PAPER REPORTENAll readings ↗

WorldVLA: Towards Autoregressive Action World Model

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

Authors: Jun Cen; Chaohui Yu; Hangjie Yuan; Yuming Jiang; Siteng Huang; Jiayan Guo; Xin Li; Yibing Song; Hao Luo; Fan Wang; Deli Zhao; Hao Chen

Affiliations: DAMO Academy, Alibaba Group; Hupan Lab; Zhejiang University

Source: 2506.21539 ↗ · Catalog record

Reading: 347 / 558 · 6 original figures & tables · ~21 min ·

1. Paper overview

In one sentence: WorldVLA shares one autoregressive generator across policy and world prediction, while masking cross-action attention to make action chunks more reliable. E03E05E06E09E10E12

At a glanceWhat to know
Research problem
Author claim

The paper targets a gap between policies that output actions without learning to interpret them, and world models that consume actions but do not directly generate them. The authors propose shared representations for both tasks and identify error propagation through earlier predicted actions as a weakness of naive autoregressive action chunking. E02E06

Core mechanism
Source description

A shared discrete autoregressive backbone supports both observation-conditioned action generation and action-conditioned future-image prediction, using modality-specific tokenizers and one vocabulary. E03E04E05

A key reported resultLIBERO four-suite policy benchmark: 81.8 at 512×512; 79.1 at 256×256.

Average task success rate (%). Table 2; all available data used for training; 50 evaluation rollouts per task with different initial states.

OpenVLA: 76.5; OpenVLA-OFT: 95.4. The higher-resolution model exceeds discrete OpenVLA by 5.3 percentage points, but not the strongest listed continuous baseline. WorldVLA lacks large-scale robot-data pretraining in this table, while still using pretrained Chameleon. E03E08E09

Reading caution
Reader analysis

Evidence is confined to LIBERO simulation. Qualitative grasp retries and coherent generated frames illustrate selected cases, not physical deployment, calibrated dynamics or a measured planning mechanism. No seed-level uncertainty accompanies the main tables. E08E09E10E11E12E13

Core contributions

  • Source description

    A shared discrete autoregressive backbone supports both observation-conditioned action generation and action-conditioned future-image prediction, using modality-specific tokenizers and one vocabulary. E03E04E05

  • Reader analysis

    An action attention mask blocks earlier actions within a chunk while retaining text/image context. Ablations separate its large contribution from the smaller incremental benefit of joint world-model training. E06E10

Figure 2. One generator is trained to act and to predict action-conditioned visual change. Original paper, p. 4 ↗

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

How to read it. Read each half from the bottom upward. On the left, the task question and M image observations enter their tokenizers; the shared WorldVLA block predicts tokens that the action detokenizer converts into K actions. On the right, the text is a fixed next-frame request, not the manipulation instruction. Current-image tokens and action tokens condition the generated image, with N indicating repeated transition examples. The orange image and blue action tokens exchange input/output roles across the two tasks. The caption and training sequences establish complementary uses of one model; they do not show a policy searching over candidate futures before choosing an action. E03E04E05

What it supports. The architectural contribution is shared action and image modeling. The world task makes actions inputs whose visual consequences must be predicted, while the policy task makes actions outputs grounded in images and language. This supports the One Model classification through a common generator, independently of whether joint training improves a particular metric.

Where the evidence stops. Figure 2 writes ΔGrip, but Section 3.2 specifies an absolute gripper state. The report follows the textual definition and preserves the conflicting label. The diagram also does not document a candidate-action scoring or planning loop.

2. Motivation

2.1 The problem and the proposed response

Author claim

The paper targets a gap between policies that output actions without learning to interpret them, and world models that consume actions but do not directly generate them. The authors propose shared representations for both tasks and identify error propagation through earlier predicted actions as a weakness of naive autoregressive action chunking. E02E06

2.2 What this reading follows

A policy normally reads an image and chooses an action; a world model reads an action and predicts its visual consequence. WorldVLA teaches one Chameleon-derived generator both mappings. The central reading question is whether these shared tasks help each other, and how much improvement instead comes from changing action decoding. The original architecture, masks, benchmark and ablations let us separate those contributions. They show stronger policy performance from joint training and a larger gain from the action mask, alongside a horizon-dependent improvement in visual prediction. They also leave implementation gaps and several source inconsistencies that matter for reproduction. E03E05E06E09E10E12

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
ArchitectureOne Model
Prediction paradigmJoint prediction
QuadrantQ1 · One Model × Joint prediction

3.1 Evidence-based assessment

Supports the recorded classification

Reader analysis

The shared Chameleon-derived autoregressive backbone and common vocabulary support One Model. Its policy and action-conditioned world functions support the recorded broad Joint prediction category. This means one model learns both outputs through distinct task sequences; it is not inverse dynamics from imagined goal frames, nor proof of simultaneous action/video emission or planning during control. E02E03E04E05

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
  • Policy: language instruction and historical image observations.
  • World prediction: image observations, corresponding actions and a fixed prediction prompt; no task instruction.
  • Policy: discretized seven-component robot actions, grouped into chunks.
  • World prediction: discrete future-image tokens decoded into visual observations.

4.2 Equations and their role

Mψ:{at=Mψpolicy(atoth:t,l),ot=Mψworld(ototh:t1,ath:t1).M_{\psi}:\begin{cases}a_t=M_{\psi}^{\mathrm{policy}}(a_t\mid o_{t-h:t},l),\\o_t=M_{\psi}^{\mathrm{world}}(o_t\mid o_{t-h:t-1},a_{t-h:t-1}).\end{cases}
Equation (3), preserving the paper's conditional-output notation: a_t is an action, o_t an image observation, l the instruction, h the history span, and ψ the unified model's parameters. The two superscripts distinguish policy and world-prediction functions of that model. E03
L=Laction+αLworld\mathcal{L}=\mathcal{L}_{\mathrm{action}}+\alpha\mathcal{L}_{\mathrm{world}}
Equation (4): the two terms are cross-entropy losses on target actions and generated images. α balances their contributions and is set to 0.04 in the experiments. E07

5. Method in detail

5.1 Make an action both a prediction target and a description of change

Reader analysis

Begin with the policy example: instruction tokens and observed images precede a target action chunk. In the world example, an action becomes part of the input and the target becomes a future image. Both examples use the same vocabulary and Chameleon-derived generator, but the world prompt is generic rather than task-specific. Training mixes the examples and adds their target-token cross-entropies, with α weighting the image objective. This construction gives the model two ways to connect visual scenes and motor commands. The authors interpret that connection as learning environmental physics and better action understanding. A narrower reading is that the two supervised tasks can transfer useful representations; the experiments do not directly measure an internal physics representation or establish a runtime search over simulated outcomes. E03E04E05E07E10

Figure 3. The proposed mask removes the route by which one predicted action conditions the next. Original paper, p. 5 ↗

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

How to read it. Treat each row as a token seeking context and each column as a possible context token. Gray cells are the permitted causal connections, as established by panel (a) and the accompanying explanation. Compare the Action₁ rows against the Action₀ columns: the gray rectangle in panel (a) becomes white in panel (b). Text and image columns stay available. Crucially, the triangular gray regions inside each action block remain, so this drawing supports independence between action groups, not unrestricted parallel prediction of all action-component tokens. Panel (c) keeps standard causal conditioning because future-image generation must use the supplied action and preceding visual context. E06E05E14

What it supports. The mask addresses a specific dependency in chunked policies: an erroneous early action can become context for a later action. Blocking that dependency lets each action group use the same observation and instruction context. The intervention changes information flow; it does not itself improve the visual tokenizer or add a planning objective.

Where the evidence stops. The prose calls actions independent and sometimes says they depend solely on visual input. Figure 3 retains text access and intra-action causal structure. Read those claims at the action-group level; the precise parallel decoding schedule is not supplied.

5.2 Distinguish independence between actions from independence inside an action

Reader analysis

A seven-token action represents several position and angle components plus the gripper state. Ordinary causal decoding lets later actions attend to all earlier action tokens, creating a route for prediction errors to affect the rest of a chunk. Figure 3 removes the cross-action rectangle but leaves triangular attention inside each action's token block. This distinction explains how an autoregressive model can support parallel action groups without the figure claiming that all scalar components are independent. Table 3 then supplies the useful comparison: hold chunking on and world training off, and change only the mask, moving average success from 54.0% to 76.6%. The error-propagation explanation is the authors' interpretation; a targeted perturbation experiment would be needed to measure that pathway directly. E04E06E10

5.3 Separate three kinds of evidence before concluding mutual benefit

Reader analysis

Policy success, generated-video quality and selected qualitative rollouts answer different questions. Table 3 measures whether robot behavior completes LIBERO tasks, while Table 4 measures image-sequence similarity on held-out paired trajectories. Figure 4 illustrates grasp retries and Figure 5 illustrates generated state changes, but neither supplies population-level failure rates. Joint training helps policy success in the reported matched row pairs and helps all four displayed visual metrics at 50 frames; the 10-frame result is mixed. Figure 7 further suggests action-conditioned prediction is more helpful than the tested instruction-conditioned video objective, although their input and target structures differ. Together these results support useful task interaction. They do not establish real-world deployment, a universally better world model, or the authors' proposed prospective planning mechanism. E08E10E11E12E13E15

5.4 Training and inference

During training

Source description

Mix policy and world-model examples under the weighted cross-entropy objective. Defaults are M=2 images, K=5 actions for Spatial/Object/Goal and K=10 for Long, N=1 world transition, and α=0.04. The weight compensates for the greater number of image tokens. E07E08

Source description

Remove unsuccessful trajectories and no-operation actions. The default split is 90% training trajectories and 10% validation; Table 2 instead uses all available training data. World-model pretraining followed by policy learning is investigated separately from the main mixed-training method. E08E17

During inference

Reader analysis

For control, provide the instruction and observed image history, decode an action chunk and detokenize its components. For visual forecasting, condition image generation on observations and supplied actions. Longer chunks delay policy adaptation. The paper motivates prospective simulation but supplies no candidate-search, scoring or imagined-rollout selection algorithm for the evaluated policy. E03E05E06E14

5.5 Implementation flow

  1. Tokenize all modalities

    Initialize from Chameleon. The VQ-GAN image tokenizer has compression ratio 16 and an 8,192-entry codebook, producing 256 tokens at 256×256 or 1,024 at 512×512. Each action dimension uses 256 bins determined from training-data ranges. Seven tokens encode three relative positions, three relative angles and one absolute gripper state. The shared vocabulary contains 65,536 entries. E03E04

  2. Select the prediction task

    Policy examples concatenate a task question and M images before K target actions. World examples combine a fixed next-frame prompt, an image and an action before a target image, repeating for N transitions. Only target action tokens or generated image tokens contribute the respective losses. This is joint capability through mixed task examples, rather than a requirement to emit both outputs on every call. E05E07

  3. Remove cross-action dependence

    Figure 3(b) removes attention from Action₁ to Action₀ while preserving text/image access and causal triangles inside each action block. Thus action groups can be decoded independently of earlier groups; the diagram does not imply independence of every component token. World prediction retains ordinary causal attention. Figure 2 labels the gripper as a delta, conflicting with the absolute state specified in Section 3.2. E03E04E06

6. Experiments & results

WorldVLA adapts Chameleon into one token-generating model with two tasks: produce robot actions from images and instructions, or predict images from observations and actions. Joint training improves LIBERO policy success and longer-horizon video prediction in the reported ablations. Its distinctive attention mask removes dependencies between successive actions in a chunk. The evidence supports shared-model learning benefits, but does not establish an inference-time planner that ranks imagined futures.

6.1 Read the original evidence

Table 2. WorldVLA improves over the listed discrete baseline, while stronger continuous-policy results remain. Original paper, p. 6 ↗

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

How to read it. Start with the lower block, which compares discrete action models. Read across each row from Spatial through Long before looking at Average; the long-horizon suite remains much harder for WorldVLA. Then inspect the upper block to place those averages among continuous policies. The pretraining flag has a narrow definition in the source caption: pretraining on large-scale robot manipulation data. A cross for WorldVLA does not erase its Chameleon initialization. Section 4.1 adds another essential condition outside the crop: Table 2 uses all available data for training, whereas the later internal ablations normally train on the 90% trajectory split. E03E08E09

What it supports. At 512×512, WorldVLA reports 81.8% average success against OpenVLA's 76.5%, a calculated 5.3-percentage-point difference. Raising WorldVLA's image resolution from 256×256 increases its average from 79.1% to 81.8%. OpenVLA-OFT reaches 95.4%, so the table supports an improvement within the discrete comparison, not overall benchmark leadership.

Where the evidence stops. The rows differ in architecture and robot-data pretraining; this is not a matched causal test of world modeling. No uncertainty intervals are supplied, and these full-data numbers must not be merged with the held-out ablation protocol.

6.2 Results and evaluation conditions

Task & protocolReported resultComparison & interpretation
LIBERO four-suite policy benchmark

Table 2; all available data used for training; 50 evaluation rollouts per task with different initial states.

81.8 at 512×512; 79.1 at 256×256.

Average task success rate (%)

OpenVLA: 76.5; OpenVLA-OFT: 95.4.

The higher-resolution model exceeds discrete OpenVLA by 5.3 percentage points, but not the strongest listed continuous baseline. WorldVLA lacks large-scale robot-data pretraining in this table, while still using pretrained Chameleon. E03E08E09

Joint training and action-mask ablation

Table 3; default 90% training split; four LIBERO suites, 50 rollouts per task.

62.8 action-only; 67.2 with world training; 54.0 naive chunking; 76.6 masked chunking; 78.1 masked chunking plus world training.

Average task success rate (%)

Joint training adds 4.4 points without chunks and 1.5 with masked chunks. Masking adds 22.6 points over naive chunking.

Masking explains the largest isolated improvement. Comparing 62.8 directly with 78.1 would combine multiple interventions. E08E10

Action-conditioned visual prediction

Table 4; paired held-out validation trajectories; 10- and 50-frame sequences.

Action World Model: 255.1 at 10 frames; 674.1 at 50 frames.

FVD (lower is better)

World Model: 250.0 and 718.6, respectively.

The 50-frame FVD reduction is about 6.2%, calculated from the table, not the introduction's 10% claim. At 10 frames FVD and SSIM worsen, while PSNR and LPIPS improve; benefits are not uniform. E08E12E20

World prediction versus instruction-conditioned video prediction

Figure 7; default evaluation protocol; compare action-only, action-video and action-world variants.

Action World Model: 67.2.

Average task success rate (%)

Action model: 62.8; Action Video Prediction Model: 63.0.

Action-conditioned supervision helps more here. The comparison also changes conditioning and sequence construction, so it does not uniquely isolate physical understanding. E08E15

World-model initialization for the policy

Table 6; separate pretraining ablation under the default data/evaluation protocol.

66.8 with world-model pretraining.

Average task success rate (%)

62.8 without world-model pretraining.

A 4.0-point gain supports transferring world-model weights; it is distinct from the joint-training result. E08E17

6.3 Ablations and diagnostic examples

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

Table 3. Pairwise row comparisons separate world-training gains from the larger mask effect. Original paper, p. 7 ↗

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

How to read it. The first four configuration columns identify what changes; the remaining columns report task success. Their suite order is Goal, Object, Spatial and Long, unlike Table 2. Compare rows 1 and 2 to add world-model training without chunking. Compare rows 3 and 4 to add the mask while keeping action chunking and omitting world-model training. Finally compare rows 4 and 5 to add world training to the masked policy. Reading only the first and last rows would combine three changes and conceal which intervention matters most. Section 4.1 supplies the default trajectory split and 50-rollout-per-task evaluation. E08E10

What it supports. The mask raises average success from 54.0% to 76.6%, a calculated 22.6-point gain. World training adds 4.4 points without chunks, from 62.8% to 67.2%, and 1.5 points with masked chunks, from 76.6% to 78.1%. The result supports complementary contributions but makes the relative sizes of those contributions clear.

Where the evidence stops. This is not a complete factorial experiment: world training with naive chunking is absent. No seed-level uncertainty is reported. The large combined improvement cannot be attributed entirely to learning environmental dynamics.

Table 4. The visual-prediction benefit is stronger at 50 frames and is not uniform at 10 frames. Original paper, p. 9 ↗

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

How to read it. Compare the two rows within a fixed horizon before comparing across horizons. The arrows mark the preferred directions: lower FVD and LPIPS, higher PSNR and SSIM. At 10 frames, the joint model has better PSNR and LPIPS but worse FVD and SSIM; the bold cells make this split visible. At 50 frames, every bold value is in the joint-model row. These are predictions evaluated against paired held-out trajectories, rather than task-completion scores. Keep SSIM and LPIPS in the table's displayed scale because the supplied metric description does not explain the scaling convention. E08E12E20

What it supports. At 50 frames, FVD falls from 718.6 to 674.1 and PSNR rises from 23.98 to 24.30. The corresponding displayed SSIM rises from 83.41 to 83.55 and LPIPS falls from 15.60 to 15.44. This is evidence for improved longer-sequence visual prediction under the reported evaluation, not a direct measurement of control success.

Where the evidence stops. The introduction claims a 10% FVD reduction; the 50-frame table values imply about 6.2%, while 10-frame FVD worsens. No uncertainty or SSIM/LPIPS scaling explanation is supplied. Generated visual coherence alone does not establish physical accuracy.

Figure 6. Masking helps across the plotted lengths, but longer open-loop chunks still carry an adaptation cost. Original paper, p. 10 ↗

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

How to read it. Read one suite at a time: chunk length is horizontal and success rate is vertical. Green triangles indicate the proposed mask, blue squares the naive mask, and the dashed gray line the no-chunk baseline. The panels use different vertical ranges, so compare values against each panel's ticks rather than comparing apparent slopes across panels. Green stays above blue at every plotted length. Several curves are nonmonotonic, while Object's naive-mask curve decreases at every tested step. In Long, the proposed-mask curve rises before falling; in Object and Spatial, the longest masked chunks fall below the no-chunk reference. This complements the discussion that a policy executing a longer chunk cannot adapt as promptly. E06E08E10E14

What it supports. Removing cross-action conditioning alleviates the naive policy's chunking problem, but does not make arbitrarily long chunks desirable. The figure supports choosing chunk length by task and considering feedback frequency alongside model throughput. It also limits the prose's broad claim of gradually decreasing success: several plotted intervals improve rather than decline.

Where the evidence stops. An apparent configuration mismatch remains: Spatial's naive curve at K=5 is near the low twenties, whereas Table 3 row 3 reports 36.7% under the stated default K=5. The source does not reconcile this; do not treat the two as an exact matched run.

7. Analysis & limitations

7.1 What the evidence leaves open

Reader analysis

Evidence is confined to LIBERO simulation. Qualitative grasp retries and coherent generated frames illustrate selected cases, not physical deployment, calibrated dynamics or a measured planning mechanism. No seed-level uncertainty accompanies the main tables. E08E09E10E11E12E13

Source description

The authors identify limited perceptual expressiveness of discrete image tokenization and reduced adaptation with excessive chunk length. Several Figure 6 curves are nonmonotonic, while LIBERO-Object's naive-mask curve decreases at every tested step. Degradation is therefore not universal at every length increment. E14E18

Reader analysis

Table 5 reports an observation-history speed/success tradeoff but does not explicitly identify its task or timing hardware. Its FPS values cannot support a hardware-independent latency claim. E16E19

7.2 Questions for discussion

  1. Does aligned action conditioning improve control beyond an equally expensive auxiliary image-generation objective?
  2. How much of the chunk-mask benefit survives when observation refresh frequency is held fixed?

8. Reproducibility audit

8.1 Requirements and known gaps

Source description

Required ingredients include the Chameleon initialization, image/action tokenizers, filtered LIBERO trajectories and two task templates. Preserve the Table 2 full-data exception. Exact no-operation thresholds, split seeds, optimizer, learning rate, batch size, training duration, compute, frozen modules and task-mixture ratio are not specified in the supplied paper. E03E04E05E08E19

Reader analysis

Proposed checks: compare causal and blocked cross-action masks at matched chunk lengths and observation-update rates; separately compare aligned versus shuffled action conditioning with matched training budgets. Report repeated-seed policy success and held-out video quality to test error propagation and action-conditioned learning, rather than assuming either mechanism from aggregate gains. E06E10E12E14E15

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: Measure the blocked error-propagation path

Reader-proposed check, not an executed experiment: train action-only causal-mask and Figure 3(b)-mask policies with the same initialization, data split, image resolution, optimization budget and chunk length. Sweep the source's lengths, first holding the observation-update schedule identical within each pair. Perturb the first predicted action's token context and measure changes to later predicted actions before executing the chunk; then measure rollout success separately. The mask should remove direct cross-action sensitivity, while within-action dependence may remain. If the reported advantage disappears after matching feedback frequency and training budget, error blocking alone would not explain the original success gap. Use repeated seeds and report uncertainty. E04E06E08E10E14

Check 2: Test whether aligned action conditioning drives joint-training gains

Reader-proposed check, not an executed experiment: compare an action-only control with joint models trained on correctly aligned versus shuffled action inputs for the world objective. Keep the same trajectories, image targets, policy examples, mask, loss weight, token budget and optimizer schedule; shuffling should preserve the action-token marginal distribution. Evaluate policy success using the same initial states and visual prediction at both 10 and 50 frames on the held-out split. A joint benefit specific to aligned actions would support learning action-dependent transitions. Similar policy gains under shuffling would weaken that explanation and favor an auxiliary-training effect. Report all four source video metrics and their exact scaling, not FVD alone. E05E07E08E10E12E15

8.3 Reading coverage

Visual audit: Visually inspected the title/authors/version on p. 1; Figure 1 on p. 2; Table 1 on p. 3; architecture and equations on p. 4; tokenization, task templates and Figure 3 on p. 5; Table 2, objective and masking on p. 6; Table 3 and all evaluation/training settings on p. 7; qualitative control examples and discussion on p. 8; Table 4 and generated sequences on p. 9; Figures 6–7 and their discussion on p. 10; Tables 5–6 and limitations on p. 11. All six final original crops were inspected for legibility, labels, legends and boundaries. Figure 2's gripper-label conflict, Figure 3's intra-action causal structure, the FVD percentage discrepancy and the apparent Spatial chunk-setting mismatch are disclosed. References on pp. 12–14 were read in the complete text but not visually inspected. No appendix is present; separate supplements were not verified. These declared pages also cover source facts underlying the proposed checks.

PDF pages inspected for this edition: 1, 2, 3, 4, 5, 6, 7, 8, 9, 10, 11. Appendix coverage: not present.

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
  • Title, author block and abstract (p. 1)
  • 1 Introduction (pp. 1–2)
  • 2 Related Works (p. 3)
  • 3 Methods; 3.1 Problem Formulation; 3.2 Architecture; 3.3 Training Strategy (pp. 4–6)
  • 4 Experiments; 4.1 Evaluation Benchmark; 4.2 Evaluation Results and Discussion (pp. 6–11)
  • 5 Conclusion and Future Work (p. 11)
  • References (pp. 11–14)

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.
  • Identity notes: the title and all 12 authors match the catalog. The inspected artifact is arXiv:2506.21539v1, stamped 26 June 2025; its title block separately says June 27, 2025. No later revision was supplied or compared.
  • The title page links the WorldVLA repository, whereas the catalog lists RynnVLA-002. Neither repository was inspected; their relationship and any later edition differences remain unverified.
  • The supplied extraction does not reconstruct figure images. This limitation was addressed by inspecting PDF pages 1–11 and all six final crops; references on pages 12–14 were read as text.
  • Separate supplemental material availability has not been fully verified. No appendix appears in the supplied 14-page PDF.
  • Code was not inspected and experiments were not reproduced.

The visual audit above records the subsequent illustrated pass.

8.4 Traceable evidence

E01PDF p. 1, title/author block, arXiv margin and date/code linesInspect

Title and 12 authors match the catalog. The margin identifies arXiv:2506.21539v1, 26 June 2025; the title block says June 27, 2025 and links WorldVLA. Affiliations are DAMO Academy, Alibaba Group; Hupan Lab; Zhejiang University.

Go to primary source ↓
E02PDF pp. 1–3, Introduction, Figure 1 and Table 1Inspect

The authors contrast action-output policies and action-conditioned world models, proposing a unified action world model with both output modalities.

Go to primary source ↓
E03PDF p. 4, Figure 2; Sections 3.1–3.2; Eqs. (1)–(3)Inspect

One model has policy and world functions and is initialized from Chameleon. The diagram routes text/images to actions and text/images/actions to future images. Its action labels include ΔGrip.

Go to primary source ↓
E04PDF p. 5, Section 3.2, tokenizer paragraphInspect

VQ-GAN compression ratio 16 and codebook 8192; 256 or 1024 image tokens; 256 action bins; seven action components including an absolute gripper state; shared vocabulary size 65,536.

Go to primary source ↓
E05PDF pp. 5–6, Section 3.3, Action Model Data and World Model Data token sequencesInspect

Policy examples contain M images and K target actions. World examples use a fixed prompt and image/action inputs to predict images over N rounds. Losses apply to generated modality tokens; world inputs do not require task instructions.

Go to primary source ↓
E06PDF p. 5, Figure 3(a–c); p. 6, Attention Mask paragraphInspect

Panel (b) blocks Action₁ rows from Action₀ columns, preserves text/image context and intra-action causal triangles. World prediction remains causally masked. Prose describes parallel generation of actions without prior-action conditioning.

Go to primary source ↓
E07PDF p. 6, Training Objective and Eq. (4); p. 7, Training SettingInspect

The objective sums action cross-entropy and α-weighted world cross-entropy; α=0.04. Defaults are M=2, K=10 for Long and 5 otherwise, N=1.

Go to primary source ↓
E08PDF pp. 6–7, Section 4.1, Benchmark, Datasets, Training Setting and MetricsInspect

LIBERO suites are evaluated after unsuccessful/no-op filtering. Default trajectory split is 90/10; Table 2 uses all available data. Policies receive 50 rollouts per task with varied initial states; video metrics use held-out paired trajectories.

Go to primary source ↓
E09PDF p. 6, Table 2, WorldVLA/OpenVLA/OpenVLA-OFT rows and Average column; p. 7, Benchmark ResultsInspect

Average SR is 79.1 for 256×256 WorldVLA, 81.8 for 512×512, 76.5 for OpenVLA and 95.4 for OpenVLA-OFT. Table pretraining denotes large-scale robot manipulation data.

Go to primary source ↓
E10PDF p. 7, Table 3, rows 1–5 and Average SR columnInspect

Rows 1–5 report 62.8, 67.2, 54.0, 76.6 and 78.1 average SR. Rows 1/2 and 4/5 isolate world training; rows 3/4 isolate the proposed mask. No world-plus-naive-chunk row is supplied.

Go to primary source ↓
E11PDF p. 8, Figure 4(a–b) and World Model Helps Action Model discussionInspect

Selected cream-cheese and wine-bottle rollouts contrast missed grasps with joint-model grasp retries. The authors attribute improved behavior to physical understanding and prospective simulation.

Go to primary source ↓
E12PDF p. 9, Table 4, both model rows and 10/50-frame columns; p. 8, Action Model Helps World ModelInspect

At 10 frames world/joint values are FVD 250.0/255.1, PSNR 29.62/29.77, SSIM 90.73/90.40, LPIPS 11.97/11.94. At 50 frames they are 718.6/674.1, 23.98/24.30, 83.41/83.55 and 15.60/15.44. Table scaling for SSIM/LPIPS is not explained.

Go to primary source ↓
E13PDF p. 9, Figure 5(a–c); p. 8, Action Model Helps World ModelInspect

Selected predicted sequences concern drawer opening, plate pushing and bowl placement; the discussion describes failed opening, disappearing objects and failed lifting in the standalone world model.

Go to primary source ↓
E14PDF p. 10, Figure 6 and action-chunk discussion continuing from p. 8Inspect

Four suite plots compare proposed and naive masks at chunk lengths 5, 10, 15 and 20 with no-chunk baselines. Proposed-mask curves exceed naive curves at every plotted length. Several curves are nonmonotonic, while LIBERO-Object's naive-mask curve strictly decreases across the four lengths. The discussion links excessive chunk length to reduced timely adaptation.

Go to primary source ↓
E15PDF p. 10, Figure 7 and World Model versus Video Prediction Model subsectionInspect

Average SR is 62.8 action-only, 63.0 action-video and 67.2 action-world. Video prediction uses task text and image history without actions and predicts K frames; world prediction is action-conditioned.

Go to primary source ↓
E16PDF p. 11, Table 5 and Historical Image Input paragraphInspect

With action chunking, one/two/four images yield SR 74.0/84.4/84.7 and FPS 3.67/3.13/2.78. The source selects two images as a tradeoff but does not explicitly label the evaluated task or timing hardware.

Go to primary source ↓
E17PDF p. 11, Table 6 and Pretrain Action Model using World ModelInspect

World-model initialization increases average SR from 62.8 to 66.8, improving all four reported suites; it is studied separately from mixing policy/world training data.

Go to primary source ↓
E18PDF p. 11, Section 5, Conclusion and Future WorkInspect

The authors propose scaling, improved unified tokenization and an auxiliary action head as future directions, identifying limited perceptual expressiveness of the current discrete tokenizer.

Go to primary source ↓
E19PDF pp. 4–7, Sections 3–4.1; p. 11, historical-input and pretraining descriptionsInspect

The supplied method/experiment descriptions give tokenization, losses and M/K/N/α defaults, but omit exact filtering thresholds, optimizer, batch size, learning rate, duration, compute, frozen-module policy, mixture ratio and split seeds.

Go to primary source ↓
E20PDF p. 2, Introduction FVD claim; p. 9, Table 4 FVD columnsInspect

The introduction states a 10% FVD reduction. Table 4 instead gives 718.6 to 674.1 at 50 frames (about 6.2% calculated reduction) and 250.0 to 255.1 at 10 frames.

Go to primary source ↓

8.5 Primary sources

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