acceptodds
Desk rejected at ICLR 2027

AcrossWAM1.0: A Modular Latent World–Action Stack for Compact Robot Policies

Desk rejected at ICLR 2027, so left out of the paper lists and the map.

Abstract

Latent world–action models avoid rendering future pixels by predicting an action-relevant visual subgoal in feature space. LaWAM established this formulation, but its original presentation left the world model, multimodal backbone, and deployment checkpoint tightly coupled. We introduce \method, a modularization and scaling study of this latent world–action stack. Rather than presenting latent subgoals as a new algorithm, we make the module boundary explicit: a policy adapter produces latent-action and action-generation contexts; a retained latent world decoder grounds the predicted transition in the current scene; and a flow-matching expert generates continuous action chunks. We further separate training-only teachers from the inference graph and provide a verifiable deployment export. On 2,000 paired LIBERO episodes, replacing a Qwen3-VL-2B backbone with Qwen3.5-0.8B yields 97.45% success versus 98.00% for the 2B model (a percentage-point difference; exact McNemar ). This does not prove equivalence, but it meets a prespecified two-point retention criterion. The compact, inference-reachable checkpoint contains 1,472.6M unique parameters, 42.4% fewer than the original 2B policy, while all retained tensors are bitwise identical to the source checkpoint. Cross-family execution is additionally checked with a MiniCPM-V adapter smoke test; closed-loop cross-family transfer remains an open evaluation. \method therefore contributes an auditable software and evaluation boundary for compact latent world–action policies, distinct from LaWAM's original latent-subgoal contribution.

Primary area:
applications to robotics, autonomy, planning
Keywords:
world action models, vision-language-action models, latent world models, robot manipulation, compact robot policies, modular architectures, flow matching

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

Building on LaWAM, this paper organizes the policy backbone, latent world decoder, and action expert into a modular system with explicit interfaces. It also separates components used during training from those required for deployment. The authors compare Qwen3-VL-2B with Qwen3.5-0.8B on the same 2,000 LIBERO evaluation episodes, achieving success rates of 98.00% and 97.45%, respectively, while reporting a 42.4% reduction in deployment parameters. The paper also provides checkpoint tensor consistency checks and forward and backward execution tests with MiniCPM-V.

Critical strengths
  1. The compact backbone study provides useful evidence on the trade-off between success rate and model size. Table 2 compares the two checkpoints on paired episodes with identical tasks and initial states. The 0.8B configuration loses only 0.55 percentage points in success rate.
  2. The deployment export includes concrete checks and provides useful implementation guidance. Section 4.5 documents the removal of modules used only during training, the removal of duplicate parameter aliases, bitwise consistency checks on 1,078 retained tensors, and a CPU loader test. Section 3.6 also distinguishes interface compatibility, closed-loop performance retention, and transfer across backbones with a frozen world decoder.
Critical weaknesses
  1. The compact-backbone study has limited coverage and lacks comparisons with existing compact policies. The controlled closed-loop comparison covers only two Qwen backbone configurations on LIBERO. This provides limited evidence on whether the framework consistently preserves policy performance when the backbone is reduced. Comparisons with existing compact policies are also missing. For example, X-VLA [1] reports a model size of 0.9B and an average LIBERO success rate of 98.1%.
  2. The value of modularity across model families lacks closed-loop validation. The MiniCPM-V forward and backward smoke test checks interface compatibility, dimension alignment, and gradient propagation. These checks do not establish whether effective robot control can be maintained after replacing the backbone. A closed-loop evaluation with at least one backbone outside the Qwen family would provide stronger support for the practical value of the modular design across model families.
  3. The parameter reduction is not linked to measured training and deployment benefits. The paper reports parameter counts and training steps, but does not provide comparable training times or GPU-hours. It also lacks measurements of end-to-end inference latency and peak GPU memory usage during deployment.

[1] Zheng, Jinliang, et al. "X-vla: Soft-prompted transformer as scalable cross-embodiment vision-language-action model." International Conference on Learning Representations. Vol. 2026. 2026.

Questions for authors
  1. The compact-backbone comparison currently covers only two Qwen configurations on LIBERO. Could the authors extend the evaluation to other model families and additional benchmarks to assess whether the observed performance retention generalizes across models and task distributions?
  2. The MiniCPM-V smoke test shows that the adapter can execute, but does not assess control performance. Could the authors provide closed-loop LIBERO results for this backbone to evaluate the practical value of the interface across model families?
  3. Table 1 reports RoboTwin success rates of 92.09% on Clean and 89.12% on Randomized for AcrossWAM1.0. However, Section 4.2 excludes inherited LaWAM results and preliminary language diagnostics from these columns, while Section 5 states that the closed-loop evidence covers only LIBERO. Please clarify the source, model configuration, and evaluation protocol for these two results, and explain the inconsistency with the text.
Reviewer_ypCBOpenReview ↗
Rating
1
Confidence
2
Informativeness
2
Presentation
2
Soundness
2
Read review
Paper summary

This is a modularisation and compact-scaling study built on LaWAM, not a new algorithm. The latent world–action formulation is inherited and stated as such. The contribution is the surrounding contract: backbone adapters, separate transition and action query groups, automatic hidden-size alignment, a training-only versus inference-reachable module split, and an exact checkpoint export. Validation is one paired LIBERO study over 2,000 episodes comparing a Qwen3-VL-2B arm with a Qwen3.5-0.8B arm, plus export accounting and a MiniCPM-V smoke test.

Critical strengths
  1. The export verification is concrete work. 1078 retained tensors are bitwise identical to the source, 321 teacher-only keys are removed, 718 alias keys are reproducible from retained ones, and the result loads through the production CPU loader. This is the level of artifact auditing that VLA papers usually skip.

  2. The paired design is the right one for a backbone swap. Both arms use the same 2000 episode keys, the McNemar test is reported with discordant counts, and the authors do not claim equivalence.

Critical weaknesses
  1. The one new claim is never tested. The paper says the module boundary is what makes backbone substitution work. There is no coupled LaWAM-style baseline at matched parameters, no ablation of the 8+8 query split, no inference without the predicted subgoal.

  2. The two headline numbers do not measure what the text says they measure. The 42.4% reduction mixes a 26.3% backbone reduction with stripping 411.1M of training-only state, and that stripping was never applied to the 2B arm. Separately, Table 1 reports RoboTwin 92.09 and 89.12 for this system, while Section 4.2, on the same page, says the RoboTwin columns are not filled because no such evaluation exists. For a paper whose selling point is auditability, that is the wrong kind of error.

Questions for authors

Q1. Please reconcile Table 1 and Section 4.2. What produced the RoboTwin entries for this system, and if they are inherited from LaWAM, will the row be corrected? On the same point, if the 2B checkpoint were exported the same way, what is the reduction.

Q2. What experiment isolates the modular boundary? A coupled baseline at matched parameters, a query-split ablation, and an inference pass with the subgoal removed would each help.

Q3. How many training seeds per arm, and what is the minimum detectable effect of the 2,000-episode design given that two suites sit at 99.2–99.4%?

Q4. Please provide the per-module parameter counts including DINOv3, the inference latency and peak memory for both arms, and one behavioural export test (same seed, identical action chunks before and after export). "Compact" needs at least one of these.

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

This paper presents AcrossWAM1.0, a modular implementation and compact-backbone study built on LaWAM’s latent world–action formulation. It defines explicit interfaces between the multimodal backbone, latent world decoder, and flow-matching action expert, while separating training-only components from the deployment checkpoint. On 2,000 paired LIBERO episodes, replacing Qwen3-VL-2B with Qwen3.5-0.8B achieves 97.45% success compared with 98.00%, meeting the authors’ predefined two-percentage-point retention criterion. The paper reports a 42.4% reduction in deployable parameters and validates checkpoint export through tensor-level equality checks and CPU loading. A MiniCPM-V forward/backward smoke test additionally demonstrates cross-family interface compatibility, although cross-family closed-loop performance remains untested. The contributions therefore center on modularization, compact-policy evaluation, and deployment auditing rather than a new latent world–action learning algorithm.

Critical strengths
  1. The paper addresses a useful deployment problem: reducing the size of latent world-action policies while retaining strong manipulation performance. The compact model achieves 97.45% success on LIBERO, close to the larger model’s 98.00%.

  2. Both policies are evaluated on the same 2,000 episodes with matched initializations. Reporting discordant outcomes and an exact paired statistical test makes the comparison more informative than aggregate success rates alone.

Critical weaknesses
  1. Confounded compact-backbone comparison. The main comparison uses Qwen3-VL-2B and Qwen3.5-0.8B, changing backbone generation and architecture alongside model size. Improvements in the newer backbone could compensate for losses from downsizing, but the current experiments do not distinguish these effects. A same-generation comparison, such as Qwen3.5-2B versus Qwen3.5-0.8B under matched training conditions, would provide a more controlled assessment. While the reported result demonstrates a useful compact configuration, this confounding limits the conclusions that can be drawn about performance retention under backbone downsizing.
Questions for authors
  1. Were any same-generation comparisons, such as Qwen3.5-2B versus Qwen3.5-0.8B, conducted but omitted from the paper? If available, could you report those results? Please also clarify whether the two reported configurations used matched training data, training budgets, and checkpoint-selection procedures. This would help assess how much of the observed performance retention can be attributed to downsizing versus differences between backbone generations.

  2. Table 2 identifies 2,555.2M as the 2B model’s deployable parameter count, whereas Table 3 labels the same number as its raw count, including training-only modules. Which interpretation is correct, and were both models processed using the same export and parameter-counting rules? This clarification is necessary to assess the reported 42.4% reduction.

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

Related papers

Discussion (0)

Sign in to comment.