acceptodds
Under review as a conference paper at ICLR 2027

Is Test-Time Video Generation Unnecessary? Adaptive Video Reasoning for World Action Models

Abstract

Recent World Action Models (WAMs) remove the future video prediction at test time, which leads to lower inference cost. This raises a question: is test-time video generation unnecessary? We empirically find that removing it can easily degrade control performance, while its benefit varies substantially across decision states and most of the performance gains can be achieved with limited predictive computation. Therefore, we propose ThinkWAM, a world action model with adaptive predictive reasoning, equipped with a router that dynamically determines whether and how much future prediction to perform for each decision state. To support adaptive predictive reasoning, we formulate a multi-intensity reasoning training scheme that enables one WAM to support both direct action prediction and predictive reasoning at different intensities. We then train the router in two stages. First, it learns the benefit of predictive reasoning; second, it learns a cost-aware allocation over reasoning intensities by balancing predicted control quality and computation cost. At test time, the router determines whether predictive reasoning is needed and, when necessary, selects an appropriate reasoning intensity. ThinkWAM achieves success rates of 98.8%, 80.5%, and 77.7% on LIBERO, LIBERO-Plus, and RoboCasa GR1, respectively and 73.8% on real-world tasks. With predictive reasoning applied to only 28% of decision states, ThinkWAM recovers up to 68% of the success-rate gain achieved by reasoning at every state.

Primary area:
applications to robotics, autonomy, planning
Keywords:
World Action Models, Video Reasoning, Adaptive Computation, Robotic Manipulation

Reviews

Among:
Shape:
1234this paper, Top 15%
Figure 1. Mean review rating of the reviewed papers at ICLR 2027. This paper averages 3.00 over 4 reviews, above 85% of the others.
Reviewer_rSXtOpenReview ↗
Rating
4
Confidence
4
Informativeness
3
Presentation
3
Soundness
3
Read review
Paper summary

ThinkWAM studies when future prediction is useful at inference time in world action models and learns to allocate it across decision states. A shared video-action backbone is trained to support direct action prediction and predictive reasoning under different denoising schedules. With the backbone frozen, a router learns whether to enable reasoning from paired action-error comparisons, and selects its intensity using a quality predictor and a cost head optimized with GRPO. The predicted future latents serve as intermediate representations without pixel decoding. Experiments on LIBERO, LIBERO-Plus, RoboCasa GR1, and four real-world tasks examine both control performance and the tradeoff between reasoning frequency and success.

Critical strengths
  1. The paper offers a useful and well-supported insight into the role of test-time prediction. I find the reasoning-ratio experiments particularly compelling because they keep the backbone fixed and compare learned gating with random routing while retaining the same intensity-selection policy. The advantage at comparable reasoning frequencies provides evidence that choosing the right states matters. On RoboCasa GR1, reasoning at 28% of decisions raises success from 72.0% to 75.8%, recovering roughly two-thirds of the improvement available at the 77.7% all-state operating point. This is a concrete result that helps explain how future prediction can be used selectively in robot control. The contribution is therefore informative beyond the final benchmark scores.

  2. The method provides a coherent way to act on this observation. Training one backbone for both direct prediction and several video-action refinement relationships makes adaptive execution possible within a shared policy. Separating the decision to reason from the allocation of reasoning intensity is sensible, and reusing the existing state representation avoids requiring an additional large model for routing. The evaluation is also substantial: it includes distribution shifts on LIBERO-Plus, a different embodiment on RoboCasa GR1, and physical manipulation. The selectively routed model reaches 77.0% on LIBERO-Plus and 75.8% on RoboCasa GR1, while the all-state version reaches 80.5% and 77.7%. The reported real-world average of 73.8%, a 17.5-point improvement over Fast-WAM, adds practical support. I also appreciate the multi-intensity training ablation and the detailed description of the routing supervision in the appendix.

I support acceptance. The combination of a clear empirical finding, a method designed around that finding, and evidence across several evaluation settings makes this a valuable contribution to adaptive computation for robot policies.

Critical weaknesses

The main remaining gap is the measurement of inference efficiency. The paper reports reasoning frequencies and FLOPs per forward, but total cost also depends on the selected action depth A and the router overhead. Consequently, a comparison at matched reasoning frequency is not necessarily a comparison at matched total computation. End-to-end latency and average compute per control decision would make the practical efficiency benefit easier to assess, particularly for the budget-factor ablation. This limits the precision of the efficiency claims, but does not undermine the central evidence that selective future prediction can improve control performance.

Questions for authors
  1. Could you report end-to-end latency, including state encoding and routing, together with the average selected action depth and total FLOPs per decision for the direct, adaptive, and all-state settings? The same measurements for the fixed-intensity ablation would help quantify the contribution of the budget factor.

  2. How were the offline routing states divided into training and validation sets, and how was the gating threshold chosen for the reported Adaptive operating points? In particular, please clarify whether the threshold was fixed using training/validation data or selected using evaluation rollouts. This would help interpret how the routing policy transfers to new states and distribution shifts.

5 comments
AuthorsOpenReview ↗

Response to Reviewer rSXt (1/3)

We greatly appreciate Reviewer rSXt’s constructive comments and insightful feedback. The suggestions have prompted us to carefully consider each point in our method, and we appreciate the opportunity to strengthen the paper accordingly. We provide detailed responses below.

1. Ablation Study on Efficiency (Weakness 1 and Question 1)

We thank the reviewer for this helpful suggestion. Tables R1–R3 report inference latency and TFLOPs per model inference and per rollout, including state encoding and routing. Adaptive reasoning improves success over Direct while reducing computation relative to All-State. For example, on RoboCasa GR1, adaptive reasoning at 28.0% of decisions achieves 75.8% success, compared with 72.0% for no reasoning (Direct) and 77.7% for 100% reasoning (All-State), while reducing rollout inference time by 38.9% and computation by 54.9%, relative to All-State. The mean selected action depth, is 4.01/4.31 on LIBERO, and 4.12/4.35 on RoboCasa GR1 for all-state/adaptive, respectively. These results show how adaptive predictive reasoning improves task success while reducing the inference time and computation required over complete rollouts. All inference evaluations are conducted on NVIDIA H20 GPU.

As defined in Sec. 3.2.3 in the main paper, the router selects reasoning intensity (A,V)(A,V), where AA is the number of joint video-action updates and VV controls the refinement rate of future-video latents. Inference terminates after AA joint updates. To analyse the quality–cost trade-off, Tables R4-R7 report results with reasoning enabled at every decision, a fixed reasoning intensity (A,V)(A,V), and the router disabled. These fix ed-intensity measurements exclude routing overhead. They include success rates, average inference latency, and computational cost per model inference. At V=25V=25, reducing AA from 5 to 2 reduces TFLOPs by 60.0% and latency by 54.8% on both LIBERO and LIBERO-Plus. Success decreases by only 0.5 and 1.1 percentage points, respectively. At A=5A=5, the V=25V=25 and V=15V=15 schedules use the same five joint updates but expose the action branch to differently refined future latents. On LIBERO-Plus, V=15V=15 achieves 81.3% success, compared with 79.8% for V=25V=25, with identical computation of 14.51 TFLOPs and similar latency. Together, these results show that AA controls computational cost, while VV affects control performance within that budget, motivating our budget factor's selection: choosing the proper reasoning intensity (A,V) based on the quality-cost trade-off.

We will update the relevant sections of our manuscript to include these comprehensive results and detailed discussions.

Table R1. Rollout-level inference efficiency on LIBERO across reasoning ratios. Latency (ms) and TFLOPs are averages per model inference. Per-rollout metrics accumulate all model inference calls within a rollout. 0% ratio denotes action-only inference; 100% denotes all-state reasoning with adaptive intensity. Intermediate ratios use gated reasoning.

Reasoning ratio (%)Success (%)LatencyTFLOPsInference time/rollout (s)TFLOPs/rollout
100.098.8163.0611.572.48175.9
85.098.7148.1510.212.25155.0
76.798.7139.949.472.11143.1
20.098.277.913.781.2058.4
13.297.869.683.031.0846.8
4.397.559.212.040.9231.7
0.097.353.921.550.8424.3

Table R2. Rollout-level inference efficiency on LIBERO-Plus across reasoning ratios. Latency (ms) and TFLOPs are averages per model inference. Per-rollout metrics accumulate all model inference calls within a rollout. 0% ratio denotes action-only inference; 100% denotes all-state reasoning with adaptive intensity. Intermediate ratios use gated reasoning.

Reasoning ratio (%)Success (%)LatencyTFLOPsInference time/rollout (s)TFLOPs/rollout
100.080.5161.1410.893.59242.8
87.179.9146.849.653.29216.0
56.178.8113.946.732.62154.8
17.777.080.943.771.9189.3
0.074.456.141.551.3738.0

Table R3. Rollout-level inference efficiency on RoboCasa GR1 across reasoning ratios. Latency (ms) and TFLOPs are averages per model inference. Per-rollout metrics accumulate all model inference calls within a rollout. 0% ratio denotes action-only inference; 100% denotes all-state reasoning with adaptive intensity. Intermediate ratios use gated reasoning.

Reasoning ratio (%)Success (%)LatencyTFLOPsInference time/rollout (s)TFLOPs/rollout
100.077.781.925.511.67112.2
91.277.575.514.921.61105.0
87.677.060.633.551.3277.2
28.075.847.922.381.0250.6
7.373.036.611.340.7728.0
0.072.032.370.970.6720.2

Due to the 5,000-character limit per reply, Tables R4–R7 are provided in Response to Reviewer rSXt (2/3).

AuthorsOpenReview ↗

Response to Reviewer rSXt (2/3)

Continuing our response to Question 1 in Response to Reviewer rSXt (1/3), we provide Tables R4–R7 below.

Table R4. Effect of reasoning depth AA on LIBERO with V=25V=25. Fixed intensity at every decision; router disabled. Success is in %; latency (ms) and backbone TFLOPs are measured per model inference.

AASpatialObjectGoalLongAvg.LatencyTFLOPs
298.699.897.696.698.271.35.80
399.299.897.296.898.3100.18.70
498.899.898.097.098.4128.911.60
599.299.897.897.898.7157.614.51

Table R5. Effect of reasoning depth AA on LIBERO-Plus with V=25V=25. Fixed intensity at every decision; router disabled. Success is in %; latency (ms) and backbone TFLOPs are measured per model inference. Columns Cam., Robot, Lang., Light, Bg., Noise, and Layout denote camera, robot, language, lighting, background, sensor-noise, and layout shifts, respectively.

AACam.RobotLang.LightBg.NoiseLayoutTotalLatencyTFLOPs
257.079.486.695.262.886.583.478.771.25.80
358.981.286.096.664.584.883.379.2100.08.70
459.881.087.096.963.886.583.579.8128.811.60
560.880.587.896.562.687.183.179.8157.614.51

Table R6. Effect of video-schedule parameter VV on LIBERO with A=5A=5. Each inference executes only the first five steps of the VV-step video-denoising schedule (e.g., 5/255/25 for V=25V=25 and 5/155/15 for V=15V=15). Fixed intensity at every decision; router disabled. Success is in %; latency (ms) and backbone TFLOPs are measured per model inference.

VVSpatialObjectGoalLongAvg.LatencyTFLOPs
599.2100.097.495.898.1157.614.51
1099.4100.098.097.498.7157.614.51
1599.699.897.497.498.6157.614.51
2099.499.898.097.098.6157.614.51
2599.299.897.897.898.7157.614.51

Table R7. Effect of video-schedule parameter VV on LIBERO-Plus with A=5A=5. Each inference executes only the first five steps of the VV-step video-denoising schedule (e.g., 5/255/25 for V=25V=25 and 5/155/15 for V=15V=15). Fixed intensity at every decision; router disabled. Success is in %; latency (ms) and backbone TFLOPs are measured per model inference.

VVCam.RobotLang.LightBg.NoiseLayoutTotalLatencyTFLOPs
563.081.888.497.562.787.083.680.7157.614.51
1061.781.688.097.464.187.683.980.7157.614.51
1563.980.790.198.365.087.883.581.3157.614.51
2060.481.887.497.363.886.583.380.0157.614.51
2560.880.587.896.562.687.183.179.8157.614.51
AuthorsOpenReview ↗

Response to Reviewer rSXt (3/3)

2. Clarification on Data Split and Threshold Selection

We thank the reviewer for this helpful question. We split the offline routing states into 90% training and 10% validation data using stratified sampling by category and task, ensuring that each category and task is represented in both subsets. For each reported Adaptive operating point, the gating threshold was selected exclusively using the validation set and fixed before evaluation. No evaluation rollouts or evaluation success rates were used for threshold selection. We will update the relevant sections of our manuscript to include these detailed discussions.

Thank you sincerely again for your meticulous review and for helping us improve the clarity of our work. We will update the manuscript to include these expanded experiments and discussion.

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

This paper investigates the necessity of future video prediction at test time in World Action Models (WAMs) and the trade-off between control performance and inference computation. To address fixed strategies that either always enable or entirely remove future prediction, the authors propose ThinkWAM, a WAM with adaptive predictive reasoning. Through multi-intensity reasoning training, a single WAM supports both direct action prediction and reasoning over future visual latents at different intensities. The authors also introduce a two-stage router: it first learns whether future prediction is needed in the current state from differences in action prediction error with and without predictive reasoning, and then uses GRPO to dynamically select the reasoning intensity by balancing predicted quality and computational cost. Experiments are conducted on LIBERO, LIBERO-Plus, RoboCasa GR1, and real-world robotic manipulation tasks. The results indicate that the benefits of future prediction are state-dependent. On RoboCasa GR1, enabling predictive reasoning for approximately 28% of decision states recovers approximately 68% of the success-rate gain achieved by reasoning at every state, demonstrating the potential of selective prediction to balance performance and computational efficiency.

Critical strengths

S1. Clear motivation and practical relevance. The paper addresses the fixed inference strategies in existing WAMs that either always enable or entirely remove future prediction, and proposes dynamically allocating predictive computation according to the decision state. This is a practically relevant research question and offers a valuable direction for balancing robot control performance and inference efficiency. S2. Clear paper structure and intuitive figures. The paper is well organized, with a coherent progression from the research motivation to the method design and experimental evaluation. In particular, Figures 1 and 2 clearly illustrate the overall idea of adaptive predictive reasoning and the model architecture, helping readers understand the proposed approach. S3. A coherent and well-motivated method design. The proposed multi-intensity reasoning training uses synchronous and asynchronous noise schedules to enable a single WAM to accommodate different intensities of future prediction. Together with the two-stage routing mechanism, it enables dynamic decisions about whether to reason and how much computation to allocate, resulting in a coherent overall approach.

Critical weaknesses

W1. The novelty relative to prior work needs to be articulated more clearly. Dynamic reasoning routing and adaptive allocation of test-time computation have been explored to varying degrees in related work such as Fast-WAM and AdaWAM. The main contribution of this paper is to combine multi-intensity reasoning over future latents with state-level budget allocation in a unified WAM framework. However, the overall approach remains somewhat incremental. I recommend clarifying the technical differences from prior methods and using more targeted experiments to demonstrate its distinctive advantages. W2. Insufficient evaluation of end-to-end inference efficiency. The paper primarily reports the computational cost of a single forward pass under different reasoning modes and task success rates as a function of the reasoning ratio. However, it does not provide key metrics such as average FLOPs per decision and actual inference latency. Since different reasoning intensities incur different computational costs, the reasoning ratio alone does not accurately quantify the practical efficiency gains. I recommend reporting trade-off curves between success rate and computational cost, as well as between success rate and inference latency, and conducting comparisons under matched computational budgets. W3. The independent contribution of future prediction to performance gains remains unclear. Enabling future video prediction also introduces additional model computation. It is therefore difficult to determine whether the gains arise from meaningful predictions of future states or from additional computation and feature interactions. I recommend compute-matched ablations, such as comparisons with an action-only baseline using more denoising steps, to more clearly establish the contribution of future prediction itself. W4. The necessity of GRPO for reasoning budget allocation is not sufficiently established. The authors already obtain a quality predictor for different reasoning intensities through offline supervised training, but do not compare against a simple selection strategy based on predicted quality and a fixed computational cost penalty. Moreover, removing the budget factor in Table 4 reduces the LIBERO-Plus success rate by only 0.6 percentage points, without a corresponding analysis of computational cost. Comparisons with budget allocation methods that do not use reinforcement learning would help clarify the actual benefits of GRPO in terms of performance or computational efficiency. W5. There is a gap between the router's supervision signal and actual control performance. The gating factor learns reasoning benefits by comparing action prediction errors with and without predictive reasoning. However, a reduction in action prediction error does not necessarily translate into an improvement in task success, particularly when multiple actions are valid. The paper does not sufficiently analyze the relationship between predicted reasoning benefits and gains in closed-loop control. I recommend a correlation analysis and an evaluation of the reliability of routing decisions under distribution shifts. W6. Inconsistent mathematical notation reduces the readability of methodological details. The paper introduces numerous mathematical symbols, some of which are reused across equations to represent different concepts. For example, (c_{ij}) in Eq. (6) denotes the ranking weight for a sample pair, (c_A(x)) in Eq. (10) denotes the state-dependent cumulative computational cost, and (c_r) in Eq. (13) denotes a fixed computational penalty coefficient. Although the subscripts distinguish these quantities, reusing the same base symbol still increases the effort required to follow the method. I recommend adopting a more consistent naming convention and providing additional explanations of the notation where appropriate.

Overall, this paper addresses a practically relevant inference efficiency problem in WAMs and reports promising results across multiple simulated and real-world robotic tasks using an adaptive predictive reasoning framework. However, end-to-end inference efficiency is not sufficiently quantified, and rigorous compute-matched experiments are missing to establish the independent contribution of future prediction. The necessity of GRPO-based budget allocation and the novelty relative to prior work also require further justification. I therefore currently lean toward Weak Reject, but would be willing to reconsider my recommendation if the authors adequately address these central concerns.

Questions for authors

See the weakness.

5 comments
AuthorsOpenReview ↗

Response to Reviewer QycY (1/4)

We greatly appreciate Reviewer QycY’s constructive comments and insightful feedback. The suggestions have prompted us to carefully consider each point in our method, and we appreciate the opportunity to strengthen the paper accordingly. We provide detailed responses below.

1. Clarification on Our Work (W1)

Thank you for the suggestion. Fast-WAM removes future prediction uniformly at inference, whereas ThinkWAM retains it selectively according to the current state. AdaWAM selects between reasoning modalities, whereas ThinkWAM controls whether and how much future-latent refinement to perform within a shared WAM backbone. The key insight motivating ThinkWAM is that test-time future prediction in WAMs is neither uniformly necessary nor uniformly removable, and its benefit varies across decision states and reasoning intensities. Building on this observation, we introduce adaptive predictive reasoning within the WAM itself, enabling a single WAM to support different levels of future-latent refinement through multi-intensity reasoning training. ThinkWAM then jointly adapts whether and how much predictive reasoning to perform according to the current decision state, allowing predictive computation to be allocated where it is most beneficial. Experiments on LIBERO, LIBERO-Plus, RoboCasa GR1, and four real-world tasks examine both control performance and the tradeoff between reasoning frequency and success.

2. Ablation Study on Efficiency (W2)

We thank the reviewer for this helpful suggestion. We conducted the experiments on NVIDIA H20 GPU, as shown in Tables Q1-Q3, which report both average cost per model inference and per rollout. These results provide a more complete characterization of the performance–computation trade-off under different reasoning budgets. The results show that ThinkWAM preserves substantial performance gains while significantly reducing inference computation. For example, RoboCasa GR1 at 28.0% reasoning achieves 75.8% success, recovering 66.7% of the All-State success-rate gain over direct prediction while reducing rollout inference time by 38.9% and computation by 54.9%. Similar trends are observed on LIBERO and LIBERO-Plus, where adaptive reasoning achieves favorable trade-offs between task success, computational cost, and inference latency.

We will update the relevant sections of our manuscript to include these comprehensive results and detailed discussions.

Table Q1. Rollout-level inference efficiency on LIBERO across reasoning ratios. Latency (ms) and TFLOPs are averages per model inference. Per-rollout metrics accumulate all model inference calls within a rollout. 0% ratio denotes action-only inference; 100% denotes all-state reasoning with adaptive intensity. Intermediate ratios use gated reasoning.

Reasoning ratio (%)Success (%)LatencyTFLOPsInference time/rollout (s)TFLOPs/rollout
100.098.8163.0611.572.48175.9
85.098.7148.1510.212.25155.0
76.798.7139.949.472.11143.1
20.098.277.913.781.2058.4
13.297.869.683.031.0846.8
4.397.559.212.040.9231.7
0.097.353.921.550.8424.3

Table Q2. Rollout-level inference efficiency on LIBERO-Plus across reasoning ratios. Latency (ms) and TFLOPs are averages per model inference. Per-rollout metrics accumulate all model inference calls within a rollout. 0% ratio denotes action-only inference; 100% denotes all-state reasoning with adaptive intensity. Intermediate ratios use gated reasoning.

Reasoning ratio (%)Success (%)LatencyTFLOPsInference time/rollout (s)TFLOPs/rollout
100.080.5161.1410.893.59242.8
87.179.9146.849.653.29216.0
56.178.8113.946.732.62154.8
17.777.080.943.771.9189.3
0.074.456.141.551.3738.0

Table Q3. Rollout-level inference efficiency on RoboCasa GR1 across reasoning ratios. Latency (ms) and TFLOPs are averages per model inference. Per-rollout metrics accumulate all model inference calls within a rollout. 0% ratio denotes action-only inference; 100% denotes all-state reasoning with adaptive intensity. Intermediate ratios use gated reasoning.

Reasoning ratio (%)Success (%)LatencyTFLOPsInference time/rollout (s)TFLOPs/rollout
100.077.781.925.511.67112.2
91.277.575.514.921.61105.0
87.677.060.633.551.3277.2
28.075.847.922.381.0250.6
7.373.036.611.340.7728.0
0.072.032.370.970.6720.2
AuthorsOpenReview ↗

Response to Reviewer QycY (2/4)

3. Compute-Matched Ablation Study (W3)

Thank you for the suggestion. To further demonstrate the contribution of future prediction, we conduct compute-matched comparisons on NVIDIA H20 GPU with action-only baselines on LIBERO-Plus and RoboCasa GR1. Specifically, we increase the number of action-only denoising steps to match the inference TFLOPs of ThinkWAM-Adaptive and ThinkWAM-All-State, while keeping future prediction disabled. All action-only baselines use the same frozen ThinkWAM backbone and only the number of action denoising steps changes. The results are presented in Tables Q4 and Q5.

On LIBERO-Plus, ThinkWAM-Adaptive (17.7% reasoning) achieves 77.0% success, outperforming the compute-matched 26-step action-only baseline (75.5%) by 1.5 percentage points. At a higher computational budget, ThinkWAM-All-State achieves 80.5%, surpassing the 93-step action-only baseline (73.6%) by 6.9 points. Similarly, on RoboCasa GR1, ThinkWAM-Adaptive (28.0% reasoning) and ThinkWAM-All-State outperform their compute-matched action-only baselines by 2.6 and 1.4 percentage points, respectively. These consistent improvements across matched computational budgets highlight the effectiveness of adaptive multi-intensity predictive reasoning in improving action generation, rather than merely increasing denoising computation.

We will update the relevant sections of our manuscript to include these comprehensive results and detailed discussions.

Table Q4. Compute-matched comparison on LIBERO-Plus. TFLOPs are measured per model inference. Success rates (%) are reported overall and under seven distribution shifts. ThinkWAM-Adaptive is enabled at 17.7% reasoning ratio; ThinkWAM-All-State reasoning is enabled at 100%. Within each budget group, the action-only baseline approximately matches ThinkWAM in TFLOPs.

TFLOPsSettingOverallCam.RobotLang.LightBg.NoiseLayout
1.55Action-only (5 steps)74.453.172.882.696.158.680.379.0
3.77ThinkWAM-Adaptive77.058.276.585.496.360.981.081.4
3.78Action-only (26 steps)75.553.474.283.594.661.382.880.3
10.89ThinkWAM-All-State80.563.081.288.095.564.287.883.3
10.88Action-only (93 steps)73.650.371.682.493.459.580.179.7

Table Q5. Compute-matched comparison on RoboCasa GR1. TFLOPs are measured per model inference. Success rate is reported in percent. ThinkWAM-Adaptive is enabled at 28.0% reasoning ratio; ThinkWAM-All-State is enabled at 100% reasoning. Within each separated budget group, the action-only baseline approximately matches ThinkWAM in TFLOPs.

TFLOPsMethodSuccess rate(%)
0.97Action-only (5 steps)72.0
2.38ThinkWAM-Adaptive75.8
2.36Action-only (23 steps)73.2
5.51ThinkWAM-All-State77.7
5.52Action-only (64 steps)76.3
AuthorsOpenReview ↗

Response to Reviewer QycY (3/4)

4. Ablation Study on GRPO (W4)

We thank the reviewer for this helpful suggestion. To further evaluate the contribution of GRPO to reasoning budget allocation, we conduct two additional ablations on LIBERO-Plus, as shown in Table Q6. It compares two variants of budget policies on LIBERO-Plus. The two variants and ThinkWAM enable future prediction at every decision (All-State) and differ only in budget selection:

  • SFT-only budget: the ThinkWAM variant that trains only the quality predictor without GRPO. At inference, it fixes A=5A=5 and selects VV based on predicted quality, without cost-aware depth allocation.
  • Fixed-cost argmax: the ThinkWAM variant with directly selecting the maximum configuration of predicted quality minus fixed computing cost from Sec. 3.4.3, without GRPO.

The ThinkWAM achieves 80.5% success, compared with 79.8% for SFT-only and 80.0% for fixed-cost argmax. It also reduces computation from 14.51 to 10.89 TFLOPs per model inference relative to SFT-only (24.9%), and from 11.60 to 10.89 relative to fixed-cost argmax (6.1%). These results demonstrate that GRPO learns a more effective state-dependent reasoning budget allocation policy, achieving a better quality–cost trade-off than both supervised allocation and direct quality-cost selection.

We will update the relevant sections of our manuscript to include these comprehensive results and detailed discussions.

Table Q6. GRPO ablation on LIBERO-Plus. Success columns report percentages. TFLOPs measure average computation per model inference. All variants enable future prediction at every decision (All-State); only budget selection changes.

Cam.RobotLang.LightBg.NoiseLayoutTotalTFLOPs
SFT-only budget60.281.387.696.463.186.783.179.814.51
Fixed-cost argmax59.480.388.197.165.087.483.380.011.60
ThinkWAM63.081.288.095.564.287.883.380.510.89
Reviewer_mmE9OpenReview ↗
Rating
3
Confidence
3
Informativeness
2
Presentation
3
Soundness
3
Read review
Paper summary

This paper proposes ThinkWAM, an adaptive world action model for robot control, which chooses whether to predict the future and what kind of prediction intensity to use according to the current decision-making state to balance task performance and reasoning expenses. Through video and action dual branching and multi-intensity training, this method makes the same model support direct action generation and action generation using the potential representation of future video at the same time; then a two-level router is introduced, and Gate judges whether future prediction is useful, and then Budget calculates according to the predicted quality and calculation. Select the specific configuration and use the GRPO training budget allocation strategy.

Critical strengths
  1. This paper focuses on the difference of demand for future prediction in different decision states, and judges whether it needs to predict through Gate, and then chooses the prediction intensity by Budget. The same model can support direct action generation and future prediction with different strength, which can save inference overhead.

  2. The evaluation was carried out on LIBERO, Libero-Plus, RoboCasa GR1 and real robot tasks, and controls such as direct prediction, full-state prediction and random gating were set up. These experiments provide support for the control effect of the whole method, and the real robot results also enhance the application of the method.

Critical weaknesses
  1. In Section 3.3, the author trains Gate with the difference between the action error of the direct prediction and the reasoning route, but it is not clear which configuration is used for the reasoning route. Therefore, Gate's judgment of the beneficial plan is not necessarily the final implementation plan. The author needs to clarify the reference configuration used by the label, and whether the final selected configuration is still better than the direct prediction.

  2. Gate's revenue label and Budget's quality label come from errors between action and demonstration. However, there may be multiple effective actions for the same task, and deviation from a certain demonstration does not necessarily mean that the control effect is worse. Although the paper reports the overall success rate, it lacks direct verification in the same state.

  3. In Table 4, the author compares the complete Budget with the fixed use ((5,25)) scheme. This experiment can reflect the overall role of the adaptive budget, but the promotion cannot be attributed to GRPO alone. According to the design of Section 3.4.3, after the quality predictor is frozen, GRPO actually only adjusts the selection probability between five candidates, and the candidate reward is already available at the time of training.

  4. In Table 4, after removing multi-intensity training, the backbone model has changed. At this time, do you regenerate the revenue and quality labels and retrain the router?

Questions for authors
  1. Clarify the reasoning configuration used to generate the Gate income label in Section 3.3, and report the budget's final selected configuration relative to the directly predicted actual income to test the consistency of Gate's supervision objectives with the final implementation decision.

  2. Supplement the same initial state, directly predict the control results of the selected configuration of Budget, and analyze the correspondence between the decrease of action error and the real control yield. At the same time, please explain the limitations of using a single demonstration action as the basis for evaluation in the case of multiple effective actions.

  3. Section 3.4.3, after the mass predictor freezes, GRPO adjusts the selection probability between the five action denoising depths, and the offline rewards of each candidate are available at the time of training. Add a simple baseline comparison with directly selecting the maximum configuration of "predicted quality minus fixed computing cost" to evaluate the additional contribution of GRPO separately.

  4. Supplement the complete training process after multi-intensity training in Table 4, clarify the generation method of income and quality labels after the backbone change, and the supporting training settings of Gate and Budget, so as to exclude the impact of the mismatch between the backbone and the router on the ablation results.

5 comments
AuthorsOpenReview ↗

Response to Reviewer mmE9 (1/2)

We greatly appreciate Reviewer mmE9’s constructive comments and insightful feedback. The suggestions have prompted us to carefully consider each point in our method, and we appreciate the opportunity to strengthen the paper accordingly. We provide detailed responses below.

1. Analysis on Gate Supervision and Executed Reasoning (Weakness 1 and Question 1)

We thank the reviewer for requesting this clarification. Gate supervision is based on the action-MSE reduction from five-step Direct inference to fixed reasoning at (A,V)=(5,25)(A,V)=(5,25), the reference configuration used to generate Gate label. And Budget-selected refers to reasoning intensity chosen by our trained budget factor for execution. We compare the gate label and final budget-selected configurations on states for reasoning from the validation set (10% of the data, following a 9:1 training–validation split), averaging over five sampling seeds. As shown in Tables M1–M2, budget-selected achieves mean gains of 2.34×10−32.34\times10^{-3} on LIBERO and 2.52×10−32.52\times10^{-3} on RoboCasa GR1, exceeding the Gate-label gains of 1.96×10−31.96\times10^{-3} and 2.24×10−32.24\times10^{-3}, respectively. Its mean MSE is also lower than Gate label on both benchmarks. Thus, the executed configurations retain and further improve the mean benefit targeted by Gate supervision.

Table M1. Comparison of Direct, Gate label, and Budget-selected on LIBERO validation split, averaged over five seeds. Direct uses five action-only denoising steps; Gate label uses fixed reasoning intensity (5,25)(5,25) for Gate supervision; Budget-selected uses the intensity selected by the trained budget factor for execution. Gain is the MSE reduction relative to Direct.

Mean action MSE ↓\downarrowMean gain over Direct ↑\uparrow
Direct5.405e−30.00e−3
Gate label3.446e−3+1.96e−3
Budget-selected3.061e−3+2.34e−3

Table M2. Comparison of Direct, Gate label, and Budget-selected on RoboCasa GR1 validation split, averaged over five seeds. Direct uses five action-only denoising steps; Gate label uses fixed reasoning intensity (5,25)(5,25) for Gate supervision; Budget-selected uses the intensity selected for execution. Gain is the MSE reduction relative to Direct.

Mean action MSE ↓\downarrowMean gain over Direct ↑\uparrow
Direct1.942e−20.00e−3
Gate label1.718e−2+2.24e−3
Budget-selected1.690e−2+2.52e−3

We will update the manuscript to include these expanded experiments and discussion.

2. Analysis of the Action Error and the Real Control Yield (Weakness 2 and Question 2)

We thank the reviewer for this helpful suggestion. We replayed the training demonstrations in simulation and conducted paired rollouts from identical states, sharing initial noise within each pair and varying it across states. Only the first decision differs: Direct uses no reasoning, whereas Budget-selected uses the router-selected reasoning intensity; both subsequently use Direct until termination. In Table M3, Budget-selected reduces predicted-action and executed-action MSE against demonstration actions by 3.4% and 9.8%, respectively, and raises success from 98.1% to 98.5%, with gains across all four suites. Thus, lower average action error accompanies improved control outcomes when the initial decision is isolated, supporting the practical utility of our error-based supervision.

Table M3. Same-state rollout evaluation on replayed LIBERO training states. Both MSE metrics use ground-truth demonstration actions; success rates are reported in percent.

Inference Action MSE ↓\downarrowExecuted Action MSE ↓\downarrowSpatialObjectGoalLongAvg.
Direct9.25e−38.25e−398.299.698.695.898.1
Budget-selected8.94e−37.44e−398.899.898.896.698.5

We thank the reviewer and agree that demonstration-based error has two limitations: it can penalize alternative valid actions, and fail to capture subsequent closed-loop outcomes. Successful demonstrations provide a reference for comparing reasoning benefits and allocating computation, enabling scalable offline supervision without exhaustive rollouts. We use demonstration error as a proxy, not an absolute criterion for action validity. Its utility is supported by learned routing outperforming random routing at matched ratios (Fig.3 in main paper), positive prediction gains on the validation set (Tables M1–M2), and improved paired-rollout success in replayed simulations (Table M3).

We will update the manuscript to include these expanded experiments and discussion.

AuthorsOpenReview ↗

Response to Reviewer mmE9 (2/2)

3. Ablation Study on GRPO (Weakness 3 and Question 3)

We thank the reviewer for this helpful suggestion. We conduct additional experiments on an ablation study of GRPO, as shown in Table M4. It compares two variants of budget policies on LIBERO-Plus. The two variants and ThinkWAM enable future prediction at every decision (All-State) and differ only in budget selection:

  • SFT-only budget: the ThinkWAM variant that trains only the quality predictor without GRPO. At inference, it fixes A=5A=5 and selects VV based on predicted quality, without cost-aware depth allocation.
  • Fixed-cost argmax: the ThinkWAM variant with directly selecting the maximum configuration of predicted quality minus fixed computing cost from Sec. 3.4.3, without GRPO.

The ThinkWAM achieves 80.5% success, compared with 79.8% for SFT-only and 80.0% for fixed-cost argmax. It also reduces computation from 14.51 to 10.89 TFLOPs per model inference relative to SFT-only (24.9%), and from 11.60 to 10.89 relative to fixed-cost argmax (6.1%). With the backbone and quality predictor frozen during GRPO, these results support the additional benefit of post-training the budget allocation policy beyond SFT and direct score maximization.

Table M4. GRPO ablation on LIBERO-Plus. Success columns report percentages. TFLOPs measure average computation per model inference. All variants enable future prediction at every decision (All-State); only budget selection changes.

Cam.RobotLang.LightBg.NoiseLayoutTotalTFLOPs
SFT-only budget60.281.387.696.463.186.783.179.814.51
Fixed-cost argmax59.480.388.197.165.087.483.380.011.60
ThinkWAM63.081.288.095.564.287.883.380.510.89

We will update the manuscript to include these expanded experiments and discussion.

4. Clarification on Multi-Intensity Training Ablation (Weakness 4 and Question 4)

We thank the reviewer for requesting this clarification. For w/o multi-intensity reasoning training, we train the backbone using only the shared synchronous schedule and then freeze it. Using a routing dataset with the same number of states as in the full model, we regenerate Gate gain labels with this new backbone by comparing five-step Direct inference with synchronous reasoning at (A,V)=(5,5), following the original paired action-error procedure with five-seed averaging, and retrain Gate on these labels. The budget factor relies on multi-intensity training to enable the backbone to support different reasoning intensities. Accordingly, in this ablation, Gate selects between Direct and a fixed reasoning configuration matched to the backbone’s synchronous training setting; the budget factor, intensity-quality label generation, and GRPO training are omitted. Gate supervision therefore corresponds to the backbone and reasoning configuration actually evaluated. We will update the relevant sections of our manuscript to include these detailed discussions.

Thank you sincerely again for your meticulous review and for helping us improve the clarity of our work. We will update the manuscript to include these expanded experiments and discussion.

The authors address my concerns in detail and help me better understand the proposed method. I will maintain my positive score.

Reviewer_U9aQOpenReview ↗
Rating
3
Confidence
3
Informativeness
2
Presentation
3
Soundness
2
Read review
Paper summary

This paper investigates whether test-time future prediction is necessary in WAMs and what level of reasoning intensity is required. The authors observe that the need for future prediction depends on the current state. Different states demand varying levels of reasoning intensity.

Motivated by this finding, the paper proposes ThinkWAM, a novel world action model equipped with adaptive predictive reasoning. First, to enable the WAM to support varying reasoning intensities, the authors train the model using a mixture of synchronous and asynchronous noise schedules. Second, they train a router to dynamically decide whether future prediction is needed and determine the appropriate reasoning intensity based on the current state. To further optimize the quality-cost trade-off, the authors employ GRPO for post-training the router, ensuring a well-balance between performance and computational cost.

The authors validate the effectiveness of ThinkWAM through extensive experiments, demonstrating the importance of adaptive reasoning. Ablation studies further confirm the necessity of the mixed noise schedule training and the budget factor. Moreover, the effectiveness of the proposed method is successfully demonstrated in real-world robotic manipulation tasks.

Critical strengths
  1. The authors make an observation that test-time future prediction in WAMs is neither universally necessary nor uniformly removable. Based on this insight, they propose ThinkWAM, which uses a router to dynamically adapt future prediction and reasoning intensity according to the current state and task. This finding is valuable for the community.

  2. The effectiveness of the proposed method is thoroughly validated through extensive experiments across multiple benchmarks as well as in real-world robotic manipulation scenarios.

Critical weaknesses
  1. As shown by the ThinkWAM-All-State results, applying reasoning across all states achieves the best performance across all benchmarks, however, it requires higher computational overhead. Could the authors provide a more detailed quality-cost comparison? This would better illustrate the specific impacts of different reasoning intensities and budget factors on both performance and computational cost. Such an analysis would not only strengthen the justification for adaptive reasoning but also clearly highlight the specific gains brought by GRPO post-training.
Questions for authors
  1. The current inference efficiency evaluation only compares the computational overhead of a single forward pass between no future prediction and predictive reasoning. Could the authors provide a comparison of the overall inference efficiency (e.g., total time per rollout) across different benchmarks or complete tasks?
  2. The color for all-state in Figure 1 appears to be inconsistent. Please double-check and fix this typo.
4 comments
AuthorsOpenReview ↗

Response to Reviewer U9aQ (1/3)

We greatly appreciate Reviewer U9aQ’s constructive comments and insightful feedback. The suggestions have prompted us to carefully consider each point in our method, and we appreciate the opportunity to strengthen the paper accordingly. We provide detailed responses below.

1. Quality–Cost Trade-offs Across Reasoning Intensities and Budget Factors (Weakness 1)

We thank the reviewer for this helpful suggestion. As defined in Sec. 3.2.3, the router selects reasoning intensity (A,V)(A,V), where AA is the number of joint video-action updates and VV controls the refinement rate of future-video latents. Inference terminates after AA joint updates. To analyse the quality–cost trade-off, Tables U1–U4 report results with reasoning enabled at every decision, a fixed reasoning intensity (A,V)(A,V), and the router disabled. These fixed-intensity measurements exclude routing overhead. They include success rates, average inference latency, and computational cost per model inference. All inference evaluations are conducted on NVIDIA H20 GPU. At V=25V=25, reducing AA from 5 to 2 reduces TFLOPs by 60.0% and latency by 54.8% on both LIBERO and LIBERO-Plus. Success decreases by only 0.5 and 1.1 percentage points, respectively. At A=5A=5, the V=25V=25 and V=15V=15 schedules use the five joint updates but expose the action branch to differently refined future latents. On LIBERO-Plus, V=15V=15 achieves 81.3% success, compared with 79.8% for V=25V=25, with identical computation of 14.51 TFLOPs and similar latency. Together, these results show that AA controls computational cost, while VV affects control performance within that budget, motivating our budget factor's selection: choosing the proper reasoning intensity (A,V) based on the quality-cost trade-off.

Table U1. Effect of reasoning depth AA on LIBERO with V=25V=25. Fixed intensity at every decision; router disabled. Success is in %; latency (ms) and backbone TFLOPs are measured per model inference.

AASpatialObjectGoalLongAvg.LatencyTFLOPs
298.699.897.696.698.271.35.80
399.299.897.296.898.3100.18.70
498.899.898.097.098.4128.911.60
599.299.897.897.898.7157.614.51

Table U2. Effect of reasoning depth AA on LIBERO-Plus with V=25V=25. Fixed intensity at every decision; router disabled. Success is in %; latency (ms) and backbone TFLOPs are measured per model inference. Columns Cam., Robot, Lang., Light, Bg., Noise, and Layout denote camera, robot, language, lighting, background, sensor-noise, and layout shifts, respectively.

AACam.RobotLang.LightBg.NoiseLayoutTotalLatencyTFLOPs
257.079.486.695.262.886.583.478.771.25.80
358.981.286.096.664.584.883.379.2100.08.70
459.881.087.096.963.886.583.579.8128.811.60
560.880.587.896.562.687.183.179.8157.614.51

Table U3. Effect of video-schedule parameter VV on LIBERO with A=5A=5. Each inference executes only the first five steps of the VV-step video-denoising schedule. Fixed intensity at every decision; router disabled. Success is in %; latency (ms) and backbone TFLOPs are measured per model inference.

VVSpatialObjectGoalLongAvg.LatencyTFLOPs
599.2100.097.495.898.1157.614.51
1099.4100.098.097.498.7157.614.51
1599.699.897.497.498.6157.614.51
2099.499.898.097.098.6157.614.51
2599.299.897.897.898.7157.614.51

Table U4. Effect of video-schedule parameter VV on LIBERO-Plus with A=5A=5. Each inference executes only the first five steps of the VV-step video-denoising schedule. Fixed intensity at every decision; router disabled. Success is in %; latency (ms) and backbone TFLOPs are measured per model inference.

VVCam.RobotLang.LightBg.NoiseLayoutTotalLatencyTFLOPs
563.081.888.497.562.787.083.680.7157.614.51
1061.781.688.097.464.187.683.980.7157.614.51
1563.980.790.198.365.087.883.581.3157.614.51
2060.481.887.497.363.886.583.380.0157.614.51
2560.880.587.896.562.687.183.179.8157.614.51

We will update the relevant sections of our manuscript to include these comprehensive results and detailed discussions.

AuthorsOpenReview ↗

Response to Reviewer U9aQ (2/3)

2. Ablation Study on GRPO (Weakness 1)

We thank the reviewer for this helpful suggestion. We conduct additional experiments on an ablation study of GRPO, as shown in Table U5. It compares two variants of budget policies on LIBERO-Plus. The two variants and ThinkWAM enable future prediction at every decision (All-State) and differ only in budget selection:

  • SFT-only budget: the ThinkWAM variant that trains only the quality predictor without GRPO. At inference, it fixes A=5A=5 and selects VV based on predicted quality, without cost-aware depth allocation.
  • Fixed-cost argmax: the ThinkWAM variant with selecting the reasoning intensity using predicted quality minus the fixed computation penalty from Sec. 3.4.3, without GRPO.

The ThinkWAM achieves 80.5% success, compared with 79.8% for SFT-only and 80.0% for fixed-cost argmax. It also reduces computation from 14.51 to 10.89 TFLOPs per model inference relative to SFT-only (24.9%), and from 11.60 to 10.89 relative to fixed-cost argmax (6.1%). With the backbone and quality predictor frozen during GRPO, these results support the additional benefit of post-training the budget allocation policy beyond SFT and direct score maximization.

Table U5. GRPO ablation on LIBERO-Plus. Success columns report percentages. TFLOPs measure average computation per model inference. All variants enable future prediction at every decision (All-State); only budget selection changes.

Cam.RobotLang.LightBg.NoiseLayoutTotalTFLOPs
SFT-only budget60.281.387.696.463.186.783.179.814.51
Fixed-cost argmax59.480.388.197.165.087.483.380.011.60
ThinkWAM63.081.288.095.564.287.883.380.510.89

We will update the manuscript to include these expanded experiments and discussion.

AuthorsOpenReview ↗

Response to Reviewer U9aQ (3/3)

3. A Comparison of The Overall Inference Efficiency (Question 1)

We thank the reviewer for this helpful suggestion. We conducted the experiments on NVIDIA H20 GPU, as shown in Tables U6-U8, which report both average cost per model inference and per rollout. Relative to the highest reasoning-ratio setting in each benchmark, LIBERO-Plus at 17.7% reasoning achieves 77.0% success, while reducing rollout inference time by 46.8% and computation by 63.2%, respectively. RoboCasa GR1 at 28.0% reasoning achieves 75.8% success, recovering 66.7% of the All-State success-rate gain over direct prediction while reducing rollout inference time by 38.9% and computation by 54.9%. These results show how adaptive predictive reasoning improves task success while reducing the inference time and computation required over complete rollouts.

Table U6. Rollout-level inference efficiency on LIBERO across reasoning ratios. Latency (ms) and TFLOPs are averages per model inference. Per-rollout metrics accumulate all model inference calls within a rollout. 0% ratio denotes action-only inference; 100% denotes all-state reasoning with adaptive intensity. Intermediate ratios use gated reasoning.

Reasoning ratio (%)Success (%)LatencyTFLOPsInference time/rollout (s)TFLOPs/rollout
100.098.8163.0611.572.48175.9
85.098.7148.1510.212.25155.0
76.798.7139.949.472.11143.1
20.098.277.913.781.2058.4
13.297.869.683.031.0846.8
4.397.559.212.040.9231.7
0.097.353.921.550.8424.3

Table U7. Rollout-level inference efficiency on LIBERO-Plus across reasoning ratios. Latency (ms) and TFLOPs are averages per model inference. Per-rollout metrics accumulate all model inference calls within a rollout. 0% ratio denotes action-only inference; 100% denotes all-state reasoning with adaptive intensity. Intermediate ratios use gated reasoning.

Reasoning ratio (%)Success (%)LatencyTFLOPsInference time/rollout (s)TFLOPs/rollout
100.080.5161.1410.893.59242.8
87.179.9146.849.653.29216.0
56.178.8113.946.732.62154.8
17.777.080.943.771.9189.3
0.074.456.141.551.3738.0

Table U8. Rollout-level inference efficiency on RoboCasa GR1 across reasoning ratios. Latency (ms) and TFLOPs are averages per model inference. Per-rollout metrics accumulate all model inference calls within a rollout. 0% ratio denotes action-only inference; 100% denotes all-state reasoning with adaptive intensity. Intermediate ratios use gated reasoning.

Reasoning ratio (%)Success (%)LatencyTFLOPsInference time/rollout (s)TFLOPs/rollout
100.077.781.925.511.67112.2
91.277.575.514.921.61105.0
87.677.060.633.551.3277.2
28.075.847.922.381.0250.6
7.373.036.611.340.7728.0
0.072.032.370.970.6720.2

We will update the relevant sections of our manuscript to include these comprehensive results and detailed discussions.

4. Color Correction in Figure 1 (Question 2)

Thank you for your valuable feedback. We sincerely appreciate that. We will revise the manuscript to correct the color inconsistency for the All-State setting.

Thank you sincerely again for your meticulous review and for helping us improve the clarity of our work. We will update the manuscript to include these expanded experiments and discussion.

Dear Reviewer U9aQ,

Thank you for your time and constructive review. Your suggestions have been very helpful in strengthening our paper.

As the discussion period draws to a close, we would be grateful to hear whether our responses have addressed your main concerns. If there are any remaining questions or points that would benefit from further clarification, please let us know. We would be happy to provide additional information.

Thank you again for your time and thoughtful review of our paper.

Comments on the paper

1 comment
AuthorsOpenReview ↗

Response to All Reviewers

We would like to sincerely thank all the reviewers for their thoughtful and detailed feedback on our manuscript. We greatly appreciate the reviewers for recognizing our work as a valuable contribution to adaptive predictive reasoning in WAMs, supported by insightful findings and thorough experimental validation. Your constructive feedback has provided valuable perspectives that guided us in refining the clarity, depth, and overall presentation of our work. We have carefully addressed each of your comments below, offering detailed explanations and outlining our plans for revisions.

To thoroughly address your constructive questions and suggestions, we have conducted extensive new simulation experiments and analyses on LIBERO, LIBERO-Plus, and RoboCasa GR1, with all inference measured on NVIDIA H20 GPUs. The main additions are summarized below.

  1. Inference Efficiency (QycY, rSXt, U9aQ; Tables Q1–Q3, R1–R3, U6–U8). We report latency and TFLOPs both per model inference and per complete rollout, across reasoning ratios from 0% to 100%. On RoboCasa GR1, adaptive reasoning at 28.0% of decisions reaches 75.8% success, compared with 72.0% for Direct. This recovers 66.7% of the All-State gain while reducing rollout inference time by 38.9% and computation by 54.9%. On LIBERO-Plus, 17.7% reasoning reaches 77.0% success (vs. 74.4% for Direct), while reducing rollout inference time by 46.8% and computation by 63.2% relative to All-State.

  2. Quality–Cost Trade-offs Across Reasoning Intensities (rSXt, U9aQ; Tables R4–R7, U1–U4). We fixed the reasoning intensity (A,V)(A,V) at every decision with the router disabled. At V=25V=25, reducing AA from 5 to 2 cuts TFLOPs by 60.0% and latency by 54.8%, while success drops by only 0.5 points on LIBERO and 1.1 points on LIBERO-Plus. At A=5A=5, changing VV keeps computation at 14.51 TFLOPs but changes success (81.3% for V=15V=15 vs. 79.8% for V=25V=25 on LIBERO-Plus). Thus AA mainly sets the cost and VV affects performance within that budget, which motivates selecting (A,V)(A,V) jointly.

  3. Compute-Matched Action-Only Baselines (QycY; Tables Q4–Q5). On the same frozen backbone, we increased the number of action-only denoising steps to match ThinkWAM's TFLOPs. On LIBERO-Plus, ThinkWAM-Adaptive outperforms the 26-step baseline by 1.5 points (77.0% vs. 75.5%), and ThinkWAM-All-State outperforms the 93-step baseline by 6.9 points (80.5% vs. 73.6%). On RoboCasa GR1, the gains are 2.6 and 1.4 points over the 23-step and 64-step baselines. The improvement therefore comes from predictive reasoning, not from extra denoising computation.

  4. GRPO Ablation for Budget Allocation (mmE9, QycY, U9aQ; Tables M4, Q6, U5). We compared ThinkWAM against an SFT-only budget and a fixed-cost argmax policy, with the backbone and quality predictor frozen. ThinkWAM reaches 80.5% success on LIBERO-Plus (vs. 79.8% and 80.0%) while reducing computation from 14.51 and 11.60 TFLOPs to 10.89 TFLOPs, reductions of 24.9% and 6.1%.

  5. Validity of Gate Supervision and Closed-Loop Impact (mmE9, QycY; Tables M1–M3, Q7–Q9). On the validation split averaged over five seeds, Budget-selected reasoning reduces action MSE beyond the Gate-label reference. In paired same-state rollouts on replayed LIBERO states, where only the first decision differs, Budget-selected reasoning lowers executed-action MSE by 9.8% and raises success from 98.1% to 98.5%.

Detailed responses to each reviewer's questions are provided in the individual threads below. Thank you again for your time and thoughtful evaluation.

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

Related papers

Loading the map…

Discussion (0)

Sign in to comment.