acceptodds
Under review as a conference paper at ICLR 2027

ATM: Why Latent World Models Can Fail to Plan

17 views

Abstract

Latent world models can achieve accurate latent prediction yet still differ substantially in downstream planning performance. We argue that a key source of this discrepancy lies in the structure of action-induced latent transitions. We formalize action-identifiability through Bayes inverse risk, characterizing how much uncertainty about an action remains after observing the transition it induces. Model-predicted transitions can become highly self-decodable while encoding a domain-specific action relationship that fails to transfer to real environment transitions. We characterize this mismatch through cross-domain inverse transfer, instantiated as the Action-Consistency Transfer Matrix (ATM), a inverse-risk matrix over real and predicted transition domains. Across TwoRoom, PushT, and OGBench-Cube, true-transition action-identifiability tracks downstream planning substantially better than standard prediction loss, while the full ATM reveals highly self-decodable yet cross-domain-inconsistent predicted transitions. Controlled interventions on the two domains further produce distinct transition structures and planning outcomes, supporting this decomposition. The same diagnostics also support lightweight model screening, reaching 98.81% pairwise ranking accuracy for candidates separated by more than 5% success.

Primary area:
applications to robotics, autonomy, planning
Keywords:
Latent World Models, Action-Identifiability, Action-Conditioned Transitions, Transition Consistency, Cross-Domain Inverse Transfer, Goal-Conditioned Planning

Reviews

Among:
Shape:
1234this paper, Top 30%
Figure 1. Mean review rating of the reviewed papers at ICLR 2027. This paper averages 2.67 over 3 reviews, above 70% of the others.
Reviewer_exuCOpenReview ↗
Rating
3
Confidence
4
Informativeness
2
Presentation
3
Soundness
2
Read review
Paper summary

This paper investigates why latent world models with similar prediction accuracy can have different planning performance. To understand this, the authors propose ATM, a diagnostic framework that evaluates action-identifiability and cross-domain consistency between real and predicted latent transitions. Following LeWM, the paper also introduces AITS and AITS-P, which add inverse dynamics supervision to real and predicted transitions, respectively. The experiments show promising correlations between ATM diagnostics and planning success, as well as improvements on several planning tasks.

Critical strengths
  1. The distinction between action-identifiability and cross-domain consistency is interesting and provides a useful perspective on why latent world models can fail to plan despite accurate predictions.
  2. The experiments are clear and show promising results. In particular, the improvement from 74% to 84% on OGBench-Cube and the strong correlation with planning success support the practical relevance of the proposed analysis.
Critical weaknesses
  1. AITS is very similar to the inverse dynamics supervision proposed in Sensorimotor World Models: Perception for Action via Inverse Dynamics (Ivashkov et al., 2026). Both methods encourage recovering actions from consecutive latent representations. The novelty of AITS is therefore unclear, and the paper should better distinguish its contribution from this existing work.
  2. Although ATM provides an interesting analysis, the additional benefit of cross-domain consistency over true-transition action-identifiability is relatively small. The Spearman correlation only improves from 0.8130 to 0.8239 after adding the cross-domain term. More evidence is needed to demonstrate why the full ATM is necessary rather than simply evaluating inverse dynamics prediction loss.
Questions for authors
  1. What is the key difference between AITS and the inverse dynamics supervision in Sensorimotor World Models? I would like to see a clear comparison with this related work.
  2. Can the authors provide stronger evidence that cross-domain consistency is necessary for evaluating planning performance, rather than simply using a true-transition inverse dynamics loss?
3 comments

We thank the reviewer for recognizing the value of distinguishing action-identifiability from cross-domain consistency and for highlighting the encouraging planning results. We address both concerns below.

1. Relationship Between AITS and Sensorimotor World Models

We appreciate the reviewer's observation regarding the similarity between AITS and Sensorimotor World Models (SMWM). We agree that both employ related supervised inverse-dynamics objectives. We also note that the two works emerged within a short period, with ATM's initial public version, including AITS, appearing slightly earlier. To preserve double-blind anonymity, we refrain from linking to the earlier public version here.

Nevertheless, their research objectives differ. SMWM uses inverse-dynamics supervision to learn action-aligned representations and mitigate representation collapse. In our work, AITS serves as a controlled intervention within the ATM framework, designed to examine how improving true-transition action-identifiability affects cross-domain structure and planning.

Crucially, we also introduce AITS-P, which supervises predicted transitions instead, and evaluate both interventions individually and jointly. This allows us to examine how supervision in different transition domains produces distinct ATM signatures and planning outcomes, without modifying the planner.

Furthermore, ATM evaluation uses two freshly trained, independent inverse probes on frozen world models, rather than the auxiliary heads used during training. This separates diagnostic measurements from intervention objectives.

Thus, our primary contribution is not the inverse-dynamics loss in isolation, but the formulation and diagnosis of cross-domain action-transfer structure, together with controlled interventions that reveal its distinct properties.

2. Why the Full ATM Matters Beyond True-Domain Inverse Risk

We acknowledge that adding the cross-domain penalty to the scalar screening score yields a modest Spearman improvement on OGBench-Cube (0.8130 to 0.8239). However, this scalar score summarizes only part of ATM; its marginal ranking improvement does not reflect the information provided by the complete bidirectional matrix.

We have added three complementary pieces of evidence.

First, the new Rope Manipulation experiments (Sec. 4.6, Fig. 3) evaluate 400 checkpoints with closed-loop CEM planning. Cross-domain inverse risk −DT,P-D_{T,P} achieves 0.786 Spearman correlation with planning quality, substantially exceeding 0.495 for −DT,T-D_{T,T}. It also achieves 0.571 Pearson, compared with 0.180 for prediction loss, indicating a much stronger linear relationship with planning quality. Fig. 3 further illustrates how prediction loss compresses high-performing checkpoints into a narrow numerical range, whereas cross-domain risk provides clearer discrimination.

Second, Table 4 demonstrates the additional information captured by the complete matrix. Two Rope checkpoints differ by only 1.34% in DT,TD_{T,T} and 0.20% in DT,PD_{T,P}, yet their reverse-transfer risks DP,TD_{P,T} differ markedly (9.750 vs. 4.749), alongside substantially different planning outcomes (CD 2.786 vs. 1.308). The better checkpoint outperforms the other on all 10 evaluation cases. The reverse-transfer direction thus reveals structural differences that are difficult to identify from true-domain inverse risk or one-way transfer alone.

Third, we provide a formal explanation in Sec. 3.3 (Eq. 6) and Appendix B.3. Under squared loss, the Bayes-optimal cross-domain excess risk satisfies

= \mathbb{E}_{\mathcal{D}_T} \left[\|h_P^{*}(\xi)-h_T^{*}(\xi)\|_2^2\right].$$ This characterizes disagreement between the inverse rules of the two domains, which cannot be determined from their in-domain risks alone. Appendix B.3 further constructs **two world models with identical prediction loss, $D_{T,T}$, $D_{T,P}$, and $D_{P,P}$, but different $D_{P,T}$ and planning errors**. This formally demonstrates that even one-way transfer can miss planning-relevant structural differences. **Together, these results establish that the full ATM provides non-redundant diagnostic information beyond true-transition inverse loss, rather than merely improving a single scalar correlation.**

Follow-up: Highlighted Revisions and Evidence for the Full ATM

Thank you again for your thoughtful review and for recognizing the value of ATM. We have uploaded an updated manuscript, with new and revised content highlighted in blue for easier reading.

We would particularly like to highlight two additions.

1. Clarification regarding SMWM

We have expanded Related Work to explicitly compare ATM with SMWM and DynaMo.

SMWM uses inverse-dynamics supervision to learn action-aligned representations and prevent collapse. In our work, AITS and AITS-P instead serve as controlled interventions to study supervision in real versus predicted transition domains. ATM's main contribution is the independent bidirectional diagnostic framework, rather than the inverse-dynamics loss alone.

2. Why the full ATM matters

Our new Rope Manipulation experiments (Sec. 4.6) evaluate 400 checkpoints with closed-loop CEM planning. Cross-domain inverse risk −DT,P-D_{T,P} achieves Spearman 0.786 with planning quality, compared with 0.495 for −DT,T-D_{T,T}.

More importantly, the new Table 4 illustrates the additional information captured by the complete matrix:

Rope checkpointDT,TD_{T,T}DT,PD_{T,P}DP,TD_{P,T}DP,PD_{P,P}CD ↓
AITS-P / 230.3411.4519.7500.0532.786
Both / 310.3361.4544.7490.0181.308

These checkpoints differ by only 1.34% in DT,TD_{T,T} and 0.20% in DT,PD_{T,P}, yet have substantially different reverse-transfer risks and planning outcomes. The latter performs better on all 10 evaluation cases.

We also added a formal risk decomposition in Sec. 3.3 (Eq. 6) and a constructive counterexample in Appendix B.3, where two models have identical prediction loss, DT,TD_{T,T}, DT,PD_{T,P}, and DP,PD_{P,P}, but different DP,TD_{P,T} and planning errors.

Thus, the full ATM provides structural information beyond within-domain inverse loss and even one-way transfer, rather than merely improving a scalar correlation.

Thank you again for your time and valuable feedback. We would be happy to discuss any remaining questions or concerns.

If you feel that our revisions and responses have sufficiently addressed your concerns, we would greatly appreciate it if you would consider increasing your rating.

I appreciate the authors' responses and summary. These address my concerns, and I acknowledge the contributions of this paper. However, since the experiments are conducted exclusively in simulated environments, the applicability of the proposed method to real-world settings remains unclear. Therefore, I will maintain my current score.

Reviewer_QAV2OpenReview ↗
Rating
3
Confidence
3
Informativeness
2
Presentation
2
Soundness
2
Read review
Paper summary

This paper studies the consistency of the action influence across real transitions and real-to-latent transitions learned by latent world models. The authors train an inverse-action probe for each case and evaluate it in these two scenarios. Their experiments include datasets including TwoRoom, PushT, and OGBench-Cube. They also introduce inverse-dynamics training losses and employ ATM measurements to rank model checkpoints as an intervention.

Critical strengths
  1. The question under investigation is interesting and less understood in the field -- testing whether the decoding relationship transfers to real transitions is more informative than checking predicted-domain action recovery alone. Indeed, the empirical evaluations shed new light on the existing latent world models' capabilities, which could be valuable to the community.

  2. The method is practical and simple. The probe is small (3-5s per checkpoint after caching the transitions) and thus practical for latent world model researchers to adopt for diagnosing their model design.

Critical weaknesses
  1. The same task pool is used for the score fitting; it is advisable to split the pool into training and evaluation subsets. Similarly, the pair count does not genuinely indicate the number of independent observations because many pairs share checkpoints.

  2. The intervention results do not clearly demonstrate the relationship between transfer and planning. For instance, predicted-transition shows a large matrix asymmetry but still boosts TwoRoom from 87% to 94% and PushT from 96% to 100%. Similarly, real-transition supervision exhibits such behavior.

Questions for authors

Please address the weak points above.

2 comments

We thank the reviewer for recognizing the practical value and efficiency of ATM. We address the two concerns below.

1. Score Fitting and Pairwise Dependence

We appreciate the reviewer's concern regarding score fitting and statistical dependence. Importantly, we deliberately restrict the fitting capacity to a single scalar coefficient λ\lambda. The scoring function SATM=−DT,T−λ∣GT→P∣S_{\mathrm{ATM}}=-D_{T,T}-\lambda|G_{T\to P}| is fixed in advance, and its underlying ATM quantities are computed independently of planning success labels. Unlike a learned ranking model, it has no flexible nonlinear mapping or checkpoint-specific fitted parameters.

Even with this highly constrained calibration, the score achieves 82.84% pairwise ranking accuracy across all 1,865 comparisons, increasing to 98.81% for the 841 pairs with success gaps exceeding 5%. These results provide meaningful evidence that ATM captures planning-relevant ordering information despite the limited fitting freedom.

We acknowledge that λ\lambda was fitted on the same candidate pool, so these figures describe in-pool ranking performance, not held-out generalization. We also agree that pairs sharing checkpoints are not independent observations: the reported pair counts represent ranking comparisons rather than independent experimental replicates, and no independent-sample significance claim is made. These qualifications concern the interpretation of the screening results, rather than the underlying diagnostic formulation.

2. Interventions, Transfer, and Planning Performance

We appreciate the reviewer's observation that AITS-P can improve planning despite substantial matrix asymmetry. This is consistent with the diagnostic hierarchy explicitly defined in Section 3.4. ATM distinguishes multiple transition properties with different diagnostic priorities: true-domain action-identifiability is primary, cross-domain transfer is secondary, and matrix asymmetry is tertiary. High asymmetry should therefore not override improvements in other, more important properties.

Indeed, Table 2 shows that AITS-P changes several diagnostics simultaneously. On TwoRoom, DT,TD_{T,T} decreases from 80.26 to 79.36 and DP,PD_{P,P} from 80 to 8, while success increases from 87% to 94%. On PushT, DT,TD_{T,T} decreases from 18.29 to 17.34 and DP,PD_{P,P} from 17 to 1, alongside a success improvement from 96% to 100%. These improvements can coexist with substantial matrix asymmetry.

The distinction becomes more apparent on OGBench-Cube: AITS achieves substantially better true-domain identifiability than AITS-P (DT,TD_{T,T}: 15.87 vs. 33.77) and higher planning success (84% vs. 78%). Thus, matrix asymmetry alone does not determine planning performance, nor is it intended to.

Our new Rope experiments further support the value of cross-domain diagnosis. Across 400 checkpoints, −DT,P-D_{T,P} achieves Spearman 0.786 with planning quality, compared with 0.495 for −DT,T-D_{T,T}. Moreover, Table 4 identifies two checkpoints with nearly identical DT,TD_{T,T} and DT,PD_{T,P}, but markedly different reverse-transfer risks (DP,TD_{P,T}: 9.750 vs. 4.749) and planning outcomes (CD: 2.786 vs. 1.308).

These findings reinforce ATM's central purpose: to characterize complementary properties of action-induced transitions, rather than treat any single matrix statistic as a necessary or sufficient predictor of planning success.

Follow-up: Highlighted Revisions and Additional Evidence

Thank you again for recognizing the value and practicality of ATM and for your constructive feedback. We have uploaded an updated manuscript, with new and revised content highlighted in blue for easier reading.

We would like to briefly emphasize two points.

1. Score fitting and statistical dependence

As clarified in our earlier response, ATM fits only a single scalar coefficient λ\lambda within a fixed scoring function, strictly limiting its fitting capacity. The original ranking results are in-pool evaluations, and the pair counts in Table 3 represent comparisons rather than independent observations. We agree with this statistical distinction, while emphasizing that the results still demonstrate meaningful ranking ability under highly constrained calibration.

2. Interventions, asymmetry, and planning performance

We particularly invite you to revisit Table 2. The following results illustrate how AITS-P improves other diagnostics and planning success despite increased matrix asymmetry:

TaskVariantDT,TD_{T,T} ↓DP,PD_{P,P} ↓Asymmetry ↓Success ↑
TwoRoomBaseline80.2680.000.9587%
TwoRoomAITS-P79.368.0072.4394%
PushTBaseline18.2917.0016.9496%
PushTAITS-P17.341.0072.83100%

ATM values follow the ×100 scaling in the original Table 2.

These results are consistent with the diagnostic hierarchy in Sec. 3.4: true-domain action-identifiability is primary, cross-domain transfer is secondary, and matrix asymmetry is tertiary. Multiple properties can change simultaneously, so increased asymmetry does not necessarily imply worse planning performance.

Furthermore, our new Rope Manipulation experiments (Sec. 4.6) show that across 400 checkpoints, cross-domain inverse risk achieves Spearman 0.786 with planning quality, compared with 0.495 for true-domain inverse risk. The new Table 4 also illustrates structural differences revealed by the full matrix but missed by one-way diagnostics.

These findings reinforce ATM's central purpose: to characterize complementary properties of action-induced transitions, rather than treating any single metric as a decisive predictor of planning success.

Thank you again for your time and thoughtful feedback. We would be happy to discuss any remaining questions or concerns.

If you feel that our revisions and responses have sufficiently addressed your concerns, we would greatly appreciate it if you would consider increasing your rating.

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

The paper proposes a method to diagnose a world model before actually deploying it for planning, and to improve the world model via an auxiliary loss term. The central hypothesis is that actions should be identifiable, meaning that we should be able to decode the action from the latent transition. This was already explored in the recent work ACID, as the authors note in the manuscript. Where the authors take one step further is in introducing cross-domain inverse transfer, which tests whether an inverse dynamics probe trained in one domain transfers to the other. The authors found that this term can improve diagnosis (better correlation with downstream planning) and training (better success rate).

Critical strengths
  • Interesting motivation: I think action identifiability, which has also been tackled in recent world model literature, is an interesting problem that is worth the community’s attention. It is directly related to how reliable a world model’s predictions are, and as shown in the paper, it can be used as a proxy for the success rate during training. What makes latent world model development hard is the lack of such probing or validation metrics that do not require costly closed-loop control, so I think this is a meaningful contribution.
Critical weaknesses
  • Methodological novelty: However, this paper seems like a mere adaptation/extension of the recently proposed ACID [1], which augments the planning cost with an action consistency term. The core motivation is quite similar, and the distinction is just using such a term for training and across real/predicted domains. Using an inverse-dynamics-like term for training reminds me of DynaMo [2], which is referenced in the paper but not well discussed.
  • Limited experiments: Experiments are only conducted on very simplistic, synthetic benchmarks, which makes readers question the effectiveness of this method. I am aware that JEPA methods are mainly validated on simple benchmarks, but the experiments presented in this paper are still not sufficient from the perspective of the JEPA literature. I recommend the authors consider more challenging settings, such as deformable object manipulation as in DINO-WM [3] and ACID, or real-robot experiments as done in SparseWM [4] and ACID.
  • Cross-domain term does not improve correlation with SR: Table 1 shows that the introduced cross-domain term only marginally helps or even degrades the correlation with success rate. I think these experimental results do not agree with the main hypothesis and motivation. Maybe the within-domain term is enough?
  • Lack of principled explanation: I think the current manuscript fails to provide a principled explanation of why the cross-domain term helps. I can agree with the intuition, but the overall presentation reads like a somewhat hand-wavy explanation. Given that ICLR is an ML conference, readers will expect a more formalized explanation of why the method works.

[1] Seo, Gawon, Dongwon Kim, and Suha Kwak. "ACID: Action Consistency via Inverse Dynamics for Planning with World Models." arXiv preprint arXiv:2607.02403 (2026). [2] Cui, Zichen J., et al. "Dynamo: In-domain dynamics pretraining for visuo-motor control." Advances in Neural Information Processing Systems 37 (2024): 33933-33961. [3] Zhou, Gaoyue, et al. "Dino-wm: World models on pre-trained visual features enable zero-shot planning." arXiv preprint arXiv:2411.04983 (2024). [4] Chun, Junha, Youngjoon Jeong, and Taesup Kim. "Sparse imagination for efficient visual world model planning." International Conference on Learning Representations. Vol. 2026. 2026.

Questions for authors

See critical weakness

3 comments

We thank the reviewer for recognizing the value of action-identifiability and efficient planning diagnostics. The revision adds formal cross-domain analysis and deformable-manipulation experiments.

1. Methodological Novelty and Related Work

We respectfully disagree that ATM is an adaptation or extension of ACID. ATM's core formulation predates ACID's first public release, including cross-domain inverse probing and the full 2×22\times2 transfer matrix. To preserve double-blind anonymity, we refrain from presenting potentially identifying chronological evidence publicly. We have raised this issue confidentially with the Area Chair.

The technical distinction is also substantive. ACID adds a real-trained inverse-consistency residual to the planning-time cost. ATM instead trains separate inverse probes on real and predicted transitions, evaluates both probes on both domains, and diagnoses action-identifiability and bidirectional transfer without changing the planner. ACID's real-to-predicted evaluation corresponds to just one direction of ATM's matrix.

DynaMo, meanwhile, learns latent forward/inverse dynamics via self-supervised pretraining without ground-truth action supervision. AITS uses observed action labels to intervene on real-transition identifiability; AITS-P intervenes on predicted transitions. They serve as controlled tests of ATM's diagnostics, not the basis of its methodological novelty.

2. More Challenging Experiments

Following the reviewer's suggestion, we added Rope Manipulation, involving deformable-object contact dynamics. We jointly trained world-model encoders and predictors under four settings (Baseline, AITS, AITS-P, Both), and measured full ATM and closed-loop CEM planning on all 400 epoch checkpoints, using final Chamfer Distance across 10 fixed cases.

The new results show that cross-domain inverse risk relates to planning quality much more strongly than within-domain identifiability on Rope, and that the full matrix exposes structural differences invisible to one-direction diagnostics. See Sec. 4.6, Fig. 3, Table 4, and App. A.4 for the protocol.

3. Why Cross-Domain Transfer Matters

The improvement from adding a transfer penalty to the scalar score in OGBench-Cube (Table 1) is indeed modest. But ATM's cross-domain contribution is not limited to that scalar score, and the new Rope results directly address whether within-domain inverse risk is sufficient.

Across 400 checkpoints, −DT,T-D_{T,T} has Spearman correlation 0.495 with planning quality, whereas −DT,P-D_{T,P} reaches 0.786. Prediction loss attains Spearman 0.780, but Pearson only 0.180, versus 0.571 for −DT,P-D_{T,P}.

Please inspect the new Fig. 3: on the original axes, high-performing checkpoints are tightly compressed in prediction-loss values, obscuring variation in planning quality. The cross-domain risk distributes these models more clearly and exhibits a much stronger linear relationship with planning quality.

The full matrix adds further information. In Table 4, two checkpoints differ by only 1.34% in DT,TD_{T,T} and 0.20% in DT,PD_{T,P}, but have very different DP,TD_{P,T} (9.750 vs. 4.749) and final CD (2.786 vs. 1.308); the better checkpoint wins on all 10 evaluation cases. Thus, within-domain risk and even one-way transfer do not capture all relevant action-transition structure.

4. Principled Explanation

We added a squared-loss decomposition in Sec. 3.3 (Eq. 6). For Bayes-optimal inverse rules,

DT,P∗−DP,P∗=EDP[∥hT∗(ξ)−hP∗(ξ)∥22],D^*_{T,P}-D^*_{P,P} =\mathbb E_{\mathcal D_P}\left[\|h_T^*(\xi)-h_P^*(\xi)\|_2^2\right],

with an analogous identity for reverse transfer. Hence cross-domain excess risk quantifies disagreement between the domains' inverse rules; the diagonal risks alone cannot determine it.

Appendix B.3 further constructs two world models with identical prediction loss and identical DT,TD_{T,T}, DT,PD_{T,P}, and DP,PD_{P,P}, but different DP,TD_{P,T} and planning errors. Together with Rope, this explains why bidirectional ATM provides diagnostic information beyond within-domain identifiability.

Follow-up: Highlighted Revisions and New Evidence

Thank you again for your constructive review. We have uploaded an updated manuscript, with all new and revised content highlighted in blue for easier reading.

We would particularly like to draw your attention to three additions addressing your concerns.

1. More challenging experiments and cross-domain evidence

Following your suggestion, we added Rope Manipulation, evaluating 400 checkpoints with closed-loop CEM planning (Sec. 4.6, Fig. 3). The correlations with planning quality are:

DiagnosticSpearman ↑Pearson ↑
Prediction loss0.7800.180
−DT,T-D_{T,T}0.4950.383
−DT,P-D_{T,P}0.7860.571

As shown in Fig. 3, prediction loss compresses high-performing checkpoints into a narrow numerical range on the original axes, whereas cross-domain inverse risk provides clearer discrimination and a substantially stronger linear association with planning quality.

2. Why the full ATM matters and its theoretical basis

The new Table 4 compares two Rope checkpoints with nearly identical DT,TD_{T,T} and DT,PD_{T,P} but markedly different reverse-transfer risks DP,TD_{P,T} and planning outcomes, illustrating the additional information captured by bidirectional transfer.

We have also added a formal squared-loss decomposition in Sec. 3.3 (Eq. 6) and a constructive planning counterexample in Appendix B.3, providing a principled explanation of why cross-domain transfer matters.

3. Methodological clarification

We have expanded Related Work to explicitly distinguish ATM and its controlled interventions from DynaMo and SMWM. As clarified in our earlier response, ATM's main contribution is its bidirectional diagnostic framework, rather than a modification of the planning-time objective.

Thank you again for your time and valuable feedback. We would be happy to discuss any remaining questions or concerns.

If you feel that our revisions and responses have sufficiently addressed your concerns, we would greatly appreciate it if you would consider increasing your rating.

Dear Reviewer,

Please reply to the authors' rebuttal.

Thanks!

AC.

Comments on the paper

1 comment
AuthorsOpenReview ↗

Final Rebuttal Summary and Manuscript Revisions

Final Rebuttal Summary and Manuscript Revisions

Dear Area Chair and Reviewers,

We sincerely thank all three reviewers for their thoughtful feedback. We particularly appreciate JEYj's recognition of the importance of action-identifiability, QAV2's acknowledgment of ATM's practical value and efficiency, and exuC's positive assessment of our diagnostic perspective and empirical results.

We have uploaded a revised manuscript, with all additions and revisions highlighted in blue. Below, we summarize how we addressed the main concerns.

1. Reviewer JEYj: Novelty, Experiments, Cross-Domain Evidence, and Theory

  • Novelty: We clarified that ACID modifies planning-time costs, whereas ATM diagnoses frozen world models through a complete bidirectional inverse-transfer matrix without changing the planner. We expanded Related Work to discuss DynaMo and SMWM. Chronological evidence concerning ACID was communicated confidentially to the Area Chair.
  • Experimental scope: Following the reviewer's suggestion, we added deformable Rope Manipulation, evaluating four training configurations and 400 checkpoints, each with closed-loop CEM planning across 10 fixed cases (Sec. 4.6, Fig. 3, Table 4, App. A.4).
  • Cross-domain evidence: On Rope, −DT,P-D_{T,P} achieves Spearman 0.786, compared with 0.495 for −DT,T-D_{T,T}. Its Pearson 0.571 also substantially exceeds prediction loss (0.180).
  • Theory: We added a squared-loss decomposition of cross-domain excess risk (Sec. 3.3, Eq. 6) and a constructive planning counterexample (App. B.3).

2. Reviewer QAV2: Score Calibration and Intervention Interpretation

  • Statistical concerns: We clarified that the screening score fits only one scalar coefficient within a fixed function. We explicitly acknowledge that the reported rankings are in-pool results, not held-out generalization, and that pair counts are comparisons rather than independent observations.
  • Interventions: We explained the diagnostic hierarchy (Sec. 3.4): true-domain identifiability is primary, cross-domain transfer secondary, and matrix asymmetry tertiary. As Table 2 illustrates, improved planning can coexist with increased asymmetry because multiple diagnostic properties change simultaneously. The new Rope experiments provide additional cross-domain evidence.

3. Reviewer exuC: SMWM Comparison and Necessity of the Full ATM

  • Related work: We added an explicit SMWM comparison. While both methods use supervised inverse dynamics, SMWM targets representation learning and collapse prevention, whereas AITS/AITS-P are controlled interventions on different transition domains. ATM independently evaluates frozen models using freshly trained probes.
  • Full-matrix evidence: The new Table 4 compares two Rope checkpoints with nearly identical DT,TD_{T,T} and DT,PD_{T,P}, but different reverse-transfer risks (DP,TD_{P,T}: 9.750 vs. 4.749) and planning outcomes (CD: 2.786 vs. 1.308). The better checkpoint wins all 10 cases.
  • Formal justification: Appendix B.3 constructs two models with identical prediction loss, DT,TD_{T,T}, DT,PD_{T,P}, and DP,PD_{P,P}, but different DP,TD_{P,T} and planning errors, demonstrating the additional information provided by bidirectional transfer.

Summary

Our revision adds (i) extensive deformable-manipulation experiments, (ii) formal risk analysis and a constructive counterexample, and (iii) clearer comparisons with related work. Together, these strengthen ATM's central contribution: distinguishing within-domain action-identifiability from cross-domain transfer and revealing planning-relevant differences that simpler diagnostics miss.

We sincerely appreciate the reviewers' and Area Chair's time, expertise, and constructive guidance. We have carefully addressed the concerns through new evidence, clarifications, and manuscript revisions, and respectfully hope these substantial additions will be considered in the final assessment.

As the discussion period draws to a close, we understand that further exchanges may not always be possible. If the paper is accepted, we would be happy to address any remaining concerns and incorporate further suggestions in the camera-ready version.

Thank you again for your time and consideration.

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

Related papers

Loading the map…

Discussion (0)

Sign in to comment.