acceptodds
Under review as a conference paper at ICLR 2027

DIWA: Decision-Influential World Abstraction for VLA-WAM Policies

Abstract

Predicting future observations can improve robot decision making, but a controller need not model every visible change with equal fidelity. We introduce Decision-Influential World Abstraction (DIWA), a framework that allocates future modeling capacity according to its effect on action generation. A lightweight estimator ranks object–time queries before an action-conditioned world decoder expands them. Its supervision measures changes in action responses, value estimates, and task progress under latent interventions, with shared randomness for stochastic action heads. A complementary regret geometry organizes representations by the relative returns of aligned candidate action rules. During inference, only selected future queries are decoded, avoiding dense future generation. Across LIBERO, RoboTwin, RoboCasa, and real-robot manipulation, DIWA achieves 73.5% aggregate success, improving over DreamVLA by 5.6 percentage points. The out-of-distribution evaluation yields 75.8% success, a 10.6-point improvement. Mean inference latency is 89 ms, compared with 213 ms for dense imagination and 241 ms for WorldVLA. Three-seed simulation results, 300 physical trials per method, budget sweeps, and intervention diagnostics distinguish the roles of influence supervision, regret geometry, and selective computation. These results support using decision sensitivity to determine which future information a robot policy should compute.

Primary area:
applications to robotics, autonomy, planning
Keywords:
World Action Models, Vision-Language-Action Models, Adaptive Computation, Robot Manipulation

Reviews

Among:
Shape:
1234this paper, Bottom 10%
Figure 1. Mean review rating of the reviewed papers at ICLR 2027. This paper averages 1.67 over 3 reviews, above 10% of the others.
Reviewer_hMbAOpenReview ↗
Rating
1
Confidence
3
Informativeness
2
Presentation
3
Soundness
1
Read review
Paper summary

DIWA extends DreamVLA with selective future prediction. A preliminary action chunk conditions query scoring and world prediction. The model selects at most 12 of 48 object-time queries before decoding their future features, which are then fused with the current context to generate the final action. Training adds intervention-based influence supervision and regret geometry based on measured candidate returns. The paper reports 73.5% aggregate success across three simulation platforms and physical manipulation, versus 67.9% for DreamVLA, with 89 ms latency versus 213 ms for dense DIWA. The main concerns are the distinct task contributions of ranking and action refinement, the data requirements of the additional supervision, and the scope of the physical evaluation.

Critical strengths
  1. Selective decoding offers a useful reported tradeoff between success and computation. Sparse DIWA reaches 73.5% aggregate success at 89 ms, versus 73.4% at 213 ms for dense DIWA (Table 4). The fixed-budget sweep also shows where stronger pruning hurts: reducing from 12 queries to six lowers success from 73.6% to 69.8% (Table 10).

  2. The independent diagnostics provide evidence that the learned quantities track their intended targets. Influence scores achieve Spearman ρ=0.71\rho=0.71 with intervention-induced action changes and 83% Top-12 recall, versus 0.48 and 65% without cross-episode swaps (Table 5). Latent-regret distance correlation is 0.68, versus 0.29 without regret geometry. These results support the learned ranking and representation objectives within the reported diagnostic setting. The finite-candidate proposition is appropriately limited to the measured, aligned candidate bank.

Critical weaknesses
  1. The paper does not show whether predicting futures improves on executing the initial action proposal directly. The proposal also guides which futures are selected and predicted, so a poor proposal may limit the information available to correct it. The results do not establish how much refinement helps or how it depends on proposal quality.

  2. DIWA uses extra supervision whose source and cost are not fully reported. Candidate returns and progress labels go beyond action demonstrations, but collection is described only for LIBERO (Appendix B). Without the label coverage and collection costs on the other platforms, readers cannot judge the required data budget or how much the gains depend on this extra supervision.

  3. The physical tasks are not described well enough to judge what robot capabilities were tested. Table 9 lists six task categories, but the setup, required actions, and trial protocol are unclear (Appendix E). The reported success rates therefore give limited evidence about the complexity and range of behaviors covered.

Questions for authors
  1. How does the final action improve on the preliminary proposal, and how does this benefit depend on proposal quality? Examples or quantitative evidence would help explain the role of predicted futures in this step.

  2. How were candidate-return and progress labels obtained on each platform, with what coverage and collection cost? Please clarify how this additional supervision affects the comparison with the baselines.

  3. Please provide real-robot demo videos showing DIWA running at inference time and executing representative tasks. Include an annotated setup figure showing the robot, gripper, cameras, and workspace. Briefly explain the task goals, control loop, and trial protocol so readers can assess the evaluation's scope and complexity.

  4. Could you clarify the repeated seed ordering and recurring score increments in Table 7, ideally using the underlying success counts or logs? Also, what explains the 142 ms latency without influence versus 89 ms for full DIWA, given that Appendix E says the scorer is still evaluated and the query budget is not expanded?

Reviewer_TbpWOpenReview ↗
Rating
2
Confidence
3
Informativeness
1
Presentation
1
Soundness
1
Read review
Paper summary

The paper proposed a VLA-WAM policy called DIWA which adds an additional selection stage to the DreamVLA world-action model. It forms object-time queries and use a lightweight score network to predict the top queries and then use those to decode by an action-conditioned world model decoder. During training, it perturbs the decoded future tokens and the score function is supervised by how much the policy's action, value and progress changes given the perturbation. The paper evaluates on LIBERO, RoboTwin, RoboCasa and real robots.

Critical strengths
  1. The paper proposes a novel way to selectively decoding the future imagination to help with the action generation while maintaining a relatively low latency.
  2. The evaluation spans three simulation benchmarks and also real world manipulation tasks.
Critical weaknesses
  1. The author proposes a complicated framework that combines future imagination with action generation. The proposed framework is optimized with eight weighted losses but the paper doesn't include clear ablations to illustrate the necessity of each term and there is also no intuition or analysis of the weight for each loss. Dense DIWA scores closely to Full DIWA and the overall gain over the baseline is not clearly attributed to the loss term.
  2. The cited baselines on LIBERO is below 80% but there are more state-of-the-art methods, e..g, pi0.5 that has higher performance on LIBERO. And the performance gain over DreamVLA is about 5% which is limited in terms of proposing this complicated framework.
Questions for authors
  1. The real world tasks are not clearly defined. In the paper table I, it reports multiple models's performance in the real world but in the Table 9 only the name of the real world tasks are defined. It is not showing the qualitative images of what the task looks like, what are the evaluation conditions etc, making the readers harder to evaluate the difficulty of the tasks.
Reviewer_hxuCOpenReview ↗
Rating
2
Confidence
3
Informativeness
1
Presentation
2
Soundness
2
Read review
Paper summary

This paper studies how to allocate future-prediction computation according to its influence on robot decisions and proposes Decision-Influential World Abstraction (DIWA). Candidate futures are organized as object–time queries, and a lightweight estimator predicts their importance before future decoding. Only selected queries are expanded to support action generation. The estimator is supervised by changes in action responses, value estimates, and predicted progress under latent-token substitutions. The method additionally regularizes representation geometry using regret vectors derived from measured candidate-action returns. Experiments across three simulation platforms and real-robot manipulation report improvements in success and robustness, while budget sweeps, component ablations, and intervention diagnostics examine efficiency and the proposed mechanisms.

Critical strengths

A clear problem formulation with selection before expensive computation. The paper distinguishes predictive detail from decision utility and selects queries before future decoding, rather than compressing fully generated futures. This provides a direct route to reducing inference computation, supported by the reported success–latency comparison between sparse and dense variants.

Critical weaknesses

1.The current comparisons do not clearly isolate the core contribution. Relative to DreamVLA, DIWA introduces not only query selection but also candidate-return supervision, progress supervision, and multiple auxiliary objectives. The success gains therefore cannot be attributed specifically to decision-influence-based selection. Moreover, matched query budgets and training conditions are not established across the ablation variants, limiting component-level interpretation. The intervention diagnostics primarily validate sensitivity to latent substitutions and do not sufficiently establish its connection to actual task utility. 2. Incomplete supervision and evaluation records limit verification of the main results. The method requires measured candidate-action returns, but their collection outside LIBERO, particularly on the physical platform, is insufficiently documented. Appendix E also identifies missing run-specific hardware, timing settings, baseline configurations, physical-system details, and diagnostic sample records. These omissions directly affect the fairness of additional data budgets, the validity of latency comparisons, and the reproducibility of the mechanism diagnostics.

Questions for authors

1.How much benefit remains when the dense, uniform-selection, and influence-selection models share candidate-return and progress supervision and the same training recipe, with matched query counts between selection variants? Which reported comparisons satisfy these controls? 2. How were candidate-action returns collected for RoboTwin, RoboCasa, and the physical platform? Please clarify state-reset procedures, candidate execution counts, additional interaction budgets, and how these supervision records correspond to the reported checkpoints. 3. Beyond correlation with latent-intervention responses, what evidence shows that high-influence queries contain information that improves actual task success? How do you rule out sensitivity to substitution artifacts or information on which the current policy incorrectly relies? 4. Could you provide the actual hardware, input configurations, sampling settings, and repeated timing statistics for the latency results, and explain the 142 ms latency of the no-influence variant? In particular, how does the switch described in the appendix—which still evaluates the estimator but zeros its logits—correspond to that variant’s actual implementation and measurement?

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

Related papers

Loading the map…

Discussion (0)

Sign in to comment.