acceptodds
Under review as a conference paper at ICLR 2027

Enfold: Folding World Model Imagination into Predictive Representations for Ultra-Efficient Embodied Control

Abstract

World generative models are typically used through their outputs: a rendered future, a video-conditioned action, or latent context from a costly generative branch. We argue that their more reusable asset is the computation that constructs a future. As a generator transforms a corrupted future into a coherent trajectory, its intermediate states organize appearance, spatial layout, and interaction across abstraction levels. Can this computation be internalized in a representation inferred from the present alone? We present Enfold, which transfers it into a representation predicted from the current visual context and language instruction. During training, multi-level states exposed while the generator processes an observed future supervise a current-only encoder. The representation conditions future generation and is read by task heads without allowing task gradients to reshape the encoder. At deployment, action prediction bypasses the generator. This asymmetry encourages the encoder to retain transition structure predictable from the present rather than stochastic, future-dependent nuisance variation. Across LIBERO, RoboTwin2.0, and real-robot tasks, Enfold supports strong control and cuts action latency by relative to Fast–WAM; Enfold-Flash reaches . Analyses show that it suppresses nuisance variation and preferentially captures changes emerging over longer horizons. When a human alters the current scene, both the generated continuation and executed actions adapt, inconsistent with fixed trajectory replay. These results recast a world generator as a source of predictive control representations: its future need not be materialized at every step if its internal structure can be enfolded into the present.

Primary area:
applications to robotics, autonomy, planning
Keywords:
world generative model, representation learning, predictive representations, efficient, embodied control

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_kS4pOpenReview ↗
Rating
2
Confidence
3
Informativeness
2
Presentation
3
Soundness
2
Read review
Paper summary

This paper presents Enfold, a framework that transfers the intermediate computation of a video world model into a current-observation representation. During training, a predictive encoder learns to match multi-level generator states extracted from observed future videos, while the resulting representation also conditions future generation and a detached action head. At inference, control uses only the encoder and action head, bypassing video generation. Experiments on LIBERO, RoboTwin 2.0, and real-robot tasks show competitive control performance with substantially reduced action latency.

Critical strengths
  1. Enfold introduces an interesting representation-level interface for world models, distilling multi-level generator states into a current-observation representation and avoiding expensive video generation during control.

  2. The paper provides broad validation across simulation, real-robot control, future-video prediction, and representation analysis, showing competitive task performance with substantially lower inference latency.

Critical weaknesses
  1. The control improvements are marginal. Enfold improves over Fast-WAM by only 0.2 points on LIBERO and performs slightly worse on RoboTwin 2.0, while Enfold-Flash differs by similarly small margins.

  2. The reported efficiency gains are not fully isolated from model size and implementation differences. Enfold has roughly half the parameters of Fast-WAM, while the 10.1× Enfold-Flash result additionally uses TensorRT acceleration that is not applied to the baselines. A matched latency comparison using the same backbone, action head, sampling configuration, and deployment backend is needed to determine how much of the speedup comes specifically from bypassing the generative pathway.

Questions for authors
  1. Can the authors provide results across multiple training seeds or confidence intervals for LIBERO and RoboTwin 2.0? The differences from Fast-WAM are only 0.2 points or less, so their statistical reliability would materially affect the assessment.

  2. Can the authors provide a matched latency comparison that controls for model size, action-head configuration, sampling steps, and deployment backend? In particular, how much speedup is obtained solely by bypassing the video generator, separately from the smaller model and TensorRT acceleration?

2 comments
AuthorsOpenReview ↗

## Response to Reviewer kS4p (W1 / Q1)

We thank the reviewer for the questions on performance reliability and inference efficiency. Enfold's contribution is a representation interface that combines competitive control with improved robustness and low deployment cost. We have added independent training runs and distribution-shift evaluations, and clarify the sources of the reported speedup below. We have also included all the experiments in the revised version of the manuscript, with the added or revised content highlighted in red.

W1 / Q1 — Control performance across training seeds and distribution shifts

Enfold’s clearest advantage is under distribution shift, with substantial gains on LIBERO-Plus and improved RoboTwin clean-to-randomized transfer. Average LIBERO-Plus success improves from 51.5% for Fast-WAM to 68.7% with Cosmos (+17.2 percentage points) and 65.2% with Wan (+13.7 points). Enfold–Cosmos improves over Fast-WAM in six of seven LIBERO-Plus categories, including +43.3 points on Camera and +23.1 on Language. The Wan variant shows the same six-category improvement pattern, extending the robustness advantage across teachers.

MethodCameraRobotLanguageLightBackgroundNoiseLayoutAvg.
Fast-WAM16.444.568.978.253.737.760.751.5
Enfold, Cosmos teacher59.761.792.089.848.850.078.868.7
Enfold, Wan teacher49.269.276.184.946.451.379.165.2

These results reveal larger differences than those on saturated standard benchmarks.

For RoboTwin C2R, all methods are trained on clean demonstrations and tested on randomized scenes, with clean performance reported for reference. Our C2R training uses no data augmentation, which may contribute to the low absolute success under randomization:

MethodCleanRandomized, OOD
Fast-WAM77.81.9
Enfold, Cosmos teacher78.13.9
Enfold, Wan teacher80.37.1

Enfold–Cosmos and Fast-WAM perform similarly on clean scenes, but randomized success increases from 1.9% to 3.9% with Cosmos and 7.1% with Wan. Absolute success remains low in this challenging setting, but both OOD evaluations show an advantage beyond comparable in-distribution control.

To assess the reliability of standard-benchmark performance, we also independently retrained Enfold with seeds 42, 43, and 44. We report success (%) and mean ± sample standard deviation.

Training seed / methodLIBERORoboTwin cleanRoboTwin randomized
4297.891.6091.94
4397.991.7892.12
4497.891.5491.86
Mean ± sample standard deviation97.83 ± 0.0691.64 ± 0.1291.97 ± 0.13
Fast-WAM (reported)97.691.8891.78

Enfold consistently maintains high success across independent training runs, with standard deviations of 0.06, 0.12, and 0.13 points. This supports the repeatability of its competitive standard-benchmark performance; its overall RoboTwin mean is 91.81%, comparable to Fast-WAM’s 91.83%.

AuthorsOpenReview ↗

## Response to Reviewer kS4p (W2 / Q2)

W2 / Q2 — Inference architecture, action-head capacity, and deployment acceleration

Enfold removes the video generator from the deployed action path and makes the action head architecturally independent. Fast-WAM already skips future-video generation but still executes a video DiT to obtain action context. Its Mixture-of-Transformers (MoT) architecture couples the action expert to the video backbone through shared attention, constraining action-head depth and compatible attention interfaces. Enfold transfers predictive supervision into a current-observation encoder, so control requires only this encoder and the action head. This decoupling allows action-head depth, hidden dimension, and attention-head configuration to be chosen independently of the video backbone.

The submitted measurements on an A100 40GB separate the reported comparisons as follows:

ComparisonLatency per action predictionObserved ratioInterpretation
Fast-WAM vs. standard Enfold493 ms vs. 134 ms3.68×End-to-end comparison of different model configurations
Standard Enfold vs. Enfold-Flash134 ms vs. 49 ms2.73×Deployment acceleration of the same Enfold model and inference path
Fast-WAM vs. Enfold-Flash493 ms vs. 49 ms10.06×Combined architecture/configuration and deployment comparison

The standard configuration delivers a 3.7× system-level speedup, and TensorRT provides a further 2.7× acceleration within Enfold, yielding the reported 10.1× overall ratio. The 3.7× comparison includes model-configuration differences; the 10.1× comparison additionally includes the deployment backend.

The Wan-teacher variant achieves 97.4% on LIBERO and the robustness results above, demonstrating that the representation interface extends beyond Cosmos. With the encoder and action head held fixed, changing the training-time teacher leaves the deployed computation graph unchanged.

Additional profiling clarifies the three sources of acceleration. We compared action-head depths and progressively enabled TensorRT. Ratios below use this new profiling run only.

ConfigurationBackendLatency per action chunk
Fast-WAMPyTorch530.26 ms
Enfold, 30-layer action headPyTorch301.87 ms
Enfold, 10-layer action head (default)PyTorch120.31 ms
Enfold, action head acceleratedPyTorch + TensorRT45.44 ms
Enfold, encoder and action head acceleratedTensorRT36.00 ms

The 30-layer Enfold control is 1.76× faster than Fast-WAM in PyTorch, reflecting the overall architectural difference rather than isolating video-backbone removal alone. Beyond removing the video backbone, decoupling permits a smaller action head, and our predictive representation supports strong control at this reduced capacity: the 10-layer head adds 2.51× acceleration while achieving 97.8% LIBERO success versus Fast-WAM’s 97.6%. Finally, TensorRT adds 3.34× deployment acceleration (120.31 → 36.00 ms), separately from these architectural gains.

Reviewer_5PF7OpenReview ↗
Rating
3
Confidence
2
Informativeness
2
Presentation
3
Soundness
3
Read review
Paper summary

This paper proposes Enfold, a framework for efficient embodied control that learns predictive representations by distilling multi-level hidden states from a video generative model. The learned representations support both future-video generation and action prediction, while eliminating the need for video generation during control inference. Experiments on simulated and real-world robotic manipulation tasks demonstrate competitive performance with substantially reduced inference latency.

Critical strengths
  1. Reasonable methodological design. Using intermediate hidden states from multiple generator layers as supervision is an interesting approach to extracting predictive information. The separation of G2R, R2G, and action prediction objectives is clearly formulated.
  2. Strong empirical results. Enfold achieves competitive performance on LIBERO and RoboTwin2.0, along with promising real-world results. The reported 3.7× inference speedup over Fast-WAM, increasing to 10.1× with TensorRT acceleration, demonstrates practical value.
  3. Useful representation analyses. The authors investigate generator features across network depths and corruption timesteps, and provide additional analyses of representation stability and future predictability. These experiments help motivate the proposed multi-level supervision strategy.
Critical weaknesses
  1. Insufficient comparison with alternative supervision signals. Although Table 4 compares generator hidden-state supervision with future-pixel and action supervision, it lacks comparisons with representations extracted from ground-truth future observations using pretrained visual encoders, such as DINOv3. Such comparisons would help establish whether the improvements specifically arise from the generative model's intermediate representations.
  2. The contribution of R2G requires further investigation. Since R2G does not directly update the predictive encoder, its role in improving downstream control remains unclear. While Table 5 demonstrates improvements in video generation, an ablation removing R2G would help clarify its necessity for the overall framework.
  3. Clarity of mathematical presentation. The methodology section introduces a considerable number of equations, but some mathematical symbols and operations are not sufficiently explained when first introduced. For example, "h" in equation 1, "LN" in equation 3, and "SmoothL1" in equation 4. This makes the technical formulation unnecessarily difficult to follow. More explicit definitions and explanations of the key equations would improve the readability and accessibility of the paper.
Questions for authors
  1. How would G2R perform if the generator hidden-state targets were replaced with future representations extracted by DINOv3 or another pretrained visual encoder, under matched training conditions?
  2. What is the effect of removing R2G entirely? In particular, does it affect downstream action prediction, or primarily future-video generation quality?
  3. How sensitive is the learned representation to the architecture/capacity of the prediction head ? Since z is not explicitly regularized, could different prediction-head capacities substantially change the information encoded in z?
2 comments
AuthorsOpenReview ↗

## Response to Reviewer 5PF7 (W1 / Q1 & W2 / Q2 & W3)

We thank the reviewer for the positive assessment. We have added the requested supervision and prediction-head comparisons, examined the role of representation feedback, and prepared clearer explanations of the mathematical formulation. We have also included all the experiments in the revised version of the manuscript, with the added or revised content highlighted in red.

W1 / Q1 — Comparison with future-encoder supervision

We have added the requested future-feature prediction baselines. They retain the same DINOv3-based current encoder, action head, gradient isolation, training data, and training steps, replacing the prediction targets and adapting the necessary output dimensions.

Supervision targetSpatialObjectGoalLongAvg.
Future DINOv3 ViT-H+/16 features97.698.697.286.094.9
Future V-JEPA2 ViT-H/16 features95.097.695.274.690.6
Generator states (Enfold–Cosmos)97.0100.096.897.497.8

Generator-state supervision improves average success by 2.9 and 7.2 points over DINOv3 ViT-H+/16 and V-JEPA2 ViT-H/16 targets, with the largest gains on Long (11.4 and 22.8 points). DINOv3 ViT-H+/16 describes future appearance, and V-JEPA2 ViT-H/16 captures temporal evolution. Under matched training conditions, these results demonstrate the advantage of the proposed supervision target. Enfold’s targets expose multi-level states used to reconstruct the transition conditioned on the current scene and instruction, connecting predictive learning to the task-conditioned transitions needed for extended manipulation.

W2 / Q2 — Effect of R2G on control and video generation

We removed only the generator’s zz conditioning while retaining its video objective, isolating R2G feedback from one-way G2R supervision.

LIBERO control success (%)

VariantSpatialObjectGoalLongAvg.
Enfold97.0100.096.897.497.8
w/o R2G98.299.497.695.497.7

Video-generation comparison (submitted Table 5)

Generator conditioningPSNR ↑ (dB)SSIM ↑LPIPS ↓
Visual/language context25.920.8850.0550
Visual/language context + detached zz27.270.9020.0483

R2G strengthens future generation and improves Long performance. One-way G2R already achieves 97.7% average control success; adding representation feedback reaches 97.8% and improves Long by 2.0 points. In the submitted future-video comparison, representation conditioning raises PSNR from 25.92 to 27.27 and SSIM from 0.885 to 0.902, while reducing LPIPS from 0.0550 to 0.0483.

Although R2G sends no direct gradient to the encoder, it affects the teacher states used in subsequent G2R updates. G2R trains zz to expose predictive interaction cues, such as whether the gripper still holds an object; feeding them back helps the generator model subsequent changes. The resulting teacher states provide an indirect route to improve representation learning. These video and Long gains justify retaining feedback at only 0.119% additional parameters, 3.63% training FLOPs, and no action-inference overhead.

W3 — Clarity of mathematical notation

We appreciate the concrete examples and we now define hℓ,tGh^G_{\ell,t} (Eq. 1) as generator block ℓ\ell’s spatiotemporal hidden-state tokens at corruption timestep tt, LN (Eq. 3) as layer normalization, and SmoothL1 (Eq. 4) as quadratic for small errors and linear for larger ones. These manuscript changes are highlighted in red.

AuthorsOpenReview ↗

## Response to Reviewer 5PF7 (Q3)

Q3 — Sensitivity to prediction-head capacity

We added a variant without the prediction head and G2R loss, and compared 2, 6, 10, and 16-layer prediction heads under the same training budget. The table reports LIBERO success (%).

Prediction-head configurationSpatialObjectGoalLongAvg.
Without future prediction (head and G2R removed)97.699.692.889.694.9
2 layers98.699.498.096.498.1
6 layers98.699.298.294.697.7
10 layers (default)97.0100.096.897.497.8
16 layers96.898.696.895.496.9

The prediction branch makes a clear contribution, while control performance is relatively insensitive to moderate head capacities on LIBERO. Removing the head and G2R reduces average success to 94.9%; even a 2-layer head achieves 98.1%. Across 2, 6, and 10 layers, average success lies within 97.7–98.1%. Thus, although zz has no separate explicit regularizer, G2R supports useful control representations without requiring a narrowly tuned head capacity.

The 16-layer head reaches 96.9%, below the shallower variants. Two possible explanations are less complete optimization under the same training budget and a shift of predictive computation into the larger head. Since G2R supervises the encoder and head jointly, a more expressive head may perform more transformations internally, reducing the pressure to make predictive relationships readily accessible in zz.

A moderate prediction-head capacity can therefore be chosen according to training budget and task difficulty. On LIBERO, 2 layers already suffice for strong average performance, while the default 10 layers achieves the highest Long success. These results support flexibility in head size, rather than a uniquely optimal depth.

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

Enfold learns predictive representations through a two-way interaction with a video world model: the model supervises the representation and distills future prediction ability to the representation, and the representation in turn helps the generator to predict future videos. An action head uses the learned representation for control, allowing action prediction without running the video generator at deployment. Experiments on LIBERO, RoboTwin2.0, and real robots demonstrate competitive control performance with reduced action inference latency.

Critical strengths
  1. The deployment interface offers a meaningful tradeoff. Enfold removes the video generator from action-time computation while retaining competitive performance on two simulation benchmarks. Real-world results provide further evidence that the representation supports nontrivial manipulation rather than only offline prediction.
  2. The presentation is clear and easy to follow.
Critical weaknesses
  1. The core contribution is not sufficiently isolated. Table 4 and Table 5 are not enough. For example, how does Enfold compare with some simpler represenation learning methods, e.g. a future-encoder-feature prediction objective using the same DINO-based current encoder? This would directly test the added value of generator-state supervision. Also, what changes in control performance when bidirectional coupling (like R2G conditioning) is removed or when the gradient-routing choices are changed? Can one-way G2R distillation obtain similar action performance? These ablations will help validate that if your design choices are necessary.
  2. Evidence for representation quality remains limited. Near-saturated benchmark results on Libero and Robotwin limit the conclusions, both in main experiments and in ablations. Existing real-world OOD tests and qualitative interventions provide some support, but more systematic distribution-shift and perturbation-recovery evaluations, such as LIBERO-Plus or equivalent settings, would better substantiate the claim that your learned representation suppresses nuisance variation.
Questions for authors

Can the real-world evaluation additionally report full-task success rate and uncertainty, alongside the staged-completion score?

6 comments
AuthorsOpenReview ↗

## Response to Reviewer eMzY (W1)

We thank the reviewer for the constructive suggestions. Following the suggestions, we have added matched future-feature prediction baselines, feedback and gradient-routing ablations, and systematic distribution-shift evaluations. These experiments directly address the contribution of generator-state supervision and the utility of the learned representation. We have also included all the experiments in the revised version of the manuscript, with the added or revised content highlighted in red.

W1 — Comparison with future-feature supervision

We added the requested baselines by replacing generator-state targets with future DINOv3 ViT-H+/16 or V-JEPA2 ViT-H/16 features. The current DINOv3 encoder, action head, gradient isolation, training data, and training steps are matched, with the necessary output-dimension adaptation. The table reports LIBERO success (%).

Supervision targetSpatialObjectGoalLongAvg.
Future DINOv3 ViTH+/16 features97.698.697.286.094.9
Future V-JEPA2 ViT-H/16 features95.097.695.274.690.6
Generator states (Enfold–Cosmos)97.0100.096.897.497.8

Generator-state supervision improves average success by 2.9 percentage points over DINOv3 ViT-H+/16 targets and 7.2 points over V-JEPA 2-H targets. Most of this advantage comes from LIBERO-Long, where the margins reach 11.4 and 22.8 points, respectively. A possible explanation lies in what the student is trained to anticipate. DINOv3 ViT-H+/16 supplies a description of future appearance, and V-JEPA2 ViT-H/16 extends that description across time. The video generator supplies intermediate states from reconstructing the observed transition given the current scene and instruction. We hypothesize that this helps the student anticipate task-relevant outcomes, such as whether a grasp secures the instructed object before the next action. This offers a plausible explanation for the larger gain on multi-stage tasks.

W1 — R2G and gradient routing

We tested feedback by removing only the generator’s zz conditioning, retaining video training. Separately, we allowed video-loss or action-loss gradients into the encoder to test gradient routing. Results are LIBERO success (%).

VariantSpatialObjectGoalLongAvg.
Enfold97.0100.096.897.497.8
w/o R2G98.299.497.695.497.7
w/o sg(R2G)97.499.496.296.297.3
w/o sg(action)98.699.696.291.696.5

G2R establishes strong control representations, while R2G improves future generation and Long performance. One-way G2R reaches 97.7% average success versus 97.8% for Enfold; feedback improves Long from 95.4% to 97.4%. It also improves video prediction (PSNR 25.92 → 27.27 dB, submitted Table 5). G2R trains zz to expose predictive interaction cues, such as whether the gripper still holds an object. Feeding these cues back helps the generator model subsequent changes; its updated states then supervise the encoder through G2R. This provides an indirect route to improve control despite the detached generation loss. R2G adds only 0.119% parameters and 3.63% training FLOPs, with no action-inference overhead.

Gradient isolation preserves generator-state prediction as the encoder’s learning objective. Allowing video gradients reduces average success by 0.5 points; allowing action gradients reduces it by 1.3 points and Long by 5.8 points. Video reconstruction also rewards appearance details, while action fitting rewards demonstration-specific outputs; both can compete with the encoder’s predictive objective. The results support letting each head learn its own task while G2R shapes the shared representation.

AuthorsOpenReview ↗

## Response to Reviewer eMzY (W2 & Q1)

W2 — Representation quality under distribution shift

To evaluate robustness beyond the saturated standard benchmarks, we added LIBERO-Plus, covering seven perturbation categories:

MethodCameraRobotLanguageLightBackgroundNoiseLayoutAvg.
Fast-WAM16.444.568.978.253.737.760.751.5
Enfold, Cosmos teacher59.761.792.089.848.850.078.868.7
Enfold, Wan teacher49.269.276.184.946.451.379.165.2

Average LIBERO-Plus success improves from 51.5% for Fast-WAM to 68.7% with Cosmos (+17.2 percentage points) and 65.2% with Wan (+13.7 points). Enfold with Cosmos outperforms Fast-WAM in six of seven categories, including +43.3 points on Camera, +23.1 on Language, and +18.1 on Layout. The Wan-teacher variant also improves in six categories, showing that these gains extend across teacher architectures. These results provide systematic evidence of robustness to viewpoint, instruction, lighting, noise, and configuration changes.

We additionally conduct a clean-to-randomized RoboTwin evaluation, retraining all methods on the same clean demonstrations and evaluating them on both clean and randomized scenes. Our C2R training uses no data augmentation, which may contribute to the low absolute success under randomization:

MethodCleanRandomized, OOD
Fast-WAM77.81.9
Enfold, Cosmos teacher78.13.9
Enfold, Wan teacher80.37.1

Enfold improves randomized-scene success from 1.9% to 3.9% with Cosmos and 7.1% with Wan, although this severe transfer setting remains challenging for all methods. Unlike the submitted Table 2, this experiment uses only clean training demonstrations. Together, LIBERO-Plus and this clean-to-randomized evaluation provide quantitative distribution-shift evidence that complements the intervention examples in the paper.

How future supervision shapes the policy representation may help explain this pattern. Fast-WAM jointly trains video and action prediction, and its action expert reads features computed from the current frame. Enfold gives the policy encoder a direct predictive objective: from the current scene and instruction, it must predict the generator’s internal states as the generator processes the observed future. This rewards zz for encoding cues that anticipate the interaction, such as which object the gripper is approaching and how that object is likely to move. Those relationships can remain informative under a new viewpoint or lighting condition, giving the action head a basis for control that can transfer across visual conditions.

Q1 — Full-task success on the real robot

We additionally report full-task success on the ID real-robot evaluation, defined as completing all three ordered stages. Each model is evaluated over 30 trials per task.

MethodFold TowelDesktopSpoonStore PlateAvg.
Motus40.046.713.343.335.8
π0.5\pi_{0.5}86.766.766.770.072.5
Fast-WAM66.766.753.360.061.7
Enfold83.376.770.086.779.2
Enfold-Flash66.760.066.780.068.3

Enfold achieves 79.2% average full-task success, compared with 61.7% for Fast-WAM and 72.5% for π0.5\pi_{0.5}. It outperforms Fast-WAM on all four tasks and achieves the highest success on three of four tasks.

Thanks for your detailed response and the new experiments. However, in your new ablation, Enfold's improvent on w/o R2G and w/o sg(R2G) is marginal, especially w/o R2G. This raises more concerns about the paper’s claim and methodological contribution.

Comments on the paper

3 comments
Reviewer_kS4pOpenReview ↗

Response

The additional LIBERO-Plus comparison is incomplete. Enfold-Cosmos improves substantially over Fast-WAM (68.7% vs. 51.5%), but Fast-WAM is among the weaker methods on this benchmark. Several methods already included in the paper’s standard LIBERO comparison have higher publicly reported LIBERO-Plus performance, including OpenVLA-OFT (69.6%), LingBot-VA (69.5%), VLA-JEPA (79.5%), Cosmos-Policy (82.2%), and π0.5\pi_0.5 (84%–86%). Thus, the new result establishes an advantage over Fast-WAM rather than competitive robustness on LIBERO-Plus. The paper should compare against these available baselines under the same released/corrected-instruction protocol. Moreover, generator-state supervision still lacks a matched non-generative multi-level feature-distillation baseline. I therefore maintain my rating of 2.

AuthorsOpenReview ↗

Response to Reviewer kS4p

Thank you for highlighting these additional baselines. We will include their published results with the corresponding evaluation protocols and training settings. We have also added the requested multi-level non-generative supervision baselines, as detailed below.

1. LIBERO-Plus: robustness and architectural efficiency.

Enfold’s architectural contribution is to combine robust control with substantially lower inference cost, while retaining optional video generation. By folding generator-state supervision into predictive representations, Enfold decouples action prediction from the video generative model. As reported in Table 1, Enfold reduces action-chunk latency from 493 ms (Fast-WAM) to 134 ms—a 3.7× speedup on the same A100 GPU. Enfold-Flash further reaches 49 ms (10.1×) with TensorRT acceleration.

Enfold uses neither an additional embodied-pretraining stage nor data augmentation, making Fast-WAM the most direct comparison under these training conditions. Enfold-Cosmos achieves 68.7% on LIBERO-Plus under the released-instruction protocol, improving over Fast-WAM by 17.2 percentage points, with gains in six of seven perturbation categories.

We summarize all five published baselines below, together with their reported training backgrounds:

MethodPublished LIBERO-Plus success (%)Relevant training background
OpenVLA-OFT69.6Inherits OpenVLA pretraining on 970K robot trajectories from Open X-Embodiment [1].
LingBot-VA69.5Pretrained on approximately 16K hours of robot manipulation data [2].
VLA-JEPA79.5Pretrained on SSv2 human videos and DROID robot data; reported dataset sizes are 220K videos and 76K trajectories, respectively [3].
Cosmos Policy82.2Directly adapts a pretrained video model to LIBERO, without a separate embodied-pretraining stage; its official training recipe uses random cropping, rotation, and color jitter [4].
π0.5\pi_{0.5}84.3 [6]Base checkpoint pretrained on 10K+ hours of robot data [5].

All listed LIBERO-Plus results use the released instructions.

2. Multi-level non-generative supervision baselines.

We have now added multi-level non-generative feature-distillation baselines using future DINOv3 and V-JEPA 2 features as supervision targets. These comparisons retain the same current-observation encoder, action head, gradient isolation, training data, and training steps, with adaptation to the target feature dimensions. All three variants use multi-level supervision.

Supervision targetSpatialObjectGoalLongAverage
Multi-level future DINOv3 features96.098.497.085.094.1
Multi-level future V-JEPA 2 features94.898.696.694.896.2
Multi-level generator states (Enfold-Cosmos)97.0100.096.897.497.8

LIBERO success rates (%); bold identifies Enfold.

Enfold achieves 97.8% average success, exceeding multi-level DINOv3 and V-JEPA 2 supervision by 3.7 and 1.6 percentage points, respectively. On LIBERO-Long, the gains are 12.4 and 2.6 points. Enfold’s advantage persists when the non-generative baselines also use multi-level targets, supporting the value of generator-state supervision for control.

References

[1] OpenVLA, pretraining data.

[2] LingBot-VA, training data.

[3] VLA-JEPA, Section 4.1, pretraining datasets and their reported sizes.

[4] Cosmos Policy, target-task adaptation; official LIBERO training configuration and augmentation implementation.

[5] Physical Intelligence: openpi, base model pretraining.

[6] La-JEPA, Table 14, released-instruction evaluation of π0.5\pi_{0.5}.

AuthorsOpenReview ↗

# Author Summary for the Area Chairs, Senior Area Chair and Program Chairs

Dear Area Chair, Senior Area Chair, and Program Chairs,

Thank you for your time and careful consideration of our submission. We appreciate the constructive discussion with all three reviewers, which has helped us strengthen the experimental evidence and clarify the contribution of Enfold.

Across the reviews, Enfold’s representation-level interface and practical deployment value were recognized as strengths. Following our responses, Reviewer 5PF7 confirmed that most of their concerns had been addressed and maintained their positive rating. We would like to highlight the additional evidence most relevant to the committee’s assessment.

Enfold makes the predictive knowledge of video world models available to an efficient control policy, while retaining optional video generation. Action prediction uses the learned representation without executing the video generator. The submitted measurements show a 3.7× end-to-end speedup over Fast-WAM, increasing to 10.1× with TensorRT acceleration. This separation between predictive supervision and deployed computation is a central contribution of the work.

During the discussion, we strengthened the evidence in three areas:

  • The value of generator-state supervision. We added the requested non-generative alternatives, including the multi-level DINOv3 and V-JEPA 2 controls requested in the follow-up discussion. With the current encoder, action head, training data, and training steps held fixed, Enfold achieves 97.8% average LIBERO success, compared with 94.1% and 96.2% for these multi-level alternatives. These results provide direct evidence for the proposed supervision target.

  • Robustness beyond standard benchmark performance. Enfold-Cosmos achieves 68.7% on LIBERO-Plus versus 51.5% for Fast-WAM, improving in six of seven perturbation categories, without additional embodied pretraining or data augmentation. We additionally included all five stronger published baselines raised during the discussion, together with their training backgrounds.

  • The roles of the design choices and the reliability of the results. We added feedback and gradient-routing ablations, prediction-head capacity comparisons, and three independent training runs. These experiments clarify the contributions of representation feedback and gradient isolation, and demonstrate repeatable control performance across LIBERO and RoboTwin.

We believe this strengthened evidence supports a positive assessment of Enfold’s methodological contribution, practical efficiency, and control performance. We will incorporate the additional experiments and clarifications into the revised manuscript.

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

Related papers

Loading the map…

Discussion (0)

Sign in to comment.