- Rating
- 1
- Confidence
- 3
- Informativeness
- 1
- Presentation
- 1
- Soundness
- 2
Read review
Paper summary
This paper addresses the lack of object-level semantics and fine-grained spatial structure in VAEbased perception for World Action Models (WAMs), and proposes P-WAM, which preserves the original VAE-based generative video-action backbone and prediction objectives while adding a discriminative perception branch based on SigLIP and DINOv3 and alternately injecting fused perception tokens into later DiT cross-attention layers. Experiments show that P-WAM improves policy success rates in RoboCasa simulation and real-world robot tasks while maintaining comparable SSIM/FID for future-frame prediction. It demonstrates the complementary effectiveness of discriminative visual perception for generative world-action modeling.
Critical strengths
- Well-motivated and practical architectural integration. P-WAM adds pretrained discriminative perception to a shared video–action backbone while retaining the original VAE pathway, denoising targets, and prediction objectives. The reported overhead is modest—approximately 12% additional FLOPs and 4% higher inference latency— while future-frame prediction quality remains comparable on the reported SSIM and FID metrics. 【pp. 4–6, Table 1】
- Consistent gains over a strong baseline. P-WAM outperforms Cosmos Policy across all five reported RoboCasa evaluation seeds, raising average success from 65.7% to 67.7% (t=8.16, p<0.01). Real-world success improves from 64.0% to 73.7%, with gains on five of seven tasks. 【Table 5; pp. 7–8, Table 3】
- Useful controls addressing alternative explanations. Additional training of Cosmos Policy reaches only 65.6%, and random-initialized encoders reach 66.3%. These controls provide evidence that pretrained perception contributes to the gains beyond additional training steps or architectural capacity alone. 【p. 8】
- Complementary evaluation of representations and behavior. The paper combines frozen video-action DiT representation probing, simulation, real-world deployment, and architectural ablations. The probe shows improved foreground mIoU, supporting more decodable perceptual information in the shared representation. 【Fig. 3, Tables 2–4】
Critical weaknesses
- Limited conceptual novelty for ICLR. P-WAM primarily integrates established components—pretrained visual encoders, feature pooling and fusion, and cross-attention conditioning—into an existing world–action backbone. The paper acknowledges precedents for separate perception and generation interfaces and alternating condition injection. Although practically useful, the integration provides limited new modeling principles or mechanistic insight, leaving its conceptual contribution somewhat modest for ICLR. 【pp. 1, 3–5】
- Mechanistic validation and preservation of world modeling remain insufficient. Concurrent improvements in segmentation probing and policy success do not establish that enhanced semantic and spatial perception causes the policy gains. The pretraining control does not isolate this specific mechanism. Likewise, unchanged SSIM and slightly improved FID do not adequately establish preserved action-conditioned dynamics, contact behavior, or physical consistency. Since later backbone layers and output projections are updated, retaining the generative pathway does not guarantee preservation of its functionality. 【pp. 5, 7–8, Fig. 3, Table 1; p. 18, Table 9】
- Statistical robustness and ablation conclusions remain limited. The five reported random seeds demonstrate evaluation stability, but whether each seed involves independent retraining is not clearly documented. Most ablations lack multiple seeds or confidence intervals, making small differences—such as 0.6 percentage points over SigLIP alone and 1.0 point over concatenation—insufficient to establish robust designsuperiority. Although App. A provides per-task results, the main paper lacks separate ID/OOD breakdowns and variance estimates for the key ablations, limiting assessment of the distribution and reliability of the gains. 【p. 7, Table 4; p. 16, Tables 5–6; p. 21】
- Generality is not yet established beyond the evaluated setting. The framework is tested only on the Cosmos Policy backbone, while real-world evaluation covers one robot platform and seven tasks, with 25 trials per task. Benefits are uneven: simulated articulated-object tasks show a slight decrease, and real-world pouring and clothfolding tasks remain unchanged. These results support effectiveness in the tested settings, but broader benefits across WAM architectures and manipulation skills remain unproven. 【p. 7, Table 3; p. 16, Table 6; pp. 19–21】
Questions for authors
Q1: What is the key technical novelty beyond integrating existing components, particularly relative to Cosmos3 and prior perception–generation or hybrid VLA/WAM approaches? Q2: What evidence links improved semantic/spatial perception to policy gains? Could targeted interventions distinguish this mechanism from general pretraining benefits? Q3: How is preserved world-modeling capability established beyond SSIM/FID—for example, through action-conditioned prediction, contact behavior, physical consistency, or long-horizon consistency?