ×
67.62
64.47
90.47
76.77
90.92
83.41
18.87
71.62
66.65
Draft-only (1.3B)
4.26 ×
66.95
64.10
88.83
66.40
85.42
72.00
16.60
67.13
69.69
Refine-all (3-step)
1.08 ×
67.46
63.71
91.05
74.75
90.79
82.83
19.81
71.26
66.89
SDVG-style
1.51 ×
67.50
63.89
87.71
67.04
85.29
73.97
17.08
67.36
65.86
SpecWM (3-step)
1.21 ×
66.78
64.03
91.39
75.27
90.35
83.32
19.21
71.29
67.62
Table 3: LongSpecBench results on the 80-case evaluation split with 509-frame, multi-turn rollouts. Scores use [0,100] ; generation-loop Speedup is relative to Target-only. WMQ and QES aggregate the displayed scores as defined in Sec. 5.4 . Bold and underline mark the best and second-best values per column.
Video world models predict future observations conditioned on historical observations and control signals, enabling long-horizon generation through autoregressive state transitions. Unlike conventional video generation models that primarily capture visual appearance and motion, video world models learn the underlying dynamics governing environment evolution under agent actions, providing a foundation for embodied AI and interactive simulation. Recent progress has largely relied on adapting pretrained video generation models through post-training or distillation. Although effective, these approaches often require complex training pipelines, substantial computational resources, and suffer from the mismatch between bidirectional pretraining and causal streaming inference. Recent studies have shown that training autoregressive video world models from scratch is feasible and scalable. However, the community still lacks a lightweight, transparent, and fully reproducible baseline trainable end-to-end with modest computational resources. We present MiniWorld, a reproducible framework for training streaming video world models from scratch. MiniWorld employs a block-causal Video Diffusion Transformer trained with Flow Matching in the latent space of a pretrained Video VAE. Building on Diffusion Forcing, it adopts a chunk-wise non-decreasing noise schedule and two-stage continued training to improve temporal modeling and stability. During inference, MiniWorld combines a rolling KV cache with pipelined asynchronous denoising for efficient streaming generation under bounded computation. The entire model can be trained within several days on a single 8-GPU server. By releasing the training and inference codebase and pretrained checkpoints, we hope MiniWorld will facilitate future research on video world modeling.
Autoregressive video world models enable interactive, long-horizon exploration, but flexible control remains challenging. Exploring a source video from new viewpoints requires the generated rollout to remain synchronised with the recorded event, place observed content in the requested view, plausibly complete newly exposed regions, and recover previously generated appearance on revisits. Existing methods typically address these requirements through task-specific modules or additional training. We present World in World, a training-free inference-time interface that converts heterogeneous control evidence into camera- and time-labelled clean visual states, which are read through the native self attention of a frozen causal video model. The evidence comprises source-video observations, target-view scene projections, geometry renderings that guide completion of newly exposed subject regions, and retrieved generated states beyond the rolling cache. Each evidence source carries token-level support and its own availability schedule. A correspondence router combines persistent point identities with geometry to establish token correspondences, guiding supported queries towards matching source-video tokens. Evidence-wise attention CFG (EWA) then independently regulates each auxiliary channel's additional contribution using attention responses from the same denoising forward pass. The shared interface supports camera-controlled rerendering, long-horizon revisiting, and human-motion transfer with the same frozen backbone. We evaluate World in World on camera-controlled video rerendering under diverse viewpoint changes, assessing perceptual quality, temporal consistency, and camera-following accuracy.
We present EditWorld, a video world model for precise editing and flexible referencing in interactable worlds. Existing video world models primarily focus on navigation, letting users explore generated worlds but offering limited control over how existing world content is modified. EditWorld extends world modeling from exploration to precise modification by streaming editing instructions and reference images during autoregressive generation. To support these capabilities, EditWorld introduces Gated Causal Attention for temporally varying editing conditions and reference images, together with a Sparse Context mechanism that maintains a bounded historical context for long-horizon inference. We further adopt joint autoregressive and bidirectional training with annealed self-resampling, and construct a dedicated data synthesis and annotation pipeline that provides supervision for world editing. We also present WBench-Editing to systematically evaluate streaming world editing capabilities. EditWorld achieves the best overall performance on WBench-Editing with an overall score of 73.8 and an editing score of 80.0, substantially outperforming existing methods on editing-related metrics. https://github.com/leoisufa/EditWorld