acceptodds
Under review as a conference paper at ICLR 2027

Vision-Language-Action or World-Action? Just Action Models

12 views

Abstract

The vision-language-action (VLA) and world-action model (WAM) communities debate whether control policies should inherit pretrained knowledge from vision-language models or from visual generators. The debate, however, bundles three separable choices: backbone initialization, the objective applied to future observations, and whether the machinery for that future loss is kept at inference. A gap between a VLA and a WAM therefore cannot simply be attributed to any one of them. We show that decoupling these choices yields a common framework for both families, which we term Just Action Model (JAM): a pretrained visual backbone with an action expert, trained with an action loss and a foresight loss. The foresight loss trains current-observation features to predict frozen encodings of the recorded future through a prediction head discarded after training, so no visual generation is involved in either training or inference, and the backbone can be chosen freely. We first compare objectives within fixed policy architectures, then evaluate the foresight recipe across backbone families. The resulting JAM policies outperform their matched FastWAM and ImageWAM counterparts on LIBERO and LIBERO-Plus, and achieve similar or slightly higher success rates on RoboTwin. Compared with FastWAM, JAM improves the success rates from 97.0%/55.3%/91.2% to 98.2%/64.5%/91.4%, respectively. With cached foresight targets, JAM reduces estimated online training FLOPs per sample per update by 73% relative to FastWAM. Our results show that future supervision improves control without visual generation or additional inference machinery, and strong foresight performance is attainable with either generator-initialized or vision-language backbones.

Primary area:
applications to robotics, autonomy, planning
Keywords:
vision-language-action models, world-action models, representation learning

Reviews

Among:
Shape:
1234this paper, Top 15%
Figure 1. Mean review rating of the reviewed papers at ICLR 2027. This paper averages 3.00 over 3 reviews, above 85% of the others.
Reviewer_W7DZOpenReview ↗
Rating
3
Confidence
2
Informativeness
2
Presentation
3
Soundness
2
Read review
Paper summary

JAM replaces the visual-generation objective of an action policy with prediction of frozen future representations from its existing backbone features. The prediction head and target branch are discarded after training, preserving the corresponding action-only deployment path. The paper first compares objectives within FastWAM and ImageWAM architectures, then applies the recipe to self-supervised and vision-language backbones. Experiments cover LIBERO, a shared LIBERO-Plus subset, RoboTwin, target and head ablations, statistical comparisons, and training-cost estimates.

Critical strengths
  1. The fixed-architecture comparisons address a consequential question: whether generators need to retain visual-generation supervision during policy fine-tuning. Table 1 separates the deployed architecture from the training objective and target representation. The best configurations improve LIBERO-Plus from 55.31 to 64.54 percent for Wan and from 65.27 to 79.95 percent for Ovis-U1, without adding inference machinery. The current-offset, future-offset, random-anchor, and VAE-direct controls are particularly useful: they show that future target information matters and that a large pretrained target transformer is not always necessary. This is a meaningful empirical contribution even though future-feature alignment itself has substantial precedent.

  2. The main improvements have stronger support than isolated benchmark averages. Table 20 reports paired outcomes and task-bootstrap intervals, and Table 21 repeats the selected JAM configurations across three independent training runs with low observed variability. The paper also provides action-only Qwen and DINO controls, distinguishes local evaluations from published scores, explicitly identifies the 1,506-task OOD subset, and limits cross-backbone conclusions rather than claiming to isolate pretraining alone. These details make the central conclusion credible within the evaluated settings. The main value is the controlled evidence for simplifying generator-based policy training, rather than a fundamentally new alignment loss.

Critical weaknesses
  1. The headline training-cost reduction mixes the objective change with a caching advantage that may also be available to the generative baselines. In Table 9, FastWAM's 15.52 TFLOPs include 6.54 TFLOPs of online VAE encoding, while cached JAM uses only 0.93 TFLOPs for its current frame. Under the same fixed-preprocessing assumptions, the generative baseline's frozen VAE encodings of recorded futures also appear cacheable; flow-time and noise sampling could still be performed online. The paper transparently reports online and cached JAM costs, but does not provide an equivalently cached FastWAM/ImageWAM reference. Thus, 73 percent is supported for the specific reported pipelines, but does not isolate the saving attributable to replacing visual generation. This is a fixable limitation of an important efficiency claim, rather than evidence against the policy-performance gains.
Questions for authors
  1. Can the generative baselines cache their frozen future-frame VAE encodings under the same preprocessing assumptions as JAM? Please report the online cost and measured update time with caching available to both methods, or explain what prevents this. Reporting one-time cache construction and storage would also clarify the practical trade-off.

  2. How were the target type, readout offsets, and loss hyperparameters selected before the final LIBERO-Plus evaluation? Were they selected using an independent validation set? Table 20 characterizes fixed-checkpoint uncertainty and Table 21 repeats selected JAM configurations; please clarify whether the generative baselines also have independent training-run estimates. This would distinguish the stability of the training comparison from uncertainty conditional on the selected models and fixed OOD manifest.

1 comment
AuthorsOpenReview ↗

Response to Reviewer W7DZ (1/1):

We sincerely thank Reviewer W7DZ for the recognition of our work. The further questions are addressed as follows.

W1, Q1: Can the generative baselines cache their future encodings?

A: Yes, and with both methods cached, foresight still cuts the estimated FLOPs by 57.7% for Wan and 31.9% for Ovis-U1. The abstract, introduction and conclusion now state these figures in place of 73%, and Table 9 adds the cached baselines. A cached FastWAM gives bit-identical latents and losses to online training. For Wan, replacing generation saves the same 5.72 TFLOPs per sample with or without caching: caching cuts FastWAM's VAE cost from 6.54 to 0.93 TFLOPs but leaves 8.93 TFLOPs of trainable visual computation, against 3.17 for JAM. FLOPs per sample and GPU time per update, profiled on one GH200 as in Table 9, each JAM row against its baseline in the same setting:

PolicyTargetsTFLOPs ↓Reduction ↑GPU time (s / step) ↓Reduction ↑
FastWAMonline15.52–0.789–
JAM (Wan)online9.8036.9%0.53232.6%
FastWAMcached9.91–0.512–
JAM (Wan)cached4.1957.7%0.25849.6%
ImageWAM (Ovis-U1)online5.81–0.473–
JAM (Ovis-U1)online4.8815.9%0.40314.8%
ImageWAM (Ovis-U1)cached5.39–0.436–
JAM (Ovis-U1)cached3.6731.9%0.29432.6%

Building the Wan LIBERO cache once takes 22 minutes on eight GH200s and 16.2 GB of storage. Caching saves 0.27 s of GPU time per update for either method, so the cache is repaid within a quarter of one 20k-update run; a cached FastWAM reads the same cache, so this cost is shared by both methods.

Q2: Were the configurations selected on an independent validation set?

A: Following FastWAM [R1] and ImageWAM [R2], we did not hold out any part of LIBERO or LIBERO-Plus; LIBERO alone returns the selected configurations, and so do 1,246 LIBERO-Plus tasks that were not used in the selection. The baselines keep their published settings. Wan's loss weight, depth and head and Ovis-U1's λ_F=0.3 with [1] were selected on LIBERO, the latter before any LIBERO-Plus evaluation. Targets and readouts are compared in Table 1, and LIBERO picks the same two: each of the three training runs of VAE-direct [1,2] (98.00 to 98.25) and of Yak [1] (96.60 to 96.90) is above every alternative on LIBERO (at most 97.65 and 95.75). We also re-evaluated the generative and foresight checkpoints of Table 1 on a second manifest drawn with seed 43, without re-selection. LIBERO-Plus success (%), new Table 22:

PolicyFuture anchorReadout offsetsPaper manifest ↑Seed-43 manifest, all tasks ↑Seed-43 manifest, tasks outside the paper manifest ↑
FastWAM––55.3155.4456.02
JAM (Wan)Pretrained Wan[0]44.6241.7041.01
[1,2]55.5858.1059.87
[0,1,2]54.1854.9156.02
Random Wan[1,2]58.6360.3661.08
Pretrained VAE[1,2]64.5465.3466.45
ImageWAM (Ovis-U1)––65.2763.9464.13
JAM (Ovis-U1)Pretrained Yak[0]69.6569.1969.74
[1]79.9577.0977.53
[0,1]72.5871.1871.99
Random Yak[1]72.5170.3269.98
Pretrained VAE[1]76.4372.7873.27

Selecting on the 1,246 tasks outside the paper manifest returns the same two configurations, and there they lead the generative baselines by 10.43 and 13.40 points. Across 10,000 random splits of the 2,752 distinct tasks of the two manifests into validation and test halves, the validation half selects VAE-direct [1,2] and Yak [1] every time.

Q2: Independent training runs for the generative baselines.

A: We added two runs of each baseline. Table 20 measures uncertainty at fixed checkpoints; these runs measure training variability. Mean ± SD over three training runs (revised Table 21):

PolicyLIBERO ↑LIBERO-Plus ↑
FastWAM96.87±0.3753.45±2.31
JAM (Wan)98.13±0.1364.61±0.37
ImageWAM (Ovis-U1)92.50±0.6964.14±2.49
JAM (Ovis-U1)96.75±0.1579.92±0.43

JAM leads its baseline on LIBERO-Plus by 11.16 points for Wan and 15.78 for Ovis-U1, above the single-run leads of the submission. Every JAM run beats every baseline run on both benchmarks (Table 21): on LIBERO-Plus the weakest JAM run is 8.97 points above the strongest FastWAM run and 13.61 above the strongest ImageWAM run.

References

[R1] Yuan, Tianyuan, et al. "Fast-WAM: Do World Action Models Need Test-time Future Imagination?" arXiv:2603.16666, 2026.

[R2] Zhang, Yuyang, et al. "ImageWAM: Do World Action Models Really Need Video Generation, or Just Image Editing?" CoRL, 2026.

Reviewer_c8KXOpenReview ↗
Rating
3(+1)
Confidence
3
Informativeness
2
Presentation
2
Soundness
2
Read review
Paper summary

This paper proposes Just Action Model (JAM), a framework that investigates whether visual generation objectives are necessary for learning effective robot control policies from future observations. The authors disentangle three key factors distinguishing Vision-Language-Action (VLA) models from World-Action Models (WAMs): backbone initialization, future supervision objectives, and inference architecture. JAM replaces visual generation objectives with a foresight loss that encourages current-observation features to predict frozen representations of future observations, eliminating the need for image or video generation during policy training and inference.

The authors evaluate JAM on existing architectures, including FastWAM and ImageWAM, and further extend the framework to generative, self-supervised, and vision-language backbones. Experiments demonstrate improved success rates on LIBERO and LIBERO-Plus, substantial reductions in training computation, and comparable or slightly improved performance on RoboTwin 2.0.

Critical strengths
  1. Meaningful research question and simple design. The paper investigates whether computationally expensive visual generation objectives are truly necessary for effective robot policy learning, addressing a practically important problem in VLA and WAM research. The proposed solution is simple and general: it introduces future supervision through a training-only prediction head and foresight loss, without modifying the original action generation mechanism or introducing additional inference overhead.

  2. Well-controlled experimental comparisons. The authors conduct relatively fair comparisons by keeping the policy architecture, training data, optimizer, and training budget consistent when replacing the generation loss with the foresight loss in FastWAM and ImageWAM. This experimental design strengthens the validity of the main findings. Furthermore, caching future representations significantly reduces online training FLOPs, demonstrating the computational benefits of the proposed approach.

  3. Comprehensive ablation studies. The paper investigates several important design choices, including future anchors, supervision offsets, backbone initialization, loss weights, and prediction head architectures. The additional statistical analyses and independent training runs further strengthen the experimental evaluation and provide useful insights into the effectiveness of the proposed framework.

Critical weaknesses
  1. Limited methodological novelty. The core idea of JAM is to leverage a frozen future encoder to provide latent representation supervision, which is closely related to existing methods such as VLA-JEPA, JEPA-WAM, and FRAPPE. The main contribution appears to lie in applying established future representation prediction techniques to existing WAM architectures and empirically demonstrating their effectiveness and efficiency, rather than introducing a fundamentally new learning mechanism.

  2. Insufficient comparisons with closely related methods. Although the authors discuss JEPA-WAM, VLA-JEPA, and other related approaches, direct experimental comparisons under a unified evaluation protocol are missing. Consequently, it remains difficult to identify the specific advantages of JAM over existing future representation learning methods, beyond its demonstrated improvements over generative WAM baselines.

  3. Unclear whether the improvements originate from learning future dynamics. The experiments show that a randomly initialized future anchor can still provide effective supervision, while VAE latent targets outperform pretrained transformer features for the Wan backbone. These observations suggest that the performance gains may partly arise from auxiliary representation regularization rather than meaningful modeling of future dynamics. The current experiments do not sufficiently distinguish between these explanations.

  4. Limited analysis of the learned foresight representations. The authors employ future-frame reconstruction probes and attention visualizations to analyze the learned representations. However, JAM with future-offset supervision improves future-frame reconstruction PSNR by only approximately 0.14 dB over ImageWAM. Such a marginal improvement provides limited evidence for explaining the substantially larger gains in policy success rates.

  5. Confounding factors in cross-backbone comparisons. The evaluated backbones differ in architecture, parameter count, pretraining data, language conditioning, and action expert design. Therefore, these comparisons demonstrate the applicability of JAM across different backbone families but cannot isolate the contributions of different pretraining paradigms. Furthermore, Table 3 reports only backbone sizes rather than complete deployed policy sizes. For instance, JAM (Qwen3.5-2B) has approximately 3.60B deployed parameters, which may lead to a misleading interpretation of model scale and computational efficiency.

  6. Insufficient explanation of future anchor selection. VAE targets achieve the best performance for Wan, whereas pretrained transformer targets work better for Ovis-U1. Although the authors provide possible explanations for this discrepancy, the paper lacks a systematic investigation of which backbone properties determine the effectiveness of different future representations. This limits the practical guidance for applying JAM to new architectures.

  7. Limited improvements on RoboTwin and lack of real-world validation. While JAM achieves substantial improvements on LIBERO-Plus, its gains over matched generative baselines on RoboTwin 2.0 are generally below one percentage point. This suggests that the benefits may vary considerably across benchmarks and task settings. Moreover, all experiments are conducted in simulation, leaving the effectiveness and generalization of JAM under real-world visual variations, contact dynamics, and complex manipulation scenarios unverified.

Questions for authors
  1. Novelty relative to existing methods. What are the fundamental methodological differences between JAM and closely related approaches such as JEPA-WAM, VLA-JEPA, and FRAPPE? Could the authors provide more direct comparisons under a unified evaluation protocol to better establish the unique contributions of JAM?

  2. Future dynamics learning versus representation regularization. How can the authors verify that the improvements introduced by the foresight loss arise from learning meaningful future state transitions rather than from additional representation regularization? Would temporally shuffled or mismatched future targets provide a useful control experiment?

  3. Effectiveness of random future anchors. Why does a randomly initialized Wan future anchor outperform its pretrained counterpart in the reported experiments? Does this suggest that semantic or temporal knowledge acquired during pretraining is not essential for effective foresight supervision?

  4. Limited improvements on RoboTwin. Why does JAM achieve substantial improvements on LIBERO-Plus but only marginal gains on RoboTwin 2.0? Given that RoboTwin evaluates diverse bimanual manipulation tasks under randomized conditions, potentially providing a more challenging test of generalization and real-world-relevant robustness, could the authors further investigate the reasons behind this discrepancy?

  5. Real-world validation. Could the authors evaluate JAM on real robotic systems, particularly on tasks involving complex object interactions, occlusions, and contact dynamics? Such experiments would help establish whether the improvements observed in simulation translate to practical robotic manipulation.

6 comments
AuthorsOpenReview ↗

Response to Reviewer c8KX (1/4):

We sincerely thank Reviewer c8KX for the recognition of our work. The further questions are addressed as follows.

W1, Q1: Novelty relative to existing methods.

A: JAM and the related methods make opposite arguments. VLA-JEPA [R1], JEPA-WAM [R2] and FRAPPE [R3] start from a policy and add a latent world model to it. JAM starts from a world-action model and removes its visual generation, and identifies what remains, supervision from the recorded future, as the part that carries the benefit. The policy generates no image or video in training and predicts no future at deployment. With the deployed action path, data and training budget of FastWAM and ImageWAM fixed, LIBERO-Plus rises from 55.31 to 64.54 for Wan and from 65.27 to 79.95 for Ovis-U1.

Once visual generation is removed, a WAM and a VLA differ in the backbone they start from, not in how they are trained: both are action models trained with the same recipe (Table 3).

Methodologically, JAM differs from the closest methods in four ways, each measured inside one recipe (Table 1 and the table under W2):

  • Deployment is unchanged. A separate training-only head supervises existing visual-backbone tokens, so the deployed network is exactly the baseline's. FLARE [R4] mixes future-prediction tokens with state and action tokens inside the action DiT, and FRAPPE [R3] retains its parallel prefix-and-LoRA streams at inference. Moving JAM's supervision to FLARE-style action queries lowers LIBERO-Plus by 7.1 points.
  • JAM does not need an external encoder or extra pretraining. Its targets come from the policy's own frozen model, its VAE or a copy of its backbone. VLA-JEPA [R1] and the concurrent JEPA-WAM [R2] predict in a V-JEPA space, VLA-JEPA with additional human-video pretraining, and FRAPPE aligns with multiple visual encoders. Swapping JAM's target for V-JEPA 2 lowers LIBERO-Plus by 11.8 points.
  • Only future positions are supervised. JEPA-WAM uses dense joint current–future targets; adding the current position to JAM's targets lowers LIBERO-Plus by 7.4 points for Ovis-U1.
  • Generation is removed, and one recipe trains every backbone. The fixed targets can be cached: with caching available to both methods, estimated training FLOPs fall by 57.7% against FastWAM and 31.9% against ImageWAM. The same recipe trains generative, self-supervised and vision-language backbones, including ones that cannot generate.

W2, Q1: Direct comparisons with closely related methods.

A: We added two direct comparisons under one evaluation protocol (new Table 25). First, we ran target and placement swaps inside the Ovis-U1 recipe, keeping the policy, budget, layer 11, λ_F=0.3 and readout [1]. To swap the target, we replace Yak with VideoMAE-B or V-JEPA 2, the encoder family behind the targets of JEPA-WAM [R2] and VLA-JEPA [R1]. To swap the placement, we keep Yak but supervise FLARE-style future queries in the action expert [R4] instead of the visual tokens. Every variant beats the generative baseline on both benchmarks.

PolicyFuture anchorSupervised tokensLIBERO ↑LIBERO-Plus ↑
ImageWAM (Ovis-U1) action-only––90.2061.35
ImageWAM (Ovis-U1)––92.9065.27
JAMPretrained Yakvisual tokens96.7579.95
Pretrained VideoMAE-Bvisual tokens96.4572.24
Pretrained V-JEPA 2 ViT-Lvisual tokens95.0568.19
FLARE-style control (Ovis-U1)Pretrained Yakaction queries94.9072.84

JAM's own target and placement score highest. Both external video encoders trail the policy's own Yak target, but both still beat action-only training on both benchmarks. V-JEPA 2 is the weakest target: it predicts masked content in feature space rather than reconstructing pixels, so its targets lose the spatial detail that control depends on. Action queries are a weaker placement: they sit in the action expert, so the foresight signal reaches the visual tokens only indirectly, whereas JAM supervises the visual tokens that the action tokens read. They also have to stay at inference, while JAM's head is discarded after training.

PolicyLIBERO-Plus ↑
VLA-JEPA, released checkpoint76.96
JAM (Ovis-U1)79.95
JAM (Qwen3.5-2B)81.14

Second, we ran the released VLA-JEPA checkpoint with its official settings under our protocol, with the same episodes, step limits and LIBERO-Plus manifest: despite its human-video pretraining, it scores below JAM (Ovis-U1) and JAM (Qwen3.5-2B) on LIBERO-Plus. Of JEPA-WAM, VLA-JEPA and FRAPPE, only VLA-JEPA has released LIBERO weights: FRAPPE's released weights cover eight RoboTwin tasks only, and we found no public release of JEPA-WAM. Under their own protocols, JEPA-WAM and VLA-JEPA report 79.2 and 79.5 on LIBERO-Plus; JAM (Ovis-U1) and JAM (Qwen3.5-2B) reach 79.95 and 81.14 without an external target encoder or extra pretraining.

AuthorsOpenReview ↗

Response to Reviewer c8KX (2/4):

W3, Q2: Future dynamics versus representation regularization.

A: Using the recorded future from the same demonstration gives the best LIBERO-Plus result for both backbones. Compared with a future from another demonstration of the same task at the same stage, it improves success by 12.22 points for Wan (52.32% to 64.54%) and about 2.7 points for Ovis-U1 (77.22% to 79.95%). The mismatched future already improves over action-only training for both backbones (45.55% to 52.32% for Wan, 61.35% to 77.22% for Ovis-U1), suggesting that task- and stage-level cues also contribute.

More specifically, we ran the mismatched-target control proposed in Q2 (new Table 23): only the target changes, while the anchor, head, loss weight and budget stay fixed. The real future is the one recorded in the same demonstration; the mismatched future is taken from another demonstration of the same task, drawn at random, at the same fraction of the episode (t′ in the table), so it shows the same task stage but not the motion that actually follows the current frame. For the Ovis-U1 controls at [1], Yak receives frame t or t−16 in place of t+16, so the readout position stays the same.

PolicyFuture anchorReadout offsetsTarget framesTime stepsLIBERO ↑LIBERO-Plus ↑
FastWAM action-only––––95.2545.55
FastWAM––––97.0055.31
JAM (Wan)Pretrained VAE[1,2]real futuret+4, t+8, …, t+3298.1564.54
[0]current framet89.9534.53
[1,2]mismatched futuret′+4, t′+8, …, t′+3297.6552.32
ImageWAM (Ovis-U1) action-only––––90.2061.35
ImageWAM (Ovis-U1)––––92.9065.27
JAM (Ovis-U1)Pretrained Yak[1]real futuret+1696.7579.95
[0]current framet94.5569.65
[1]current framet94.4567.33
[1]mismatched futuret′+1695.0577.22
[1]past framet−1692.2573.84

Current- and past-frame targets fall below both future targets on both benchmarks. For Wan, the current-frame VAE target scores 34.53 on LIBERO-Plus, below action-only (45.55). For Ovis-U1, a current-frame Yak target improves over action-only (61.35 to at most 69.65), but by much less than the real future (79.95). The past frame (73.84) also stays below both future targets.
A mismatched future beats the current frame but not the real future. For both backbones, it improves over action-only training and over the current-frame target (52.32 against 34.53 for Wan, 77.22 against at most 69.65 for Ovis-U1), but stays below the real future.

Regularization alone does not explain the gain: with the same anchor, head and loss weight, a current-frame target does far worse than either future target.

W4: Analysis of the learned foresight representations.

A: JAM's features hold the future as well as a generator's, without any pixel objective. ImageWAM is trained to generate the future frame; JAM matches it on the future-frame probe, 19.57 versus 19.43 dB, both far above the 16.13 dB of copying the current frame. A probe of the robot state at t+16 (Appendix D.6) gives the same result: from frozen current-frame features, the instruction and the current state, both policies reach an R² near 0.95, against 0.75 for keeping the current state unchanged.

PolicyFuture-frame PSNR (dB) ↑Future-state R² ↑LIBERO-Plus ↑
ImageWAM (Ovis-U1)19.430.9507±0.000265.27
JAM (Ovis-U1)19.570.9531±0.000479.95

The state probe is fit on 20,000 windows and tested on 2,000 windows from held-out episodes; ± is the SD over three probe seeds.

The probes therefore show that JAM acquires the ability to anticipate the future without generation. They show what is decodable, not what drives the policy [R5, §4.3], so we also examine attention, robustness and target interventions:

  • What the action expert attends to. In Figures 5 and 8, on the same backbone and observation, ImageWAM attends to the gripper and the held object, whereas JAM attends to the instructed target where the object must go: the second mug, the rack, the plate, the microwave.
  • Where the gain appears. The probes use unperturbed LIBERO frames, where the two policies also differ less in success, 92.90 versus 96.75. The gain appears under perturbation and is largest for camera and sensor noise, 57.9 to 83.8 and 68.6 to 91.9 in Table 6, with 71 JAM-only against 9 baseline-only camera successes in Figure 6.
  • Effect of the target frame. At the same Yak readout [1], LIBERO-Plus success is 67.33% with the current frame, 77.22% with a stage-matched future from another demonstration, and 79.95% with the recorded future. These controls show that the choice of target matters, although they do not establish which features the policy uses for control.
AuthorsOpenReview ↗

Response to Reviewer c8KX (3/4):

W5: Confounding factors in cross-backbone comparisons.

A: These differences come from the backbones themselves; one rule builds every policy from its backbone and keeps everything else fixed.

  • Action expert: it mirrors the backbone block for block and reads the same-layer backbone tokens, as in FastWAM and ImageWAM; for DINOv2 and Qwen3.5 it is initialized from the backbone's own blocks. Actions read the observation, but not the other way around.
  • Language: the instruction enters through the backbone's own interface. DINOv2 does not have one, so a frozen Qwen3-4B passes it to the action expert only.
  • Target: a frozen copy of the backbone, read at about 60% of its depth, or at the final block for DINOv2, where its self-supervised losses act.
  • Everything else: data, views, optimizer, budget, λ_F, head and readout are shared and were carried over from Ovis-U1 without re-tuning.

Table 3 therefore shows that one rule works for every family; it does not rank them. Whether foresight helps is tested within each backbone, where only the loss changes (Tables 1 and 14). We added two repeats of each Qwen and DINOv2-L setting. LIBERO-Plus success (%), revised Table 14:

PolicyRun 1 ↑Run 2 ↑Run 3 ↑Mean ± SD ↑
Qwen3.5-2B action-only77.0379.7578.2278.33±1.36
JAM (Qwen3.5-2B)81.1482.4780.7681.46±0.90
DINOv2-L action-only68.2667.5365.8767.22±1.22
JAM (DINOv2-L)70.3969.7270.2570.12±0.35

Foresight leads in every run for both backbones, and its runs vary less.

The revised Table 3 now reports complete policy sizes in two new columns; Appendix B adds Wan and FLUX.2:

PolicyTrainable policyTotal with frozen inference encoders
JAM (Wan)6.02B11.85B
JAM (Ovis-U1)2.17B4.64B
JAM (FLUX.2)4.52B8.57B
JAM (Qwen3.5-2B)3.60B3.60B
JAM (DINOv2-B)221M4.24B
JAM (DINOv2-L)767M4.79B

Totals exclude training-only heads and anchors. DINOv2 runs only 221M or 767M parameters at evaluation because its Qwen3-4B instruction features are cached; its totals include that 4.02B encoder. JAM (Wan) and JAM (Ovis-U1) deploy exactly their baselines' parameters, so their cost savings come from training alone.

Q3: Effectiveness of random future anchors.

A: For Wan, the target needs the pretrained VAE's encoding of the real future, not a pretrained transformer on top of it. When the target VAE itself is randomized, LIBERO-Plus drops from 64.54 to 45.75, the action-only level (45.55). The random Wan anchor is a fixed random transformer applied to pretrained-VAE latents of the recorded clip, so its targets still encode the real future, and random teachers give useful, input-dependent targets [R6]. These latents already show how the scene changes across the clip, so the transformer on top is not needed: the latents alone score highest (64.54), above both the random transformer (58.63) and the pretrained one (55.58), whose features were learned for denoising. Yak's input is not a clip but two still images, the current frame and one future frame, and what relates the two is its pretrained image-editing layers, so for Ovis-U1 the pretrained anchor matters: 79.95 against 72.51 with random weights.

W6: Explanation of future anchor selection.

A: The choice of target depends on what the anchor sees. Wan's VAE encodes a strided window of future frames, so its latents already trace how the scene changes: the VAE alone (64.54) beats the pretrained and random transformer anchors (55.58 and 58.63). Ovis-U1's anchor sees a single future frame, which its pretrained layers relate to the current observation and instruction: the pretrained Yak copy (79.95) beats the VAE (76.43) and a random copy (72.51). For a new backbone: use the VAE latents when the anchor sees a multi-frame window, and a frozen copy of the pretrained backbone when it sees a single future frame or there is no VAE, as for Qwen3.5 and DINOv2. In every case the target comes from the policy's own frozen model: external video encoders trail both own-model targets of Ovis-U1 (68.19–72.24). The readout and loss weight carry over: Qwen3.5 and DINOv2 use Ovis-U1's readout [1] and λ_F = 0.3 without re-tuning. On a second LIBERO-Plus manifest drawn with a different seed, the same two targets remain the best of the five foresight configurations of each backbone.

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

The paper argues that the comparison between VLAs and world-action models (WAMs) mixes three choices that should be studied separately: how the backbone is initialized, what loss is placed on the recorded future observations, and which modules survive to deployment. Building on this view, the authors train a pretrained visual backbone and a flow-matching action expert with the action loss plus a foresight loss, a cosine regression from intermediate current-frame tokens (about 60% depth) onto the features of a frozen future anchor, where the prediction head and the anchor are discarded after training and the targets can be cached. In the first stage, the generative objectives of FastWAM (Wan2.2-5B) and ImageWAM (Ovis-U1) are replaced while architecture, data, optimizer, budget and evaluator are kept fixed; with the best target per architecture, LIBERO-Plus rises from 55.31% to 64.54% and from 65.27% to 79.95%, RoboTwin 2.0 changes by less than one point, and the estimated online cost per update falls by 73.0% and 36.8% (Table 9). In the second stage, the same objective is applied to DINOv2-B/L and Qwen3.5-2B, and the Qwen policy (81.14%) ends up close to the best generator-initialized one, with action-only controls reported in Table 14. The authors conclude that learning from recorded futures does not require a visual generative loss, and that generator-initialized and vision-language backbones can both reach strong performance under this recipe.

Critical strengths
  1. This work proposes to treat backbone initialization, future objective and deployed modules as independent factors, and tests the objective factor under controlled conditions. Replacing only the loss of FastWAM and ImageWAM, while the deployed path stays identical to the baselines, means the LIBERO-Plus differences in Table 1 can be attributed to the training objective rather than to architecture or inference procedure, which is rarely possible when WAMs and VLAs are compared through published numbers.
  2. This work shows that a small training-only head with fixed targets can replace video or image denoising without loss in success rate and at a much lower training cost. The LIBERO-Plus gains of 9.23 and 14.67 points have task-bootstrap intervals that exclude zero, the paired counts are lopsided (111/250 and 89/310 baseline-only versus JAM-only successes, Table 20), and three independent JAM runs vary by less than half a point (Table 21). Since generation over a nine-frame window disappears from the trainable branch, the online estimate drops from 15.52 to 4.19 TFLOPs per sample, and the measured GPU time drops by 67.3%.
  3. This work reports target ablations that are informative beyond the main result. Future-position readouts are clearly better than the current position for Wan (44.62% to 55.58%), a frozen random Wan anchor still reaches 58.63%, a fixed lifting of the VAE latents is the best Wan target (64.54%), whereas for Ovis-U1 the pretrained Yak anchor is preferred over the VAE (79.95% versus 76.43%). Combined with the backbone and anchor initialization grid (Table 17, where a random backbone falls to 71.15% to 84.40% on LIBERO), these results indicate that the pretrained policy backbone matters much more than the pretrained target.
Critical weaknesses
  1. Lack of a control showing that the gain comes from information about the future. Several results in Table 1 point to a regularization effect that does not require future content: the random Wan anchor matches or exceeds the pretrained one (97.65 versus 97.45 on LIBERO, 58.63 versus 55.58 on LIBERO-Plus), and for Ovis-U1 the offset [0] target already lifts LIBERO-Plus from 65.27% to 69.65%. Random teachers are known to provide useful distillation targets on their own (Sarnthein et al., 2023), a work the authors cite. The missing experiment keeps the anchor, head, offsets and λF unchanged and replaces the recorded future with a mismatched one, for example the frame 16 steps ahead from another demonstration of the same task, a temporally shuffled window, or the frame 16 steps in the past. If most of the LIBERO-Plus gain survives, the result should be framed as feature regularization rather than foresight.
  2. Lack of a clear boundary between foresight and the generative objective it replaces. The strongest Wan target is not a learned representation: VAE-direct lifts each 2×2×48 latent patch by fixed coordinate repetition and scaling, which preserves cosine similarity, so the policy is effectively regressing the clean future VAE latents with a cosine loss. This is close to a deterministic, noise-free counterpart of the latent denoising objective of FastWAM, and the comparison then mostly varies the noise level and the loss function, as well as whether noisy future tokens pass through the backbone during training (294 tokens for generation versus 98 for foresight, Appendix B.2). The claim that "no visual generation is involved" is therefore less sharp than the abstract suggests, and the paper would be stronger if it located JAM within this continuum, for instance by training FastWAM with denoising restricted to small noise levels, or with an x0-prediction cosine loss on the same latents.
  3. Lack of a held-out protocol for selecting the configuration, and asymmetric tuning between JAM and the baselines. Offsets, anchor type, supervised block, head type and width, and λF are all compared on the same LIBERO and LIBERO-Plus evaluations that provide the final numbers, and the best target per architecture is selected by its LIBERO-Plus score. The recipe is also sensitive: among λF ∈ {0.4, 0.5, 0.6, 0.7} only 0.6 exceeds FastWAM on LIBERO (Table 12), block 20 drops to 95.45% (Table 11), and at 10k updates JAM trails FastWAM on LIBERO-Plus (48.47% versus 55.25%, Table 16). The baselines are run once with their default settings, and the locally trained ImageWAM (Ovis-U1) reaches only 80.2% on LIBERO-Long, which inflates the largest reported gap. Selection on a separate LIBERO-Plus manifest drawn with a different seed, and three runs for FastWAM and ImageWAM, would make the main comparison much more reliable.
  4. Lack of a matched comparison with existing future-latent objectives, which is the comparison that would establish what JAM adds. FLARE (Zheng et al., CoRL 2025), FRAPPE (Zhao et al., 2026), VLA-JEPA (Sun et al., ECCV 2026), VLAFlow (Xia et al., 2026) and JEPA-WAM (Lin et al., 2026) already supervise policy features with representations of future observations. The related-work section distinguishes JAM by where the head is attached and how the target space is chosen, but none of these alternatives is trained inside the FastWAM or ImageWAM setup. A supplementary experiment that keeps the JAM architecture and budget and swaps in a V-JEPA 2 or DINOv2 future target, and a FLARE-style future query inside the action expert, is needed to show that the specific placement and target selection of JAM matter.
  5. Lack of support for the cross-family conclusion as it is phrased in the abstract and conclusion. The backbones in Table 3 differ in parameter count (86.6M to 2.0B), action interface, text encoder and supervised depth (DINOv2 is supervised at its last block), and the paper itself acknowledges that pretraining effects cannot be isolated. Table 14 also shows that action-only Qwen3.5-2B already reaches 77.03% on LIBERO-Plus, 11.8 points above ImageWAM, while foresight adds 4.11 points for Qwen and 2.13 for DINOv2-L, with the robot and language categories (Qwen) and the language, light and background categories (DINOv2-L) favoring action-only training. Most of the difference between families therefore comes from the backbone and interface, and the evidence that foresight helps non-generative backbones rests on single runs with gains of a few points, so "strong foresight performance is attainable with either family" should be stated with more caution.
Questions for authors
  1. What happens if the future targets are replaced by mismatched futures (another episode, shuffled window, or past frame at the same offset) while everything else is kept fixed?
  2. The Yak anchor reads the current frame, the instruction and the future frame. Does the [0] readout therefore already contain future information, and is the [0] versus [1] comparison for Ovis-U1 really a current versus future comparison?
  3. Was any part of LIBERO or LIBERO-Plus held out for choosing offsets, anchors, λF and depth? If not, how much do the headline numbers change when selection is done on a separate manifest?
  4. Can you provide runs with several seeds for FastWAM and the locally trained ImageWAM, and explain the 80.2% LIBERO-Long result of the latter relative to the original ImageWAM report?
  5. FastWAM passes noisy future tokens through the backbone during training but not at deployment. Beyond the action-only control (95.25% and 45.55%), have you tested whether removing this train/deploy mismatch accounts for part of the improvement?
  6. Caching requires fixed target preprocessing. Do the generative baselines use image augmentation that JAM cannot apply to the future branch, and does this interact with the camera and lighting perturbations of LIBERO-Plus?
  7. Have you tried combining LG and LF? If the two are complementary, the claim that generation is unnecessary would need to be qualified.
5 comments
AuthorsOpenReview ↗

Response to Reviewer HDHt (1/5):

We sincerely thank Reviewer HDHt for the recognition of our work. The further questions are addressed as follows.

W1, Q1: Does the gain come from information about the future?

A: Using the recorded future from the same demonstration gives the best LIBERO-Plus result for both backbones. Compared with a future from another demonstration of the same task at the same stage, it improves success by 12.22 points for Wan (52.32% to 64.54%) and about 2.7 points for Ovis-U1 (77.22% to 79.95%). The mismatched future already improves over action-only training for both backbones (45.55% to 52.32% for Wan, 61.35% to 77.22% for Ovis-U1), suggesting that task- and stage-level cues also contribute.

More specifically, we ran the controls proposed in Q1 (new Table 23): only the target changes, while the anchor, head, loss weight and budget stay fixed. The real future is the one recorded in the same demonstration; the mismatched future is taken from another demonstration of the same task, drawn at random, at the same fraction of the episode (t′ in the table), so it shows the same task stage but not the motion that actually follows the current frame. For the Ovis-U1 controls at [1], Yak receives frame t or t−16 in place of t+16, so the readout position stays the same.

PolicyFuture anchorReadout offsetsTarget framesTime stepsLIBERO ↑LIBERO-Plus ↑
FastWAM action-only––––95.2545.55
FastWAM––––97.0055.31
JAM (Wan)Pretrained VAE[1,2]real futuret+4, t+8, …, t+3298.1564.54
[0]current framet89.9534.53
[1,2]mismatched futuret′+4, t′+8, …, t′+3297.6552.32
ImageWAM (Ovis-U1) action-only––––90.2061.35
ImageWAM (Ovis-U1)––––92.9065.27
JAM (Ovis-U1)Pretrained Yak[1]real futuret+1696.7579.95
[0]current framet94.5569.65
[1]current framet94.4567.33
[1]mismatched futuret′+1695.0577.22
[1]past framet−1692.2573.84

Current- and past-frame targets fall below both future targets on both benchmarks. For Wan, the current-frame VAE target scores 34.53 on LIBERO-Plus, below action-only (45.55). For Ovis-U1, a current-frame Yak target improves over action-only (61.35 to at most 69.65), but by much less than the real future (79.95). The past frame (73.84) also stays below both future targets.
A mismatched future beats the current frame but not the real future. For both backbones, it improves over action-only training and over the current-frame target (52.32 against 34.53 for Wan, 77.22 against at most 69.65 for Ovis-U1), but stays below the real future.
A random teacher alone gives no gain. Random teachers can give useful targets [R1], but when the target VAE itself is randomized, LIBERO-Plus drops from 64.54 to 45.75, the action-only level (45.55). The random Wan anchor of Table 1 is a different case: it is a fixed random transformer applied to pretrained-VAE latents of the recorded clip, so it still encodes the real future. These latents already show how the scene changes across the clip, so the transformer on top is not needed: the latents alone score highest (64.54), above both the random transformer (58.63) and the pretrained one (55.58), whose features were learned for denoising. Yak's input is not a clip but two still images, the current frame and one future frame, and what relates the two is its pretrained image-editing layers, so for Yak the pretrained anchor is better (79.95 against 72.51 with random weights).

Regularization alone does not explain the gain: with the same anchor, head and loss weight, a current-frame target does far worse than either future target.

AuthorsOpenReview ↗

Response to Reviewer HDHt (2/5):

Q2: Does the [0] readout already contain future information?

A: No, Yak [0] contains no future information. Yak's attention mask lets current-frame and instruction tokens attend only to each other and not to the future frame, so [0] and [1] are current and future readouts of the same frozen encoder. The frozen Wan anchor is bidirectional, so Wan's current-frame control is VAE-direct [0].

W2: Boundary between foresight and the generative objective.

A: VAE-direct foresight regresses clean future latents with a cosine loss, predicted from current-frame tokens through a training-only head. It thus differs from FastWAM's denoising in noise level, loss and where the prediction is computed. Its trainable visual branch sees only the 98 current-frame tokens, whereas FastWAM's also processes 196 noisy future tokens, for 294 in total. By "no visual generation" we mean that JAM does not use a denoising objective or a sampling path, and that future latents, noisy or clean, do not enter the trainable backbone. We ran the suggested variants and the combination asked for in Q7 (new Table 24):

Visual training objective, success (%)Wan LIBERO ↑Wan LIBERO-Plus ↑Ovis-U1 LIBERO ↑Ovis-U1 LIBERO-Plus ↑
Generation, L_G97.0055.3192.9065.27
Foresight, L_F98.1564.5496.7579.95
Small-noise generation, σ ≤ 0.192.5034.33––
x0-prediction cosine generation94.6045.29––
L_G + L_F98.5567.2694.7570.58

Noise level. Foresight's target is indeed the clean future latent, but the noise level of a denoising objective describes its input. In denoising, σ sets how much of the clean future the input still contains: all of it at σ = 0 and none at σ = 1. Foresight does not give the backbone any future latents, so its input corresponds to σ = 1, where the clean future must be predicted from the current frame. At small noise, the input already contains most of the target. Accordingly, moving FastWAM toward small noise does not approach foresight: restricted to σ ≤ 0.1, all else unchanged, FastWAM scores 34.33 on LIBERO-Plus, the level of the current-frame VAE target (34.53, W1) and 11 points below action-only training (45.55); in both, the answer is already in the input.

Loss form. Giving FastWAM foresight's loss instead, the cosine distance between the predicted clean latents x̂₀ = x_t − σ·v̂ and the recorded ones on the same 2×2×48 tokens with the foresight weight 0.6, scores 94.60 on LIBERO and 45.29 on LIBERO-Plus, the action-only level. Neither lowering the noise nor changing the loss closes the gap to foresight (64.54). The remaining difference between the two objectives is that foresight predicts the future from the current-frame tokens alone.

Q5: Train/deploy mismatch in FastWAM.

A: There is no train/deploy mismatch on the action path. FastWAM's attention mask lets the future tokens read the current frame, but neither the current-frame nor the action tokens can read the future tokens, so the future does not affect the actions. We checked this on the trained FastWAM and JAM (Wan) checkpoints: replacing the future tokens with noise, zeros or another episode's future leaves the actions bit-identical, and removing them, as at deployment, changes the actions only by floating-point rounding. Removing this mismatch therefore cannot explain the improvement.

Q7: Combining L_G and L_F.

A: The gains do not require generation: foresight alone beats generation alone on both generative backbones. We also trained the combined objective L_G + L_F (last row of the table under W2). Compared with generation alone, it is 11.95 LIBERO-Plus points higher for Wan and 5.31 higher for Ovis-U1, so foresight helps with or without generation. Compared with foresight alone, it is 2.72 points higher for Wan and 9.37 lower for Ovis-U1, while it brings back the training cost of the generative branch: an estimated 1.6 times foresight alone with online targets and 2.4 times with cached ones for Wan, and 1.4 and 1.5 times for Ovis-U1. JAM is meant to learn from recorded futures without any visual generation, so that the same recipe also serves backbones that cannot generate: Qwen3.5-2B and DINOv2 have no image or video decoder, so L_G does not exist for them.

AuthorsOpenReview ↗

Response to Reviewer HDHt (3/5):

W3, Q3: Held-out protocol for selection.

A: Following FastWAM [R2] and ImageWAM [R3], which compare their design choices on these benchmarks and do not describe a separate validation split, we did not hold out any part of LIBERO or LIBERO-Plus during development. The selection does not depend on the reporting manifest: LIBERO alone and a second LIBERO-Plus manifest both return the same two configurations.

We tuned only the settings that the foresight loss adds. The baselines keep their authors' settings; JAM's loss weight, depth and head were chosen on LIBERO, and its targets and readouts are compared in Table 1. LIBERO alone selects the same two configurations: VAE-direct [1,2] and Yak [1] also have the highest LIBERO success among the foresight rows of their architecture (98.15 and 96.75), and each of their three training runs is above every alternative (Table 21). These settings are not fragile: on LIBERO, depths 14 to 18 (Table 11) and loss weights 0.5 to 0.7 (Table 12) lie within or above FastWAM's own three runs (96.45 to 97.15), and they carry over without re-tuning, from the pretrained Wan target to VAE-direct and from Ovis-U1 to Qwen3.5-2B and DINOv2. The 10k comparison of Table 16 uses the pretrained Wan target; with the selected VAE-direct target, the 10k checkpoint of the 20k run already scores 57.97 on LIBERO-Plus, above FastWAM's final 55.31.

The second manifest uses seed 43. We re-evaluated the generative and foresight checkpoints of Table 1 on it without re-selection; its 1,246 tasks outside the paper manifest were not used in the original selection. LIBERO-Plus success (%), new Table 22:

PolicyFuture anchorReadout offsetsPaper manifest ↑Seed-43 manifest, all tasks ↑Seed-43 manifest, tasks outside the paper manifest ↑
FastWAM––55.3155.4456.02
JAM (Wan)Pretrained Wan[0]44.6241.7041.01
[1,2]55.5858.1059.87
[0,1,2]54.1854.9156.02
Random Wan[1,2]58.6360.3661.08
Pretrained VAE[1,2]64.5465.3466.45
ImageWAM (Ovis-U1)––65.2763.9464.13
JAM (Ovis-U1)Pretrained Yak[0]69.6569.1969.74
[1]79.9577.0977.53
[0,1]72.5871.1871.99
Random Yak[1]72.5170.3269.98
Pretrained VAE[1]76.4372.7873.27

The selected configurations remain unchanged on the second manifest. The two manifests give 2,752 distinct tasks. In 10,000 random validation/test halves, the validation half selects VAE-direct [1,2] among the five Wan foresight configurations of Table 1 and Yak [1] among the five Ovis-U1 ones every time; selecting on the 1,246 new tasks alone returns the same two. On those 1,246 tasks, the selected policies lead the generative baselines by 10.43 and 13.40 points, against 9.23 and 14.67 on the paper manifest.

W3, Q4: Several seeds for the baselines, and ImageWAM's 80.2% LIBERO-Long.

A: Mean ± SD over three training runs of each policy (revised Table 21):

PolicyLIBERO ↑LIBERO-Plus ↑
FastWAM96.87±0.3753.45±2.31
JAM (Wan)98.13±0.1364.61±0.37
ImageWAM (Ovis-U1)92.50±0.6964.14±2.49
JAM (Ovis-U1)96.75±0.1579.92±0.43

JAM leads its baseline on LIBERO-Plus by 11.16 points for Wan and 15.78 for Ovis-U1, both larger than the single-run gaps of Table 1, and every JAM run beats every baseline run on both benchmarks (Table 21).

The 80.2 Long score is the highest of the three ImageWAM (Ovis-U1) runs, so the paper run does not inflate the gap. The original ImageWAM paper's 98.4 Long score belongs to FLUX.2; that paper does not report a LIBERO breakdown for Ovis-U1 [R3], and our row uses its official configuration.

Comments on the paper

2 comments

Dear Reviewers,

Thank you again for your reviews. As the discussion period is coming to an end, we would like to confirm whether our responses and revised manuscript address your concerns. If your concerns have been addressed, we would be grateful if you would consider updating your scores. We are happy to answer any remaining questions.

Best regards,
The Authors

AuthorsOpenReview ↗

Summary of the Discussion

Dear Area Chair and Reviewers,

Thank you for your reviews and for the discussion. All three reviewers now recommend weak acceptance. Reviewer c8KX raised the score from 2 to 3, noting that the new comparisons and ablations "strengthen the empirical evidence supporting the proposed approach", and the ratings of Reviewers HDHt and W7DZ remain at 3. Below we summarize the main concerns and the new results that address them.

  • Future content versus regularization (HDHt, c8KX). With the anchor, head and loss weight fixed, the real future outperforms every control target (current frame, past frame, mismatched future) on both backbones (Table 23).
  • Boundary with generation (HDHt). Lowering FastWAM's noise level or training it with the foresight loss does not close the gap to foresight (34.33 and 45.29 against 64.54 on LIBERO-Plus, Table 24).
  • Related methods (HDHt, c8KX). Within the same recipe, JAM's own target and placement outperform VideoMAE-B, V-JEPA 2 and FLARE-style action queries, and JAM exceeds the released VLA-JEPA checkpoint under our evaluation protocol (Table 25).
  • Selection and training variance (HDHt, W7DZ). LIBERO alone and a second LIBERO-Plus manifest select the same configurations (Table 22), and across three training runs, every JAM run exceeds every baseline run on both benchmarks (Table 21).
  • Cross-backbone comparison (HDHt, c8KX). Foresight improves over action-only training in all three runs of Qwen3.5-2B and DINOv2-L (Table 14), and Table 3 now reports complete policy sizes.
  • Training cost (W7DZ). With both methods caching their future encodings, JAM still reduces estimated training FLOPs by 57.7% relative to FastWAM and 31.9% relative to ImageWAM (Table 9).

All of these results are in the revised paper, marked in blue, and we will keep them in the final version. Thank you again for your time and feedback.

Best regards,
The Authors

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

Related papers

Loading the map…

Discussion (0)

Sign in to comment.