Awesome-WAMs
GitHub
PAPER REPORTENAll readings ↗

AcrossWAM1.0:A Modular Latent World-Action Stack for Compact Robot Policies

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

Authors: Yafei Zhang, Nan Wu

Affiliations: Across Physical AI, Beijing, China; Institute of Automation, Chinese Academy of Sciences, Beijing, China

Source: 2608.29937 ↗ · Catalog record

Reading: 7 / 558 · 5 original figures & tables · ~20 min ·

1. Paper overview

In one sentence: AcrossWAM separates transition prediction, future-feature decoding and action generation so a compact backbone can be substituted, but the reported substitution still adapts the decoder and action stack. E02E05E07E09E11E16E17

At a glanceWhat to know
Research problem
Author claim

The authors address implementation coupling between a multimodal policy, its latent world model, and deployment checkpoint. Changing a backbone can alter hidden dimensions, query handling, image arguments, and vocabulary behavior. Their research question is whether an explicit module contract supports compact policies with retained closed-loop performance and verifiably removable training state. E02E13

Core mechanism
Author claim

An explicit adapter contract separates transition-query context, action-query context, hidden-size alignment, and training versus inference modules. The latent-subgoal algorithm itself is attributed to LaWAM. E02E05E07

A key reported resultClosed-loop LIBERO manipulation: Reported compact success: 97.45% (1,949/2,000); difference −0.55 percentage points; p = 0.2664.

Pooled success and exact paired McNemar test. Four suites, 10 tasks per suite, 50 trials per task; 2,000 matched episode keys per policy, with identical seeds, order, and initializations.

Qwen3-VL-2B: 98.00% (1,960/2,000). Discordant outcomes: 46 larger-only successes and 35 compact-only successes. The observed drop meets the prespecified two-percentage-point engineering criterion. Failure to detect a paired difference does not establish equivalence or statistical non-inferiority, particularly with one checkpoint per backbone. E08E09

Reading caution
Reader analysis

Table 1 prints AcrossWAM1.0 RoboTwin Clean/Randomized values of 92.09/89.12, whereas Sections 4.2 and 5 and Appendix C explicitly restrict its benchmark evidence to LIBERO. This unresolved source contradiction prevents accepting those entries as established AcrossWAM1.0 results. E15

Core contributions

  • Author claim

    An explicit adapter contract separates transition-query context, action-query context, hidden-size alignment, and training versus inference modules. The latent-subgoal algorithm itself is attributed to LaWAM. E02E05E07

  • Source description

    A paired evaluation compares newly trained Qwen3-VL-2B and Qwen3.5-0.8B policies using the same episode initializations. E08E09

  • Author claim

    An export audit checks retained tensors, removed training modules, duplicate aliases, and CPU loading; a separate MiniCPM-V smoke test probes the adapter boundary. E11E13

Figure 1. The interface passes a predicted transition and spatial subgoal into control. Original paper, p. 2 ↗

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

How to read it. Start in the upper panel with two observed endpoints. Frozen DINOv3 produces current and future feature tokens; the inverse-dynamics Q-Former compresses their transition into the purple teacher latent. The orange world decoder reconstructs the future feature, with state grounding supplying extra training supervision. Then follow the lower panel from current images and language. Purple ACT queries support transition prediction, while turquoise flow queries provide action context. The predicted latent and current DINOv3 tokens meet at the retained decoder. Its orange future-feature output enters the green action expert. Snowflakes indicate frozen components, and dashed paths identify supervision that is unnecessary when the policy runs. E04E05E06E07E20

What it supports. The diagram places an explicit predicted future feature between intended change and action generation. At inference, the world decoder remains active even though the future observation and inverse-dynamics teacher disappear. That distinction explains why removing training-only weights does not mean removing the world-model pathway from the deployed policy.

Where the evidence stops. A clean interface does not establish frozen-module portability. In the compact experiment, the backbone, mapper, decoder and action expert are optimized together; the figure alone cannot predict behavior after an arbitrary backbone swap.

2. Motivation

2.1 The problem and the proposed response

Author claim

The authors address implementation coupling between a multimodal policy, its latent world model, and deployment checkpoint. Changing a backbone can alter hidden dimensions, query handling, image arguments, and vocabulary behavior. Their research question is whether an explicit module contract supports compact policies with retained closed-loop performance and verifiably removable training state. E02E13

2.2 What this reading follows

A smaller vision-language backbone is useful only if the surrounding robot policy continues to work. AcrossWAM makes that surrounding structure explicit: the backbone predicts an intended transition, a retained world decoder turns it into a spatial future feature, and a flow model produces the action chunk. Its most informative experiment compares two backbone configurations on the same LIBERO episodes, then audits what remains in an exported compact checkpoint. Read the diagram to separate training supervision from inference dependencies, the paired tables to judge retained control performance, and the latent rollout as a representation diagnostic. The evidence supports one concrete scaling experiment; broad plug-and-play portability and deployment speed remain open. E02E05E07E09E11E16E17

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

Supports the recorded classification

Reader analysis

The recorded Dual-system classification is supported by explicit boundaries among the policy adapter, retained world decoder, and separately conditioned action expert; joint optimization does not make them a single predictive model. IDM is supported by the endpoint-conditioned inverse-dynamics teacher that defines the transition bottleneck, with an essential qualification: that teacher is training-only. Inference predicts a latent, decodes a future feature, and generates actions with Flow-DiT rather than executing the inverse-dynamics encoder. Latent prediction and visual-subgoal conditioning are demonstrated architectural properties; the real-time-control tag remains unverified because compact latency is absent. E03E04E05E06E07E17

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
  • Current main-view and optional wrist/auxiliary images, language instruction, and optional proprioceptive state. [E03, E05]
  • Training additionally uses future observations, demonstrated action chunks, and valid state displacements where available. [E04, E06]
  • A 32-dimensional predicted transition latent and a spatial future-feature subgoal. [E04, E05]
  • A continuous chunk of 50 actions spanning 1.2 seconds in the reported configuration. [E06, E18]

4.2 Equations and their role

p(a,h^t+Δ,zt∣ot,ℓ,st)=pθ(zt∣ot,ℓ) pω(h^t+Δ∣ht,zt)⋅pη(a∣cθ,ht,h^t+Δ,st),ht=fψ(ot).\begin{aligned}p(a,\hat h_{t+\Delta},z_t\mid o_t,\ell,s_t)&=p_\theta(z_t\mid o_t,\ell)\,p_\omega(\hat h_{t+\Delta}\mid h_t,z_t)\\&\quad\cdot p_\eta(a\mid c_\theta,h_t,\hat h_{t+\Delta},s_t),\\h_t&=f_\psi(o_t).\end{aligned}
Equation (1) separates the language-conditioned transition prior, scene-conditioned world decoder, and action expert. The explicit future feature mediates between intended change and control. E03
qϕ(zt∣ht,ht+Δ)=N(μϕ,diag⁡(σϕ2)),LS1=SmoothL1⁡β=0.1(h~t+Δ,ht+Δ)+λs∥ms⊙(gξ(st,zt)−Δst)∥22+βKLDKL ⁣(qϕ ∥ N(0,I)).\begin{aligned}q_\phi(z_t\mid h_t,h_{t+\Delta})&=\mathcal N(\mu_\phi,\operatorname{diag}(\sigma_\phi^2)),\\\mathcal L_{\mathrm{S1}}&=\operatorname{SmoothL1}_{\beta=0.1}(\tilde h_{t+\Delta},h_{t+\Delta})\\&\quad+\lambda_s\lVert m_s\odot(g_\xi(s_t,z_t)-\Delta s_t)\rVert_2^2\\&\quad+\beta_{\mathrm{KL}}D_{\mathrm{KL}}\!\left(q_\phi\,\|\,\mathcal N(0,I)\right).\end{aligned}
Equations (2) and (5) define the endpoint-conditioned posterior and its training objective. The mask m_s excludes unavailable state dimensions; λ_s and β_KL weight grounding and regularization. E04
z^t=Mρ(cθLA),h^t+Δ=Dω(ht,z^t),Ldistill=∥z^t−zt∗∥22,Lsubgoal=∥h^t+Δ−ht+Δ∥22.\begin{aligned}\hat z_t&=M_\rho(c_\theta^{\mathrm{LA}}),&\hat h_{t+\Delta}&=D_\omega(h_t,\hat z_t),\\\mathcal L_{\mathrm{distill}}&=\lVert\hat z_t-z_t^*\rVert_2^2,&\mathcal L_{\mathrm{subgoal}}&=\lVert\hat h_{t+\Delta}-h_{t+\Delta}\rVert_2^2.\end{aligned}
Equations (6)–(7) connect the policy's transition prediction to both the frozen teacher target z_t* and the observed future feature. Only the policy prediction and decoder remain necessary at inference. E05E07
aτ=(1−τ)a0+τa1,uτ=a1−a0,Lflow=E ⁣[∥vη(aτ,τ∣cθ,ht,h^t+Δ,st)−uτ∥22],LS2=Lflow+0.1Lsubgoal+0.1Ldistill.\begin{aligned}a_\tau&=(1-\tau)a_0+\tau a_1,&u_\tau&=a_1-a_0,\\\mathcal L_{\mathrm{flow}}&=\mathbb E\!\left[\lVert v_\eta(a_\tau,\tau\mid c_\theta,h_t,\hat h_{t+\Delta},s_t)-u_\tau\rVert_2^2\right],\\\mathcal L_{\mathrm{S2}}&=\mathcal L_{\mathrm{flow}}+0.1\mathcal L_{\mathrm{subgoal}}+0.1\mathcal L_{\mathrm{distill}}.\end{aligned}
Equations (8)–(10) train a velocity field between Gaussian noise a_0 and normalized demonstrated actions a_1. The displayed auxiliary weights are those of the compact configuration. E06E18

5. Method in detail

5.1 Learn change before asking the backbone to predict it

Author claim

The first stage creates a target for intended change from paired observations. DINOv3 is frozen, so the current and future patch features inhabit a stationary representation space. The inverse-dynamics Q-Former sees both endpoints and returns a Gaussian transition posterior with a 32-dimensional latent. The world decoder must reconstruct future features from the current features plus that latent. A state-displacement loss, applied only where state dimensions are valid, gives the bottleneck motion-related supervision, while KL regularization restricts the posterior. The causal idea is that the current image already explains persistent scene appearance, leaving the latent to carry the transition. That is the authors’ representation rationale; the reconstruction objective alone does not prove the latent has isolated physical action factors. E04E18E20

5.2 Replace future access with a learned transition prediction

Source description

The second stage removes the teacher’s privileged access to the future from the policy pathway. Current images and language enter a backbone adapter with two groups of eight queries. One group feeds a mapper that predicts the transition latent; the other supplies context to the flow-based action expert. Distillation aligns the predicted latent with the frozen teacher, and subgoal supervision aligns the decoder output with actual future features. During inference, current DINOv3 tokens and the predicted latent are sufficient to produce the spatial subgoal. The flow expert then transforms Gaussian noise into a 50-action chunk using both current and predicted future features. The architectural separation is explicit, but the compact training run updates the decoder as well as the new backbone, mapper and action expert. E05E06E07E18

5.3 Separate the control experiment from the artifact audit

Reader analysis

The scaling claim combines two kinds of evidence that answer different questions. Matched LIBERO episodes compare behavior: the compact configuration loses eleven successes out of 2,000, with the largest suite decrease in Long. The checkpoint audit compares stored parameters: changing the backbone reduces the raw count, and exporting removes additional training-only state. My reading is that these results establish a promising compact configuration, while leaving the cost of obtaining and deploying it incompletely characterized. The paper supplies one checkpoint per configuration, no measured latency or memory comparison, and an inconsistent label for the larger parameter baseline. The auxiliary latent rollout helps explain what is represented, but cannot fill those experimental gaps. A faithful reproduction should preserve these distinctions when reporting retention, footprint and visual prediction quality. E09E10E11E12E16E17

5.4 Training and inference

During training

Source description

Stage 1 combines future-feature reconstruction, masked state-displacement prediction, and KL regularization. Frozen DINOv3 supplies a stationary feature target; the authors argue that the bottleneck and state supervision encourage action-relevant transitions rather than shared appearance. E04

Source description

Stage 2 combines flow matching, future-subgoal supervision, and latent distillation. In the compact experiment, DINOv3 and the teacher are frozen, while the backbone, mapper, world decoder, and action expert are optimized. The decoder therefore adapts during backbone scaling. E05E06E07

Source description

The compact run uses 25,000 training steps. Appendix A specifies two noise samples per batch, gradient clipping at 1.0, AdamW peak learning rates of 10^-4 for the VLM/flow expert and 3×10^-4 for the decoder. The action expert has hidden size 1,024, 16 heads, and 16 layers. E08E18

During inference

Source description

Current observations and language drive latent prediction, one-pass world decoding, and action generation with 10 integration steps; optional state appears in the action-conditioning formulation. The inverse-dynamics teacher, state-grounding head, and future observation are absent. E03E06E07

Reader analysis

The reported LIBERO success measures closed-loop simulated control. The auxiliary RGB decoder used for the qualitative latent rollout is outside policy inference and success evaluation. The supplied text does not establish how many predicted actions execute before observations are refreshed. E06E08E16E18

5.5 Implementation flow

  1. Learn the transition representation

    Frozen DINOv3 encodes current and future observations into patch tokens. A one-query inverse-dynamics Q-Former observes both endpoints and produces a Gaussian transition posterior. A decoder uses latent-conditioned adaptive normalization and gated transformer updates to predict future tokens from current tokens. A masked state-displacement objective grounds the latent in robot motion when state labels exist. E04

  2. Expose backbone-specific processing through an adapter

    The backbone processes images, instruction, and two disjoint groups of eight learned queries. Latent-action queries supply transition context; flow queries supply action-generation context. The adapter exposes these slices and accommodates backbone-specific image arguments and hidden dimensions. E05E13E18

  3. Predict and decode a policy subgoal

    A one-query cross-attention mapper compresses latent-action query states into the predicted transition latent. During training it imitates the frozen inverse-dynamics teacher. At inference the retained decoder combines that prediction with current DINOv3 tokens to produce a spatial future feature without receiving the actual future frame. E05E07

  4. Generate continuous actions

    A Flow-DiT expert conditions on backbone context and concatenated current/predicted-future visual tokens. Its learned vector field transforms Gaussian action noise into an action chunk through numerical integration. Thus the future-feature prediction participates directly in action generation. E06E07

6. Experiments & results

AcrossWAM1.0 studies whether LaWAM's latent-subgoal policy can support explicit backbone interfaces, a smaller multimodal backbone, and an auditable deployment export. A predicted transition latent is decoded into future visual features that condition continuous action generation. The strongest evidence is a paired LIBERO comparison: the compact policy achieves 97.45% versus 98.00% for the larger configuration. Cross-family evidence is limited to adapter execution, and inconsistent RoboTwin entries and parameter labels require clarification.

6.1 Read the original evidence

Table 2. Matched episodes show a small pooled change when the backbone shrinks. Original paper, p. 7 ↗

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

How to read it. Read the two rows as complete policy configurations rather than isolated language models. The four middle columns contain suite success percentages; Pooled includes both the percentage and successful-episode count. Each configuration is evaluated on the same 2,000 episodes, with 500 per suite. The compact row reaches 1,949 successes, compared with 1,960 for the larger row. Long and Spatial decline, Object increases slightly, and Goal is unchanged. Keep the rightmost parameter column separate from the success comparison: Table 3 identifies different checkpoint accounting categories, so its relationship to the label “Deploy params” here needs scrutiny. E08E09E12

What it supports. The compact configuration attains 97.45% pooled success versus 98.00%, a decrease of 0.55 percentage points over matched episodes. That is within the paper’s prespecified two-point retention tolerance. It supports retained performance for these two trained checkpoints and evaluation episodes, with a modest suite-dependent pattern.

Where the evidence stops. This is one checkpoint per configuration, so the comparison does not establish training-seed robustness. The 2,555.2M figure is called deployment parameters here but raw policy parameters in Table 3; a deployment-to-deployment reduction is therefore not cleanly established.

Table 3. Backbone scaling and export pruning contribute different parameter reductions. Original paper, p. 8 ↗

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

How to read it. The first row supplies the denominator for both percentages: the larger backbone’s raw policy contains 2,555.2 million unique parameters. The second row changes to the compact backbone while retaining a raw checkpoint, producing 1,883.8 million. The final row additionally exports the compact policy, leaving 1,472.6 million. Thus the final percentage combines a backbone change with removal of training-only state. The accompanying audit reports 1,078 retained tensors with identical values and identifies 321 removed training-only keys. Reconstructible aliases are tracked separately, which matters because counting checkpoint keys is not equivalent to counting unique learned parameters. E11E12

What it supports. The compact raw checkpoint is reported as 26.3% smaller than the larger raw checkpoint; its deployment export is 42.4% smaller than that same baseline. The export audit provides evidence about retained tensor values and checkpoint structure. It does not itself measure GPU memory, inference latency or executed robot behavior.

Where the evidence stops. The baseline is a raw checkpoint, despite Table 2 labeling the same count as deployment parameters. Export verification is reported on CPU; the paper does not supply a corresponding GPU or simulator parity test or a measured deployment speedup.

6.2 Results and evaluation conditions

Task & protocolReported resultComparison & interpretation
Closed-loop LIBERO manipulation

Four suites, 10 tasks per suite, 50 trials per task; 2,000 matched episode keys per policy, with identical seeds, order, and initializations.

Reported compact success: 97.45% (1,949/2,000); difference −0.55 percentage points; p = 0.2664.

Pooled success and exact paired McNemar test

Qwen3-VL-2B: 98.00% (1,960/2,000). Discordant outcomes: 46 larger-only successes and 35 compact-only successes.

The observed drop meets the prespecified two-percentage-point engineering criterion. Failure to detect a paired difference does not establish equivalence or statistical non-inferiority, particularly with one checkpoint per backbone. E08E09

Suite-level LIBERO retention

The same paired comparison, with 500 episodes per suite.

Compact: Long 95.0%; Object 99.4%; Goal 96.8%; Spatial 98.6%.

Success rate

Larger model: Long 96.6%; Object 99.2%; Goal 96.8%; Spatial 99.4%.

The largest observed decline is in Long, at 1.6 percentage points. The aggregate result does not imply uniform retention across task types. E10

Checkpoint compactness

Unique-parameter accounting across backbone replacement and deployment export.

Reported compact raw checkpoint: 1,883.8M; compact deployment export: 1,472.6M; reported reduction against the 2B reference: 42.4%.

Unique parameters

The 2,555.2M reference is labeled deployable in Table 2 but raw in Table 3.

The compact export count is consistently reported. The reference's raw/deploy labeling needs clarification before treating the percentage as independently reconciled inference-only accounting. Parameter reduction does not establish latency or memory improvement. E11E12E17

Deployment export verification

Author-reported state-dictionary audit and production CPU loading.

1,078 retained tensors bitwise identical; 321 training-only keys representing 411.1M parameters removed; 718 removed alias keys reconstructible bitwise.

Tensor preservation and removal accounting

Exported state compared with the source checkpoint.

This supports checkpoint-state preservation and CPU materialization. It does not establish exported-policy behavior on an accelerator or in the simulator; those checks were not included. E11E17

Cross-family adapter execution

MiniCPM-V 4.6 forward/backward smoke test.

Reported compatible image/embedding inputs, query-dependent outputs, automatic hidden-size alignment, and decreasing training loss.

Input compatibility, query influence, hidden-size alignment, and optimization behavior

Exercises the adapter contract beyond the two Qwen configurations.

This tests software execution. It supplies no cross-family manipulation success rate or evidence of frozen-decoder transfer. E13

Directional language adherence

Two-task RoboTwin diagnostic changing left versus right; adherence measured among completed conflicting-instruction placements.

Guidance variant adherence: 93/185 (50.3%), versus 52/190 (27.4%); Fisher exact p = 1.0×10^-5.

Directional adherence and collapse

Conflicting-instruction collapse: 38.3% versus 36.7%, p = 0.74. Non-conflicting collapse rises from 1% to 10%.

Improved adherence among completed placements accompanies worse completion behavior on non-conflicting instructions. This conditional diagnostic is not an aggregate RoboTwin benchmark result. E14

6.3 Ablations and diagnostic examples

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

Figure 2. The latent rollout preserves scene layout while arm detail drifts. Original paper, p. 8 ↗

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

How to read it. Read downward within a column before moving through time. The first row is the actual simulator observation; the second decodes its ground-truth latent through an auxiliary RGB decoder. Their difference reveals the visualization decoder’s own reconstruction limits. The third row instead decodes the model’s autoregressive predicted latent. Compare that row with the second to isolate additional prediction drift. Across the four displayed steps, the arm becomes less sharply localized while the table and several objects remain recognizable. The bottom row visualizes the magnitude of predicted latent changes, with warmer regions near the moving arm and manipulated area. It has no numerical color scale. E16E21

What it supports. The visualization supports a narrow representational observation: the predicted latent retains substantial scene structure across the displayed rollout, while fine arm appearance degrades. The change map is consistent with localized motion information, but it does not quantify object tracking accuracy, physical consistency or the quality of the actions generated from those features.

Where the evidence stops. The RGB decoder is an auxiliary visualization tool and is absent from policy inference and success evaluation. This single held-out trajectory is not a closed-loop control trial, and the uncalibrated heatmap should not be read as a probability.

Table 4. The suite breakdown makes the pooled retention claim more interpretable. Original paper, p. 12 ↗

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

How to read it. Use the delta column to compare the compact configuration with the larger configuration within each suite. A negative value means fewer compact-model successes. Long has the largest decrease at 1.6 percentage points; Spatial decreases by 0.8, Object improves by 0.2, and Goal is unchanged. The final column reports the paper’s paired tests, using the matched episode outcomes rather than treating all successes as independent between models. The pooled test combines the suites; the main text gives 46 episodes solved only by the larger configuration and 35 solved only by the compact one. Those discordant outcomes determine the paired comparison. E09E10

What it supports. The aggregate decrease is not uniform across task families. Long-horizon tasks contribute the largest visible suite loss, while the two other object/goal categories remain close. The pooled p-value of 0.266 does not supply evidence of a significant difference under the reported test, but it also does not prove equality or universal interchangeability.

Where the evidence stops. A non-significant paired test is not an equivalence test. These figures characterize evaluation variation for the reported checkpoints; repeated training runs and a prespecified noninferiority analysis would be needed to strengthen a general retention claim.

7. Analysis & limitations

7.1 What the evidence leaves open

Reader analysis

Table 1 prints AcrossWAM1.0 RoboTwin Clean/Randomized values of 92.09/89.12, whereas Sections 4.2 and 5 and Appendix C explicitly restrict its benchmark evidence to LIBERO. This unresolved source contradiction prevents accepting those entries as established AcrossWAM1.0 results. E15

Reader analysis

The experiment changes both Qwen generation and nominal backbone size while retraining the decoder and other policy modules. It establishes performance for these configurations, rather than isolating parameter count or frozen-interface transfer as the causal factor. E07E09

Reader analysis

No reported intervention removes the subgoal, disables distillation, or freezes the decoder under an otherwise matched comparison. Consequently, these experiments do not newly establish the control benefit of the inherited latent-subgoal mechanism. E02E09E13E17

Source description

The authors identify missing training-seed replication, compact-policy latency and peak-memory measurements, cross-family closed-loop evaluation, and accelerator/simulator testing of the export. Published LaWAM robot and latency results are not new AcrossWAM1.0 evidence. E07E17

Reader analysis

The auxiliary-decoded latent rollout is qualitative. Its caption's claims about preserved structure and accumulating blur could not be independently assessed from the supplied text. E16

7.2 Questions for discussion

  1. How much compact-policy performance depends on adapting the world decoder, and would a frozen decoder survive a genuinely different backbone family? [E07, E13]
  2. Would directional adherence improve when evaluated jointly with completion, rather than conditioned on completed placements? [E14]
  3. Which artifacts reconcile the contradictory RoboTwin row and the larger checkpoint's raw/deploy labels? [E12, E15]

8. Reproducibility audit

8.1 Requirements and known gaps

Source description

The authors describe accompanying code, configurations, evaluators, paired statistics, and export-verification artifacts. Their availability and behavior were not checked in this reading, and no experiment was reproduced. E19

Reader analysis

Reproduction needs the exact backbone and Stage-1 checkpoints, training data and preprocessing, action normalization, query/vocabulary handling, and matched LIBERO episode keys. The supplied text gives losses and several optimizer settings, but omits data mixture/size, batch size, hardware budget, and action-execution/replanning schedule. E04E08E13E18E19

Reader analysis

A focused portability test would compare frozen versus trainable world decoders under matched training and evaluation, then repeat across backbone families and training seeds. Separately, running the exported checkpoint through the accelerator and simulator would test the deployment boundary beyond tensor equality. E07E11E17

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 world decoder really transfer across backbones?

Reader-proposed experiment: repeat the compact backbone substitution with matched data, training steps, initialization policy and evaluation episodes under two conditions: freeze the retained world decoder, or fine-tune it as in the paper. Repeat training seeds and report paired task success plus future-feature error. If freezing preserves the success distribution and subgoal accuracy, the interface supports stronger module portability. A reproducible drop restricted to the frozen condition would show that decoder adaptation is an essential part of the reported substitution. E05E07E09E13

Check 2: Does the exported checkpoint preserve executed behavior?

Reader-proposed experiment: load the compact raw checkpoint and its deployment export on the same GPU with identical precision, preprocessing, random action noise and controller settings. Compare retained tensors, generated action chunks, peak device memory and end-to-end latency before running the same paired LIBERO episodes. Export should preserve actions within a declared numerical tolerance and retain episode outcomes; a systematic deviation would falsify behavioral parity even if tensor names and counts pass the CPU audit. Report raw-to-raw and export-to-export parameter comparisons separately. E08E11E12

8.3 Reading coverage

Visual audit: Visually inspected Figure 1, Table 2, Figure 2, Table 3 and Table 4 in the immutable source PDF, then inspected all five final PNG crops. Tables 2 and 3 retain the source’s inconsistent raw/deployment baseline labels; Figure 2 is an auxiliary RGB visualization. Table 1’s RoboTwin/body-text inconsistency remains unresolved and is not used as an accepted result here. This pass does not execute code, verify external repository contents, or add training seeds.

PDF pages inspected for this edition: 2, 7, 8, 12. 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
  • 3.1 Model Identity and Scope
  • 3.2 Problem Formulation
  • 3.3 Stage 1: Learning an Action-Bearing Latent World Model
  • 3.4 Stage 2: Predicting Latent Actions from Language and Vision
  • 3.5 Subgoal-Conditioned Flow-Matching Action Expert
  • 3.6 Inference, Module Boundaries, and Evidence Levels
  • 4.1 Experimental Setup
  • 4.2 Recent Benchmark Context
  • 4.3 Compact-Backbone Study
  • 4.4 Qualitative LIBERO Prediction Visualization
  • 4.5 Deployment Reachability and Exact Export
  • 4.6 Adapter Contract Validation
  • 4.7 Language Grounding Diagnostic
  • 5 Limitations
  • 6 Conclusion
  • Ethics Statement
  • Reproducibility Statement
  • AI Use Statement
  • References
  • A Implementation Details
  • B Compact LIBERO Results by Suite
  • C Evidence Taxonomy and Explicit Non-Claims

Outside the original text pass

  • Figure images were not supplied or visually inspected; only their extracted text and captions were read.
  • Merged cells in Table 1 were not reconstructed. Its legible AcrossWAM1.0 RoboTwin entries conflict with the surrounding prose and are withheld as accepted benchmark results.
  • Separate supplemental material availability remains unverified.
  • No repository, checkpoint, external reference, or machine-readable experimental artifact was inspected.
  • 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 visual audit above records the subsequent illustrated pass.

8.4 Traceable evidence

E01PDF p. 1, title, author block, and arXiv version lineInspect

The title and authors Yafei Zhang and Nan Wu match the catalog identity. The supplied artifact identifies itself as arXiv:2608.29937v1, dated 30 August 2026.

Go to primary source ↓
E02PDF pp. 1–3, Abstract, Section 1 contributions, and Section 3.1Inspect

The paper attributes latent-subgoal learning to LaWAM and presents its own contribution as modular interfaces, backbone scaling, and deployment verification.

Go to primary source ↓
E03PDF p. 3, Section 3.2, Eq. (1)Inspect

The formulation specifies observations, language, optional state, a transition latent, frozen visual features, a future-feature decoder, and a conditioned action expert.

Go to primary source ↓
E04PDF p. 4, Section 3.3, Eqs. (2)–(5)Inspect

Stage 1 uses frozen DINOv3, a one-query Gaussian inverse-dynamics posterior with latent dimension 32, latent-modulated decoding, masked state grounding, future-feature loss, and KL regularization.

Go to primary source ↓
E05PDF pp. 4–5, Section 3.4, Eqs. (6)–(7)Inspect

Eight latent-action and eight flow queries have separate roles. A cross-attention mapper predicts the teacher-aligned latent, which the retained decoder converts into a supervised future feature.

Go to primary source ↓
E06PDF p. 5, Section 3.5, Eqs. (8)–(10)Inspect

Flow matching interpolates Gaussian noise and normalized action chunks. The expert conditions on backbone and current/future visual context; the configuration uses 50 actions, a 1.2-second horizon, 10 integration steps, and auxiliary weights of 0.1.

Go to primary source ↓
E07PDF p. 5, Section 3.6 and opening of Section 4; PDF p. 6, Section 4.1Inspect

Inference excludes future frames, the inverse-dynamics encoder, and state-grounding head. Compact training freezes DINOv3 and the teacher while optimizing the decoder and policy modules. LaWAM's published control and latency results are explicitly inherited references.

Go to primary source ↓
E08PDF p. 6, Section 4.1, Models and training, Benchmarks, and StatisticsInspect

The compact model trains for 25,000 steps. Evaluation uses 50 trials for each of 10 tasks in four suites, identical episode keys and initializations, a two-point retention criterion, and one checkpoint per backbone.

Go to primary source ↓
E09PDF p. 7, Table 2, pooled columns; PDF p. 6, Section 4.3Inspect

The larger and compact policies achieve 1,960/2,000 and 1,949/2,000 successes. Larger-only and compact-only successes number 46 and 35; the reported exact McNemar p-value is 0.2664.

Go to primary source ↓
E10PDF p. 12, Appendix B, Table 4, Long/Object/Goal/Spatial rowsInspect

Larger versus compact success rates are 96.6/95.0, 99.2/99.4, 96.8/96.8, and 99.4/98.6 percent, respectively, with 500 episodes per suite.

Go to primary source ↓
E11PDF p. 7, Section 4.5; PDF p. 8, Table 3Inspect

The export audit reports 1,078 unchanged retained tensors, 321 removed training-only keys covering 411.1M parameters, 718 reconstructible aliases, and successful CPU loading. Table 3 lists compact raw/export counts of 1,883.8M/1,472.6M and a 42.4% reduction against the larger reference.

Go to primary source ↓
E12PDF p. 7, Table 2, Qwen3-VL-2B deploy-parameters cell; PDF p. 8, Table 3, first row and captionInspect

Table 2 calls 2,555.2M the larger model's deployable count. Table 3 labels the same count a raw policy and defines raw as including training-only modules.

Go to primary source ↓
E13PDF pp. 7–8, Section 4.6Inspect

MiniCPM-V 4.6 passes an author-reported adapter forward/backward test with query influence, compatible image arguments, hidden-size alignment, and decreasing loss. Its vocabulary requires explicit resizing/freezing treatment; no manipulation result is claimed.

Go to primary source ↓
E14PDF pp. 8–9, Section 4.7; PDF p. 9, Section 5Inspect

The two-task language diagnostic reports adherence of 52/190 versus 93/185, Fisher p=1.0×10^-5, conflicting collapse of 36.7% versus 38.3% with p=0.74, and non-conflicting collapse increasing from 1% to 10%.

Go to primary source ↓
E15PDF p. 7, Table 1, AcrossWAM1.0 row; PDF p. 6, Section 4.2 final paragraph; PDF p. 9, Section 5; PDF p. 12, Appendix CInspect

The table assigns AcrossWAM1.0 RoboTwin values of 92.09 and 89.12. The surrounding text says those columns are not filled with inherited or diagnostic results and restricts verified closed-loop benchmark evidence to LIBERO.

Go to primary source ↓
E16PDF pp. 6–7, Section 4.4; PDF p. 8, Figure 2 caption, text onlyInspect

A held-out trajectory is autoregressively predicted for 33 steps and rendered with an auxiliary RGB decoder solely for interpretation. The decoder is absent from policy inference and success evaluation; the caption describes increasing long-horizon blur.

Go to primary source ↓
E17PDF p. 9, Section 5; PDF p. 12, Appendix CInspect

Explicit gaps include training-seed replication, frozen cross-family decoder transfer, cross-family control, compact latency and peak memory, and accelerator/simulator export validation.

Go to primary source ↓
E18PDF p. 12, Appendix AInspect

Implementation settings include hidden size 1,024, latent dimension 32, eight queries of each type, a 16-layer/16-head expert, two noise samples per batch, clipping at 1.0, and module-specific AdamW peak learning rates.

Go to primary source ↓
E19PDF p. 9, Reproducibility StatementInspect

The authors state that an accompanying repository contains model and training code, configurations, evaluators, paired statistical tests, and machine-readable export verification. These external materials were not supplied for inspection.

Go to primary source ↓
E20PDF p. 2, Figure 1; original diagram visually inspectedInspect

The upper panel shows paired frozen DINOv3 features, inverse-dynamics teacher, world decoder and training supervision. The lower panel shows separate ACT/flow query contexts, a predicted latent, retained decoder, current features and Flow-DiT action expert. Legend distinguishes frozen components and training-only paths.

Go to primary source ↓
E21PDF p. 8, Figure 2; original image grid visually inspectedInspect

Columns display steps 0, 11, 22 and 33. Rows distinguish simulator observations, auxiliary-decoded ground-truth latents, auxiliary-decoded predicted latents and latent-change magnitude. Predicted arm detail blurs over time while scene layout remains recognizable. The heatmaps have no numerical color scale.

Go to primary source ↓

8.5 Primary sources

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