Faster-WAM: Do World Action Models Need Deep Action Modules?
1. Paper overview
In one sentence: Faster-WAM lets a one-layer action head read all 30 video-transformer layers, reducing reported inference latency while trading some RoboTwin accuracy for a stronger LIBERO-Plus result. e-interfacee-fusione-ropee-inferencee-latencye-robotwine-discrepancies
| At a glance | What to know |
|---|---|
| Research problem | Author claim The authors argue that mirroring a deep video transformer in the action module duplicates computation unnecessarily. They ask whether a shallow head can recover the useful hierarchy of video representations without maintaining a corresponding deep action stack. e-problem |
| Core mechanism | Source description Dock of Transformers makes a video transformer a representation hub with flexible head interfaces. Faster-WAM tests this general proposal with one action head; the additional head types drawn in the conceptual diagrams are not separate demonstrated applications. e-probleme-interfacee-boundaries |
| A key reported result | Complete action-chunk inference: 66.5 ms Reported latency. A 24GB consumer GPU; 32-action horizon and ten denoising steps for Faster-WAM. The reported timing includes encoding, backbone processing and action denoising. Fast-WAM: 211.7 ms, approximately 3.2× slower by the reported values; π0: 68.2 ms; π0.5: 71.4 ms; X-VLA: 105.3 ms. Fast-WAM’s time sums separately measured T5 and model execution. LingBot-VA and Motus are unmeasured because of memory capacity. The GPU model and timing dispersion are unspecified, and a chunk latency is not a measured physical control frequency. e-inferencee-latency |
| Reading caution | Source description No physical robot deployment is reported. The conclusion identifies real-world validation as future work, so simulated robustness should not be presented as demonstrated physical control. e-boundaries |
Core contributions
- Source description
Dock of Transformers makes a video transformer a representation hub with flexible head interfaces. Faster-WAM tests this general proposal with one action head; the additional head types drawn in the conceptual diagrams are not separate demonstrated applications. e-probleme-interfacee-boundaries
- Source description
The demonstrated action module has approximately 30M head parameters plus 20M fusion parameters. The complete model still contains roughly 5.05B trainable parameters, including the video backbone. e-implementation
Figure 2. Give a shallow head a learned view of the full video hierarchy. Original paper, p. 4 ↗
Excerpt from the authors’ paper; cropped without altering the figure or table.
How to read it. Read the video hub as the common source of intermediate representations. The docking interface takes keys and values from its transformer layers and converts them into a form a task head can use. The expanded KV-Fusion block separates channel projection from aggregation across layers; these address feature-space compatibility and depth access respectively. In the Faster-WAM instantiation, the relevant task head is a single action-transformer layer. It reads fused video context alongside its own action tokens. The additional head branches illustrate the general DoT architecture. They should not be counted as independently trained or evaluated Faster-WAM applications. e-interfacee-fusione-ropee-implementation
What it supports. The computational reduction is on the action side: roughly 30M head parameters and 20M fusion parameters accompany a retained 30-layer video hub. The complete model has about 5.05B trainable parameters. A small head can therefore have broad depth access without making the whole policy a small model.
Where the evidence stops. The diagram abstracts the positional conversion. The text specifies inverse 3D rotation, fusion, key normalization and 1D re-rotation before action attention. Values are not rotary-encoded. The backbone is jointly trained, not frozen.
2. Motivation
2.1 The problem and the proposed response
The authors argue that mirroring a deep video transformer in the action module duplicates computation unnecessarily. They ask whether a shallow head can recover the useful hierarchy of video representations without maintaining a corresponding deep action stack. e-problem
2.2 What this reading follows
A deep video model can contain useful features without requiring an equally deep action expert. Faster-WAM tests that idea by keeping the full video backbone and shrinking its action head to one layer. The crucial interface gives that layer access to the entire video hierarchy, projects keys and values into its feature space, and aligns rotary positions before attention. The deployment path then uses only the current frame, with ten action-denoising steps. The most useful reading connects this interface to the cumulative ablation and the timing protocol: the reported speedup is substantial, but control gains vary by benchmark and two numerical inconsistencies remain in the source. e-interfacee-fusione-ropee-inferencee-latencye-robotwine-discrepancies
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 dimension | Recorded classification |
|---|---|
| Major category | WAMs |
| Architecture | Dual-system |
| Prediction paradigm | Other mechanisms |
| Quadrant | Outside quadrants |
3.1 Evidence-based assessment
Insufficient evidence to decide
The efficient-control WAM classification is supported by joint optimization of a pretrained video backbone and an action head. Other mechanisms and Outside quadrants are defensible for the deployment path: it generates actions from current-frame features without generated-future conditioning or joint future-video/action sampling. The recorded Dual-system architecture depends on whether separately parameterized hub/head modules count as separate systems; they form one jointly trained policy. The paper does not settle that catalog boundary, so the full snapshot receives an unclear assessment. e-interfacee-implementatione-inference
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
| Inputs | Outputs |
|---|---|
|
|
4.2 Equations and their role
5. Method in detail
5.1 Preserve the representation hierarchy while reducing the action stack
Faster-WAM keeps the expensive video hub: a 30-layer Wan2.2 transformer initialized from a video foundation model. The change is that the action pathway no longer mirrors every layer. A single transformer layer reads a fused representation built from the entire hierarchy. This separates two quantities that are easy to confuse: the depth of the action network and the amount of backbone information it can access. Approximately 50M parameters cover the head and fusion module, while the full trainable model remains about 5.05B. At deployment, the video hub processes only the current conditioning frame. The action head then runs for ten denoising steps to produce a 32-action chunk. Neither the retained video depth nor the repeated sampling disappears merely because the action head has one layer. e-probleme-interfacee-implementatione-inference
Figure 1. Depth access and action-network depth become separate design choices. Original paper, p. 3 ↗
Excerpt from the authors’ paper; cropped without altering the figure or table.
How to read it. Compare where the action pathway can obtain video information in each panel. MoT pairs video and action layers in an ordered correspondence. H-Bridge relaxes the full-depth pairing through selected connections. DoT instead treats the video transformer as a hub and puts representation selection inside a docking interface. The diagram’s important change is connectivity, rather than a claim that early or late backbone features are dispensable. For Faster-WAM, all 30 video layers remain available to one action layer. The head learns how to combine them, and the video hub retains its own attention computation. e-probleme-interfacee-ablatione-boundaries
What it supports. The figure motivates a testable hypothesis: preserving access to a deep representation hierarchy may matter more than duplicating its depth in an action expert. The cumulative ablation supports this particular implementation, but does not establish that one layer is sufficient for every robot task or every backbone.
Where the evidence stops. This is a conceptual comparison, not a measured compute profile or a full head-depth sweep. The general multi-head DoT drawing exceeds the scope of the single action-head experiments reported here.
5.2 Align features and positions before computing cross-modal attention
KV-Fusion first addresses feature compatibility. Separate learned projections map video keys and values into the head’s space, and a learned matrix for each attention head mixes the 30 backbone layers into one action-layer representation. The same layer coefficients apply to keys and values, preserving their pairing; the channel projections differ. Position requires another operation because video keys carry three-dimensional rotary encoding while action queries use one dimension. The interface undoes the video rotation, fuses the canonical features, normalizes the resulting keys and applies the head’s rotary basis. Values have no rotary transform. Appendix A explains the incompatible cross term and the restored one-dimensional relative-position form. That algebra establishes positional compatibility, not a general theorem that the policy preserves physical geometry or will succeed under every scene change. e-fusione-ropee-layer-weights
5.3 Read the efficiency claim alongside its accuracy and protocol boundaries
My reading is that the paper demonstrates a useful action interface, with a benchmark-dependent tradeoff. The 66.5 ms timing is a substantial reduction from the reported Fast-WAM value, but the latter combines separately timed T5 and model execution. LIBERO and LIBERO-Plus success improve, whereas RoboTwin’s mean falls by 2.47 percentage points. The ablation narrows the architectural explanation: all four shallow-head variants share one layer, and fusion, alignment and removal of direct text cross-attention produce successive gains. It does not isolate depth from routing in the first transition, nor show that learned fusion weights cause robustness. Finally, retain the source’s numerical disagreements. The sensor prose differs from Table 2, and the baseline prose differs from Figure 3(a); silent correction would make the experiment appear more internally settled than it is. e-latencye-liberoe-robotwine-shiftse-ablatione-discrepancies
5.4 Training and inference
During training
The backbone, action head, KV-Fusion and proprioceptive encoder are jointly optimized, while the video VAE is frozen. The model starts from Wan2.2-TI2V-5B; the absence of additional embodied pretraining does not mean training from scratch. e-implementatione-training
The training recipe follows Fast-WAM objectives and uses AdamW with learning rate 1e-4, weight decay 0.01, cosine decay and logit-normal timesteps. A 32-action horizon corresponds to nine video frames after factor-four temporal downsampling. The paper does not fully restate the inherited loss and masking specification. e-training
LIBERO training uses 2,000 demonstrations for ten epochs at global batch 128. RoboTwin uses 27,500 demonstrations for five epochs at global batch 1024. These settings and the pretrained backbone must accompany any claim of data efficiency. e-training
During inference
Only the observed conditioning frame passes through the video backbone. Its keys and values support action denoising, without sampling future video. The one-layer head is evaluated repeatedly for ten denoising steps per 32-action chunk, at guidance scale 1.0. e-inference
The inference path is direct action generation from video features. It neither solves an explicit inverse-dynamics problem from generated future frames nor establishes a one-step action sampler. Removing direct text cross-attention also does not remove language conditioning from the overall model. e-interfacee-implementatione-inference
5.5 Implementation flow
- Expose the full video hierarchy
Run the 30-layer video DiT and retain keys and values associated with the conditioning frame from every layer. The lightweight head can access this hierarchy without matching its depth. e-interfacee-fusione-implementation
- Project and fuse keys and values
Undo video RoPE on the keys, merge heads for separate key/value channel projections, then mix backbone layers using a learned matrix for each attention head. Keys and values share the layer-mixing coefficients, while their channel projections differ. e-fusion
- Restore a compatible positional basis
Normalize fused keys and apply one-dimensional RoPE compatible with action queries. The appendix derives why leaving the keys in a three-dimensional video basis introduces a mismatched cross term. Values have no rotary position transform. e-rope
- Denoise through mixed-context attention
The action query reads concatenated action and fused-video keys/values in its single transformer layer. The video hub and head retain separate attention computations. Direct action text cross-attention is removed, but language still conditions the video pathway. e-interfacee-ropee-implementatione-inference
6. Experiments & results
Faster-WAM couples a pretrained 30-layer video backbone to a single-layer action head through learned cross-layer key/value fusion and positional alignment. Deployment uses current-frame features and ten action-denoising steps without generating future video. Its reported 66.5 ms latency accompanies a LIBERO-Plus improvement over Fast-WAM, but RoboTwin success falls and two result descriptions conflict with printed figures or tables.
6.1 Read the original evidence
Table 1. The fastest measured row does not dominate every success column. Original paper, p. 7 ↗
Excerpt from the authors’ paper; cropped without altering the figure or table.
How to read it. First separate the LIBERO block from RoboTwin. LIBERO contains four suite scores, their average and a latency column; RoboTwin has clean, randomized and average success. The pretraining indicator describes additional embodied pretraining, not whether a pretrained foundation model is used. Compare Faster-WAM with Fast-WAM across all three summaries: 98.5% versus 97.6% on LIBERO, 66.5 versus 211.7 ms, and 89.17% versus 91.64% on RoboTwin. Retain both footnotes. The two N/A latency entries exceeded the 24GB memory limit, while Fast-WAM’s number is the sum of separately measured T5 and model times. e-liberoe-latencye-robotwine-evaluatione-inferencee-implementation
What it supports. The reported latency ratio is approximately 3.2× in Faster-WAM’s favor, with a 0.9 percentage-point LIBERO improvement. RoboTwin moves in the other direction: its mean is 2.47 points below Fast-WAM and 3.03 below LingBot-VA. This is evidence for an efficiency tradeoff that depends on the benchmark.
Where the evidence stops. The paper identifies a 24GB consumer GPU but not its exact model or latency dispersion. Timing operations vary slightly across methods. A 32-action chunk’s generation time is not a measured physical control rate, and missing timings are not measured slowdowns.
Table 2. Robustness improves relative to Fast-WAM, with visible category-specific weaknesses. Original paper, p. 8 ↗
Excerpt from the authors’ paper; cropped without altering the figure or table.
How to read it. Read each perturbation column vertically to compare policies under the same shift. Faster-WAM improves over Fast-WAM in all seven categories, with camera success rising from 16.4% to 67.9% and language success from 68.9% to 92.1%. Then read its row horizontally: background changes and robot perturbations remain much harder than language or lighting changes. The rightmost average is the paper’s aggregate over its benchmark evaluation, not an instruction to average the seven displayed categories equally. These results use the same checkpoints trained on LIBERO and evaluated on 10,030 perturbed instances. e-evaluatione-shiftse-discrepanciese-boundaries
What it supports. The reported overall increase is 23.5 percentage points, from 51.5% to 75.0%. That recovery does not make Faster-WAM the highest-scoring listed policy: π0.5 reaches 85.7% and GEACT 80.3%. The table therefore supports a strong improvement within this WAM comparison while retaining a broader robustness gap.
Where the evidence stops. The printed sensor cell is 82.3%, whereas Section 4.2.3 says 82.7%. This edition preserves the table value and records the disagreement; it does not infer which underlying evaluation number is correct. No confidence intervals are supplied.
6.2 Results and evaluation conditions
| Task & protocol | Reported result | Comparison & interpretation |
|---|---|---|
| Complete action-chunk inference A 24GB consumer GPU; 32-action horizon and ten denoising steps for Faster-WAM. The reported timing includes encoding, backbone processing and action denoising. | 66.5 ms Reported latency | Fast-WAM: 211.7 ms, approximately 3.2× slower by the reported values; π0: 68.2 ms; π0.5: 71.4 ms; X-VLA: 105.3 ms. Fast-WAM’s time sums separately measured T5 and model execution. LingBot-VA and Motus are unmeasured because of memory capacity. The GPU model and timing dispersion are unspecified, and a chunk latency is not a measured physical control frequency. e-inferencee-latency |
| LIBERO manipulation 40 tasks, 50 trials per task, four suites; 2,000 training demonstrations. | 98.5% Mean success rate | Fast-WAM: 97.6%; LingBot-VA: 98.5%; π0.5: 96.9%. Faster-WAM’s Long suite is 97.8%, versus Fast-WAM 95.2%. The 0.9 percentage-point aggregate gain over Fast-WAM accompanies a 2.6-point Long-suite gain, but does not establish a statistically resolved advantage over a tied or nearly saturated baseline. e-evaluatione-liberoe-boundaries |
| RoboTwin 2.0 manipulation 50 tasks, 100 trials per task in clean and randomized conditions; 27,500 training demonstrations. | 89.70% clean; 88.64% randomized; 89.17% average Success rate | Fast-WAM: 91.42%, 91.86%, 91.64%; LingBot-VA: 92.90%, 91.50%, 92.20%. The mean is 2.47 percentage points below Fast-WAM and 3.03 below LingBot-VA. The efficiency result therefore has a task-dependent accuracy tradeoff. These are simulator evaluations. e-evaluatione-robotwine-boundaries |
| LIBERO-Plus robustness The same LIBERO-trained checkpoints, evaluated on 10,030 perturbed instances across seven categories. | 75.0% Reported average success rate | Fast-WAM: 51.5%, a 23.5 percentage-point gap; π0.5: 85.7%; GEACT: 80.3%. Every listed perturbation category improves over Fast-WAM, but the strongest listed aggregate VLA baselines remain higher. Background and robot changes still yield only 57.0% and 49.0%. e-evaluatione-shifts |
| LIBERO-Plus cumulative design ablation Sequential changes; all four Faster-WAM variants use a one-layer head. | 60.3% final-layer-only → 66.8% fused KV → 71.3% aligned RoPE → 75.0% without action text cross-attention Success rate in Figure 3(a) | Figure 3(a) and Table 2 show Fast-WAM at 51.5%; Section 4.3 instead says 49.5%. The conditional additions are +6.5, +4.5 and +3.7 percentage points. The initial depth/interface change combines factors, and the cumulative design cannot establish each component’s independent contribution or resolve the conflicting baseline text. e-ablatione-discrepancies |
6.3 Ablations and diagnostic examples
Read component removals and qualitative examples within their stated evaluation conditions.
Figure 3(a–b). Follow the conditional improvements, then inspect what the learned mixer contains. Original paper, p. 8 ↗
Excerpt from the authors’ paper; cropped without altering the figure or table.
How to read it. Panel (a) begins with Fast-WAM at 51.5%. The remaining four points all use a one-layer head: final-layer-only context reaches 60.3%, fused keys and values 66.8%, RoPE alignment 71.3%, and removal of direct action text cross-attention 75.0%. Each point adds to the preceding configuration, so a step’s gain is conditional on earlier choices. Panel (b) shows the learned layer coefficients for 24 attention heads across 30 backbone layers. Its signed color scale matters: these values are parameters of the fusion operator, not nonnegative attention probabilities. Strong middle-layer coefficients coexist with contributions elsewhere in the hierarchy. e-ablatione-layer-weightse-discrepanciese-fusion
What it supports. The plot supports the usefulness of the proposed interface in this training configuration: adding fusion gives 6.5 points, alignment another 4.5, and removing the action text branch another 3.7. The heatmap is consistent with using more than the final video layer, but coefficient magnitude alone cannot establish causal importance.
Where the evidence stops. Section 4.3 calls the baseline 49.5%, conflicting with the plot’s 51.5% and Table 2. The first transition changes depth and interface together, and the experiment is not a full factorial test of depth, fusion, position and language access.
7. Analysis & limitations
7.1 What the evidence leaves open
No physical robot deployment is reported. The conclusion identifies real-world validation as future work, so simulated robustness should not be presented as demonstrated physical control. e-boundaries
Two source inconsistencies remain unresolved: Table 2 prints sensor success as 82.3%, whereas the prose says 82.7%; Figure 3(a) and Table 2 use a 51.5% Fast-WAM baseline, whereas the ablation prose says 49.5%. Results here follow the original table and plotted labels. e-discrepancies
Baseline success rates are drawn from cited studies, while latency operations differ slightly and Fast-WAM’s time is assembled from two separate measurements. No exact GPU SKU, precision, confidence intervals or training-seed dispersion is supplied. These omissions limit a strictly matched efficiency or significance comparison. e-evaluatione-latencye-boundaries
The cumulative ablation and signed fusion-weight heatmap support the proposed interface but do not isolate every factor or measure causal layer importance. The broader multi-head DoT design is represented conceptually; only the action application is evaluated. e-probleme-ablatione-layer-weightse-boundaries
7.2 Questions for discussion
- Would the one-layer advantage survive a matched depth-by-routing experiment that holds RoPE alignment, language access and training budget fixed?
- Does the speed advantage persist with identical end-to-end timing boundaries and hardware, and how much RoboTwin success must be traded for each latency reduction?
8. Reproducibility audit
8.1 Requirements and known gaps
A reconstruction needs the Wan2.2-TI2V-5B initialization, the stated datasets and preprocessing, a 30-layer video hub, a one-layer 24-head action module, and separate key/value projections with shared layer-mixing coefficients. Preserve the current-frame-only inference boundary and ten-step action schedule. e-fusione-implementatione-traininge-inference
The inherited Fast-WAM loss/masking recipe, exact rotary position assignment and normalization conventions, training hardware, inference precision and timing methodology require implementation-level verification. This reading did not inspect code or external supplements and cannot certify those missing details. e-traininge-ropee-latency
Before reproducing the ablation or sensor result, obtain the underlying evaluation artifacts to reconcile the two printed inconsistencies. Until then, retain the contradictory locations explicitly rather than silently replacing one number. e-discrepancies
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: Separate head depth from information access
Reader-proposed: cross one, two and 30 action layers with final-layer-only versus all-layer video access, and with aligned versus unaligned rotary keys. Hold the video initialization, demonstrations, training updates, denoising schedule and language path fixed; repeat training seeds and report success with end-to-end latency. A persistent one-layer advantage at matched routing would support the depth hypothesis. Gains that disappear after matching fusion or position would instead attribute the original improvement mainly to the interface. Check both LIBERO-Plus and RoboTwin so that a robustness gain cannot conceal a manipulation loss. e-implementatione-fusione-ropee-ablatione-robotwin
Check 2: Reconcile the numbers under one end-to-end timing protocol
Reader-proposed: obtain the evaluation artifacts for the sensor column and Fast-WAM ablation baseline, then rerun the compared policies on the same GPU, precision and checkpoint inputs. Time the full path from received observation and instruction to a completed 32-action chunk, including text encoding consistently; declare any text-cache policy, warmup, synchronization and memory limit, and report median and tail latency. If the approximately 3.2× ratio persists with matched boundaries while success remains within uncertainty of the reported values, the efficiency conclusion strengthens. Otherwise identify whether timing composition or accuracy changes explain the difference. This is a proposed verification, not a reproduced result. e-latencye-inferencee-evaluatione-discrepancies
8.3 Reading coverage
Visual audit: Read the complete 15-page source text in four consecutive chunks, including the bibliography and Appendix A. Visually inspected the title block on page 1, Figures 1–3, Tables 1–2 including latency footnotes, the main method equations and the Appendix A derivation on pages 14–15. Final crops of Figures 1, 2 and 3(a–b) and Tables 1–2 were individually reopened after extraction from 300-DPI renders. The sensor and ablation-baseline text/table disagreements were checked against the original page images. Remaining PDF pages were read as text. No external supplements, code or cited manuscripts were inspected, and no experiments were run.
PDF pages inspected for this edition: 1, 3, 4, 5, 6, 7, 8, 14, 15. 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 Dock of Transformers
- 3.1.1 KV-Fusion Docking Mechanism
- 3.1.2 Video-Action RoPE Alignment
- 3.2 Faster-WAM
- 4.1 Implementation Details
- 4.2 Benchmarks and main results
- 4.3 Ablation Study
- 5 Conclusion
- References, as supplied bibliography only
- Appendix A: RoPE alignment derivation
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.
- Read all 15 pages of extracted text, including the bibliography and Appendix A. Visually inspected PDF pages 1, 3, 4, 5, 6, 7, 8, 14 and 15; remaining pages were read as text.
- All three numbered figures and both numbered tables were visually inspected. Final crops were reopened individually after extraction from 300-DPI source renders.
- No external supplemental material, project implementation or cited paper was inspected. No experiments were reproduced.
- Title, identifier and authors were checked against the retained PDF title page, which is version 1 dated 3 August 2026.
The visual audit above records the subsequent illustrated pass.
8.4 Traceable evidence
e-identityPDF p. 1, title block and version stamp
The title is Faster-WAM: Do World Action Models Need Deep Action Modules? Authors are Liheng Ma, Rui Heng Yang, Zhanguang Zhang, Mateo Clemente, Ziwen Hu, Tongtong Cao and Yingxue Zhang. The affiliations are Huawei Noah’s Ark Lab, Huawei Celia Team and Department of Foundation Model, 2012 Labs. The stamp reads arXiv:2608.02365v1 [cs.AI], 3 Aug 2026.
Go to primary source ↓e-problemPDF pp. 1–3, Abstract, Introduction and Figure 1
The authors question matching the action expert to the depth of a pretrained video transformer. Figure 1 contrasts MoT one-to-one layer pairing, H-Bridge connections to selected layers, and DoT docking heads with access to the backbone hierarchy. The multi-head diagram is a general architecture; Faster-WAM instantiates an action head.
Go to primary source ↓e-interfacePDF p. 4, Section 3.1, Figure 2 and unnumbered mixed-attention expression
The docking interface exposes video keys and values and fuses them into head-compatible representations. An action query attends over concatenated action and fused-video keys/values. The video backbone is the representation hub; its depth need not equal that of the head.
Go to primary source ↓e-fusionPDF p. 5, Section 3.1.1, Equations (1)–(3)
Current-conditioning-frame keys and values from all video layers enter KV-Fusion. Keys have video RoPE removed. Separate learned channel projections W_K and W_V map merged-head features into the action feature space. A learned per-head matrix A_h of shape L_a by L_v mixes layers with the same coefficients for keys and values. Faster-WAM uses L_a=1 and L_v=30.
Go to primary source ↓e-ropePDF pp. 5–6, Sections 3.1.1–3.1.2, Equations (4)–(5); Appendix A, pp. 14–15, Equations (6)–(20)
Video keys initially carry 3D spatiotemporal RoPE, while action queries use 1D RoPE. Inverse rotation, fusion, key normalization and 1D re-rotation put keys in the head basis; values have no RoPE. Appendix A derives the mixed-basis cross term and the aligned relative-position form. The hub and heads use separate attention computations rather than MoT shared attention.
Go to primary source ↓e-implementationPDF p. 6, Section 4.1, Architecture and Training
The video backbone is Wan2.2-TI2V-5B with 30 transformer layers. The one-layer action head has hidden size 1024 and attention dimension 3072=24×128. It has about 30M parameters; KV-Fusion adds about 20M, for about 5.05B trainable parameters. Action text cross-attention is removed. The backbone, action head, KV-Fusion and proprioceptive encoder are jointly optimized; the video VAE is frozen.
Go to primary source ↓e-trainingPDF pp. 6–7, Sections 3.2 and 4.1–4.2
The paper follows Fast-WAM training objectives and inference procedure without additional embodied pretraining. Training uses AdamW, learning rate 1e-4, weight decay 0.01, cosine decay and logit-normal flow timesteps. The action horizon is 32; factor-four temporal downsampling yields nine training video frames per chunk, and multiview images are spatially concatenated before the VAE. LIBERO uses 2,000 demonstrations, ten epochs and global batch 128. RoboTwin uses 27,500 demonstrations, five epochs and global batch 1024.
Go to primary source ↓e-inferencePDF p. 6, Sections 3.2 and 4.1, Inference
At inference the model processes only the conditioning frame, without generating future video. Its current-frame video KV conditions action denoising. A 32-step action chunk uses ten denoising steps with classifier-free guidance scale 1.0. Reported latency includes encoding, backbone forward execution and action denoising, with some variation in operations across methods.
Go to primary source ↓e-evaluationPDF p. 7, Section 4.2, Benchmarks
LIBERO evaluates 40 tasks with 50 trials per task. LIBERO-Plus applies seven perturbation categories to the same LIBERO-trained checkpoints, yielding 10,030 instances. RoboTwin 2.0 evaluates 50 tasks with 100 trials per task under clean and randomized conditions. Baseline success rates are drawn from cited works.
Go to primary source ↓e-liberoPDF p. 7, Table 1, LIBERO columns
Faster-WAM achieves Spatial 98.4%, Object 100.0%, Goal 97.0%, Long 97.8% and average 98.5%. Fast-WAM reports 98.2%, 100.0%, 97.0%, 95.2% and 97.6%. LingBot-VA also averages 98.5%; π0 averages 94.1% and π0.5 96.9%.
Go to primary source ↓e-latencyPDF p. 7, Table 1 including both footnotes and Section 4.2.1; p. 6, Section 4.1
On a 24GB consumer GPU, Faster-WAM latency is 66.5 ms, versus Fast-WAM 211.7 ms, π0 68.2 ms, π0.5 71.4 ms and X-VLA 105.3 ms. Fast-WAM’s value sums separately measured T5 and model latency. LingBot-VA and Motus have N/A because they exceed the stated memory capacity. The exact GPU model, precision and latency dispersion are not specified.
Go to primary source ↓e-robotwinPDF p. 7, Table 1, RoboTwin 2.0 columns
Faster-WAM reports 89.70% clean, 88.64% randomized and 89.17% average. Fast-WAM reports 91.42%, 91.86% and 91.64%; LingBot-VA reports 92.90%, 91.50% and 92.20%. Faster-WAM exceeds the listed VLA averages but trails these two WAMs.
Go to primary source ↓e-shiftsPDF p. 8, Table 2 and Section 4.2.3
Faster-WAM success across background, camera, language, light, object, robot and sensor perturbations is 57.0%, 67.9%, 92.1%, 94.3%, 82.7%, 49.0% and 82.3%, with a reported 75.0% average. Fast-WAM reports 53.7%, 16.4%, 68.9%, 78.2%, 60.7%, 44.5%, 37.7% and 51.5%. π0.5 averages 85.7% and GEACT 80.3%.
Go to primary source ↓e-discrepanciesPDF p. 8, Table 2 versus Section 4.2.3, and Figure 3(a) versus Section 4.3
The sensor cell in Table 2 is 82.3%, while the prose says 82.7%. Figure 3(a) labels the Fast-WAM baseline 51.5%, agreeing with Table 2, while Section 4.3 calls that baseline 49.5%. This report preserves the printed table/plot numbers and records both prose discrepancies without inferring corrected experimental data.
Go to primary source ↓e-ablationPDF p. 8, Figure 3(a); pp. 8–9, Section 4.3
The plotted LIBERO-Plus sequence is Fast-WAM 51.5%, one-layer Final Only 60.3%, adding fused KV 66.8%, adding RoPE alignment 71.3%, and removing action text cross-attention 75.0%. The last four points all use a one-layer head. The sequence is cumulative rather than a full factorial experiment.
Go to primary source ↓e-layer-weightsPDF p. 8, Figure 3(b); p. 9, Section 4.3
The heatmap displays 30 backbone-layer columns and 24 attention-head rows, with signed layer-fusion weights. Middle layers contribute strongly in the plotted weights, while information is distributed across the hierarchy. These are learned fusion parameters, not per-input routing probabilities or controlled measures of causal importance.
Go to primary source ↓e-boundariesPDF p. 9, Section 5, Conclusion; pp. 7–8, Tables 1–2 and Figure 3
All reported control benchmarks are simulated; deployment on real robots is stated as future work. The reported tables and ablation plot do not provide confidence intervals or training-seed variation. The broad DoT proposal is demonstrated through a single action-head application.
Go to primary source ↓8.5 Primary sources
Faster-WAM: Do World Action Models Need Deep Action Modules? ↗
PDF · 7,258 extracted words
Source fingerprint
c244f0d03968783f48341bb34891a81d61e2c04ef04df2cf7f5b15f08c2ca96e