PAPER REPORTENAll readings ↗

Next Forcing: Causal World Modeling with Multi-Chunk Prediction

English reading report: Method, equations, original figures, experiments and reproducibility.

Authors: Gangwei Xu; Qihang Zhang; Jiaming Zhou; Xing Zhu; Yujun Shen; Xin Yang; Yinghao Xu

Affiliations: Robbyant; HUST; HKUST; HKUST (GZ)

Source: 2606.11187 ↗ · Project page ↗ · Catalog record

Reading: 174 / 558 · 6 original figures & tables · ~19 min ·

1. Paper overview

In one sentence: Multi-chunk supervision improves a causal video/action backbone, while optional two-chunk generation trades deployment accuracy against a claimed reduction in video-denoising cost. e02e05e06e09e10e12e15

At a glanceWhat to know
Research problem
Author claim

Teacher-forced denoising can exploit the visual similarity between adjacent chunks, particularly at high frame rates. The authors call this myopic supervision: reconstructing local appearance need not learn longer temporal dynamics. Their intervention changes prediction targets while retaining the causal video/action framework. Appearance copying is the proposed causal explanation, rather than a directly measured representation property. e02e03

Core mechanism
Source description

Three temporally shifted flow-matching targets supply auxiliary supervision through chained, lightweight predictors and multi-layer backbone features. e04e05e07

A key reported resultRoboTwin 50-task manipulation: Clean 94.1; Random 93.5

Average task success rate (%). Table 1; simulated bimanual tasks, Clean fixed configurations and Random object/scene randomization. Comparator checkpoint/frame-rate choices are not individually specified.

LingBot-VA 92.9/91.5; Fast-WAM 91.9/91.8 (Clean/Random). Highest among listed methods; Table 1 gains over LingBot-VA are 1.2/2.0 percentage points. Do not replace these baseline entries with Table 5 values. e08e10

Reading caution
Reader analysis

MCP adds training cost, which the authors identify as the main limitation. Step-count convergence gains therefore do not establish equal-compute or wall-clock efficiency. e16e11

Core contributions

  • Source description

    Three temporally shifted flow-matching targets supply auxiliary supervision through chained, lightweight predictors and multi-layer backbone features. e04e05e07

  • Source description

    The same trained checkpoint supports baseline-architecture deployment or parallel generation of two video chunks using the first auxiliary module. e09

  • Source description

    The paper evaluates simulated manipulation, physical-law video generation, and general-video pretraining, separating action-bearing and video-only experiments. e08e13e14

Figure 2. Future-chunk losses reach the main video backbone through fused intermediate features. Original paper, p. 3 ↗

Excerpt from the authors’ paper; cropped without altering the figure or table.

How to read it. Start at the lower left: clean context and a noisy current chunk enter the 30-block main model. Four outgoing arrows collect layers 4, 12, 20, and 30 into the fusion MLP. Follow its output into the first Concat box, where it joins the embedded noisy next chunk. A projection and three transformer blocks produce the first future prediction. The horizontal arrows then carry features into the next two modules; these are a chain, not three independent heads. Each FM loss supervises its corresponding temporal target. Equations (7)–(8) confirm these feature directions and the path for gradients back into the main model. e04e05e06e07e09

What it supports. MCP makes future prediction a training responsibility of the backbone: the auxiliary losses reach several main-model depths instead of supervising only a final output. The default design uses three future horizons, while the reported accelerated inference mode retains only the first MCP module.

Where the evidence stops. This diagram depicts the video mechanism, not the full action-decoding architecture. Section 4.4 supplies the unified video/action model. The three training branches should not be read as three extra chunks generated in the evaluated inference mode.

2. Motivation

2.1 The problem and the proposed response

Author claim

Teacher-forced denoising can exploit the visual similarity between adjacent chunks, particularly at high frame rates. The authors call this myopic supervision: reconstructing local appearance need not learn longer temporal dynamics. Their intervention changes prediction targets while retaining the causal video/action framework. Appearance copying is the proposed causal explanation, rather than a directly measured representation property. e02e03

2.2 What this reading follows

A robot world model can learn to reconstruct the next view without learning much about what changes over time. Next Forcing tackles that concern by asking lightweight auxiliary modules to denoise several future chunks during training. Their errors reach intermediate video features, while inverse dynamics still turns predicted visual futures into actions. The key reading distinction is between what improves the trained representation and what remains active during deployment. The figures below trace both paths, then connect them to simulated manipulation results, controlled design changes, and the accuracy cost of retaining an auxiliary predictor. The timing claims require more caution than the success-rate measurements. e02e05e06e09e10e12e15

3. Research context

We place the paper in the collection through its world–action interface. The catalog labels and the reading’s assessment are shown separately.

Catalog dimensionRecorded classification
Major categoryWAMs
ArchitectureDual-system
Prediction paradigmIDM
QuadrantQ4 · Dual-system × IDM

3.1 Evidence-based assessment

Conflicts with the recorded classification

Reader analysis

IDM is directly supported by Equation (9). However, the recorded Dual-system/Q4 architecture conflicts with the explicit unified Mixture-of-Transformers implementation and cross-modal attention at every layer. My assessment favors One Model × IDM on architecture evidence, not merely joint training. Auxiliary video predictors do not constitute an independent action system. The supplied catalog snapshot is preserved unchanged. e05e06

This is the collection’s architectural analysis, not a new related-work survey. Benchmark comparisons and their protocols appear in Section 6.

4. Problem formulation

4.1 Inputs and outputs

InputsOutputs
  • VAE-encoded observation chunks and language instruction
  • Past actions for the video/action factorization
  • Training-only noisy, temporally shifted future-video targets
  • Predicted video chunks
  • Inverse-dynamics action sequences conditioned on predicted future observations

4.2 Equations and their role

xt=(1t)x0+tϵ,v=ϵx0x_t=(1-t)x_0+t\epsilon,\qquad v^*=\epsilon-x_0
Here x_0 is clean data, epsilon is Gaussian noise, t is the noise level, and v* is the flow-matching velocity target. Sampling integrates the learned velocity from t=1 to t=0. e03
xi+1pθ(xi,a<i,),aigψ(xi+1,a<i,)x_{i+1}\sim p_\theta(\cdot\mid x_{\le i},a_{<i},\ell),\qquad a_i\sim g_\psi(\cdot\mid x_{\le i+1},a_{<i},\ell)
Equation (9): p_theta predicts future video and g_psi decodes actions. x_{≤i} and a_{<i} are observation/action histories; ell is the instruction. The action decoder receives the predicted future x_{i+1}. e06
L=Lvideo+Laction+k=13wkLkMCP\mathcal{L}=\mathcal{L}_{\mathrm{video}}+\mathcal{L}_{\mathrm{action}}+\sum_{k=1}^{3}w_k\mathcal{L}^{\mathrm{MCP}}_k
Video, action, and each depth-k MCP term are flow-matching losses. The future-depth weights w_1,w_2,w_3 are 0.5,0.2,0.1; replicated boundary chunks do not contribute to MCP loss. e07e08
σ~i=sσi1+(s1)σi\widetilde{\sigma}_i=\frac{s\sigma_i}{1+(s-1)\sigma_i}
Appendix C transforms 1,000 base noise levels sigma_i in [0,1] by shift s, then samples schedule indices uniformly. Larger s emphasizes noisier examples; MCP uses 10 and the main stream 5. e18

5. Method in detail

5.1 Make a future target harder to solve locally

Source description

The basic flow-matching task starts with a clean video latent, mixes it with Gaussian noise, and trains a velocity prediction toward the clean sample. Under teacher forcing, clean earlier chunks provide strong context. Next Forcing keeps that task but also shifts targets one, two, and three chunks forward. Each horizon gets independent noise and a matching temporal position. Sequence-end replication preserves the tensor layout, but the replicated targets are removed from the auxiliary loss. Appendix C then makes a consequential change: the MCP noise shift is 10 while the main model uses 5. This biases the auxiliary tasks toward noisier inputs. The authors’ explanation is that weaker target information makes the modules rely more on backbone features; improved success rates support the design, but do not directly measure this reliance. e03e04e07e08e18e12

Figure 5. The same video-token mask preserves causal access in the main model and MCP modules. Original paper, p. 17 ↗

Excerpt from the authors’ paper; cropped without altering the figure or table.

How to read it. Read each row as a query Q and each column as a key K; green permits attention and white blocks it. In the upper-left quadrant, clean queries see clean chunks at the same or earlier index. The upper-right quadrant is entirely blocked, so noisy tokens cannot influence clean context through this attention operation. The lower-left quadrant is strictly causal: a noisy query cannot read the clean token at its own index. The lower-right quadrant allows only its own noisy chunk. The right panel applies the same pattern to shifted MCP targets. The caption and Appendix A bullets agree with these cells and directions. e17e04e07e06

What it supports. The shared pattern blocks direct access to the clean target while retaining usable history and within-chunk denoising. Reusing it across the main model and all MCP depths avoids constructing a different video-attention mask for every future horizon during each training step.

Where the evidence stops. Action tokens are explicitly omitted. This figure cannot establish the full cross-modal mask. Correct target shifts, positional offsets, and exclusion of padded losses still require the separate rules in Sections 4.2–4.5.

5.2 Follow the gradient to the video model, then follow video to action

Reader analysis

The auxiliary modules do not merely supervise the final video output. Features from four backbone depths are concatenated and compressed into one fused representation. The first MCP module combines that representation with its own noisy future input; later modules use the preceding module’s features. Because auxiliary losses backpropagate through these connections, the main model can retain a benefit after every MCP component is removed. Action decoding follows a different dependency. Equation (9) first predicts a future observation, then decodes actions conditioned on that prediction and the available history. Video and action streams communicate at every transformer layer inside a unified Mixture-of-Transformers architecture. This explicitly supports inverse dynamics, while the architecture evidence motivates questioning the catalog’s Dual-system label. Training multiple targets does not itself turn the action decoder into joint future/action prediction. e05e06e07e09

5.3 Separate denoising throughput from closed-loop control

Reader analysis

In standard deployment, the auxiliary fusion and predictors disappear, leaving the main autoregressive model. In the accelerated mode, the first MCP module stays: current and next video chunks emerge within one denoising trajectory. The module still consumes main-model features, so parallel chunk generation does not mean every network block is independent. The deeper two training modules are unused. Table 4 compares the resulting success rates using the same checkpoint, and its lower-frame-rate losses show that the choice can affect control quality. Section 3.2 describes refreshing context with actual observations after action execution, but the PDF does not provide an end-to-end timing breakdown for the accelerated loop. Reader interpretation: a deployment decision needs both measured latency and task success; the two-chunk mechanism alone cannot establish a doubled physical control rate. e03e05e09e15

5.4 Training and inference

During training

Source description

Multi-embodiment pretraining precedes RoboTwin post-training on 2,500 Clean plus 25,000 Random demonstrations, up to 50k steps on 64 GPUs. Ablations instead use 2,500 Clean demonstrations, 25 fps, 20k steps, and 16 GPUs. e08

Source description

Chunk size is sampled from 1–4 frames. Main/MCP noise shifts are 5/10; noisy history augmentation occurs with probability 0.5. MCP weights initialize from the last few backbone layers. The paper does not explicitly specify whether the pretrained VAE is frozen. e03e08

During inference

Source description

Integrate flow from noise to data, predict video, decode actions, and refresh the visual context with actual observations after action execution. Removing every MCP component preserves the baseline main-model architecture; its gains then originate in training. e03e06e09

Source description

Alternatively retain depth 1: the main model and MCP module produce current/next chunks within one denoising trajectory. Depths 2–3 are unused. The claimed 2× acceleration describes two-chunk advancement; the source supplies no end-to-end timing breakdown. e09e15

5.5 Implementation flow

  1. Construct future targets

    Shift video by one, two, or three chunks, replicate the terminal chunk for padding, inject independent noise, and shift temporal RoPE positions. Exclude padded targets from the MCP losses. e04e07

  2. Fuse and chain representations

    Concatenate hidden states from layers 4, 12, 20, and 30 of the 30-layer video backbone and compress them through a two-layer MLP. Each MCP depth combines preceding features with its own noisy target embedding, projects them, and applies three transformer blocks. Gradients return through the fusion to the backbone. e05

  3. Preserve causal access

    Main and MCP video tokens share a mask: noisy queries see strictly earlier clean chunks and their own noisy chunk; clean queries see same/earlier clean chunks and no noisy tokens. Appendix A omits action-token masking. e17

  4. Decode actions

    The video stream predicts a future observation; the action stream conditions on that future and history. Layerwise cross-modal attention connects the streams within a unified Mixture-of-Transformers model. MCP supervises video directly and benefits action decoding indirectly. e06

6. Experiments & results

Next Forcing trains a causal video/action model with auxiliary losses for three future video chunks. Chained predictors teach the video backbone across temporal horizons; actions still come from inverse dynamics inside a unified transformer. RoboTwin gains are largest early in high-frame-rate training. Auxiliary modules can be removed or one retained for two-chunk generation, with an accuracy tradeoff and unmeasured end-to-end speedup.

6.1 Read the original evidence

Table 1. Next Forcing leads the listed RoboTwin comparison in both evaluation settings. Original paper, p. 7 ↗

Excerpt from the authors’ paper; cropped without altering the figure or table.

How to read it. Read across each row to compare methods within one RoboTwin setting. Clean uses fixed initial configurations; Random changes object poses and scene layouts. Each value is the average success percentage across 50 simulated bimanual manipulation tasks. The last column gives Next Forcing, and the adjacent LingBot-VA column is the architectural baseline. Section 5.1 describes training on combined Clean and Random demonstrations; the table does not give a complete matched-data or matched-compute protocol for every listed comparator. Keep these entries attached to Table 1: the appendix reports a separate checkpoint series with slightly different LingBot-VA endpoint values. e08e10e11

What it supports. The displayed comparison is 94.1 versus 92.9 on Clean and 93.5 versus 91.5 on Random, gains of 1.2 and 2.0 percentage points over LingBot-VA. This supports an improvement on the paper’s simulated task suite, with higher success than every method listed in this table.

Where the evidence stops. Appendix B lists LingBot-VA at 92.8/91.8% for 12 fps and 50k steps, rather than this table’s 92.9/91.5%. The source does not reconcile the difference. No uncertainty estimates accompany either set.

Figure 1. The largest early training benefit appears in the high-frame-rate setting. Original paper, p. 2 ↗

Excerpt from the authors’ paper; cropped without altering the figure or table.

How to read it. Read the four panels as two frame rates crossed with two evaluation settings. Red squares are Next Forcing and blue circles are LingBot-VA. The horizontal axis counts training steps; it does not show elapsed time. Because vertical ranges differ, compare labeled values within a panel before comparing visual gap sizes across panels. In the rightmost panel, the vertical annotation identifies the 5k-step Random gap. The horizontal annotation compares Next Forcing near 20k steps with the baseline near 45k steps. Appendix B supplies the exact values behind the curves and should be used for numerical comparisons. e11e02e08e16

What it supports. At 50 fps and 5k steps, Random success rises from 31.9% to 61.6%, a 29.7-point gain or approximately 93.1% relative improvement. At 20k steps, Next Forcing reaches 85.0%, exceeding the baseline’s 84.5% at 45k; the step ratio is about 2.3.

Where the evidence stops. MCP adds training computation, so the horizontal annotation is not measured wall-clock acceleration. The curves support faster progress per step; they do not directly measure appearance copying or prove why the representations improve.

Table 4. Retaining the first MCP module changes the accuracy profile across frame rates. Original paper, p. 10 ↗

Excerpt from the authors’ paper; cropped without altering the figure or table.

How to read it. For each frame-rate group, compare Standard and MCP-accelerated within the same Clean or Random column. The entries are task success percentages, even though the row label includes 2×. The fps headings describe the evaluated video frame rate, not measured throughput. Section 4.6 explains the operational difference: standard inference removes the auxiliary modules; accelerated inference retains depth 1 to produce the next chunk during the current chunk’s denoising trajectory. Depths 2 and 3 are unused. These two modes share a trained checkpoint, so this comparison tests a deployment choice rather than a separate training recipe. e09e15e08

What it supports. At 12 fps, Random success falls from 93.5% to 90.6%, a 2.9-point cost. At 25 fps, it falls from 91.4% to 89.8%. At 50 fps, it rises from 90.5% to 91.3%. The observed tradeoff therefore depends on frame rate and evaluation setting.

Where the evidence stops. The source provides no latency, hardware-specific throughput, or uncertainty measurements here. Two-chunk advancement motivates the 2× label, but this table alone cannot verify a twofold reduction in end-to-end robot control latency.

6.2 Results and evaluation conditions

Task & protocolReported resultComparison & interpretation
RoboTwin 50-task manipulation

Table 1; simulated bimanual tasks, Clean fixed configurations and Random object/scene randomization. Comparator checkpoint/frame-rate choices are not individually specified.

Clean 94.1; Random 93.5

Average task success rate (%)

LingBot-VA 92.9/91.5; Fast-WAM 91.9/91.8 (Clean/Random).

Highest among listed methods; Table 1 gains over LingBot-VA are 1.2/2.0 percentage points. Do not replace these baseline entries with Table 5 values. e08e10

High-frame-rate training convergence

RoboTwin, 50 fps, matched training-step checkpoints in Table 5.

At 5k: 70.2/61.6 Clean/Random; at 50k: 91.8/90.5.

Average task success rate (%) and step ratio

LingBot-VA: 45.5/31.9 at 5k and 88.6/85.2 at 50k.

The early Random gain is 29.7 points, approximately 93.1% relative. Next Forcing reaches 85.0% at 20k versus baseline 84.5% at 45k: approximately 2.3× fewer steps, not measured wall-clock acceleration. e08e11

MCP design ablation

Clean-only RoboTwin, 25 fps, 20k steps on 16 GPUs.

Default MCP 85.8; no multi-layer fusion 83.6; MCP shift 5 gives 83.2.

Average task success rate (%)

Default baseline 75.6; one-block MCP 86.5 versus default three-block 85.8.

MCP adds 10.2 percentage points under this restricted protocol. Three blocks are retained for fewer generated artifacts, an unquantified quality criterion. e08e12

PhyWorld combinatorial generalization

Video-only models with action streams removed; out-of-template (OOT) and in-template (IT).

OOT/IT: FVD 4.7/3.2; abnormal ratio 8/2%.

Fréchet Video Distance (FVD; lower better); abnormal ratio (%, lower better)

LingBot-VA: FVD 5.3/3.5; abnormal ratio 12/3%.

Supports improved physical-video fidelity on these settings, not robot execution or comprehensive physical understanding. e13

General-video pretraining

Approximately 3.5M in-house clips; video-only training on 32 GPUs. Two held-out sets of 1,024 clips each; 50k steps.

Human activity 94; camera-driven scenes 97.

FVD (lower better)

LingBot-VA 225 and 204, respectively.

Reported reductions are 58% and 52%. The private data and distinct test distributions limit direct reproduction and cross-benchmark FVD comparisons. e14

Parallel-chunk inference accuracy

Same checkpoint, standard versus depth-1 MCP-accelerated inference; RoboTwin at 12, 25, and 50 fps.

Accelerated Clean/Random: 93.5/90.6, 91.0/89.8, 92.2/91.3, respectively.

Average task success rate (%)

Standard: 94.1/93.5, 92.6/91.4, 91.8/90.5.

The 12-fps Random drop is 2.9 points; 50-fps Random increases by 0.8. No uncertainty or runtime measurements establish significance or actual end-to-end 2× speedup. e09e15

6.3 Ablations and diagnostic examples

Read component removals and qualitative examples within their stated evaluation conditions.

Table 2. Feature fusion and noisier auxiliary targets help under the Clean-only ablation protocol. Original paper, p. 7 ↗

Excerpt from the authors’ paper; cropped without altering the figure or table.

How to read it. Treat the left and right halves as separate experiments, not paired interventions across a row. Both use the Clean-only protocol: 2,500 demonstrations, 25 fps, 20k steps, and 16 GPUs. The left half varies the baseline, whose default main noise shift is 5 with noisy history augmentation. The right half varies the MCP design, whose defaults are shift 10, multi-layer fusion, initialization from the main model, and three transformer blocks per depth. Compare each changed row with its own default. Here s_main and s_mcp control noise scheduling, not physical video frame rate or prediction depth. e08e12e05e18

What it supports. Adding default MCP raises success from 75.6% to 85.8%, a 10.2-point improvement. Removing fusion loses 2.2 points, and lowering the MCP shift loses 2.6 points. One block scores 86.5%, but the authors retain three blocks because they report fewer visual artifacts.

Where the evidence stops. The artifact advantage is not quantified, and no seed-level uncertainty is given. This table does not isolate chained versus independent heads or the number of future horizons, despite the architecture’s emphasis on a causal chain.

7. Analysis & limitations

7.1 What the evidence leaves open

Reader analysis

MCP adds training cost, which the authors identify as the main limitation. Step-count convergence gains therefore do not establish equal-compute or wall-clock efficiency. e16e11

Source description

Table 1 gives LingBot-VA 92.9/91.5%, while Table 5 gives 92.8/91.8% at 12 fps and 50k steps. The source does not reconcile these entries; both are preserved in their own protocols. e10e11

Reader analysis

Success tables lack seed-level uncertainty. The ablations support feature fusion and noise scheduling but do not isolate the causal chain or count of prediction depths. The appearance-shortcut explanation remains an interpretation of outcomes. e02e12e15

7.2 Questions for discussion

  1. Does MCP still help when physical prediction horizons, rather than chunk counts, are matched across frame rates?
  2. Does the one-block versus three-block artifact tradeoff predict closed-loop failures?

8. Reproducibility audit

8.1 Requirements and known gaps

Source description

Required ingredients include the LingBot-VA/Wan2.2 initialization, pretrained VAE, multi-embodiment data, RoboTwin demonstrations, and exact target shifting, padding exclusion, RoPE, masks, and shifted noise schedules. e03e04e07e08e17e18

Reader analysis

The source refers implementation settings to LingBot-VA but does not supply GPU models, optimizer/batch/learning-rate details, software versions, evaluation trial counts, or denoising-step settings. Its in-house general-video data are not enumerated. These gaps prevent a self-contained reproduction from this PDF. e08e09e14

Reader analysis

Proposed checks: compare chained and independent MCP heads with matched training compute; measure actual latency and success for both inference modes on identical rollouts. These would test mechanism and deployment claims separately. e05e08e09e12e15

8.2 Proposed reproduction checks

The following checks are proposals motivated by the paper. They have not been run as part of this reading.

Check 1: Does the causal chain help beyond auxiliary loss and compute?

Reader-proposed experiment, not performed: reproduce the 25-fps Clean-only setting with identical demonstrations, initialization, noise schedules, fusion layers, loss weights, masks, and evaluation seeds. Compare the default three-depth chain with three independent heads, each receiving the same fused backbone features, while keeping the per-head architecture and parameter count matched. Include the no-MCP baseline. Evaluate with all auxiliary modules removed, reporting success both at matched optimizer steps and matched accelerator-hours. Record uncertainty across training seeds. If independent heads match the chain within uncertainty, the evidence would favor multi-horizon supervision without establishing a distinct benefit from chaining; a consistent chain advantage at matched compute would support that mechanism. e05e07e08e12e17e18

Check 2: Is two-chunk generation faster at acceptable task success?

Reader-proposed experiment, not performed: use one checkpoint and identical Clean/Random evaluation seeds at 12, 25, and 50 fps. Compare standard inference with depth-1 MCP retention on the same hardware, holding the solver and denoising-step count fixed and documenting observation refresh and action execution schedules. Time video denoising, action decoding, observation updates, and the complete control cycle separately; report latency distributions, memory, and success uncertainty. The falsifiable claim is a near-twofold improvement in measured video-generation throughput, with an explicitly chosen acceptable success loss. A smaller end-to-end gain would show that other costs or control scheduling limit the practical benefit, even if denoising itself accelerates. e03e06e09e15

8.3 Reading coverage

Visual audit: The title, all method and experiment pages, and Appendices A–C were rendered and visually inspected. This includes Figures 1–5, Tables 1–5, the full convergence table, noise-schedule equations, and all retained implementation and compute details. All six final original crops were separately viewed; the table crops were tightened to remove caption fragments. Figure 2 arrows were checked against Section 4.3 and Eqs. (7)–(8); Figure 5 cells and Q/K directions agree with Appendix A. Table 1 and Table 5 baseline differences remain explicit. Figure 3 is not cropped: its blue boxes appear in both generated rows although its caption describes baseline deviations. Reference-only pages 11–16 were read in text, not visually inspected. No separate supplement or code was inspected.

PDF pages inspected for this edition: 1, 2, 3, 4, 5, 6, 7, 8, 9, 10, 17, 18. Appendix coverage: reviewed.

Original figures and tables remain the work of the source’s authors. Extractions preserve their scientific content; any HTML wrapper layout is disclosed with each figure. The surrounding reading notes are our own.

Text reading scope & known omissions
  • Abstract
  • 1 Introduction
  • 2 Related Work
  • 2.1 World Action Models
  • 2.2 Autoregressive Video Generation
  • 3 Preliminaries
  • 3.1 Flow Matching
  • 3.2 Autoregressive Video Generation with Teacher Forcing
  • 4 Next Forcing
  • 4.1 Overview
  • 4.2 Multi-Chunk Prediction Objective
  • 4.3 Chained MCP Modules
  • 4.4 Joint Video-Action Architecture
  • 4.5 Training Objective
  • 4.6 Inference
  • 5 Experiments
  • 5.1 Experimental Setup
  • 5.2 Main Results
  • 5.2.1 Comparison with State-of-the-Art
  • 5.2.2 Training Convergence Analysis
  • 5.2.3 PhyWorld Benchmark
  • 5.2.4 Generality on Video Pretraining
  • 5.3 Ablation Studies
  • 5.4 Inference Acceleration
  • 6 Conclusion
  • References
  • A Attention Mask Details
  • B Detailed Training Convergence Results
  • C Timestep Shift Formulation

Outside the original text pass

  • Text extraction does not reconstruct figure images; inspect the retained PDF for figures and equation/table layout.
  • Separate supplemental material availability has not been fully verified.
  • The extraction-only figure omission was addressed by inspecting the original PDF: all five figures and all five tables were visually reviewed, including Appendices A–C.
  • Identity and version scope: the inspected title page identifies arXiv:2606.11187v1, 9 June 2026. Title and all seven authors agree with the catalog; no revision or edition difference was observed. Only this supplied version was reviewed. Its affiliation abbreviations are verified in metadata, whereas the catalog did not identify affiliations.
  • No separate supplemental artifact was supplied beyond the appendix inside the PDF. References were read as part of the supplied text, without opening the cited works. Code and project pages were not inspected; no experiments were reproduced.

The visual audit above records the subsequent illustrated pass.

8.4 Traceable evidence

e01PDF p. 1, title block and arXiv margin stampInspect

The title matches the catalog. The seven authors, in order, are Gangwei Xu, Qihang Zhang, Jiaming Zhou, Xing Zhu, Yujun Shen, Xin Yang, and Yinghao Xu. The affiliations are Robbyant, HUST, HKUST, and HKUST (GZ). The stamp identifies arXiv:2606.11187v1, 9 June 2026.

Go to primary source ↓
e02PDF pp. 1–2, Section 1, appearance-shortcut and myopic-supervision discussionInspect

The authors argue that adjacent video chunks, especially at high frame rates, are similar enough for local copying to weaken dynamics learning. MCP adds future-horizon targets; this is the proposed explanation for improved learning.

Go to primary source ↓
e03PDF p. 4, Sections 3.1–3.2, Eqs. (1)–(3)Inspect

Flow matching interpolates clean data and Gaussian noise, predicts noise minus clean data, and integrates from noise to data at inference. A pretrained VAE supplies chunked video latents. Teacher forcing uses clean past chunks and language; deployment replaces generated frames with observations after action execution.

Go to primary source ↓
e04PDF p. 5, Section 4.2, Eqs. (4)–(6)Inspect

Each MCP target shifts the video by k chunks, replicates the final chunk at boundaries, adds independent noise, and uses temporal RoPE position i+k. Higher MCP timestep shifts emphasize noisier targets.

Go to primary source ↓
e05PDF p. 3, Figure 2; p. 5, Section 4.3, Eqs. (7)–(8)Inspect

Hidden states from video backbone layers 4, 12, 20, and 30, including clean and noisy tokens, are concatenated and fused by a two-layer MLP. The first MCP module uses this fusion; later modules use preceding-depth features. Each concatenates the noisy future embedding, projects, and applies three transformer blocks. MCP gradients backpropagate to the backbone.

Go to primary source ↓
e06PDF p. 6, Section 4.4, Eq. (9)Inspect

Video prediction precedes inverse-dynamics action decoding conditioned on the predicted future and history. Video and action streams interact through cross-modal attention at every layer inside a unified Mixture-of-Transformers architecture. MCP acts on the video stream.

Go to primary source ↓
e07PDF p. 6, Section 4.5, Eqs. (10)–(13)Inspect

Video and action have flow-matching losses. Each MCP depth has a shifted-target flow-matching loss with the last k padded chunks excluded. The objective adds the weighted MCP losses to both main losses.

Go to primary source ↓
e08PDF p. 7, Section 5.1, benchmark and implementation paragraphs; Table 2 captionInspect

The backbone has 30 Wan2.2 layers. Main/MCP shifts are 5/10; noisy history augmentation probability is 0.5; MCP weights are 0.5, 0.2, 0.1; chunk size is sampled from 1–4. MCP initialization copies the last few main layers. Multi-embodiment pretraining precedes RoboTwin training with 2,500 Clean and 25,000 Random demonstrations, up to 50k steps on 64 GPUs. Ablations use only 2,500 Clean demonstrations, 25 fps, 20k steps, and 16 GPUs. RoboTwin evaluates 50 simulated bimanual tasks.

Go to primary source ↓
e09PDF pp. 6–7, Section 4.6Inspect

One checkpoint supports removing all MCP components or retaining depth 1 to generate the next video chunk alongside the main current chunk in one denoising trajectory. Depths 2 and 3 are unused in this accelerated mode. The authors infer 2× acceleration from advancing two chunks and describe extra computation as nearly free; no latency measurements accompany this explanation.

Go to primary source ↓
e10PDF p. 7, Table 1, Clean and Random rowsInspect

Next Forcing reports 94.1/93.5% average success, LingBot-VA 92.9/91.5%, and Fast-WAM 91.9/91.8%, respectively. These are the comparison-table entries, not the distinct 50k-step baseline entries of Table 5.

Go to primary source ↓
e11PDF p. 2, Figure 1 and caption; p. 8, Section 5.2.2; p. 18, Appendix B, Table 5Inspect

At 50 fps and 5k steps, Next Forcing reports 70.2/61.6% Clean/Random versus 45.5/31.9%. At 20k, Random success is 85.0%, exceeding the baseline 84.5% at 45k. At 50k, the corresponding rates are 91.8/90.5% versus 88.6/85.2%. Table 5 gives 12-fps baseline 50k values of 92.8/91.8%, unlike Table 1.

Go to primary source ↓
e12PDF p. 7, Table 2, both ablation panels; pp. 9–10, Section 5.3Inspect

The default baseline/MCP scores are 75.6/85.8%. MCP with shift 5 scores 83.2%; without multi-layer fusion 83.6%; without weight initialization 83.8%; one block 86.5%; five blocks 85.0%. Main shift 10 gives a baseline 78.4%; removing noisy history gives 69.8%. Three blocks are retained for fewer visual artifacts, without a quantified artifact metric. The table contains no chain-versus-independent-head or prediction-depth-count ablation.

Go to primary source ↓
e13PDF p. 8, Figure 3 and Section 5.2.3; p. 9, Table 3Inspect

PhyWorld removes both models’ action streams. OOT/IT FVD is 4.7/3.2 for Next Forcing versus 5.3/3.5 for LingBot-VA; abnormal ratios are 8/2% versus 12/3%. Figure 3 shows selected five-frame trajectories; blue boxes appear on both generated-model rows, although the caption describes baseline deviations.

Go to primary source ↓
e14PDF p. 9, Figure 4 and Section 5.2.4Inspect

Pure-video pretraining uses approximately 3.5 million in-house clips, each 5–10 seconds, on 32 GPUs. Two held-out sets have 1,024 samples each: human activity and camera-driven scenes. At 50k steps FVD is 94 versus 225 and 97 versus 204, reported as 58% and 52% reductions.

Go to primary source ↓
e15PDF p. 10, Table 4 and Section 5.4Inspect

Standard versus MCP-accelerated Clean/Random success is 94.1/93.5 versus 93.5/90.6 at 12 fps; 92.6/91.4 versus 91.0/89.8 at 25 fps; and 91.8/90.5 versus 92.2/91.3 at 50 fps. The row is labeled 2×, but the table reports success rates rather than measured execution times.

Go to primary source ↓
e16PDF p. 10, Section 6, limitation statementInspect

The authors identify additional MCP training cost as the main limitation.

Go to primary source ↓
e17PDF p. 17, Appendix A, Figure 5, caption and four attention rulesInspect

Green indicates allowed attention, rows are queries, and columns are keys. Clean tokens attend to same/earlier clean chunks and never to noisy tokens. Noisy tokens attend to strictly earlier clean chunks and their own noisy chunk. Main and MCP modules reuse this mask structure. Action tokens are explicitly omitted.

Go to primary source ↓
e18PDF p. 18, Appendix C, Eqs. (14)–(15)Inspect

A schedule of 1,000 uniform base noise levels in [0,1] is transformed by s sigma divided by 1+(s−1)sigma. Uniformly sampling schedule indices gives a distribution biased toward higher noise for larger s. Main and MCP shifts are 5 and 10.

Go to primary source ↓

8.5 Primary sources

Scroll across the image to inspect details. Press Esc to close.