acceptodds
Under review as a conference paper at ICLR 2027

Acting through Registers

Abstract

World Action Models (WAMs) use generative visual representations to condition robot actions, yet how these representations organize action-relevant information remains poorly understood. We identify persistent high-norm spatial outliers in WAM representations and find that their action relevance persists even after normalization and K/V projection attenuate their extreme magnitude. Motivated by this observation, we introduce RegisterWAM, which adds learnable non-spatial registers to the world model and uses them directly for action conditioning. Registers alone do not fully eliminate spatial outliers; pairing them with observation flow-matching (FM) loss supervision suppresses spatial outliers while concentrating high-norm responses in registers. Beyond this reorganization, observation FM loss supervision also improves action learning without registers, motivating an additional noised-observation FM objective that learns current-scene recovery from corrupted inputs. The resulting model achieves 81.64% success on LIBERO-Plus without additional embodied pretraining data or foundation models, reaches 93.0% average success on RoboTwin 2.0, and improves real-world manipulation performance. Code and models will be made public.

Primary area:
applications to robotics, autonomy, planning
Keywords:
World Action Models, embodied Intelligence

Reviews

Among:
Shape:
1234this paper, Top 49%
Figure 1. Mean review rating of the reviewed papers at ICLR 2027. This paper averages 2.33 over 3 reviews, above 51% of the others.
Reviewer_TnNROpenReview ↗
Rating
3
Confidence
3
Informativeness
2
Presentation
3
Soundness
2
Read review
Paper summary

This paper studies high-norm spatial tokens in World Action Models and finds that their influence on predicted actions persists after normalization and K/V projection attenuate their extreme magnitudes. It proposes RegisterWAM, combining non-spatial registers with observation flow-matching supervision and using register states to condition action generation. An additional noisy-observation objective trains current-scene recovery. The method is evaluated on LIBERO-Plus, RoboTwin 2.0, real-world manipulation, and an ImageWAM backbone.

Critical strengths
  1. The representation analysis distinguishes activation magnitude from downstream action relevance. Tracing tokens through the K/V interface and measuring masked-token action sensitivity provides useful evidence beyond norm visualizations alone.

  2. The method is motivated through informative controls. Registers alone do not eliminate final-layer outliers, while observation supervision and registers have complementary effects across layers. The comparison between register-only and combined conditioning also helps justify the interface choice.

  3. Evaluation covers several complementary settings. The full LIBERO-Plus evaluation supports robustness across distribution shifts, while real-world experiments and ImageWAM transfer provide evidence beyond a single simulation benchmark.

Critical weaknesses
  1. The explanatory depth of the representation findings remains limited. Masked-token sensitivity establishes that a token affects predicted actions, but does not establish what information it carries. Likewise, moving high-norm responses from spatial tokens to registers does not directly demonstrate a corresponding transfer of information. The relationship between this reorganization and improved policy performance therefore remains unresolved, particularly because observation supervision also helps without registers.

  2. The methodological novelty is incremental, which makes the empirical evidence especially important. Registers and compact latent interfaces are established designs, while the paper does not report performance variability across training runs. This limits confidence in smaller gains, such as the RoboTwin improvement, although the broader LIBERO-Plus results are encouraging.

Questions for authors
  1. Please specify the exact targets and weighting of the clean- and noisy-observation FM losses, particularly at σ=0\sigma=0. Does the zero-noise implementation use the standard velocity target ϵ−x\epsilon-x?

  2. How was the 20% LIBERO-Plus evaluation subset selected? Please clarify the sampling unit, selection rule, task and perturbation coverage, and provide the instance list or generation procedure. Does Table 3 use this subset or the full benchmark?

  3. How many independent training runs and evaluation trials support the principal comparisons? What variability is observed for the key ablation, RoboTwin, and ImageWAM results?

  4. Beyond token norms and masked-token sensitivity, is there evidence distinguishing the proposed non-local representational role of outliers from the behavior of ordinary background tokens? Such evidence would strengthen my assessment of the empirical contribution.

3 comments
AuthorsOpenReview ↗

Summary and Q4/W1

We thank the reviewer for recognizing the representation analysis and the breadth of our evaluation. We have added local/global generation interventions in FastWAM and OpenWAM[1](extra WAM), clarified the flow-matching implementation and evaluation protocol, and supplemented the empirical evidence with additional training and evaluation results on LIBERO-Plus and ImageWAM.

About Q4/W1: The non-local role of spatial outliers and policy performance

Our additional experiments support a non-local computational role for spatial outliers, consistently across FastWAM and OpenWAM. We mask observation-token K/V contributions in layers 20-27 and compare future-video generation with the unmasked output using the same initial noise. Local and global generation-change MSE measure the spatial extent of the intervention, while PSNR changes characterize its effect on reconstruction fidelity (new Figures 13-14).

Compared with ordinary-token masking, outlier masking produces smaller changes at its corresponding spatial location but larger changes across the generated video. The median per-observation outlier-to-ordinary local/global MSE ratios are 0.0144/6.54 for FastWAM and 6.45×10−6/10.636.45\times10^{-6}/10.63 for OpenWAM. The global PSNR distributions also show a broader response to outlier masking.

This distinction persists within the same padding region. In OpenWAM, we compare the outlier with the other 19 tokens in its native padding region. Relative to their mean perturbation effect within each observation, the outlier exhibits median local/global MSE ratios of 0.0137/30.92, demonstrating substantially stronger global influence despite sharing the same padding region. These analyses use 200 observations from distinct episodes spanning all 40 LIBERO tasks.

Together, these interventions provide evidence beyond action sensitivity that spatial outliers exert influence beyond their corresponding image regions. This motivates dedicated non-spatial registers, together with observation FM supervision to constrain spatial representations. Figure 15 shows their complementary effects: registers attenuate intermediate-layer spatial peaks, while observation supervision suppresses residual terminal-layer peaks. The component ablations further connect this design to task performance (Appendix C.2, Table 9).

We interpret the norm redistribution as representation reorganization; the generation interventions establish its non-local functional motivation, and the component controls assess the resulting policy design.

[1] Wang et al. OpenWAM. https://arxiv.org/abs/2609.07398

About Q3/W2: Contribution, repeated training, and performance variability

Our contribution is the identification and functional analysis of spatial outliers in Video-DiT world-action models, together with the resulting register-based action interface and observation supervision.

We have supplemented the empirical evaluation with additional trained Full models, repeated additional component ablations, and an additional ImageWAM model. The current descriptive summaries are:

Model/evaluationSuccess rate (%)Training results
Full, LIBERO-Plus: 10,030 instances81.67±1.2481.67\pm1.243
Full, fixed subset: 2,008 instances81.51±1.6381.51\pm1.633
No registers, fixed subset79.98±0.7779.98\pm0.772
No clean FM, fixed subset79.28±0.1479.28\pm0.142
ImageWAM + Ours, fixed subset85.55±0.1985.55\pm0.192

The key component controls each include two separately trained checkpoints. The no-register results are 79.43% and 80.53%, and the no-clean-FM results are 79.38% and 79.18%, evaluated on the same fixed 2,008-instance subset.

Two ImageWAM training runs achieve 85.42% and 85.68%, yielding 85.55±0.19%85.55\pm0.19\%. Both runs use the original ImageWAM implementation and directly adopt the 20% LIBERO-Plus subset provided with its source code.

Given the substantial computational cost of repeating the RoboTwin training and evaluation pipeline, we prioritized supplementary experiments on LIBERO-Plus and ImageWAM.

All the above changes have been included in the latest revision.

About Q1: Observation FM targets and weighting at σ=0\sigma=0

At σ=0\sigma=0, the input is the clean observation latent xox_{o}, and we supervise the prediction with −xo-x_{o}. This is the conditional expectation of the FM velocity target ϵ−xo\epsilon-x_{o} at the clean endpoint, since the noise is independent and zero-mean: E[ϵ−xo∣xo]=−xo.\mathbb{E}[\epsilon-x_o\mid x_o]=-x_o.

Under squared-error regression, this deterministic target has the same optimal predictor as the sampled velocity target at this endpoint. The loss averages over latent channels and spatial positions. The clean-observation term shares the visual-loss weight of 2 and the valid-frame normalization with future frames.

For the noisy-observation branch, the input is 0.5xo+0.5ϵ0.5x_{o}+0.5\epsilon and the target is ϵ−xo\epsilon-x_{o}, with loss weight 2w(500)≈4.15282w(500)\approx4.1528. We have clarified these targets and weights in Appendix D.1.

About Q2: Construction of the 20% subset and Table 3

The sampling unit is a benchmark instance (suite, task_id). We partition all 10,030 LIBERO-Plus instances into 28 suite ×\times perturbation-category strata, sort buckets and task IDs, and use a random generator initialized with sampling seed 42 to sample max(1, round(0.2*n)) without replacement per bucket. This yields 2,008 instances. This generation procedure specifies the instance selection reproducibly; the fixed instance list is retained for all component-ablation comparisons. Each selected instance has one rollout; this gives suite/category coverage rather than exact balance over task difficulty.

Table 1 uses the full 10,030-instance benchmark; Table 3's ImageWAM+Ours directly uses the 20% LIBERO-Plus subset (2,008 instances) provided with the original ImageWAM code, where our method achieves 85.55±0.19%85.55\pm0.19\% using the original ImageWAM implementation. The baseline of 83.10% is taken from the ImageWAM paper.

We will release the subset-generation code and the fixed 2,008-instance evaluation list alongside our code release to enable exact reproduction of the evaluation subset.

All the above changes have been included in the latest revision.

Reviewer_1jamOpenReview ↗
Rating
2
Confidence
3
Informativeness
2
Presentation
2
Soundness
2
Read review
Paper summary

The paper identifies an interesting behavior of spatial tokens in WAM representations. Motivated by this observation, it introduces RegisterWAM, which adds learnable non-spatial registers to the world model and uses them directly for action conditioning. Combined with observation flow-matching supervision, RegisterWAM suppresses spatial outliers while concentrating high-norm responses in the registers. Experiments demonstrate improvements in task success over the baseline models.

Critical strengths
  1. The paper identifies an interesting phenomenon involving high-norm spatial-token outliers in the middle-to-late layers. The authors propose a plausible hypothesis that these tokens handle information beyond their corresponding local image regions.

  2. The use of registers provides dedicated non-spatial capacity that may accommodate information otherwise represented by spatial outlier tokens. Combined with observation FM supervision, this design produces more uniform spatial-token norms.

  3. The method achieves strong performance on several benchmarks without additional embodied-policy pretraining or external priors.

Critical weaknesses
  1. The relationship between spatial outliers and action-learning performance remains unclear. Figure 4 shows that the outlier tokens have relatively high, but not dominant, action sensitivity. However, sensitivity alone does not establish whether these tokens improve or degrade action quality. Masking ordinary spatial tokens can also change the predicted actions, and such changes do not indicate whether performance becomes better or worse. The motivation for suppressing these outliers therefore needs further clarification.

  2. The residual outliers in the final layers are insufficiently explained. Figure 12 shows that registers alone substantially reduce spatial outliers across much of the network, while elevated spatial norms remain in the final one or two layers. It is unclear whether these residual outliers reflect the same mechanism as the earlier-layer peaks or arise from a different process. Although register-count ablations are provided, the paper does not systematically show whether increasing the number of registers can eliminate these residual outliers without observation FM supervision.

  3. The register-count sweep in Figure 6 is conducted with observation FM supervision, whereas adding registers alone produces only a small improvement over the baseline. Figure 14 shows that selected register–head pairs attend to task-relevant interaction regions, including the gripper and objects. However, the connection between these attention patterns and the limited standalone performance gain is unclear. It would be helpful to clarify how observation supervision enables the register interface to become more effective, and how much of the overall improvement should be attributed to registers rather than the auxiliary losses.

Questions for authors
  1. Can you provide evidence connecting spatial outliers to action quality or task success, beyond changes in predicted actions? For example, how does task performance change when masking outliers compared with masking matched non-outlier tokens? Otherwise, why suppressing these outliers important?

  2. Why do spatial outliers remain in the final layers after adding registers (Figure 12)? Could you report layer-wise outlier statistics?

  3. How does a model with both clean- and noised-observation FM losses, but without registers, perform? Table 5 introduces registers, clean-observation FM, and an interface change together, making their individual contributions difficult to separate.

  4. Does the same outlier behavior occur in other WAM architectures? The representation analysis primarily focuses on FastWAM, so additional evidence would help establish the scope of the findings.

9 comments
AuthorsOpenReview ↗

Summary and W1/Q1

We thank the reviewer for the positive assessment of our empirical findings on spatial outliers, the clarity of our architectural motivation, and the comprehensive evaluation across LIBERO-Plus, RoboTwin 2.0, and real-world tasks. In response to your suggestions, we have extended our analysis to OpenWAM [1], added local/global future-video generation interventions, and conducted separate component-removal ablations. We address your specific questions below.

About Q1/W1: Task-Success Impact of Outlier Masking and the Motivation for Dedicated Non-Spatial Capacity

1. Direct intervention on task success (Table 8). To test how spatial outliers affect policy execution, we mask the outlier's K/V contribution to action attention (and compare against masking a uniformly sampled non-outlier token) across 2,000 ID rollouts on LIBERO and the fixed 2,008-instance OOD subset on LIBERO-Plus:

ModelBenchmarkBaselineOutlier MaskRandom Non-Outlier Mask
FastWAMLIBERO (ID)96.95%97.45%96.95%
FastWAMLIBERO-Plus (OOD)51.00%51.05%51.74%
OpenWAMLIBERO (ID)99.10%99.40%99.15%
OpenWAMLIBERO-Plus (OOD)69.77%69.12%69.37%

On ID LIBERO, masking the outlier yields small numerical gains (+0.50% on FastWAM and +0.30% on OpenWAM). On OOD LIBERO-Plus, however, test-time single-token masking does not produce a consistent improvement across models or controls. Thus, test-time masking alone does not establish a uniform policy gain under distribution shifts, which motivates examining the outlier's role inside the world model itself.

2. Non-local computational role in future-video generation (new Figures 13–14). To examine what these tokens compute beyond action sensitivity, we mask observation K/V contributions in layers 20–27 during future-video generation (with fixed initial noise) across 200 observations spanning all 40 LIBERO tasks. Relative to masking ordinary spatial tokens, masking the outlier causes much smaller local change at its own spatial patch (DLD_L median ratio: 0.0144 in FastWAM, 6.45×10−66.45 \times 10^{-6} in OpenWAM) but substantially larger global change across the entire generated video (DGD_G median ratio: 6.54 in FastWAM, 10.63 in OpenWAM).

3. Motivation: Functional decoupling is the goal; outlier suppression is its observable manifestation. Our motivation is not to treat outlier suppression as an end in itself, nor to view spatial outliers merely as noise to be erased post hoc. Rather, these outliers serve as a diagnostic signature that spatial patch tokens are being overloaded with non-local computation, entangling local visual encoding and scene-wide dynamics within the same spatial grid.

This motivates providing dedicated non-spatial registers to accommodate non-local computation while applying observation FM supervision to keep spatial tokens anchored to local visual prediction. As shown in Figure 15, registers suppress intermediate-layer spatial peaks while observation FM supervision suppresses residual terminal-layer peaks—and this representational separation translates into consistent policy gains on LIBERO-Plus (Table 9).

About Q2/W2: Residual Terminal-Layer Outliers and Register Capacity

1. Why terminal-layer outliers remain without observation supervision. Adding registers alone suppresses intermediate-layer spatial outliers, but elevated norms persist in the final layers (layers 28–29, Figure 15). This indicates that providing extra non-spatial slots does not by itself constrain how spatial tokens behave at the output interface—without explicit visual reconstruction targets on the spatial tokens, residual non-local specialization can still remain in the final spatial layer. Introducing clean-observation FM supervision explicitly regularizes the dense spatial outputs, eliminating these terminal peaks while concentrating high-norm responses in the registers (Figure 5b and Figure 15).

Regarding why terminal peaks appear after WAM adaptation in the first place, we note that pretrained Wan2.2 exhibits intermediate-layer outliers without pronounced terminal peaks (Figures 2 and 10). We hypothesize that this contrast is related to the action supervision introduced during world-action adaptation, though isolating the exact training dynamics responsible for terminal-layer formation remains an open question.

2. Scaling register capacity (new Figure 16). We further evaluated 8, 16, and 64 registers without clean-observation FM supervision across 10 fixed observations (Appendix B.3). Increasing register count monotonically attenuates—but does not eliminate—the terminal-layer norm ratio (SoutlS_{\text{out}}^l):

Number of RegistersLayer 28 SoutlS_{\text{out}}^lLayer 29 SoutlS_{\text{out}}^l
83.513.02
163.252.83
642.602.29

This confirms that capacity scaling alone is insufficient to fully separate spatial and non-spatial roles without observation-side supervision.

About Q3/W3: Isolating Observation Supervision from the Register Interface

We added the requested component-removal controls in Appendix C.2 (Table 9), evaluating each component separately while keeping Noised Obs. FM fixed across all runs (2 independent training runs per control on the fixed 2,008-instance LIBERO-Plus subset):

ConfigurationRegistersClean Obs. FMAction InterfaceLIBERO-Plus (%)
Baseline0—Raw78.59
No registers0✓Raw79.98±0.7779.98 \pm 0.77
No clean FM8—Reg.79.28±0.1479.28 \pm 0.14
Full (RegisterWAM)8✓Reg.81.51±1.6381.51 \pm 1.63
  • Clean Obs. FM alone (No registers) improves success from 78.59% to 79.98% under the raw interface.
  • Registers alone (No clean FM) achieve 79.28% while compressing the action-conditioning context from 98 spatial tokens down to 8 registers.
  • Combining both (Full) reaches 81.51%, showing clear complementarity: registers provide a compact non-spatial interface, while clean-observation FM regularizes the spatial tokens so that global action-relevant representations concentrate into the registers (consistent with Figure 15).

About Q4: Generalization of Outlier Behavior to Other WAM Architectures

Yes. We extended our investigation to the official OpenWAM model [1] (Appendix B.1–B.2, Figures 12 and 14):

  1. Pronounced spatial outliers: OpenWAM exhibits strong later-layer observation-token outliers (Soutl>30S_{\text{out}}^l > 30 in layers 20–29), localized in the native image padding region (Figure 12).
  2. Consistent non-local pattern: Masking OpenWAM's outlier produces negligible local change at its own patch (DLD_L median ratio: 6.45×10−66.45 \times 10^{-6}) but large global change across the predicted future video (DGD_G median ratio: 10.63), matching our observations in FastWAM.

[1] Wang et al. OpenWAM. https://arxiv.org/abs/2609.07398

Reviewer_HyDDOpenReview ↗
Rating
2
Confidence
3
Informativeness
2
Presentation
1
Soundness
2
Read review
Paper summary

Paper points out the presence of high norm spatial tokens especially in middle layers of FastWAM studies the influence of these spatial tokens on the actions predicted by FastWAM. Authors then introduce register tokens which do not have any spatial relevance to attenuate the norm in the middle and final layers. Further analysis show it does not solve the problem entirely and authors introduce a novel FM loss to attenuate attention. The resulting policy is then evaluated on downstream tasks in sim on Libero Plus and Robotwin benchmarks and on real tasks.

Critical strengths
  1. Main observation of the authors, regarding influence o norm outliers and their influence on action generation is particularly interesting. Although high-norm spatial tokens are established in vision transformers, examining their behavior after adaptation for action prediction provides a useful research question.
  2. Method shows strong performance in downstream performance across Robotwin and Libero Plus benchmarks as well a suite of 6 real world tasks .
Critical weaknesses
  1. Paper is really hard to follow. The lack of related works section in main paper removes any context for potential readers. The experiments to analyze a specific hypothesis are in general no coherent which makes the paper a hard read.
  2. The presentation of the paper and lack of related works makes it really hard to establish the unique contribution of the paper and how the contributions follow or distinguished from Darcet et al.'s Vision Transformers Need Registers[R1]. Register-based compression for control also has a precedent in Driving on Registers [R2]. The methodological novelty therefore depends on establishing the specific value of this WAM adaptation and its additional supervision which is not well motivated.
  3. Claims made in (182-189) regarding interpretation of high norm features are not supported by any experiment apart from that they influence actions. 4.The final-model ablations do not isolate the principal methodological additions in Table 5.
  4. Mechanistic conclusions made exceed the evidence shown in paper as the experiments are primarily focused on FAST WAM making it hard to draw more general insights.

[R1]-Vision Transformers Need Registers, Darcet et al [R2] Driving on Registers, 2026, Kirby et al

Questions for authors
  1. Details regarding implementation in general and specifically FM loss are present.
  2. What evidence distinguishes the proposed nonlocal computational role from locally informative tokens that also influence actions? 3
8 comments
AuthorsOpenReview ↗

Summary and W1/W2

Thank you for the constructive comments. We have clarified the relationship to prior register methods and added direct evidence of the spatial role of outliers in FastWAM and OpenWAM[4] (extra WAM).

About W1/W2: Presentation and distinction from prior register methods

We thank the reviewer for highlighting the need for clearer context and contribution positioning. Note that a complete Related Work section was already included in Appendix A of our initial submission due to main-text page limits; in the revision, we have moved the pointer to Related Work to the beginning of the Introduction's second paragraph (Related work is discussed in Appendix A) and expanded Appendix A to distinguish our setting from both ViTs and image DiTs.

Our study concerns generative Video DiTs for world-action modeling, where visual representations jointly support future-video generation and action conditioning. This setting differs from the ViT representations studied in Vision Transformers Need Registers [1] and the ViT-based control interface in Driving on Registers [3]. It also differs from image DiTs, which model individual images rather than future video sequences coupled to actions.

This distinction is supported by the observed behavior. Recent image-DiT work [2] reports no analogous patch-token outliers in the models examined. In contrast, we identify pronounced spatial outliers in Video-DiT WAMs and demonstrate their non-local influence on future-video generation in both FastWAM and OpenWAM (new Figures 12-14).

The behavior of registers also motivates a setting-specific design: register insertion alone attenuates intermediate-layer outliers but leaves residual terminal-layer peaks, whereas observation FM supervision suppresses these remaining peaks (Figure 15). This finding motivates studying register conditioning together with observation supervision.

Our contribution therefore combines an empirical finding in Video-DiT WAMs, a functional analysis of its non-local effects, and an action interface with complementary observation supervision. The revised presentation connects these findings to the component ablations in Table 9.

[1] Darcet et al. Vision Transformers Need Registers. https://arxiv.org/abs/2309.16588

[2] Starodubcev et al. Registers Matter for Pixel-Space Diffusion Transformers. https://arxiv.org/abs/2605.16147

[3] Kirby et al. Driving on Registers. https://arxiv.org/abs/2601.05083

[4] Wang et al. OpenWAM. https://arxiv.org/abs/2609.07398

About Q2/W3: Evidence for a non-local computational role

We have added direct interventions on future-video generation to distinguish non-local influence from local information that also affects actions. For each observation, we individually mask observation-token K/V contributions to video queries in layers 20-27 and compare the generated video with the unmasked output using the same initial noise. We measure local change at the token's spatial location and global change over the generated future video; the corresponding PSNR changes quantify effects on reconstruction fidelity (new Figures 13-14).

ModelLocal MSE ratioGlobal MSE ratio
FastWAM0.01446.54
OpenWAM6.45×10−66.45\times10^{-6}10.63

Each value is the median, over 200 observations from distinct episodes spanning all 40 LIBERO tasks, of the outlier's change MSE divided by the mean change MSE of the other individually masked observation tokens. In both models, outlier masking has a much smaller local effect but a larger global effect. The global PSNR distributions additionally show that these interventions affect whole-video reconstruction. Together, these measurements provide direct evidence of a predominantly non-local computational influence in the tested layers, beyond the action sensitivity reported in the original submission.

Therefore, this contrast between weak local perturbation and strong global video change provides a practical criterion to distinguish non-local computational tokens from locally informative ones.

About W4: Isolating the principal methodological components

We have added separate component-removal controls in Appendix C.2, Table 9, complementing the cumulative ablation in Table 5. All configurations below retain noised-observation FM.

ConfigurationRegistersClean FMInterfaceSuccess (%)
Baseline0Raw78.59
No registers0✓Raw79.98±0.7779.98\pm0.77
No clean FM8Reg.79.28±0.1479.28\pm0.14
Full8✓Reg.81.51±1.6381.51\pm1.63

The no-register control retains both observation FM objectives and uses observation/future tokens for action conditioning (Raw). It evaluates the register-based interface under the same observation objectives. The no-clean control retains the register interface (Reg.) while removing clean-observation supervision. Each component-removal control includes two training runs.

These controls separately examine the register interface and clean-observation supervision. Figure 15 provides complementary representation-level evidence: registers attenuate intermediate-layer spatial peaks, whereas observation supervision suppresses residual terminal-layer peaks.

About W5: Evidence beyond FastWAM

We have extended the analysis to the official OpenWAM model [4]. Figure 12 shows pronounced observation-token outliers across network depth, together with feature maps overlaid on three observations. Figure 14 further characterizes their spatial influence through the same local/global generation interventions used for FastWAM. OpenWAM exhibits the same qualitative pattern: outliers have weak influence at their own spatial locations but substantially stronger global influence than ordinary spatial tokens. These additional results establish that the observed phenomenon and its non-local functional signature also occur in a second WAM architecture.

About Q1: Observation flow-matching implementation

At σ=0\sigma=0, the input is the clean observation latent xox_{o}, and we supervise the prediction with −xo-x_{o}. This is the conditional expectation of the FM velocity target ϵ−xo\epsilon-x_{o} at the clean endpoint, since the noise is independent and zero-mean: E[ϵ−xo∣xo]=−xo.\mathbb{E}[\epsilon-x_o\mid x_o]=-x_o.

Under squared-error regression, this deterministic target has the same optimal predictor as the sampled velocity target at this endpoint. The loss averages over latent channels and spatial positions. The clean-observation term shares the visual-loss weight of 2 and the valid-frame normalization with future frames.

For the noisy-observation branch, the input is 0.5xo+0.5ϵ0.5x_{o}+0.5\epsilon and the target is ϵ−xo\epsilon-x_{o}, with loss weight 2w(500)≈4.15282w(500)\approx4.1528. We have clarified these targets and weights in Appendix D.1.

[4] Wang et al. OpenWAM. https://arxiv.org/abs/2609.07398

From OpenReview, CC BY 4.0. Authors as signed there; shown unedited.

Related papers

Loading the map…

Discussion (0)

Sign in to comment.