Completion Aware Guidance for World Action Models
World Action Models (WAMs) jointly predict visual futures and robot actions, yet they often hallucinate locally plausible trajectories that never execute the critical state transition needed to finish the task. This paper identifies "task-incomplete imagination" as a failure mode arising from the mismatch between long-horizon video pretraining and short-horizon control generation, and proposes Completion Aware Guidance (CAG), a training-free sampling method that up-weights instruction tokens already attended to by the model. Across two WAM architectures, CAG boosts success rates and cuts incomplete-imagination failures, offering a lightweight inference-time fix for a structural objective mismatch.
The paper makes a crisp conceptual contribution by isolating task-incomplete imagination as distinct from video corruption or action misalignment. However, its empirical gains are modest and uneven: while DreamZero zero-shot success rises from 69% to 75%, RoboTwin results are mixed, with several tasks regressing. The theoretical objective involving an oracle phase-completion indicator $E_t(z_t) \in \{0,1\}$ is elegant but the practical method substitutes an attention-based heuristic, and the paper does not validate that this proxy faithfully approximates the ideal gradient.
Quote: "WAMs exhibit task-incomplete imagination, generating visually coherent, instruction-relevant futures while omitting the state transitions required for task completion" (Section 1)
The diagnosis of objective mismatch is convincing and experimentally grounded. Figure 2(a) shows that extending the generation horizon from 1.6 seconds to 5 seconds recovers task-completing transitions, supporting the claim that the backbone is capable but short-horizon chunking suppresses this behavior. The training-free design of CAG is practically appealing, and the cross-attention intervention is well-motivated by the guidance objective.
Quote: "the native horizon produces the required completion event, whereas short-chunk autoregressive generation remains visually coherent but repeatedly omits it" (Section 3)
First, the RoboTwin 2.0 results are inconsistent: Open Microwave drops from 33% to 30%, Place Can Basket falls from 63% to 53%, and other gains are thin. Averaging across tasks masks these regressions and the small sample size (30 rollouts per task) limits statistical confidence. Second, the practical implementation of CAG relies on a sparse binary gating function $\psi$ applied to attention scores, but the paper gives no threshold or ablation for this discretization, and the ideal phase-completion indicator $E_t(z_t)$ is unavailable at test time. The method therefore optimizes a plausible proxy rather than the stated objective. Third, because CAG amplifies instruction tokens already receiving high attention, it may simply increase the temperature of existing behavior rather than steer toward truly novel transitions.
Quote: "Place Can Basket ... 63 ... 53" (Table 1)
Quote: "CAG is currently limited to sampling-time intervention, and extending completion aware guidance to WAM training is a promising direction for future work" (Section 5, Limitations)
The evidence partially supports the central claim. The failure-mode analysis on DreamZero is the strongest evidence, showing task-incomplete imagination incidence drops from 79% to 40%. However, comparisons to $\pi_{0.5}$ and GO-1 in Table 1 are only provided as context and do not represent controlled comparisons under identical conditions. The paper also does not compare against other test-time guidance techniques such as classifier-free guidance scaling or prompt engineering, making it difficult to isolate whether CAG's token-specific scaling offers unique benefits over simpler baseline interventions.
Quote: "the fraction attributable to task-incomplete imagination decreases from 79% to 40% with CAG" (Section 4)
Quote: "We follow the corresponding evaluation protocols, changing only inference-time guidance between the baseline and CAG" (Section 4)
Reproducing the results would be challenging. The paper mentions no code release, and the experiments require existing WAM implementations (DreamZero and Fast-WAM) that are themselves recent and complex. While the hyperparameter $\alpha$ is fixed at 0.15, the sparse binary gating function $\psi$ is not fully specified, and the set of layers $\mathcal{L}$ used for attention aggregation is not detailed. Independent reproduction would need to infer multiple implementation choices.
Quote: "fix $\alpha=0.15$ across all tasks and backbones without per-task tuning" (Section 4)
Quote: "where $\psi$ is a sparse binary gating function" (Section 3)
World Action Models (WAMs) predict visual futures and robot actions, yet they remain susceptible to task-incomplete imagination, where plausible, action-consistent predictions omit the transition needed for task completion. In this paper, we show that this failure is not inherent to the world model backbone, but emerges when adapted for short-chunk control, which can repeatedly favor plausible local continuations over task-completing transitions. To address this, we introduce Completion Aware Guidance (CAG), a training-free sampling method that guides generation toward task completion. Across representative WAMs, CAG improves success from 64% to 70% on a RoboTwin 2.0 subset and from 69% to 75% in zero-shot simulation, while reducing task-incomplete imagination from 79% to 40%.
Pick a starting point or write your own. Challenges run in the background, so you can keep reading while the AI investigates.
No challenges yet. Disagree with the review? Ask the AI to revisit a specific claim.