acceptodds
Withdrawn from ICLR 2027

Correct Action Timing Is Not Action Fidelity in World Models

Withdrawn from ICLR 2027, so left out of the paper lists and the map.

Abstract

World models used for planning must predict how changing an action changes the future. Under non-anticipation and a nonzero response moment, we prove that an external randomized code identifies action timing from its rightmost nonzero code-transition moment. Even with exact timing, a restricted predictor at the population-risk optimum can estimate action effects less accurately than predicting no effect. An exact paired-risk identity gives an architecture-independent sufficient prediction-accuracy condition at the action-effect scale. Controlled visual experiments recover all 18 offset families, while only 7 of 36 aligned models beat zero response after 12 epochs. This distinction persists with longer controlled training and larger models on public Walker/Walk. Response weighting, motivated by the identity, improves eleven of twelve controlled comparisons and both Walker architectures under matched labels.

Primary area:
learning on time series and dynamical systems
Keywords:
World Models, System Identification, Action Fidelity, Counterfactual Prediction

Reviews

Among:
Shape:
1234this paper, Bottom 10%
Figure 1. Mean review rating of the reviewed papers at ICLR 2027. This paper averages 1.67 over 3 reviews, above 10% of the others.
Reviewer_F6uqOpenReview ↗
Rating
1
Confidence
3
Informativeness
1
Presentation
1
Soundness
2
Read review
Paper summary

This paper focuses on actions in world models in two ways. Firstly authors propose a method to identify a temporal shift between actions and visual using small perturbations. Secondly, authors show that under certain assumptions, the action effect estimated by a model too simple can be worse than no effect at all, and provide a numerical constraint as a condition to ensure that action effect is more accurate than no effect. Authors run experiments in simple environments such as pendulum and cart-pole with a synthetic, constant, action lag using different model architectures, as well as a larger experiment on the Walker/Walk simulator. Results show that the action lag is identified correctly and that while most models’ prediction beat copying the next frame, they struggle to model the difference between actions accurately.

Critical strengths
  • World models typically focus their evaluation more heavily on visuals. It is nice to see work that focuses on actions and how action modelling impacts world models beyond visual quality.
  • The work comprises a lot of experiments that are very thorough in the setup considered. For example, experiments comprise multiple seeds and multiple architectures.
  • The work is theoretically grounded in toy scenarios.
Critical weaknesses
  • My main concern is the presentation of this work. The paper is extremely difficult to read and understand. The main paper is full of jargon and equation variables which are not introduced or explained. For example, variables \pi, and U in equation (1) are never introduced. Variable q is in equation (1), yet its purpose is never discussed, only the challenges associated with using q are discussed in section 3.3. Most of the methodology has to be inferred from context. While the appendix is more readable to some extent and provides some clarifications, this is far from enough to make the paper readable and understandable. The abstract and Figure 1 are great examples of this: they cannot be understood until the full paper has been read.
  • The action lag experiments are sensible but don’t seem broadly applicable beyond a toy environment. It would require access to the data collection procedure to add perturbation and assume action offsets are constant throughout the trajectories. A concrete example of how this would happen and be used in practice would be useful.
  • The action fidelity experiments have very narrow findings. There is no evidence that the models evaluated here have been trained sufficiently. It is uncertain whether this is a problem of undertraining, lack of capacity, insufficient data, or if action fidelity is poor, despite the model having the ability to generate high quality visuals.
  • The core message of the paper seems to be that fixing action timing does not guarantee action fidelity, which doesn’t seem like a groundbreaking finding? Temporally aligned data does not guarantee that complex dynamics can be learned accurately.
Questions for authors
  • Can authors reframe their contributions and core findings in simpler, understandable terms and provide intuition?
  • How could the action lag experiment be used in practice on real data?
  • Is the main objective of action fidelity experiments to show that actions can be inaccurate while visuals are good quality (thus highlighting the important of evaluating action quality), or are authors aiming for a different message?
Reviewer_9HLAOpenReview ↗
Rating
2
Confidence
3
Informativeness
2
Presentation
2
Soundness
3
Read review
Paper summary

This paper asks whether world models learn the effects of actions once actions and video frames are correctly aligned in time. The authors propose a timing alignment method and show that standard prediction loss can still lead to poor action responses, even with correct timing. Experiments in simulated environments support this finding. Giving more weight to action differences reduces response error, but benefits for planning and control remain untested.

Critical strengths
  1. The question matters for action-controlled video generation. A video can look reasonable while showing the wrong effect of an action. Comparing outcomes under different actions helps reveal errors that standard prediction metrics may miss.

  2. The analysis offers useful insight into training objectives. The paper separates errors shared by both predictions from errors in the difference between them. Comparing training losses with the same paired data also helps show the value of directly supervising action effects.

Critical weaknesses
  1. The experiments are still far from practical video world models. Most tests use low-resolution images and one-step feature prediction, without long videos or complex interactions. The Walker experiments broaden the evidence, but it remains unclear how strongly these findings apply to larger generative models. The theoretical example shows that failure is possible, but does not establish why the tested networks fail.

  2. Lower response error has not yet been linked to better physical predictions or control. More accurate visual differences do not necessarily mean more accurate contact, motion, or long-term behavior. On Walker, response weighting still does not beat the zero-response baseline, and there are no planning experiments to show practical benefits.

Questions for authors
  1. Does poor action response persist with stronger models and more training data? What evidence separates a problem with the prediction loss from limits in model capacity or training?

  2. Does lower response error lead to better results in practice? Could visual examples or control experiments show whether the model predicts action outcomes more accurately and helps complete tasks?

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

The paper explores why world models may fail to accurately predict how actions affect future observations, even when actions are correctly aligned with video frames. The assumption here is that the training data may have incorrect action–video correspondence and fixing the action timing may help world model accurately predict how actions affect future observations. The papers explores this from the perspective of two challenges: 1) action timing, which identifies when an action occurs in to the video, and 2) action fidelity, which measures whether the model correctly predicts the consequences of that action. The method explores two approaches to address these challenges. For action timing correction, random perturbations are introduced into the action during data collection, and uses correlations with visual motion to correct the action timesteps. Next, the world model is trained to predict the outcome two different actions from the same initial state and error is minimized using the difference between those outcomes. The results section discusses that the method correctly recovers all 18 action-timing offsets across three controlled environments. However, only 7 of 36 correctly aligned models predict action effects better than a baseline that assumes actions have no effect. Overall the contribution and key takeaway of the paper is that correctly aligning actions with video frames does not guarantee that a world model may learn accurate consequences given an action and the current visual state.

Critical strengths
  1. The paper clearly shows that accurate visual prediction given an action and aligning actions with observations are two distinct capabilities of the model, and the accuracy of action timing may not guarantee that world world models accurately predict the effects of those actions.

  2. The proposed right-boundary registration method recovers all 18-action offset configurations and there is a interesting finding that only 7 of 36 correctly action-video aligned models outperform the baseline, showing that a world model can predict future frames well but may still fail to understand how actions affect the environment.

Critical weaknesses
  1. There are some limitations with the method of action-timing correction. First, the method requires injecting random perturbations into actions during data collection. However, since many world models have been trained on existing action-video datasets, there is a concern of how difficult or easy it is for the action-timing correction to be incorporated on naturally collected datasets. Second, the method makes as assumption that action timestamps have a constant offset. However, if the world model is run in a real-world environments, delays can vary over time. That raises a concern whether the proposed method can address variable or even unpredictable delays in the real-world.

  2. The evaluation has been conducted on relatively simple environments, making it unclear how well the proposed method would perform in complex or real-world scenarios. The evaluation has been conducted on three controlled environments (Pendulum, CartPole, and PointMass2D) using small CNN-GRU and Transformer models. Although the paper also shows evaluation on Walker/Walk with larger models, the method can be tested on more complex scenarios or real-world scenarios.

  3. The response-aware training requires the dataset which has the information what happens when two different actions are applied to exactly the same initial state and the dataset in this paper has been collected by resetting a simulator. However, there is a concern that such paired counterfactual data can be difficult to collect for real-world environments and it shows limited practical application of the training method.

Questions for authors
  1. Would training the models for more epochs or using larger models improve action fidelity without requiring additional response-aware training?
  2. Would improving action fidelity also help world models make better decisions in planning and control tasks?

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

Related papers

Discussion (0)

Sign in to comment.