acceptodds
Under review as a conference paper at ICLR 2027

LaFree World Model: Label-Free Pretraining for Action-Conditioned JEPA World Models

33 views

Abstract

Learning world models from observation sequences could reduce dependence on synchronized action labels, but learning to explain observed transitions does not necessarily yield dynamics that support downstream control. We introduce LaFreeWM, an end-to-end action-conditioned JEPA world model that learns dynamics using discrete latent actions inferred from consecutive observations. A quantized action interface, together with latent transition objectives, couples action discovery and representation learning without action annotations. A small action-labeled subset subsequently adapts the pretrained world model to the environment's control inputs for planning. Across control benchmarks, LaFreeWM improves planning with limited action supervision while retaining a compact efficient planner. Ablations show the benefit of end to end training and reveal a prediction–planning mismatch: removing quantization substantially lowers held-out latent prediction error while degrading downstream planning. The results demonstrate the value of label-free pretraining for counterfactual planning with action-conditioned world models and highlight the importance of evaluating inferred-action representations through the control they enable.

Primary area:
unsupervised, self-supervised, semi-supervised, and supervised representation learning
Keywords:
self-supervised learning, world models, latent actions, joint-embedding predictive architectures, label-efficient learning, vector quantization, model-based planning

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_4oDvOpenReview ↗
Rating
1
Confidence
4
Informativeness
1
Presentation
2
Soundness
1
Read review
Paper summary

This work presents LaFreeWM, an extension to LeWM that can leverage action-free data for training the world model using learned discrete latent actions. The authors propose 3 additional losses for shaping the latent actions and an additional training stage for grounding the latent actions that enables planning in the true action space. LaFreeWM is compared to other JEPA world model variants including the base LeWM on 4 simulated manipulation and navigation environments.

Critical strengths

Thorough related work section on latent action models.

Detailed Appendix.

Critical weaknesses

Evaluation protocol is weak and results are not substantial.

Evaluating goals with a fixed offset implies they are chosen from a dataset. Assuming they are sampled from the training dataset, this does not test the most basic form of generalization. Furthermore, limiting the horizon of the goals to at most 15 steps cannot be considered ``long horizon’’ in my opinion. The fact that none of the methods tested in this study achieve good performance on some of these tasks makes it hard to attribute any gains in performance to fundamental properties of said methods.

In terms of reported results in Table 1, with 10% labels, LeWM’s performance does not degrade within a single standard deviation from the 100%-label results in all but TwoRoom. LaFreeWM only outperforms it beyond a single standard deviation on 2 of the 4 tasks. These results do not justify the complexity added to LeWM as I see it. With 1% labels the gains are more clear but this severely limits the data available to LeWM and thus the results are not surprising in this extreme scenario.

Ablation study is limited and not thoroughly explained.

The ablation study is limited in the sense that it focuses on a single task and label budget. The variants are not thoroughly explained other than their short names in Table 2 and are not motivated. The results of "Without JEPA loss’" and "Detached prediction gradient" are within a standard deviation of the best performing variant and are thus not justified. Overall, LaFreeWM adds significant complexity on top of LeWM but it is hard to tell which of the additions contribute to better performance, if at all.

Questions for authors
  • L_pred is not defined anywhere. How is it different from L_LAM?
  • Why is L_JEPA over the actions necessary? It seems that you are encouraging the latent action to represent a normalized delta between latent states. As I see it, the standard latent action formulation already encourages this and your ablation study does not support its significance for planning.
  • The fact that LaFreeWM with 10% action labels performs better than LeWM with the full data budget on some tasks suggests that something in the training objective improves performance regardless of the amount of action-labeled data available. What do you think that is?
9 comments

Thank you for your review.

Weaknesses: We understand the assumption, as previous approaches like LeWorldModel do sample from the training dataset. We also found that the difference in performance between training and test start and goal pairs is insignificant. However, in our case the start and goal positions are not sampled from the training dataset. We agree that this should be clarified in the paper.

LeWorldModel rolls out the world model for 5 prediction steps. We use 15, three times as many, and prediction errors compound over a longer rollout. SAGE [1] shows that the planning success of LeWM is sensitive to this horizon. On PushT it drops from 89.3% at the standard goal offset of 5 prediction steps to 56.0% at 10 steps and 8.7% at 20 steps. This could explain the lower performance of LeWorldModel and the other models in our evaluation. We agree that 15 steps is not necessarily long horizon and will drop that wording. The low numbers are concentrated at the hardest setting, h = 15 on PushT. At shorter horizons PushT is solved.

We also found the problem with the TwoRoom environment, an increased sensitivity to the variance regularisation (SIGReg) loss term. After re-evaluating all models, LaFreeWM reaches 96.7 ± 1.9 and 99.3 ± 0.9 on TwoRoom with 1% and 10% labels, against 36.7 ± 5.0 and 85.3 ± 2.5 for LeWM. We now outperform LeWM clearly in 3 out of 4 environments (with Cube being less significant). We believe that this and the ablations together justify the additional complexity.

We agree that label-only LeWM is expected to suffer with 1% of the labels, and that scarcity is the setting we target. The informative comparison is with methods that see the same unlabelled data. Pretraining on it with a constant action gives 26.7% on PushT at 10% labels, against 59.3% for LaFreeWM (TwoRoom 76.0 vs 99.3, Reacher 67.3 vs 70.7, Cube 80.0 vs 82.0), and Garrido et al. gives 13.3% at 1% labels, against 33.3%. With 1% labels LaFreeWM is also on par with LeWM trained on all labels on PushT (33.3 vs 30.0) and Reacher (67.3 vs 62.7).

We agree that the ablation text should better explain and motivate each variant, and we will revise it. We also expanded the ablations to the 3 other environments, and the change in success relative to the full model (500 test episodes) is shown below and will be put in the appendix.

EnvNo-JEPADetached Prediction gradientBoth
Reacher−2.6 ± 2.0−1.6 ± 1.9−3.4 ± 2.0
Cube−0.2 ± 1.30.0 ± 1.3−2.6 ± 1.4
TwoRoom−0.8 ± 0.6−1.4 ± 0.8−1.8 ± 0.7

Questions: Question 1: L_pred is shown in Figure 2(a) and appears in the full objective (Eq. 7) as the next-CLS MSE, but we never wrote out its definition. L_pred is the main learning loss of the latent world model (encoder, dynamics model and IDM). It minimizes the distance between the prediction of the dynamics model and the future latent states. L_LAM, in contrast, is purely for action shaping and is only seen by the IDM. It is the prediction loss of a separate network that takes the latent and the latent action as input and outputs the next latent.

Question 2: The two look similar, but the difference is in what each predictor gets to see. The LAM decoder receives the current state together with the latent action and predicts the next latent, so the latent action only has to be informative in combination with the state. The JEPA predictor receives only the latent action, without the state, and has to predict the normalized difference between the two latents, so the latent action on its own is encouraged to carry the direction of change. The first one encourages predictability of the future, the other one predictability of change. We also measured this on PushT, the average cosine similarity between the actual direction of latent state change and the mean direction of its assigned code is 0.58 for LaFreeWM and 0.28 for the No-JEPA ablation, so with the JEPA loss the code alone predicts the direction of change much better. On planning, the effect is small but consistent, removing the JEPA loss lowers success in all four environments, by 0.2 to 2.6 points, and lowers effective codebook use from 61.5 to 45.0 codes.

Question 3: Indeed, if you were able to perfectly predict the actions, you would expect more or less the same results as LeWM. The Separate IDM ablation shows that decoupling the encoder brings us closer to the LeWM performance. One hypothesis is that coupling the encoder with the IDM acts as an additional regularisation on the encoder, encouraging representations that encode more of the action information, which makes it easier for the dynamics model to predict the future latent. Figure 4 in the appendix indeed shows that the LaFreeWM encoder responds more clearly to the position of the pusher than the LeWM encoder does.

References [1] Cheng et al. SAGE: Subgoal-Conditioned Action Generation for Latent World Model Planning. arXiv:2607.17973, 2026.

I appreciate your effort in writing this response, answering my questions and running additional experiments.

I do not see a new revision of the ablation text so it is still hard to understand and review the ablation study. The additional ablation results you presented show marginal improvements.

Overall the paper applies latent action objectives to LeWM to improve its performance when less action labels are available. The novelty and significance of the additional losses is not made clear by the results and text.

The benefits of latent actions in learning from limited action-labeled data is well established by previous work. My opinion is that adding latent action objectives to LeWM is a limited contribution and therefore I maintain my score.

Thank you for reading our response and for the follow-up.

Our apologies for the delayed revision. We have now uploaded a revised PDF. The ablation text in Section 5.2 is rewritten, Table 1 now reports the re-tuned TwoRoom results, and the additional results are in Appendix C.4 and C.5.

Significance of the additional losses. The VQ bottleneck and the LAM loss are what make the latent actions useful. Without either of them, pretraining on the unlabelled data gives no gain over the constant action ablation (18.0% and 26.7% vs 26.7%). Without LAM the codebook is still fully used (61.2 effective codes), so the codes are there but they do not help planning. The JEPA loss has a different role. Its effect on planning is small (0.2 to 2.6 points) however positive for all environments, it mainly changes what the codes represent: the cosine similarity between a code and the direction of latent change goes from 0.28 to 0.58, and codebook usage from 45.0 to 61.5.

Novelty. We agree that latent actions are known to help with limited labels in policy learning and video generation. Our results show that this does not carry over by default to planning with an end-to-end JEPA world model. On the same unlabelled data, a LAPO-style IDM ends at the level of the constant action ablation, which uses no latent actions at all, on both PushT (23.3% vs 26.7%) and TwoRoom (76.7% vs 76.0%), where LaFreeWM reaches 60.7% and 99.3% (Table 6, Appendix C.4). On PushT the same holds for our adaptation of Garrido et al. (16.0%) and a continuous latent action (18.0%). So the result is not obtained by adding an existing latent action objective to LeWM.

The second finding is the coupling between encoder and IDM. Concurrent work [1, 2, 3] shows that an inverse-dynamics loss on the real actions keeps the encoder of a JEPA world model sensitive to actions, which helps planning. These methods need the label of every transition. We find a similar effect without labels: when the encoder is trained through the latent action IDM, success goes from 40.7% to 59.3%, and the encoder responds more clearly to the pusher position (Figure 4 in the appendix). This is also why we think the contribution is more than label efficiency: with 10% of the labels, LaFreeWM is above LeWM trained on all labels in all four environments (PushT 60.7 vs 30.0, TwoRoom 99.3 vs 92.7, Reacher 70.7 vs 62.7, Cube 82.0 vs 77.3; Table 1).

Third, the continuous variant has a 60 times lower prediction error and still plans worst, because the IDM can pass the future observation on to the predictor. Lower prediction error is therefore not a reliable signal when latent actions are learned together with the world model.

[1] Ivashkov et al. Sensorimotor World Models. arXiv:2606.20104, 2026. [2] Zhang et al. Delta-JEPA. arXiv:2606.31232, 2026. [3] Strohm et al. Keeping JEPA World Models Plannable When Little of the Frame Moves. arXiv:2610.03137, 2026.

Reviewer_3SucOpenReview ↗
Rating
2
Confidence
3
Informativeness
1
Presentation
1
Soundness
2
Read review
Paper summary

This paper proposes LaFreeWM, a label-free pretraining framework for action-conditioned JEPA world models. The method uses an inverse dynamics model and vector quantization to infer discrete latent actions from consecutive observations, together with auxiliary objectives for transition prediction. A small amount of labeled data is subsequently used to adapt the pretrained model for raw-action planning. Experiments on four simulated control environments show improvements in label efficiency and planning performance over several baselines. The authors also investigate the effects of different latent-action learning components.

Critical strengths
  1. Meaningful research direction. Learning action-conditioned world models from observation-only trajectories is a relevant problem, particularly for reducing reliance on expensive action annotations.

  2. Promising label-efficiency results. LaFreeWM outperforms the label-only LeWM baseline across all four environments. On PushT, it improves planning success from 26.0% to 60.7% using 10% action labels.

  3. Interesting prediction–planning discrepancy. The observation that lower latent prediction error does not necessarily translate into better planning performance raises a relevant question about learning action-conditioned representations for control.

Critical weaknesses
  1. Limited methodological novelty. Learning latent actions from unlabeled videos using IDM and vector quantization has already been explored in Genie [1], while Garrido et al. [2] jointly train an IDM and world model for downstream planning. LaFreeWM mainly extends these approaches through end-to-end encoder adaptation and additional prediction objectives. However, the paper does not convincingly demonstrate that these incremental modifications constitute a substantial technical contribution.

  2. Unclear contribution of latent-action pretraining. The learned IDM and VQ codebook are bypassed during action-labeled fine-tuning, where the encoder and predictor are further optimized using raw actions. Although the constant-action ablation provides partial evidence, it remains unclear whether the improvements come from latent-action discovery or general representation pretraining. A matched action-free JEPA pretraining baseline would help isolate this contribution.

  3. Insufficient evidence for key methodological claims. In Table 2, removing the JEPA loss reduces planning success only from 59.3% to 58.0%, while detaching the prediction gradient yields 56.0%, with differences small relative to the reported variability. Moreover, Section 5.3 primarily shows that inferred codes contain transition information, rather than demonstrating consistent control semantics. The reported prediction–planning discrepancy also compares latent MSE across independently learned representations, making the results difficult to interpret without a common evaluation space.

[1] J. Bruce et al., Genie: Generative Interactive Environments., Proceedings of the 41st International Conference on Machine Learning (ICML), PMLR 235:4603–4623, 2024.

[2] Q. Garrido, T. Nagarajan, B. Terver, N. Ballas, Y. LeCun, and M. Rabbat., Learning Latent Action World Models in the Wild., Proceedings of the 43rd International Conference on Machine Learning (ICML), PMLR 306:34113–34150, 2026.

Questions for authors
  1. Methodological novelty. The core latent-action learning and planning framework is closely related to Genie [1] and Garrido et al. [2]. Could the authors clarify the key technical novelty of LaFreeWM beyond combining existing components, and explain what fundamentally distinguishes the proposed approach from these prior methods?

  2. Contribution of latent-action pretraining. Since the IDM and VQ codebook are bypassed during raw-action fine-tuning, can the authors compare LaFreeWM against an action-free JEPA pretrained on the same unlabeled data, followed by identical fine-tuning? This would help isolate the benefits of latent-action discovery from general self-supervised pretraining.

  3. Effectiveness of the proposed components. Table 2 shows only modest improvements from JEPA transition shaping and end-to-end gradient coupling, with differences relatively small compared with the reported variability. Can the authors provide stronger evidence that these components consistently improve downstream planning across different environments?

3 comments

Thank you for your review.

Response to critical weaknesses, comments and questions: Question 1: We agree that both mentioned approaches are conceptually similar. However, Genie learns a generative world model to predict pixels, and Garrido et al. study how to learn actions and evaluate them on offline planning, but they do not demonstrate closed-loop control. Garrido et al. is the closest, which is why it is one of our baselines. The differences are specific. Genie learns discrete latent actions to generate video in pixel space and is not used for planning. Garrido et al. plan in latent space, but with continuous latent actions on top of a frozen V-JEPA 2 encoder. LaFreeWM trains the encoder jointly with the IDM and predictor, and passes the latent action through a discrete bottleneck. Our ablations (Table 2) show that these two changes are what makes label-free pretraining work for planning. Training the IDM as a separate network on detached encoder states, so that it no longer shapes the encoder (Separate IDM), lowers success from 59.3% to 40.7% and collapses the codebook to 3.7 effective codes. Replacing the discrete bottleneck with a continuous one lowers success to 18.0%, below the label-only variant of LeWM trained on a subset of the data, even though its prediction error is 60 times lower. In the direct comparison, LaFreeWM outperforms our adaptation of Garrido et al. in all four environments at both label budgets (10% labels: PushT 60.7 vs 16.0, TwoRoom 99.3 vs 92.7, Reacher 70.7 vs 61.3, Cube 82.0 vs 71.3). We also add two action shaping losses. The LAM loss raises success from 26.7% to 59.3%. The JEPA loss raises effective codebook use from 45.0 to 61.5 codes and also improves planning success in all four environments according to the new ablations. The paper studies the framework as a whole, shows two important aspects for closed-loop planning, and adds the novel JEPA loss for further latent action shaping.

Question 2: The constant action ablation in Table 2 shows the contribution of latent action pretraining: with constant actions, i.e. no latent actions, we get the same performance as LeWM. We expanded this ablation to the different environments to further demonstrate the benefit of latent actions. The results show that this ablation underperforms compared to LaFreeWM in all 4 environments, and we will add the 3 remaining environments to the appendix.

methodPushTTwoRoomReacherCube
Constant action26.7 ± 4.176.0 ± 7.567.3 ± 5.080.0 ± 3.3

Question 3: We agree that the benefit of both components should be shown better and we have therefore expanded the ablation. We ablated each component, i.e. we removed the JEPA loss on the latent actions and detached the prediction gradient of the dynamics model, as well as the combination of the two, on the remaining environments and, to decrease variability, increased the number of evaluated test episodes to the maximum of 500. We will put these extra ablations in the appendix.

EnvNo-JEPADetached Prediction gradientBoth
Reacher−2.6 ± 2.0−1.6 ± 1.9−3.4 ± 2.0
Cube−0.2 ± 1.30.0 ± 1.3−2.6 ± 1.4
TwoRoom−0.8 ± 0.6−1.4 ± 0.8−1.8 ± 0.7

We also did an additional experiment as further evidence of the prediction–planning discrepancy. We shuffle the inferred latent actions across transitions and measure how much prediction quality drops, as R². R² is relative to each model's own target variance, so it doesn't depend on the scale of the latent space. For No-VQ, R² falls from 0.9999 to −0.38, so its predictor collapses. For LaFreeWM, R² goes from 0.992 to 0.984, showing a smaller reduction. So the continuous No-VQ code carries the future observation itself, and its low MSE comes from copying that information rather than from modelling dynamics.

Regarding Section 5.3: we do not claim consistent control semantics. The codes are never used for control, as grounding replaces the IDM and codebook with raw actions. Their role is to make the predictor action-conditioned during pretraining and to learn the environment dynamics without labeled actions, which the constant action ablation confirms (26.7% vs 59.3% on PushT). We will reword Section 5.3 accordingly.

Note on TwoRoom Environment We also identified the problem with the TwoRoom environment mentioned in the paper, which was due to an increased sensitivity to the variance regularisation (SIGReg) loss term. We have re-evaluated all models with their updated regularisation hyperparameter, using the same training/validation/test split. These are the values that will be updated in Table 1:

MethodLabelsTest success
LaFreeWM1%96.7 ± 1.9
LaFreeWM10%99.3 ± 0.9
Garrido et al. (Noisy)1% / 10%90.0 ± 5.9 / 92.7 ± 5.0
LeWM1% / 10%36.7 ± 5.0 / 85.3 ± 2.5
PLDM100%80.0 ± 0.0
DINO-WM100%99.3 ± 0.9

Thank you again for your review. We have uploaded a revised version of the paper that contains the changes we promised in our response:

  • The ablation text in Section 5.2 now explains each variant and what it shows.
  • The constant action ablation in all four environments is in Table 6 (Appendix C.4), together with a LAPO-style IDM.
  • The ablations of the JEPA loss and the prediction gradient in the three other environments are in Table 7 (Appendix C.4), followed by the shuffled-action experiment.
  • Section 5.3 is also updated with respect of consistent control semantics for the codes.
  • Table 1 now reports the updated TwoRoom results.

With respect to Genie and Garrido et al., the related work and Section 5.2 now also discuss the coupling between encoder and IDM. Concurrent work (e.g. Ivashkov et al., arXiv:2606.20104; Strohm et al., arXiv:2610.03137) shows that an inverse-dynamics loss on the real actions keeps the encoder of a JEPA world model sensitive to actions. We find a similar effect with latent actions, without action labels, which is not possible with a frozen encoder.

We hope that the revision addresses your concerns and that you would consider updating your score. We are available for further discussion.

Thank you for the additional experiments and clarifications. While I appreciate the authors' efforts, I still have some concerns about the methodological novelty. Therefore, I prefer to maintain my original score.

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

This paper describes LaFreeWM, a JEPA based world model who uses an inverse dynamics model for learning discrete latent action from consecutive observations. The method outperforms existing methods on three of four planning benchmarks.

Critical strengths

The paper is relevant to ICLR and presents a label free approach for world model learning. Further, the analysis of effective codebook usage is good.

Critical weaknesses
  • In appendix 4- the authors state "The action-label-free world model and fully supervised baselines train for two epochs. But Label-only LeWM trains for 10 epochs matching the fine-tuning schedule of LaFreeWM". Why is this discrepancy?

  • The paper has similarities to Learning to Act without Actions https://iclr.cc/virtual/2024/poster/17672 in that they both watch consecutive frames and use a network to guess between steps - but it has not been cited and neither compared. even if the forward model for using to predict next frame is different a comparison should have been provided.

  • As the authors describe in appendix A.4 - "Table 1 uses different K values per environment as the size of the information bottleneck is optimal for different environment" - however, this specific hyper parameter selection seems to not have been provided to label only LeWM. To get a full sense of table 1's claim all baselines should receive similar hyper parameter tuning benefits.

Questions for authors

How does your work compare to the Learning to Act without Actions https://iclr.cc/virtual/2024/poster/17672

2 comments

Thank you for your review.

Response to critical weaknesses, comments and questions: The two schedules are matched rather than different. Label-only LeWM is the variant that is trained only on the 1% or 10% labelled subset, and its training is matched to LaFreeWM during fine-tuning. The two epochs refer to pretraining, i.e. the label-free pretraining of LaFreeWM and the constant action ablations. The LeWM variant that is also trained on the rest of the data without labels is the constant action ablation in Table 2. We will make this clearer in the appendix.

Thank you for pointing this out. Conceptually, Learning to Act without Actions (LAPO) [1] is indeed similar and we will add it to the related work section. A direct comparison is however more difficult, because LAPO learns a pixel-to-pixel dynamics model and thus cannot plan in the latent space like LaFreeWM and the baselines we compare to. Its framing is also different, in that LAPO is meant to learn an IDM to help policy learning in the RL setting, whereas we learn a latent world model with the help of an IDM. We did however run an experiment where we incorporated the LAPO-learned IDM in our world model to see if our IDM works better. We find that LaFreeWM consistently outperforms this approach, and we will add the additional results to the appendix. Here are the results:

labelsmethodPushTTwoRoomReacherCube
10%LAPO IDM23.3 ± 5.276.7 ± 9.066.7 ± 8.475.3 ± 7.4

K is the bottleneck of the codebook and is thus only relevant for the LaFreeWM framework; LeWM has no codebook and thus no value K. This bottleneck is dependent on the environment, as different environments have different amounts of complexity with respect to their actions. However, the gain does not depend on it: for example, with a single K = 16 for all environments, LaFreeWM stays above label-only LeWM at 10% labels in all four (PushT 60.7 vs 26.0, TwoRoom 99.3 vs 85.3, Reacher 66.7 vs 60.7, Cube 79.3 vs 76.7; Table 4 in the appendix). K is also the only hyperparameter we selected per environment. All hyperparameters that LaFreeWM shares with LeWM (learning rate, weight decay, batch size and SIGReg weight) are kept at the defaults of the official LeWM implementation, so they were tuned for the baseline rather than for our method. The only one we changed is the SIGReg weight on TwoRoom, because of the sensitivity described below, and there all methods received the same tuning. We will clarify this in the appendix.

Note on TwoRoom Environment We also identified the problem with the TwoRoom environment mentioned in the paper, which was due to an increased sensitivity to the variance regularisation (SIGReg) loss term. We have re-evaluated all models with their updated regularisation hyperparameter, using the same training/validation/test split. These are the values that will be updated in Table 1:

MethodLabelsTest success
LaFreeWM1%96.7 ± 1.9
LaFreeWM10%99.3 ± 0.9
Garrido et al. (Noisy)1% / 10%90.0 ± 5.9 / 92.7 ± 5.0
LeWM1% / 10%36.7 ± 5.0 / 85.3 ± 2.5
PLDM100%80.0 ± 0.0
DINO-WM100%99.3 ± 0.9

References [1] Schmidt, D. and Jiang, M. Learning to Act without Actions. ICLR 2024. https://iclr.cc/virtual/2024/poster/17672

Thank you again for your review. We have uploaded a revised version of the paper that contains the changes we promised in our response:

  • The training schedules of label-only LeWM and LaFreeWM, the role of K and the shared hyperparameters are clarified in Appendix A.4.
  • LAPO is added to the related work, and the results with a LAPO-style IDM are in Table 6 (Appendix C.4), next to the constant action ablation in all four environments.
  • Table 1 now reports the updated TwoRoom results.

We hope that the revision addresses your concerns and that you would consider updating your score. We are available for further discussion.

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

Related papers

Loading the map…

Discussion (0)

Sign in to comment.