acceptodds
Withdrawn from ICLR 2027

Enforcing Causal Routing in Embodied World Models

Withdrawn from ICLR 2027, so left out of the paper lists and the map.

Abstract

Embodied world models predict how a robot's environment responds to commanded actions, enabling policy evaluation and planning with varied actions. We show that current models produce such predictions partly by peeking at the future. Frames generated at early timesteps can be corrupted by subsequent actions and the goal, a phenomenon we term **causal leakage**. Existing benchmarks typically replay recorded actions, which match the ground-truth video, so leakage does not lower scores and goes undetected. We introduce Causeway, a world model that routes information along causal paths while preserving visual quality. United Causal Attention aligns the attention structure with this causal order while retaining semantic guidance from language; Biomimetic Sampling Memory sustains prediction fidelity over episode-length rollouts by anchoring the initial observation alongside recent context. To make causal leakage measurable, we introduce the Causal Leakage Probe, to our knowledge the first black-box test of whether subsequent actions affect earlier frames or the goal affects the video directly. Causeway's future-action leakage is 63% below the best baseline's, with no measurable goal leakage under the tested interventions. It leads on six of twelve WorldArena 2.0 metrics and reduces FVD on full-episode DROID rollouts from 113.6 to 8.9 over the same backbone. Code and weights will be released.

Primary area:
applications to robotics, autonomy, planning
Keywords:
world models, embodied AI, autoregressive video generation

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_sQ2rOpenReview ↗
Rating
3
Confidence
4
Informativeness
2
Presentation
2
Soundness
2
Read review
Paper summary

Causeway targets causal leakage in action-conditioned embodied world models through the observation-to-video attention path, a model predicting pre-contact frames can see post-contact scene content. It intergrates Unified Causal Attention which partitions attention keys by modality and timestamp and licenses each query by role. Moreover, BioSaM adds geometric anchor/recent memory, a conditioning-noise cap and per-frame diffusion timesteps. Finally, the Causal Leakage Probe injects contradictory action and goal pairs and compares debiased energy in a forbidden window against a licensed one.

Critical strengths
  1. UCA is effective. Partitioning keys into scene, goal, action, and video, and licensing each query by role and timestamp, fits inside a standard attention kernel. For example, with weights and seeds held fixed, removing UCA leaves frame metrics almost unchanged but nearly doubles FVD and drops retention from 0.87 to 0.76.

  2. The ablation study is very comprehensive. Each BioSaM component is removed individually with the same weights and seeds, and the Cosmos3-Nano comparison uses the same backbone, making it a recipe comparison rather than a cross-model one.

  3. The paper is well-presented. The paper is easy to follow, with dedicated figures and tables. The structure is well-organized and the methods make sense.

Critical weaknesses
  1. The paper does not measure UCA's effect on leakage. Table 5 reports PSNR, SSIM, LPIPS, FVD, retention, context length, and sampling time, so it shows only fidelity effects. However, the authors do not test the w/o-UCA checkpoint. Since it is the key contribution of this paper, I think it should be included.

  2. The input conditioning still carries the answer and the probe cannot tell. CLP contradicts the action class or adds a “keep still” goal prefix, but the scene text and object locations remain, so a model can locate what to manipulate without using the action, and the probe cannot find that. Therefore, it's hard to tell if the proposed methods really work.

  3. Leakage is not tested in real-world robots. The paper does not test the probe's calibration on real-robot data or a large-scale simulator, where it makes its fidelity and ablation claims, and no single dataset shows both the leakage statistic and the effect of removing UCA. Whether the probe transfers is therefore untested.

Questions for authors
  1. Please run CLP on the w/o-UCA checkpoint from Table 5 on the dataset used there.

  2. Since scene tokens are generated from the task description, and the goal text contains object locations, both are conditioning inputs visible under UCA. Can you run the probe with scene captions generated from a contradictory task, or with the goal stripped of object locations? I think this experiment can better demonstrate the effectiveness of your methods.

  3. What does the language conditioning contribute? The w/o-language row is 24.09 PSNR and 8.9 FVD against 24.29 and 8.9 with language, and the text notes that most of the improvement over the official checkpoint is already present without it. Is the VLM memory doing work these metrics don't capture, and can you show it on a task where language matters?

  4. The paper claims the first black-box test for causal leakage. On what dimension does CLP differ from MiraBench and FACT, both of which are cited once and never compared?

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

This paper investigates causal leakage in action-conditioned embodied world models: future actions may influence earlier predicted frames, and task goals may affect generated videos even when commanded actions remain fixed. It argues that evaluations based on recorded action replay can overlook these dependencies because future actions and instructions are consistent with the reference outcomes.

The proposed model, Causeway, introduces United Causal Attention to restrict direct cross-modal information paths, separating scene descriptions from goals and routing goal information through action representations. Biomimetic Sampling Memory maintains initial observations and selected historical context for long-horizon prediction. The paper also introduces the Causal Leakage Probe (CLP), a black-box evaluation that measures responses to controlled future-action and goal interventions.

Experiments on RoboTwin 2.0 and DROID evaluate intervention-specific leakage, video quality, and full-episode prediction, with additional ablations and a downstream policy-training experiment. Causeway reports a 63% lower action-leakage ratio than the best evaluated learned baseline and no measurable goal leakage under the tested interventions, while improving several video-quality and long-horizon prediction metrics.

Critical strengths
  • A useful diagnosis of an evaluation blind spot. The paper highlights that accurate reconstruction under recorded actions does not establish whether a world model respects temporal action dependencies. CLP makes this distinction testable through controlled future-action and goal interventions. Matched sampling seeds, physical-time evaluation windows, action-response gates, and renderer and temporal-shift controls strengthen the interpretability of the measurements. This is a consequential contribution because action-faithful prediction is central to using world models for planning and policy evaluation.

  • A practical model design with complementary empirical support. Causeway combines restricted cross-modal attention with a bounded memory mechanism for episode-length prediction. Under the tested interventions, it achieves lower measured leakage while remaining competitive on video quality and long-horizon fidelity. The DROID ablations provide evidence that attention routing and historical context contribute to prediction quality; for example, removing UCA increases FVD from 8.9 to 16.9. These results make the proposed design relevant beyond the diagnostic metric itself and demonstrate a useful balance between information-routing constraints, prediction fidelity, and context efficiency.

Critical weaknesses
  • The evidence for goal-leakage suppression is limited by the intervention design and the unvalidated scene–goal decomposition. The CLP evaluation uses 32 probe points from 11 episodes and four sampling seeds. Its goal interventions consist of three “hold still” prefixes appended to the original task context, covering a relatively narrow semantic contrast and short evaluation window. Moreover, the scene description is generated by a language model that also receives the task description, but the paper does not audit whether task-dependent information enters the scene field, which remains directly accessible to the video tower. Holding this field fixed during goal interventions would not test that potential pathway. Although the paper appropriately qualifies its conclusion by the tested interventions, the evidence remains insufficient to establish how reliably the proposed decomposition preserves useful semantics without introducing goal dependence.

  • The contribution of the individual routing constraints to leakage reduction is not sufficiently isolated. UCA retains bidirectional communication within the action and target-video streams, leaving indirect paths through which future-action information can reach earlier predictions, as the paper acknowledges. In Table 4, removing the temporal action mask changes CLP_A from 0.034 to 0.044, but no paired uncertainty estimate for this difference is reported. The video-feedback comparison is conducted with the temporal mask already disabled, leaving component interactions unresolved. Meanwhile, the stronger within-tower causal variants in Table 5 report prediction quality without corresponding CLP results. Consequently, the experiments do not clearly establish which constraints drive the leakage reduction or fully characterize the claimed quality–causality trade-off.

Questions for authors
  • Scene–goal decomposition and intervention scope. Since Qwen receives both the initial image and the task description, what checks were performed to prevent intended actions or outcomes from entering the scene field? During CLP_L evaluation, was the scene text held fixed while only the goal span was modified? Please clarify whether the reported result tests only the post-decomposition goal pathway or also covers potential task dependence introduced during preprocessing.

  • Interpretation of the zero goal-leakage result. Please report the unrounded CLP_L estimate, its confidence interval, and the absolute goal-response energy. Were the generated videos identical across goal interventions under matched seeds, or did the estimator yield a near-zero value after centering and averaging? Uncentered, same-seed video differences would help distinguish output invariance from a small mean response. Please also clarify whether goal interventions beyond the three “hold still” prefixes were evaluated.

  • Attribution to individual routing constraints. Can the authors provide a paired bootstrap estimate and confidence interval for the CLP_A difference between the full model and the variant without the temporal action mask? Are CLP results available for the causal action-self-attention and causal target-video-self-attention variants already evaluated in Table 5? These results would help establish the contribution of the temporal mask and whether the prediction-quality costs of stronger restrictions are accompanied by measurable leakage reductions.

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

The paper introduces Causeway, an embodied world model that reduces causal leakage through attention routing and sustains long rollouts with sparse visual memory. It also proposes CLP an intervention-based test of future-action and goal dependence. Experiments on RoboTwin 2.0, WorldArena 2.0, and DROID show reduced measured leakage and improved video prediction, with component ablations supporting the design.

Critical strengths
  • The strongest part for me is the problem they identify: a model can produce a convincing replay while still getting the causal dependencies wrong. CLP gives a concrete way to catch this, which matters if we want to use these models to compare actions or plan.
  • The improvements are not limited to the new leakage metric. The model also produces better long rollouts, and the ablations suggest that both routing and memory help. That makes the approach more convincing.
Critical weaknesses
  • I am not fully convinced that three separate towers are necessary. Since the actions are supplied rather than generated, could a simpler action encoder provide comparable conditioning? A comparison would help justify the separate action tower.
  • UCA blocks certain direct information paths, but it does not guarantee causal outputs. Action and generated-video self-attention remain bidirectional, leaving possible paths such as A3 -> A1 -> V1 or A3 -> V3 -> V1 across layers
  • UCA and BioSaM seem to address largely separate problems: information leakage and long-horizon consistency. The individual ablations are useful, but the paper does not clearly establish whether there is a meaningful interaction between the two, beyond combining two independently useful components.
  • The paper should report per-tower parameter counts and inference timings under matched settings to clarify the additional cost. If the authors can address the above questions, I would be open to raising my score.
Questions for authors
  • When BioSaM selects historical video latents, which historical action tokens are included in the context? Are actions sampled alongside each selected memory frame, or is a larger action history retained? How is their temporal alignment maintained?
  • How does BioSaM compare with memory mechanisms from prior work under a matched context or compute budget? The current comparisons with uniform sampling, full history, and initial-frame-plus-overlap context do not establish its advantage over other published memory methods.

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

Related papers

Discussion (0)

Sign in to comment.