acceptodds
Under review as a conference paper at ICLR 2027

GlanceWAM: Sparse Test-Time Imagination for World-Action Models

Abstract

Video generative models provide rich physical priors for robot learning, yet ex- isting world-action models (WAMs) face a fundamental trade-off: synchronous video generation at control rate is latency-prohibitive, while abandoning test- time visual imagination sacrifices task success. We show that visual imagina- tion achieves both real-time inference and superior success rates when generated asynchronously off the critical path and consumed directly in latent space. We introduce GlanceWAM, which decouples imagination from control on a single shared video DiT backbone: an asynchronous proposer glances ahead on a slow clock to imagine a single lookahead frame seconds into the future in the back- ground, while an action head decodes action chunks at control rate (48 ms) purely in latent space without blocking. Enabled by a non-interfering attention mask that isolates video representations and staleness-robust horizon training that accom- modates asynchronous lookahead aging, GlanceWAM breaks the speed–success dilemma. Trained purely on demonstrations, it attains 72.2% on the 24-task Robo- Casa kitchen benchmark (vs. 67.1% for synchronous Cosmos Policy) and 99.0% on LIBERO while cutting per-chunk control latency 24×relative to synchronous world-action models (48 ms on one A100). In single-arm and bimanual real-robot manipulation, it achieves higher average success than π0.5 without any robot-data pretraining.

Primary area:
applications to robotics, autonomy, planning
Keywords:
World-Action Models, Visual Foresight, Robot Manipulation, Diffusion Transformers, Imitation Learning, Flow Matching, Asynchronous Inference, Latent Dynamics, Visuomotor Control

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_VUjwOpenReview ↗
Rating
2
Confidence
2
Informativeness
2
Presentation
2
Soundness
2
Read review
Paper summary

This paper introduces GlanceWAM, a world-action model that separates a slow asynchronous visual-lookahead generator from a fast action-chunk policy. The method uses one shared latent video DiT, a three-class attention mask intended to prevent future-lookahead leakage, and random lookahead horizons with explicit offset conditioning to handle aging predictions. Experiments report strong results on RoboCasa, LIBERO, and five real-robot tasks, alongside low hold-phase action latency.

Critical strengths
  • The paper is well motivated, communicates the distinction between co-training with a clean, teacher-forced lookahead token and deployment with an asynchronously generated latent lookahead.
  • The paper makes a useful systems-level point that a world-model-derived visual target need not be recomputed on every control cycle.
Critical weaknesses
  • The main simulation experiments in Section 4.1 pause the simulator during lookahead generation, effectively assuming zero refresh delay. Although the authors demonstrate asynchronous real-robot deployment and report a pipelined experiment in Section 4.5, these results do not fully characterize the effects of delayed, missed, or stale lookahead updates. Since the action head explicitly conditions on the remaining temporal offset, delayed refreshes may introduce inconsistencies between the lookahead representation and its associated offset.
  • During training, the action head is conditioned on clean lookahead frames extracted from demonstration trajectories. At inference, however, it receives lookahead latents generated by the video model, which may contain inaccurate object positions, incorrect spatial relationships, or other prediction errors. This issue is particularly important because Table 3a demonstrates that the policy relies heavily on lookahead conditioning. The authors should investigate the relationship between lookahead prediction quality and downstream task success, and evaluate whether training with generated or deliberately perturbed lookaheads improves robustness.
Questions for authors
  • In the actual asynchronous implementation, what happens when the new lookahead is not ready at the scheduled refresh time? Does the system hold the previous lookahead, fall back to no-lookahead mode, or block?
  • How does performance change if the action head is trained with generated lookaheads, or a mixture of clean and generated lookaheads, rather than only teacher-forced demonstration frames?
  • Does the bidirectional-mask model permit clean lookahead tokens to influence future-target prediction during training? Can the authors include a leakage-free alternative mask ablation that isolates the effect of representation isolation from a severe train-test availability mismatch?
4 comments
AuthorsOpenReview ↗

Reply to Reviewer VUjw

We thank the reviewer for the thoughtful comments. We address each point below.

1) Delayed, missed or stale lookahead updates (W1, Q1)

Delayed and stale lookaheads. Please see item 1 of the global response. We now emulate asynchronous execution in simulation: each lookahead is generated from the observation 0.5 s before the refresh at which it is first used, and the action head is conditioned on the correspondingly shorter offset, so the lookahead and its offset stay consistent. Under this protocol success is unchanged (RoboCasa 71.6% vs. 71.7% with the simulator paused; LIBERO 98.7% vs. 98.8%), and Tables 1–2 now report these numbers. The lead-time study (Sec. 4.5, Tables 4–5) also covers longer delays of 0.75 s and 1 s.

Lookahead not ready at the refresh. In both the real-robot and the simulation setups, the machine runs only this one model, so the generation time varies very little. Generation takes 339 ms at 10 Euler steps on an RTX 4090 (Table 3b), and the 0.5 s lead time leaves enough buffer on top of it. A lookahead that misses its refresh therefore does not arise in our deployment, and we did not need a fallback for it.

2) Clean training lookaheads vs. generated lookaheads at test time (W2, Q2)

Lookahead quality vs. success. Sec. 4.5 varies lookahead quality through the number of sampler steps (Table 3b). From 1 to 30 Euler steps, image sharpness changes visibly (Figure 12), yet success stays within 1.9% (71.2% at 1 step, 71.7% at 10). The policy reads the low-frequency spatial layout of the lookahead, which the generated latent already gets right, so the gap between clean and generated lookaheads costs little.

Training with generated lookaheads. We ran this experiment (Table 3a, "Training-recipe variants" in Sec. 4.4). Mixing self-generated lookaheads into training is challenging: running the full 10-step denoising for each training sample would slow training down substantially. In the spirit of the chain-of-forward training of Epona [1], we instead generate the lookahead online with a single Euler step from the current model, and use it in place of the clean lookahead for 50% of the samples, after a 1,000-step warm-up with clean lookaheads only. This adds little training cost, and the sampler sweep above shows that a 1-step lookahead is already sufficient at test time. The retrained model reaches 68.7%, against 71.7% with clean lookaheads only. This is one recipe, and the rebuttal period did not allow us to explore others; we conclude that the idea is not ruled out, but it does not work out of the box.

[1] Zhang, Kaiwen, et al. "Epona: Autoregressive Diffusion World Model for Autonomous Driving." ICCV 2025.

3) Leakage-free alternative mask ablation (Q3)

Yes. In the bidirectional model the future-prediction queries can attend to the clean lookahead tokens during training, so that ablation mixes two effects: leakage into the video objective, and loss of the isolation between observation and lookahead.

We added the leakage-free variant you suggested (Table 3a, "Training-recipe variants" in Sec. 4.4). Future queries are still blocked from the lookahead, as in the full mask; only the observation queries are allowed to attend to it. This retrain reaches 55.6% on RoboCasa kitchen, against 71.7% for the model with the full mask and 47.0% for the bidirectional model. Removing the observation–lookahead isolation alone, with no leakage into future prediction, therefore already costs 16.1%, and the bidirectional model loses a further 8.6%.

We value your feedback and are eager to address any further questions or concerns you may have. If you have had a chance to review our response and have additional thoughts, we would greatly appreciate your input.

Reviewer_VUjwOpenReview ↗

Response for rebuttal

Thank you for the detailed response and the additional experiments. These additions substantially alleviate my main concerns.

The updated asynchronous evaluation and new ablations have alleviated my main concerns. Robustness to missed updates and inaccurate lookaheads remains an open question, which should be acknowledged as a limitation. I have no further questions.

Thank you for carefully reading our response and the new experiments. We are glad that they address your main concerns.

Since these concerns are now resolved, we would be very grateful if you would consider updating your score to reflect your current assessment. It would mean a great deal to us for this submission.

Following your suggestion, we have added both open questions, robustness to missed updates and to inaccurate lookaheads, to Sec. 5 (Limitations and Future Work) of the uploaded revision. Thank you again for the feedback, which has made the paper better.

Reviewer_qx18OpenReview ↗
Rating
3
Confidence
3
Informativeness
2
Presentation
3
Soundness
3
Read review
Paper summary

To address the speed–success dilemma in world-action models, this paper argue that this dilemmaarises from the architectural choice to synchronously couple imagination with control, rather than from visual lookahead itself. In this regard, the method GlanceWAM is poposed to decouple visual imagination from real-time control. GlanceWAM includes three key components: 1) Asynchronous future generation: A slow-clock asynchronous generator produces a single lookahead latent frame every (Hf≈3H_f \approx 3) seconds without decoding it into RGB. Meanwhile, a fast-clock action head decodes 16-step action chunks directly in latent space, with a latency of 48 ms per chunk. 2) Non-interfering attention masks: Three types of prefix-LM masks, together with dual-path VAE encoding, prevent the lookahead frame from contaminating the representations of observations and future predictions. 3) Training for robustness to stale lookahead: The lookahead offset (u∼U(0,Hf]u \sim U(0,H_f]) is randomized, and (\Delta) is explicitly provided as a conditioning variable, enabling the policy to tolerate aging lookahead information. Expeiments on RoboCasa, LIBERO and Real-world tasks demonstrate its superior performance while keeping a high controling rate.

Critical strengths
  1. The problem formulation is clear, and the paper’s position is supported by evidence. GlanceWAM clearly distinguishes itself from both opposing approaches—synchronous WAMs and imagination-free methods—and uses controlled internal experiments to show that test-time lookahead itself provides a +7.8 percentage-point gain (from 64.4% to 72.2%). This directly addresses Fast-WAM’s claim that imagination is unnecessary. This constitutes the paper’s substantive contribution to the debate in the field.

  2. Ablation study is well designed and validates the key components, which is also inspiring for WAM design. (1) Zeroing out the lookahead tokens reduces the success rate from 71.5% to 61.6%, which is lower than the no-lookahead baseline of 64.4%. This indicates that the policy has genuinely learned to rely on lookahead information. Fixing the offset (\Delta) likewise reduces performance to 61.8%, demonstrating that the action head indeed reads the temporal conditioning. (2) Removing the prefix-LM masks causes performance to collapse to 47.0% after retraining, directly confirming that the necessity of the masking scheme. (3) Lookahead-horizon sweep (Fig. 7) and sampling-step sweep provide a scientifically meaningful and generalizable finding: the policy reads only low-frequency spatial layout information.

Critical weaknesses
  1. The conceptual novelty of the individual components is limited: Asynchronous high-level/low-level decoupling already has precedents in hierarchical robot learning—for example, π\pi series models use a high-level video planner that refreshes every few seconds together with a (\pi_0) low-level policy, which the authors have cited. Offset conditioning is conceptually similar to AHA-WAM’s phase-offset training, while the prefix-LM causal masks share the same underlying idea as the isolation masks used in Fast-WAM and AHA-WAM.

  2. GlanceWAM generates lookahead predictions at a fixed, low frequency to break speed–success dilemma, which may limit its ability to handle highly dynamic scenarios. Performance in dynamic scenarios is not evaluated.

  3. Generalization of the proposed method is not evaluated on unseen scenarios such as with variation of background, lighting and object colors. Therefore, the impact of world-model prediction drift on action generation remains unclear. Authors are suggested to evaluated also on some benchmarks like LIBERO-Plus.

Questions for authors
  1. Under execution deviations or dynamic perturbations—for example, when an object’s position changes after a failed grasp—how does the currently held 3-second lookahead become invalid, and how is it refreshed or recovered?

  2. Without staleness-robust horizon randomization and keep it at HfH_f, what's the performance?

  3. What's the performance on LIBERO-Plus?

1 comment
AuthorsOpenReview ↗

Reply to Reviewer qx18

We thank the reviewer for the positive assessment and the helpful suggestions. We address each point below.

1) Novelty relative to the π-series, AHA-WAM and Fast-WAM (W1)

We agree that decoupling a slow branch from a fast one is not new in itself. We see our contribution as bringing this idea into a world-action model, which differs from each of the works you mention.

The π-series [1] pairs a video planner with a VLA policy. GlanceWAM instead runs both clocks on one video DiT, so the policy benefits from the video model in two ways: representation shaping during co-training, and an explicit lookahead that the action head conditions on at test time, in the manner of inverse dynamics. Realizing a slow/fast design inside a WAM raises different challenges from the VLA case. The π-series video planner is not publicly released; to our knowledge, GlanceWAM would be the first open-source work of this kind.

Fast-WAM [2] removes test-time imagination altogether, which is the choice our paper argues limits performance. AHA-WAM [3] does decouple action execution from video-token generation, but, like Fast-WAM, it does not use an imagined future at inference. On our stack, removing the lookahead lowers RoboCasa success from 71.7% to 64.4% (Table 3a).

[1] Physical Intelligence. "π0.7: a Steerable Generalist Robotic Foundation Model with Emergent Capabilities." arXiv:2604.15483, 2026.

[2] Yuan, Tianyuan, et al. "Fast-WAM: Do World Action Models Need Test-time Future Imagination?" arXiv:2603.16666, 2026.

[3] Cai, Jisong, et al. "AHA-WAM: Asynchronous Horizon-Adaptive World-Action Modeling with Observation-Guided Context Routing." arXiv:2606.09811, 2026.

2) Dynamic scenes and recovery after a failed grasp (W2, Q1)

A held lookahead is not refreshed by events: it stays in use until the next scheduled refresh, which regenerates it from the current observation. After a failed grasp, the lookahead can therefore be out of date for at most one refresh period (about 3 s), which can slow recovery. We do not evaluate dynamic scenes. We now state this as a limitation in the new Sec. 5, together with the extension we consider most natural: an adaptive horizon that is long during uneventful motion and short at critical moments.

3) Performance on LIBERO-Plus (W3, Q3)

We evaluated the released LIBERO checkpoint zero-shot on the full LIBERO-Plus benchmark (10,030 episodes). GlanceWAM reaches 70.4%, above plain co-training on the same data (67.9%) and above Fast-WAM (53.8%) and Fast-WAM-Joint (67.7%). The lookahead improves robustness to camera, robot, background and layout perturbations; it is hurt by the noise perturbation, which blurs the proposer's input, so we add a simple gate that drops the lookahead when the input frame is too blurred. Without the gate the total is 68.2%.

Table R3. LIBERO-Plus, success rate (%). Fast-WAM rows as reported by Zhou et al. (arXiv:2609.34981).

MethodCameraRobotLanguageLightBackgroundNoiseLayoutTotal
Fast-WAM21.049.573.377.552.046.763.753.8
Fast-WAM-Joint36.464.994.894.954.754.878.667.7
plain co-training (no lookahead)58.450.184.586.248.277.170.167.9
GlanceWAM61.363.484.485.752.058.372.768.2
GlanceWAM + sharpness gate61.263.284.284.952.773.372.170.4

We also found that LIBERO-Plus builds each instruction from the task filename, so outside the language dimension the instruction carries extra perturbation tokens. Stripping this suffix raises every model we tested by about 12%, and GlanceWAM stays ahead of plain co-training (Table R4).

Table R4. LIBERO-Plus with the instruction suffix stripped, success rate (%).

MethodCameraRobotLanguageLightBackgroundNoiseLayoutTotal
plain co-training (no lookahead)72.967.586.089.658.293.183.379.2
GlanceWAM74.181.684.189.463.477.686.079.8
GlanceWAM + sharpness gate74.181.684.189.463.490.086.081.7

Details are in the new Appendix F of the revised paper.

4) Training with the lookahead horizon fixed at H_f (Q2)

Retraining with the offset fixed at H_f, with no randomization, gives 64.2% on RoboCasa kitchen, against 71.7% for the full model (−7.5%). Without randomization the action head only ever sees a lookahead that is exactly H_f ahead, whereas at deployment a held lookahead ages from H_f towards zero between refreshes; randomizing the offset in training covers this whole range. The result is added to Table 3a.

Reviewer_2w8YOpenReview ↗
Rating
2
Confidence
2
Informativeness
2
Presentation
3
Soundness
2
Read review
Paper summary

GlanceWAM separates visual lookahead generation from action decoding. A shared video-DiT backbone generates a future latent frame approximately three seconds ahead, which is reused across action chunks. A separate flow-matching action head reads current-observation and lookahead features together with the remaining lookahead offset. Two-pass VAE encoding and a structured attention mask prevent future-information leakage, while randomized training offsets and lookahead dropout address aging or missing conditions. The paper reports 72.2% success on RoboCasa, 99.0% on LIBERO, and 79.2% average success across five real-robot tasks, together with a 48 ms optimized hold-phase action path on an A100.

Critical strengths
  1. The paper addresses a concrete deployment problem with a coherent architecture: expensive future generation need not run at every action query, provided the policy can interpret an aging visual target. The separation of generative targets from teacher-forced lookahead conditions is carefully implemented, including protection against leakage through the causal video VAE as well as transformer attention. The supplied training and deployment code makes these implementation choices substantially more inspectable than a high-level architectural description alone.

  2. The internal comparisons provide meaningful evidence that the lookahead interface is useful. On RoboCasa, the single-layer lookahead model improves over plain co-training from 64.4% to 71.5%, with additional runs near 71%; zeroing the lookahead or providing an incorrect time offset substantially reduces success. The horizon sweep also compares against longer-horizon co-training, rather than only a short-horizon baseline. Physical experiments cover both single-arm and bimanual tasks, report a failure case where the method loses to pi0.5, and use the same task demonstrations for the compared policies. These are substantive strengths supporting the usefulness of the method.

Critical weaknesses
  1. The principal success and latency results are not yet demonstrated together under the same sparse asynchronous schedule. Section 4.1 states that the simulator is paused while lookaheads are generated for the main benchmark results. The pipelined experiment in Section 4.5 instead refreshes every chunk, whereas the proposed economical schedule refreshes about every three seconds and reuses a lookahead across three or four chunks. Appendix B measures the 48 ms hold-phase call with the lookahead already available; it does not report the latency distribution while the proposer is active. This distinction matters on a shared GPU. The supplied goal_worker.py loads a separate policy copy in a worker process and yields between Euler steps, so queueing, scheduling, and memory overhead should be measured rather than assumed absent. The real-robot deployment supports feasibility, but the stronger claim that generation never enters or affects the control path needs an end-to-end measurement under the actual sparse schedule.

  2. The causal interpretation of the masking and staleness results is stronger than the controls establish. Blocking lookahead-to-observation attention ensures forward-pass isolation for fixed weights; it does not ensure that observation representations stay identical across separately trained models, because the lookahead-conditioned action loss updates the shared backbone. Likewise, zeroing a condition shows learned dependence, but does not by itself establish that the benefit specifically requires an accurate future rather than an extra goal-like feature stream. Finally, random offsets on ground-truth demonstration frames address temporal offset variation but do not directly reproduce an old generated lookahead that has become inconsistent with the policy's executed trajectory. The large gain over retrained plain co-training is persuasive evidence for the overall design; the narrower causal and staleness claims would benefit from targeted controls.

Questions for authors
  1. Please evaluate the final checkpoint with the claimed approximately three-second refresh interval, realistic generation delay, and uninterrupted environment execution. Report success, median and tail observation-to-action latency, deadline misses, proposer completion time, and peak GPU memory while both processes are active. How are startup and an expired lookahead handled?

  2. Was the -0.5-point pipelined result obtained with a real background worker or an emulated delay, and what delay distribution and refresh interval were used? Please give a paired confidence interval for the difference; p=0.72 alone does not establish equivalence. Does the same conclusion hold when refreshing only every three to four chunks?

  3. Can you separate offset randomization from offset conditioning in an ablation and compare predicted lookaheads with a temporally mismatched or current-observation condition under the same interface? This would help distinguish useful future information from condition dependence and show whether offset training handles trajectory divergence as well as aging. Please also qualify the fixed-weight scope of the representation-isolation argument in Section 3.2.

2 comments
AuthorsOpenReview ↗

Reply to Reviewer 2w8Y

We thank the reviewer for the detailed and constructive review. We address each point below.

1) Success under the asynchronous schedule (W1, Q1, Q2)

This is a key question for the paper. Following our real-robot stack, we now emulate asynchronous execution in simulation: each lookahead is generated from the observation 0.5 s before the refresh at which it is first used. Under this protocol success is unchanged (RoboCasa 71.6% vs. 71.7% with the simulator paused; LIBERO 98.7% vs. 98.8%), and Tables 1–2 now report these numbers. Generation takes 339 ms at 10 Euler steps on an RTX 4090 (Table 3b), within the 0.5 s lead time. The lead-time study replaces the pipelined-variant sentence in Sec. 4.5. Details are in the global response and in Sec. 4.5 (Figure 9, Tables 4–5) of the revised paper.

2) Scope of the representation-isolation argument in Sec. 3.2 (W2, Q3)

We agree that the original sentence overstated what the mask guarantees. We have rewritten it in Sec. 3.2. The revised text claims only that observation tokens do not attend to the lookahead, so the observation features are independent of the lookahead and are computed the same way in training and at deployment; the lookahead influences the policy only through the lookahead tokens read by the action head. We no longer state that the representations are identical to plain co-training or that all gains stem strictly from the lookahead channel.

3) Accurate future vs. an extra goal-like feature stream (W2, Q3)

We added the control you suggested to Table 3a: under the same interface, the lookahead is replaced by the current observation. Success falls to 61.3%, the same level as zeroing the lookahead (61.6%) and below the model trained without lookahead (64.4%). The temporally mismatched condition (correct lookahead, offset pinned) gives 61.8%. A condition that carries scene content but no future information therefore recovers none of the gain.

4) Offset randomization vs. offset conditioning (Q3)

We separated the two with two retrains (Table 3a, "Training-recipe variants" in Sec. 4.4; baseline 71.7%). Fixing the training offset at H_f, with no randomization, gives 64.2% (−7.5%). Keeping the randomization but removing the offset embedding from the action head gives 68.2% (−3.5%). Randomization accounts for most of the effect; explicit offset conditioning adds a smaller part.

We value your feedback and are eager to address any further questions or concerns you may have. If you have had a chance to review our response and have additional thoughts, we would greatly appreciate your input.

Comments on the paper

1 comment
AuthorsOpenReview ↗

Global response

We are grateful to all reviewers for their careful and constructive comments. We are glad that reviewers found: (1) the problem is a concrete deployment problem (Reviewer 2w8Y), clearly formulated (Reviewer qx18) and well motivated (Reviewer VUjw); (2) the architecture is coherent and carefully implemented (Reviewer 2w8Y), and makes a useful systems-level point that a world-model-derived visual target need not be recomputed on every control cycle (Reviewer VUjw); (3) the gain over retrained plain co-training is persuasive (Reviewer 2w8Y) and the ablation study is well designed and validates the key components (Reviewer qx18). Each comment from the reviewers has been replied to individually. Here we first respond to common questions.

1) Asynchronous execution: how it is run, and success under realistic lookahead delay (2w8Y, VUjw)

Reviewers 2w8Y and VUjw both ask how success holds up when lookaheads are generated asynchronously instead of with the simulator paused. We thank them for raising this; the submission did not give enough detail here. To settle it, we added a simulation protocol that reproduces the execution order of our real-robot stack.

On the robot, the lookahead latent is refreshed roughly every 3 s (every 4 action chunks on RoboCasa). About 0.5 s before each refresh, a separate process starts generating the next lookahead from the observation available at that moment, so at the refresh it is already there and the action path never waits. In simulation we emulate this by generating each lookahead from the observation recorded a fixed lead time before the refresh, and by conditioning the action head on the correspondingly shorter offset. The lead time is adjustable. We use 0.5 s because lookahead generation on an RTX 4090 completes well within 0.5 s, and because the study below shows that 0.5 s costs almost nothing.

We measured the sensitivity to this lead time on RoboCasa kitchen and LIBERO, averaging three runs per setting. A lead time of 0 s is the paused-simulator protocol of the submission; we additionally test 0.5 s, 0.75 s and 1 s. At 0.5 s the success rate is essentially unchanged (−0.1 pt on both benchmarks); a clear drop appears only from 0.75 s. Tables 1 and 2 now report the 0.5 s results, so the comparisons with baselines there are made under emulated asynchronous execution.

Table R1. RoboCasa kitchen, success rate (%).

Lead timeAvg
0 s71.7
0.5 s71.6
0.75 s69.8
1 s67.7

Table R2. LIBERO, success rate (%).

Lead timeSpatialObjectGoalLongAvg
0 s99.399.599.396.998.8
0.5 s99.599.798.996.598.7
0.75 s99.299.598.195.398.0
1 s99.399.793.793.396.5

2) Summary of revisions

  • Main results (Abstract, Sec. 1, Sec. 4.2, Sec. 6, Tables 1–2, Figs. 1 and 8): now reported under the asynchronous protocol with a 0.5 s lead time, mean of three runs.
  • Sec. 3.2: reworded the statement of what the attention mask guarantees.
  • Sec. 4.1: updated the description of the simulation protocol.
  • Sec. 4.4, Table 3a: added the new ablation rows and a paragraph, "Training-recipe variants".
  • Sec. 4.5: added the paragraph "Asynchronous execution", Figure 9 and Tables 4–5 (lead-time study); removed the pipelined-variant sentence; added lookahead generation time to Table 3b.
  • Sec. 5 (new): Limitations and Future Work.
  • Appendix F (new): LIBERO-Plus evaluation, Table 7.

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

Related papers

Loading the map…

Discussion (0)

Sign in to comment.