acceptodds
Under review as a conference paper at ICLR 2027

ActTime: Action-Conditioned World Modeling for Time Series

34 views

Abstract

Time-series models increasingly condition on known future inputs. In controlled systems, some of these inputs are actions that can be changed, compared, and optimized before execution. A model must therefore do more than predict the realized future: it must predict how outcomes change under different actions from the same observed history and support reliable comparison of candidate plans. We introduce ActTime, an action-conditioned world model for time series. Over the observed history, ActTime updates variable-specific latent states through an action-conditioned transition and corrects them using observations. It then rolls the same transition forward under candidate actions to generate future trajectories. Training uses only factual trajectories, without paired alternative-action outcomes or planning supervision. We construct a simulator-grounded evaluation suite across building energy, traffic, greenhouse climate, and inventory management, covering factual prediction, intervention-effect prediction, offline decision making, and closed-loop planning. ActTime ranks first in 12 of 16 domain-task settings. Across eight methods, the mean pairwise model-ranking reversal rate is 17.8% between factual prediction and intervention fidelity, and rises to 28.4% between intervention fidelity and decision ranking. These results show that factual accuracy, intervention fidelity, and decision quality are related but not interchangeable. Code, checkpoints, and data are available at https://anonymous.4open.science/r/ActTime-9EC7.

Primary area:
learning on time series and dynamical systems
Keywords:
time series, world models, action-conditioned modeling, intervention prediction, decision making

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 3 reviews, above 85% of the others.
Reviewer_ykhKOpenReview ↗
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
  1. 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.

  2. 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.

  3. 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
  1. What specific technical limitations of existing action-conditioned world models does ActTime address?

  2. Why are the EIDOS features averaged across target variables and broadcast to all variable states instead of retaining variable-specific temporal representations?

  3. 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?

  4. Do the reported model-ranking reversals persist when considering only strong competing models, such as ActTime, TFT, RSSM, and MTS3?

  5. Do q and w defined in Appendix A.2 correspond to the Queue and Stopped Waiting metrics reported in Table 11, respectively?

  6. How does the computational cost of ActTime in closed-loop planning compare with that of the main world-model baselines?

8 comments
AuthorsOpenReview ↗

Response to Reviewer ykhK [Part 1]

Thank you for the detailed review and thoughtful questions.

W1 & Q1: Methodological novelty and limitations addressed relative to RSSM, AcRKN, and MTS3.

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‾\underline{\text{Section 2 of the revised paper}} and reported the comparisons in Tables 2, 4, and 15 of the revised paper‾\underline{\text{Tables 2, 4, and 15 of the revised paper}}.

W2: Action concatenation on SUMO, inconsistent EIDOS gains, and capacity-matched shared-state comparison.

(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‾\underline{\text{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‾\underline{\text{Table 16 of the revised paper}}.

The table shows the average relative change after removing EIDOS. Positive values indicate worse performance.

DatasetFactual MAEEffect NMAEOffline RegretOnline Cost
CityLearn+1.67%+3.05%+0.05%+0.93%
SUMO+1.73%−2.88%+1.66%+0.87%
Inventory+2.94%+0.43%−1.62%+0.12%
GreenLight+7.16%+2.35%+2.56%+0.48%

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.

DatasetModelFactual MAE ↓Effect NMAE ↓Offline Regret ↓Online Cost ↓
CityLearnActTime0.25100.06880.29310.8323
Shared State0.24330.12390.34440.8662
SUMOActTime0.11070.51540.02470.3024
Shared State0.17130.91740.08540.3781
GreenLightActTime0.05270.21640.10900.2948
Shared State0.11020.35670.14790.3035
InventoryActTime0.08000.22400.15183.8884
Shared State0.11010.39210.16963.9932

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‾\underline{\text{Table 4 and Appendix Table 15 of the revised paper}} accordingly.

W3: Need for mean and standard deviation over independent runs.

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‾\underline{\text{Table 2 of the revised paper}}, with mean and standard deviation in Appendix Tables 8–11 of the revised paper‾\underline{\text{Appendix Tables 8–11 of the revised paper}}.

ActTime still achieves the best mean performance in 12 out of 16 settings.

AuthorsOpenReview ↗

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.

DomainMetricOriginalVariant AVariant B
CityLearnFactual MAE ↓0.248410.248640.24856
Effect NMAE ↓0.070690.071290.07197
Offline Regret ↓0.287600.280800.29038
Online Cost ↓0.832470.832580.83444
SUMOFactual MAE ↓0.107680.107660.10796
Effect NMAE ↓0.484550.489180.49617
Offline Regret ↓0.024610.025190.02626
Online Cost ↓0.308480.312460.31407
GreenLightFactual MAE ↓0.049760.049800.05027
Effect NMAE ↓0.208380.208450.20840
Offline Regret ↓0.140340.133720.14857
Online Cost ↓0.296120.295130.29536
InventoryFactual MAE ↓0.079080.079100.07927
Effect NMAE ↓0.223310.224590.22342
Offline Regret ↓0.153140.154550.15615
Online Cost ↓3.886833.884083.89603

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 Section 3.2 of the revised paper‾\underline{\text{Section 3.2 of the revised paper}} and added the comparison in Appendix B.4 and Table 17 of the revised paper‾\underline{\text{Appendix B.4 and Table 17 of the revised paper}}.

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.

DomainMetricOriginalJoint from scratchJoint fine-tuning
CityLearnFactual MAE ↓0.244320.250990.25366
Effect NMAE ↓0.068750.076730.07105
Offline Regret ↓0.283090.291450.29189
Online Cost ↓0.832310.840720.83831
SUMOFactual MAE ↓0.108710.112180.11134
Effect NMAE ↓0.515400.555080.53579
Offline Regret ↓0.024700.033310.02979
Online Cost ↓0.302360.311170.31374
GreenLightFactual MAE ↓0.052720.054330.05552
Effect NMAE ↓0.216430.207220.19372
Offline Regret ↓0.119020.182180.14118
Online Cost ↓0.293750.294420.29607
InventoryFactual MAE ↓0.080000.082620.08368
Effect NMAE ↓0.223980.225110.22816
Offline Regret ↓0.151620.168560.17205
Online Cost ↓3.888433.976163.99782

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 Section 3.3 of the revised paper‾\underline{\text{Section 3.3 of the revised paper}} and added the comparison in Appendix B.4 and Table 18 of the revised paper‾\underline{\text{Appendix B.4 and Table 18 of the revised paper}}.

AuthorsOpenReview ↗

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.

ModelsPrediction → EffectEffect → Decision
All ten models18.6%29.4%
ActTime, TFT-Action, RSSM, MTS326.6%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 Section 4.3 and Appendix B.2 of the revised paper‾\underline{\text{Section 4.3 and Appendix B.2 of the revised paper}} 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 qq and ww in Appendix A.2 correspond to the normalized Queue and Stopped Waiting metrics in Table 11, respectively. We have clarified this correspondence in Appendix A.2 of the revised paper‾\underline{\text{Appendix A.2 of the revised paper}}.

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.

ModelCityLearnSUMOInventoryGreenLight
ActTime119.07107.3888.79193.20
RSSM114.0990.3772.53215.38
MTS3301.64230.02249.61673.95
AcRKN119.24106.12100.07195.05

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 Appendix B.1 and Table 12 of the revised paper‾\underline{\text{Appendix B.1 and Table 12 of the revised paper}}.

Reviewer_q4UXOpenReview ↗
Rating
3
Confidence
3
Informativeness
3
Presentation
2
Soundness
2
Read review
Paper summary

The paper focuses on an interesting and important topic. Then, it proposes a novel model. The experiments are well designed and cover four domains. The paper writing can be improved; for example, in the model design parts, some insights can be included.

This paper distinguishes three levels of evaluation: factual prediction, Intervention-effect prediction, and decision ranking. The authors argue that accurate prediction of realized trajectories does not necessarily mean accurate prediction of the effects of alternative actions, and accurate effect prediction does not necessarily ensure correct ranking of candidate plans. This article empirically quantifies these differences.

The authors propose an action-conditioned world model with shared dynamics. First, the latent states are iteratively calibrated using observed measurements during historical processing. In inferring the future, the model simulates different candidate action sequences based on the same inferred historical state. Both training and model selection rely entirely on factual trajectories. The experiments cover factual prediction, intervention effect prediction, offline decision-making, and closed-loop planning in four simulation areas (building energy management, traffic signal control, greenhouse climate control, and inventory management).

Critical strengths
  1. Novelty. This article takes an interesting perspective and introduces a model with some architectural novelty. It systematically distinguishes between factual prediction, intervention effect prediction, and decision quality. Through model rank inversion analysis, it quantifies the limitations of relying solely on upstream predictors for selecting downstream decision-making models.

  2. Complete experimental verification. The four domains cover continuous and discrete actions, as well as control tasks with different action dimensions. The evaluation covers factual prediction, intervention effects, offline program selection, and closed-loop planning. Within each domain, the methods use the same fact data, offline candidate sets, and closed-loop search budgets, making comparisons easier to interpret.

Critical weaknesses

In the introduction part of the model, most of it is straightforward. I can't understand why it is designed like this. Some insights could be added. For example, why inject the pre-trained temporal prior only at the end of the history and freeze the dynamics in the second stage?

Questions for authors
  1. All experiments were conducted in a simulated environment, and of course the authors have acknowledged this limitation. Are there suitable real-world data sets in which control operations are recorded against which the model can be evaluated? Where feasible, could intervention recordings or prospective controlled experiments be used to assess action responses and decision quality?

  2. Are the main results obtained with multiple independent training seeds?

3 comments
AuthorsOpenReview ↗

Response to Reviewer q4UX

Thank you for the helpful comments and suggestions.

W1: Insufficient motivation for the model design.

ActTime is designed to compare alternative actions from the same observed history. During the history, an action-conditioned transition models the state evolution, while new observations correct the predicted states. The same transition is then used to simulate future trajectories under different candidate actions.

EIDOS provides temporal features but is not trained to model action effects. We therefore use it to refine the terminal history state, where all candidate rollouts begin, rather than incorporating it into the controlled dynamics. In the second stage, we freeze the predictor so that the adapter can use this prior without changing the learned dynamics.

We also compared joint training from scratch and joint fine-tuning after Stage A. Two-stage training achieves the best result in 15 of the 16 settings. We have clarified the motivation in Sections 3.2 and 3.3 of the revised paper‾\underline{\text{Sections 3.2 and 3.3 of the revised paper}} and reported the comparison in Appendix B.4 and Table 18 of the revised paper‾\underline{\text{Appendix B.4 and Table 18 of the revised paper}}.

Q1: Can real-world action records support evaluation of intervention responses and decision quality?

We have added experiments on two real-world datasets with recorded actions and sensor measurements. The setup is described in Appendix B.6 of the revised paper‾\underline{\text{Appendix B.6 of the revised paper}}.

(1) Real-world datasets.

AGC 2019. This dataset contains control settings and climate measurements from six greenhouse compartments. We use the recorded actions to predict future temperature, humidity, and CO₂. We also test how accurately the models predict the observed responses when the controls change.

Causal Chambers Wind Tunnel. This dataset contains real experiments in which the intake fan is increased and then returned to its original setting. We use the applied actions to predict the changes in pressure and fan speed. The test set contains 3,000 physical intervention trials.

The greenhouse data allow us to test predictions under real operating conditions. The wind tunnel data allow us to test responses to controlled interventions.

(2) Evaluation and results.

We evaluate four metrics. Forecast NMAE measures future prediction error, while Response NMAE measures the error in the predicted response after an action. For decision-related evaluation, we use the predicted trajectories to estimate the cost of each recorded run. Cost MAE measures the cost prediction error, and Logged Ranking measures how often the recorded runs are ranked in the correct cost order. We use temperature tracking for AGC and pressure tracking for Wind Tunnel.

We compare ActTime with five strong baselines, including TFT, RSSM, MTS3, AcRKN, and CEDAR-StageI. We also include simple reference methods. The results are shown below and in Table 20 of the revised paper‾\underline{\text{Table 20 of the revised paper}}.

DatasetMetricTFTRSSMMTS3AcRKNCEDAR-IPersistenceAvg. Resp.ActTime
AGCForecast NMAE ↓0.35000.36300.32170.41380.39850.2787—0.2668
Response NMAE ↓0.42830.48360.41250.50950.48570.4084—0.3684
Cost MAE (°C) ↓0.82580.92570.59810.92080.57210.5674—0.5482
Logged Ranking ↑89.34%87.09%88.35%86.64%89.36%90.00%—91.71%
Wind TunnelForecast NMAE ↓0.051990.060340.070810.058120.055200.684730.071930.05137
Response NMAE ↓0.037420.044110.039710.045860.039850.789250.071790.03515
Cost MAE (Pa) ↓0.395930.419070.351260.392440.414896.585800.594830.29887
Logged Ranking ↑81.08%84.23%84.33%84.09%83.53%77.19%81.64%85.28%

Logged Ranking reports the percentage of correctly ranked pairs of recorded runs. Avg. Resp. is the average-response baseline for Wind Tunnel.

ActTime achieves the best results on all four metrics for both datasets. The cost results are based on recorded runs, not new actions selected by the models. We use the simulator experiments to evaluate action selection and closed-loop planning.

Q2: Are the main results obtained with multiple independent training seeds?

The original results were based on a single seed. 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‾\underline{\text{Table 2 of the revised paper}}, with mean and standard deviation in Appendix Tables 8–11 of the revised paper‾\underline{\text{Appendix Tables 8–11 of the revised paper}}.

ActTime still ranks first in 12 out of 16 settings.

Thanks for your response. I maintain my rating.

AuthorsOpenReview ↗

Thank You for Your Feedback

Thank you for taking the time to review our response. We appreciate your constructive feedback and suggestions, which have helped us improve the paper.

Reviewer_LLHWOpenReview ↗
Rating
3
Confidence
3(+1)
Informativeness
2
Presentation
3
Soundness
2
Read review
Paper summary

This paper proposes ActTime, action-conditioned time-series modeling for prediction and planning using factual trajectories without paired alternative-action outcomes. The model maintains variable-specific latent states, alternates action-conditioned transitions with observation-based correction over the history, and reuses the transition for open-loop prediction under candidate actions. A frozen EIDOS backbone is used for providing a residual history prior through a separately trained adapter. Experiments are conducted on four simulation environments. Evaluation metrics cover factual prediction, intervention-effect prediction, offline action selection, and closed-loop MPC. Experimental results demonstrate first-rank performance in 12 of 16 settings. Ablation studies also show the effectiveness of the proposed designs.

Critical strengths

The valid architectural contribution is the combination of variable-specific latent states, cross-variable interactions, variable-specific action representations, and observation-based correction within a shared transition model. This provides a structured alternative to introducing future actions only through a forecasting readout: actions influence the evolution of the latent system, while observed measurements repeatedly update its inferred state. The resulting predictor can evaluate multiple action plans and downstream objectives without paired alternative-action supervision or objective-specific training. This implementation is well motivated for heterogeneous numerical time series. Its strong factual prediction and candidate-ranking results across four domains provide meaningful evidence of practical usefulness.

Critical weaknesses
  1. The paper discusses CEDAR, ThermoForce, and DoFlow, but its experiments compare against general forecasting models and RSSM/MTS3 rather than these recent approaches. Although the cross-domain evaluation is valuable, the methodological advance over the closely related literature remains insufficiently established. The component ablations demonstrate utility within ActTime, but do not resolve this comparative novelty question. A representative comparison under compatible supervision, together with a clearer account of methodological differences, would strengthen the contribution beyond its broader evaluation scope.

  2. The proposed method’s dependence on the training data is unclear. Training includes extra trajectories with specially chosen actions to cover more control settings, as depicted in Appendix A.1. The experiments do not show how much the gains in intervention prediction and planning depend on these examples.

Questions for authors
  1. What is ActTime’s principal methodological distinction from CEDAR and the other cited action-conditioned approaches? Could relevant methods be evaluated under the same factual-only protocol, or do specific structural assumptions or supervision requirements prevent a fair comparison?

  2. What proportion of the training data consists of these extra trajectories? Keeping the total training size fixed, does ActTime still outperform the strongest baselines when some are replaced with trajectories from the behavior policies?

5 comments
AuthorsOpenReview ↗

Response to Reviewer LLHW [Part 1]

Thank you for the careful review and constructive suggestions.

W1 & Q1: Differences from CEDAR, ThermoForce, and DoFlow, and comparisons under compatible supervision.

We have added CEDAR-StageI and AcRKN to the revised paper and conducted additional comparisons with ThermoForce and DoFlow.

(1) How is ActTime different from these methods?

CEDAR is closely related to our work. Its Stage-I Transformer also learns action-conditioned transitions and supports rollouts under planned actions. The main difference lies in how the system dynamics are represented. CEDAR models transitions through interleaved state and action tokens. In contrast, ActTime maintains a latent state for each observed variable, uses variable-specific action representations, and models interactions among variables within a shared transition. This design explicitly accounts for how different variables respond to the same action and influence one another.

ThermoForce uses HVAC-specific physical structure to model control responses, while DoFlow relies on a causal graph for intervention modeling. ActTime does not require either assumption. We have clarified these differences in Section 2 of the revised paper‾\underline{\text{Section 2 of the revised paper}}.

(2) Comparisons under the same factual-only setting.

Following your suggestion, we evaluated all four methods on our benchmarks. They use the same factual training data, available observations and actions, and evaluation protocol, without paired counterfactual outcomes or planning supervision.

We have added CEDAR-StageI and AcRKN to Table 2 of the revised paper‾\underline{\text{Table 2 of the revised paper}}. AcRKN retains its action-conditioned Kalman dynamics, while CEDAR-StageI retains the Action-Interleaved Transformer. We use Stage I because CEDAR's Stage II requires external event and text information unavailable in our benchmarks. The baseline adaptations are detailed in Appendix A.4 of the revised paper‾\underline{\text{Appendix A.4 of the revised paper}}.

For ThermoForce and DoFlow, we also evaluated adapted versions. Their original methods rely on HVAC-specific physical assumptions and a known causal graph, respectively, which are not available across our four domains. We report these adaptations as supplementary comparisons below.

DomainMetricActTimeCEDAR-IThermoForce (adapt.)AcRKNDoFlow (adapt.)
CityLearnFactual MAE0.2479 ± 0.00330.2706 ± 0.00510.3916 ± 0.00030.2860 ± 0.01280.2816 ± 0.0064
Effect NMAE0.0688 ± 0.00410.0991 ± 0.00880.1730 ± 0.00870.1051 ± 0.01090.0914 ± 0.0065
Offline Regret0.2842 ± 0.00560.2871 ± 0.05340.6603 ± 0.00350.2962 ± 0.03800.3813 ± 0.0208
Online Objective0.8346 ± 0.00260.8277 ± 0.00320.8362 ± 0.00130.8363 ± 0.01270.8533 ± 0.0035
SUMOFactual MAE0.1088 ± 0.00100.2259 ± 0.00320.4120 ± 0.00120.1882 ± 0.00200.1703 ± 0.0043
Effect NMAE0.5118 ± 0.03310.5894 ± 0.01290.9312 ± 0.00340.5530 ± 0.00310.5171 ± 0.0181
Offline Regret0.0251 ± 0.00270.0950 ± 0.00240.4642 ± 0.00070.0685 ± 0.00500.0584 ± 0.0048
Online Objective0.3087 ± 0.00750.3492 ± 0.02230.3472 ± 0.00390.3243 ± 0.00430.3245 ± 0.0059
GreenLightFactual MAE0.0522 ± 0.00210.1435 ± 0.01110.3338 ± 0.00790.1248 ± 0.00130.1491 ± 0.0050
Effect NMAE0.2192 ± 0.00890.5277 ± 0.03940.9404 ± 0.00970.5463 ± 0.01490.4706 ± 0.0286
Offline Regret0.1330 ± 0.01740.4898 ± 0.05820.9375 ± 0.23300.2725 ± 0.08530.1361 ± 0.0189
Online Objective0.2936 ± 0.00130.3021 ± 0.00670.3617 ± 0.01100.3071 ± 0.00410.3013 ± 0.0013
InventoryFactual MAE0.0798 ± 0.00070.1771 ± 0.00690.4366 ± 0.00240.1435 ± 0.00080.1589 ± 0.0030
Effect NMAE0.2280 ± 0.00330.3542 ± 0.00750.9071 ± 0.00680.3580 ± 0.00800.2368 ± 0.0100
Offline Regret0.1506 ± 0.01660.0868 ± 0.00770.3920 ± 0.02460.1277 ± 0.00360.0842 ± 0.0133
Online Objective3.9111 ± 0.03523.9735 ± 0.03254.6910 ± 0.03733.9418 ± 0.01313.8923 ± 0.0224

Results are mean ± std over three seeds. CEDAR-I denotes the adapted Stage-I model. ThermoForce and DoFlow are adapted versions of the original methods.

Among these five methods, ActTime achieves the lowest mean factual and intervention errors in all four domains. It also performs best on three offline and two online objectives. CEDAR-StageI performs better on CityLearn online planning, while the DoFlow adaptation performs better on Inventory decision tasks.

We also extended the cross-level alignment and candidate-action ranking analyses to include AcRKN and CEDAR-StageI. The ranking disagreements persist in the ten-method comparison, while ActTime achieves the highest candidate-cost Spearman correlation in all four domains. These results are reported in Sections 4.3 and 4.4 and Appendices B.2 and B.3 of the revised paper‾\underline{\text{Sections 4.3 and 4.4 and Appendices B.2 and B.3 of the revised paper}}.

AuthorsOpenReview ↗

Response to Reviewer LLHW [Part 2]

W2 & Q2: Proportion of additional training snippets and results with behavior-policy replacements.

(1) Proportion of additional snippets.

The additional snippets account for 14.6%, 29.1%, 8.9%, and 11.7% of the unique training windows in CityLearn, SUMO, GreenLight, and Inventory, respectively. We have added these details to Appendix A.1 of the revised paper‾\underline{\text{Appendix A.1 of the revised paper}}.

(2) Comparison with fewer snippets.

We replaced half or all of the additional snippets with new data generated by the behavior policies. We kept the total training size unchanged and retrained ActTime and five strong baselines at each level. The baselines include TFT, RSSM, MTS3, CEDAR-StageI, and AcRKN.

The results are listed below and in Appendix B.5 and Table 19 of the revised paper‾\underline{\text{Appendix B.5 and Table 19 of the revised paper}}. Each entry shows ActTime / the best baseline for that metric under the same setting.

DomainRetainedEffect NMAE ↓Offline Regret ↓Online Objective ↓
CityLearn100%0.0726 / 0.0511 TFT0.2826 / 0.2866 TFT0.8332 / 0.8230 RSSM
50%0.0785 / 0.0864 TFT0.2908 / 0.3009 CEDAR0.8375 / 0.8287 TFT
0%0.0811 / 0.0898 TFT0.3225 / 0.3378 TFT0.8391 / 0.8432 TFT
SUMO100%0.5011 / 0.4278 TFT0.0240 / 0.0324 TFT0.3073 / 0.3116 TFT
50%0.5234 / 0.4961 TFT0.0280 / 0.0304 TFT0.3103 / 0.3039 TFT
0%0.5381 / 0.5989 TFT0.0284 / 0.0372 TFT0.3140 / 0.3145 TFT
GreenLight100%0.2124 / 0.2608 TFT0.1360 / 0.1409 RSSM0.2938 / 0.2946 TFT
50%0.2432 / 0.2981 TFT0.1307 / 0.1765 RSSM0.2943 / 0.2979 MTS3
0%0.3774 / 0.4050 RSSM0.1866 / 0.1610 RSSM0.3002 / 0.3044 MTS3
Inventory100%0.2286 / 0.3078 TFT0.1502 / 0.0821 CEDAR3.8780 / 3.9307 TFT
50%0.3110 / 0.3555 RSSM0.1998 / 0.1688 CEDAR3.9502 / 3.9946 TFT
0%0.4383 / 0.4578 RSSM0.3462 / 0.2633 CEDAR4.0943 / 4.0768 RSSM

Replacing the additional snippets generally reduces performance, particularly on GreenLight and Inventory. However, ActTime still performs best in 8 of 12 settings with half the snippets retained and 9 of 12 settings when all are replaced. It also achieves the lowest Effect NMAE in all four domains without these snippets. The additional snippets are helpful, but ActTime's advantage over the baselines does not depend solely on them.

AuthorsOpenReview ↗

Follow-up on Our Rebuttal

Dear Reviewer LLHW,

Thank you again for your constructive comments. Following your suggestions, we added comparisons with closely related methods and experiments on the role of additional training snippets.

As the discussion period is coming to an end, we would appreciate hearing whether our revisions have addressed your concerns. We would be happy to answer any further questions.

Best regards, The Authors

Comments on the paper

1 comment
AuthorsOpenReview ↗

Summary of Revisions

We thank all three reviewers for their careful reviews and constructive suggestions.

ActTime studies a key challenge in controlled time series: accurate prediction under observed actions does not guarantee reliable predictions under alternative actions or effective decision making. We propose an action-conditioned world model trained only on factual trajectories and evaluate it from factual prediction to closed-loop planning.

We have revised the paper and conducted additional experiments to address the reviewers' concerns. The major revisions are summarized below.

  • Comparison with closely related methods (LLHW, ykhK): We added AcRKN and CEDAR-StageI to the main experiments, along with supplementary comparisons with adapted ThermoForce and DoFlow. We also clarified the methodological differences, particularly how ActTime models variable-specific states and action responses within a shared transition.
  • Results over multiple training seeds (q4UX, ykhK): We repeated the main experiments with three independent seeds and now report means and standard deviations. ActTime still ranks first in 12 of 16 settings across ten methods.
  • Evaluation on real-world datasets (q4UX): We added experiments on AGC 2019 and Causal Chambers Wind Tunnel, covering factual prediction, observed action responses, and decision-related costs from recorded runs. ActTime achieves the best results on all eight reported metrics.
  • Additional ablations and training comparisons (q4UX, ykhK): We added a parameter-matched shared-state ablation, repeated the EIDOS ablation over three seeds, and compared alternative feature aggregation and training strategies. The results support the variable-specific state design, while the original two-stage training performs best in 15 of 16 settings.
  • Dependence on additional training snippets (LLHW): We reported the proportion of additional snippets in each domain and tested replacing them with behavior-policy data while keeping the training size fixed. ActTime remains best in 8 of 12 settings when half are replaced and 9 of 12 when all are replaced.
  • Further analysis of decision quality and efficiency (ykhK): We extended the model-ranking analysis to ten methods and repeated it using the four strong models suggested by the reviewer. Among these four models, the ranking reversals persist at 26.6% and 35.0%. We also added MPC decision latency comparisons with RSSM, MTS3, and AcRKN.

The revisions are highlighted in blue in the updated manuscript. We look forward to the reviewers' feedback and would be happy to address any further questions.

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

Related papers

Loading the map…

Discussion (0)

Sign in to comment.