- 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
-
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.
-
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
-
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.
-
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
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 on LIBERO with . Fixed intensity at every decision; router disabled. Success is in %; latency (ms) and backbone TFLOPs are measured per model inference.
| Spatial | Object | Goal | Long | Avg. | Latency | TFLOPs | |
|---|---|---|---|---|---|---|---|
| 2 | 98.6 | 99.8 | 97.6 | 96.6 | 98.2 | 71.3 | 5.80 |
| 3 | 99.2 | 99.8 | 97.2 | 96.8 | 98.3 | 100.1 | 8.70 |
| 4 | 98.8 | 99.8 | 98.0 | 97.0 | 98.4 | 128.9 | 11.60 |
| 5 | 99.2 | 99.8 | 97.8 | 97.8 | 98.7 | 157.6 | 14.51 |
Table R5. Effect of reasoning depth on LIBERO-Plus with . 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.
| Cam. | Robot | Lang. | Light | Bg. | Noise | Layout | Total | Latency | TFLOPs | |
|---|---|---|---|---|---|---|---|---|---|---|
| 2 | 57.0 | 79.4 | 86.6 | 95.2 | 62.8 | 86.5 | 83.4 | 78.7 | 71.2 | 5.80 |
| 3 | 58.9 | 81.2 | 86.0 | 96.6 | 64.5 | 84.8 | 83.3 | 79.2 | 100.0 | 8.70 |
| 4 | 59.8 | 81.0 | 87.0 | 96.9 | 63.8 | 86.5 | 83.5 | 79.8 | 128.8 | 11.60 |
| 5 | 60.8 | 80.5 | 87.8 | 96.5 | 62.6 | 87.1 | 83.1 | 79.8 | 157.6 | 14.51 |
Table R6. Effect of video-schedule parameter on LIBERO with . Each inference executes only the first five steps of the -step video-denoising schedule (e.g., for and for ). Fixed intensity at every decision; router disabled. Success is in %; latency (ms) and backbone TFLOPs are measured per model inference.
| Spatial | Object | Goal | Long | Avg. | Latency | TFLOPs | |
|---|---|---|---|---|---|---|---|
| 5 | 99.2 | 100.0 | 97.4 | 95.8 | 98.1 | 157.6 | 14.51 |
| 10 | 99.4 | 100.0 | 98.0 | 97.4 | 98.7 | 157.6 | 14.51 |
| 15 | 99.6 | 99.8 | 97.4 | 97.4 | 98.6 | 157.6 | 14.51 |
| 20 | 99.4 | 99.8 | 98.0 | 97.0 | 98.6 | 157.6 | 14.51 |
| 25 | 99.2 | 99.8 | 97.8 | 97.8 | 98.7 | 157.6 | 14.51 |
Table R7. Effect of video-schedule parameter on LIBERO-Plus with . Each inference executes only the first five steps of the -step video-denoising schedule (e.g., for and for ). Fixed intensity at every decision; router disabled. Success is in %; latency (ms) and backbone TFLOPs are measured per model inference.
| Cam. | Robot | Lang. | Light | Bg. | Noise | Layout | Total | Latency | TFLOPs | |
|---|---|---|---|---|---|---|---|---|---|---|
| 5 | 63.0 | 81.8 | 88.4 | 97.5 | 62.7 | 87.0 | 83.6 | 80.7 | 157.6 | 14.51 |
| 10 | 61.7 | 81.6 | 88.0 | 97.4 | 64.1 | 87.6 | 83.9 | 80.7 | 157.6 | 14.51 |
| 15 | 63.9 | 80.7 | 90.1 | 98.3 | 65.0 | 87.8 | 83.5 | 81.3 | 157.6 | 14.51 |
| 20 | 60.4 | 81.8 | 87.4 | 97.3 | 63.8 | 86.5 | 83.3 | 80.0 | 157.6 | 14.51 |
| 25 | 60.8 | 80.5 | 87.8 | 96.5 | 62.6 | 87.1 | 83.1 | 79.8 | 157.6 | 14.51 |
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.
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), where A is the number of joint video-action updates and V controls the refinement rate of future-video latents. Inference terminates after A 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), 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=25, reducing A 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=5, the V=25 and V=15 schedules use the same five joint updates but expose the action branch to differently refined future latents. On LIBERO-Plus, V=15 achieves 81.3% success, compared with 79.8% for V=25, with identical computation of 14.51 TFLOPs and similar latency. Together, these results show that A controls computational cost, while V 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.
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.
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.
Due to the 5,000-character limit per reply, Tables R4–R7 are provided in Response to Reviewer rSXt (2/3).