- Rating
- 3(+1)
- Confidence
- 2
- Informativeness
- 2
- Presentation
- 3(+1)
- Soundness
- 3(+1)
Read review
Paper summary
This paper proposes ActTime, an action-conditioned world model for time-series forecasting and decision making. The framework combines variable-specific latent states, action-conditioned transitions, observation-based state correction, and a pretrained temporal prior to learn controlled dynamics from factual trajectories and simulate future outcomes under candidate actions. Experiments across four simulation environments evaluate factual prediction, intervention-effect prediction, offline decision making, and closed-loop planning. The results demonstrate competitive performance across multiple tasks and highlight discrepancies among factual accuracy, intervention fidelity, and decision quality. However, the methodological novelty over existing world models remains unclear, while the necessity of key architectural components and the reliability of the reported improvements are insufficiently established.
Critical strengths
The paper provides a systematic evaluation across four simulation environments, covering factual prediction, intervention-effect prediction, offline decision making, and closed-loop planning. By simulating different candidate actions from the same historical state, the evaluation directly assesses intervention fidelity and decision quality. The analysis of model-ranking discrepancies across evaluation criteria also provides useful empirical insights into the relationship among factual accuracy, intervention fidelity, and decision quality.
Critical weaknesses
-
Unclear methodological novelty. The related work does not clearly differentiate ActTime from existing world models such as RSSM, AcRKN, and MTS3. Moreover, the proposed methodology mainly combines established components without demonstrating distinctive technical innovations, leaving its methodological novelty unclear.
-
Insufficient justification of architectural choices. The ablation results do not consistently support the proposed components. Simple action concatenation outperforms the full model on SUMO intervention prediction, while the EIDOS prior provides inconsistent benefits. The shared-state ablation also lacks capacity-matched comparisons.
-
Insufficient experimental validation. The authors should report the mean and standard deviation over multiple independent runs rather than single-run results to better establish the reliability of the reported performance improvements.
Questions for authors
-
What specific technical limitations of existing action-conditioned world models does ActTime address?
-
Why are the EIDOS features averaged across target variables and broadcast to all variable states instead of retaining variable-specific temporal representations?
-
Why does the framework adopt two-stage training with a frozen predictor and a separately trained EIDOS adapter instead of end-to-end joint training?
-
Do the reported model-ranking reversals persist when considering only strong competing models, such as ActTime, TFT, RSSM, and MTS3?
-
Do q and w defined in Appendix A.2 correspond to the Queue and Stopped Waiting metrics reported in Table 11, respectively?
-
How does the computational cost of ActTime in closed-loop planning compare with that of the main world-model baselines?
8 comments
Response to Reviewer ykhK [Part 2]
Q2: Why are the EIDOS features averaged across target variables and broadcast to all variable states instead of retaining variable-specific temporal representations?
ActTime already maintains separate latent states for all observed variables. We average the EIDOS features to provide a shared temporal prior to these states, including those without their own EIDOS features. The adapter also uses variable embeddings, allowing each state to receive a different update.
We also compared two alternatives:
- Variant A: Each target state uses its own EIDOS features, while non-target states use the averaged features.
- Variant B: Each target state uses its own EIDOS features, while non-target states receive no EIDOS prior.
All three variants use the same learned dynamics, adapter size, and training settings. The results are shown below.
| Domain | Metric | Original | Variant A | Variant B |
|---|---|---|---|---|
| CityLearn | Factual MAE ↓ | 0.24841 | 0.24864 | 0.24856 |
| Effect NMAE ↓ | 0.07069 | 0.07129 | 0.07197 | |
| Offline Regret ↓ | 0.28760 | 0.28080 | 0.29038 | |
| Online Cost ↓ | 0.83247 | 0.83258 | 0.83444 | |
| SUMO | Factual MAE ↓ | 0.10768 | 0.10766 | 0.10796 |
| Effect NMAE ↓ | 0.48455 | 0.48918 | 0.49617 | |
| Offline Regret ↓ | 0.02461 | 0.02519 | 0.02626 | |
| Online Cost ↓ | 0.30848 | 0.31246 | 0.31407 | |
| GreenLight | Factual MAE ↓ | 0.04976 | 0.04980 | 0.05027 |
| Effect NMAE ↓ | 0.20838 | 0.20845 | 0.20840 | |
| Offline Regret ↓ | 0.14034 | 0.13372 | 0.14857 | |
| Online Cost ↓ | 0.29612 | 0.29513 | 0.29536 | |
| Inventory | Factual MAE ↓ | 0.07908 | 0.07910 | 0.07927 |
| Effect NMAE ↓ | 0.22331 | 0.22459 | 0.22342 | |
| Offline Regret ↓ | 0.15314 | 0.15455 | 0.15615 | |
| Online Cost ↓ | 3.88683 | 3.88408 | 3.89603 |
The differences are generally small. The original design performs best in 11 of 16 settings, including all four Effect NMAE results. Neither alternative shows a consistent advantage. We have clarified the design choice in and added the comparison in .
Q3: Why does the framework adopt two-stage training with a frozen predictor and a separately trained EIDOS adapter instead of end-to-end joint training?
EIDOS provides temporal features from the observed history, but does not model how the system responds to actions. We want to use these features to refine the final history state without changing the learned action-conditioned dynamics. We therefore train the predictor first, then freeze it and train only the EIDOS adapter. Since training uses only factual trajectories, joint training could also change the learned dynamics without directly supervising alternative-action responses.
We also tried two alternatives. The first trains the predictor and adapter jointly from scratch. The second starts from the same Stage-A checkpoint but updates both the predictor and adapter. The EIDOS backbone is frozen in all settings.
| Domain | Metric | Original | Joint from scratch | Joint fine-tuning |
|---|---|---|---|---|
| CityLearn | Factual MAE ↓ | 0.24432 | 0.25099 | 0.25366 |
| Effect NMAE ↓ | 0.06875 | 0.07673 | 0.07105 | |
| Offline Regret ↓ | 0.28309 | 0.29145 | 0.29189 | |
| Online Cost ↓ | 0.83231 | 0.84072 | 0.83831 | |
| SUMO | Factual MAE ↓ | 0.10871 | 0.11218 | 0.11134 |
| Effect NMAE ↓ | 0.51540 | 0.55508 | 0.53579 | |
| Offline Regret ↓ | 0.02470 | 0.03331 | 0.02979 | |
| Online Cost ↓ | 0.30236 | 0.31117 | 0.31374 | |
| GreenLight | Factual MAE ↓ | 0.05272 | 0.05433 | 0.05552 |
| Effect NMAE ↓ | 0.21643 | 0.20722 | 0.19372 | |
| Offline Regret ↓ | 0.11902 | 0.18218 | 0.14118 | |
| Online Cost ↓ | 0.29375 | 0.29442 | 0.29607 | |
| Inventory | Factual MAE ↓ | 0.08000 | 0.08262 | 0.08368 |
| Effect NMAE ↓ | 0.22398 | 0.22511 | 0.22816 | |
| Offline Regret ↓ | 0.15162 | 0.16856 | 0.17205 | |
| Online Cost ↓ | 3.88843 | 3.97616 | 3.99782 |
The original method performs best in 15 of 16 settings. On GreenLight, both joint variants achieve lower Effect NMAE, with joint fine-tuning performing best. However, both variants have higher offline regret and online cost across all four domains. We have clarified the training rationale in and added the comparison in .
Response to Reviewer ykhK [Part 3]
Q4: Do the reported model-ranking reversals persist when considering only strong competing models, such as ActTime, TFT, RSSM, and MTS3?
Yes. We repeated the analysis using only the four models suggested by the reviewer, under the same evaluation protocol. We also updated the full comparison to include AcRKN and CEDAR-StageI, expanding it from eight to ten methods. The results are shown below.
| Models | Prediction → Effect | Effect → Decision |
|---|---|---|
| All ten models | 18.6% | 29.4% |
| ActTime, TFT-Action, RSSM, MTS3 | 26.6% | 35.0% |
Both reversal rates increase when considering only these four competitive models. This shows that ranking disagreements persist even among strong models, rather than being limited to comparisons involving weaker baselines. We have updated accordingly.
Q5: Do q and w defined in Appendix A.2 correspond to the Queue and Stopped Waiting metrics reported in Table 11, respectively?
Yes. The quantities and in Appendix A.2 correspond to the normalized Queue and Stopped Waiting metrics in Table 11, respectively. We have clarified this correspondence in .
Q6: How does the computational cost of ActTime in closed-loop planning compare with that of the main world-model baselines?
We compared ActTime with RSSM, MTS3, and AcRKN. We measured the time needed for one complete MPC decision, including history encoding, candidate rollouts, and action selection. All models use the same planning settings and hardware. Here are the results.
| Model | CityLearn | SUMO | Inventory | GreenLight |
|---|---|---|---|---|
| ActTime | 119.07 | 107.38 | 88.79 | 193.20 |
| RSSM | 114.09 | 90.37 | 72.53 | 215.38 |
| MTS3 | 301.64 | 230.02 | 249.61 | 673.95 |
| AcRKN | 119.24 | 106.12 | 100.07 | 195.05 |
Average latency per MPC decision (ms).
ActTime is faster than MTS3 in all four domains and has similar latency to AcRKN. RSSM is faster on CityLearn, SUMO, and Inventory, while ActTime is faster on GreenLight. We have added these results to .
Response to Reviewer ykhK [Part 1]
Thank you for the detailed review and thoughtful questions.
RSSM, AcRKN, and MTS3 already learn action-conditioned dynamics, but they do not explicitly maintain separate latent states for individual observed variables. This matters in multivariate controlled systems, where the same action may affect different variables differently, and these variables may also influence one another.
ActTime addresses this by maintaining a latent state for each observed variable, with variable-specific action representations to capture different responses. A shared transition allows information exchange across states, modeling these responses together with cross-variable interactions.
We have added AcRKN to the main comparison alongside RSSM and MTS3. ActTime achieves lower factual MAE in all four domains and lower online cost in three. We also evaluated a parameter-matched shared-state variant, which yields higher intervention error, offline regret, and online cost across all four domains. We have clarified the differences in Section 2 of the revised paper and reported the comparisons in Tables 2, 4, and 15 of the revised paper.
(1) Simple action concatenation on SUMO.
As you noted, simple action concatenation achieves lower Effect NMAE on SUMO, 0.3526 compared with 0.5154 for ActTime. However, its online cost is higher, 0.3210 compared with 0.3024. The improvement in intervention prediction therefore does not translate into better closed-loop planning. On the other three datasets, ActTime achieves lower Effect NMAE, and its online cost is lower across all four datasets. We have clarified the SUMO result in Appendix B.4 of the revised paper.
(2) Inconsistent benefits of the EIDOS prior.
We agree that EIDOS does not improve every metric. To better assess its contribution, we repeated the ablation across three training seeds. For each seed, we evaluated the same trained model with and without the EIDOS adapter, keeping the learned dynamics unchanged. The results are also reported in Table 16 of the revised paper.
The table shows the average relative change after removing EIDOS. Positive values indicate worse performance.
Removing EIDOS worsens performance in 14 of the 16 settings on average across three seeds. The only exceptions are SUMO Effect NMAE and Inventory Offline Regret. Notably, factual MAE and online cost increase in all four domains. Although the gains vary across metrics, EIDOS provides consistent benefits for factual prediction and closed-loop planning.
(3) Shared-state ablation.
We conducted a parameter-matched shared-state comparison. Both models have the same number of trainable parameters and retain the variable embeddings. In the shared-state variant, we average the dynamic states across variables after initialization, transition, and observation correction.
The shared-state variant achieves a slightly lower factual MAE on CityLearn, but has higher Effect NMAE, offline regret, and online cost on all four datasets. These results support maintaining separate dynamic states for individual variables. We have updated Table 4 and Appendix Table 15 of the revised paper accordingly.
Following your suggestion, we have rerun the main experiments with three independent training seeds. The mean results are reported in Table 2 of the revised paper, with mean and standard deviation in Appendix Tables 8–11 of the revised paper.
ActTime still achieves the best mean performance in 12 out of 16 settings.