acceptodds
Under review as a conference paper at ICLR 2027

Can We Check Whether a World Model Follows Actions Without a Simulator?

Abstract

World models predict how an environment changes when an agent acts, and agents plan by comparing these predictions across actions. This only works if changing the action moves the prediction the way the real environment would move. Many existing evaluations check this with cheap probes that need no simulator. We test whether these probes notice when a model’s prediction moves the wrong way. We use simulators that can return to a saved state and run again under a different action. This gives the true effect of each action change. We study 92 robot world models that we trained and 143 released models, including models that predict Atari frames. We also reflect each model’s response to actions, which reverses its direction and keeps its size. We prove that probes that look only at the size of the response cannot detect this, and in our experiments they do not. A popular probe, logged action retrieval, even scores the reflected models higher. Training models on its criterion improves their retrieval score while their true direction gets worse. We also propose the matched probe, which compares the model’s response with pairs of similar situations in the recorded data. It tracks the true direction in all three groups of models. We recommend testing any action probe on reflected models before trusting it.

Primary area:
reinforcement learning
Keywords:
world models, action-conditioned world models, model-based reinforcement learning, evaluation, action faithfulness, counterfactual evaluation, probe validity, model probing, robot learning, formal verification

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_4WkjOpenReview ↗
Rating
3
Confidence
3
Informativeness
2
Presentation
3
Soundness
3(+1)
Read review
Paper summary

The paper asks whether we can evaluate how well a world model predicts the effects of different actions without using a simulator. The authors use simulators as ground truth and test several existing metrics on models from robotics, control tasks, and Atari. They also create corrupted models by reversing or rotating their predicted action effects. They show that several metrics fail to detect these errors, and action retrieval sometimes even prefers the corrupted models. They propose a new metric, the matched probe, which compares similar recorded states with different actions. It agrees better with the simulator in many settings. They also use it as a training loss, with mixed results.

Critical strengths

I like the idea of testing whether evaluation metrics actually measure what they claim. The corrupted models are a simple but convincing experiment, and the theoretical results support the findings. The evaluation is broad, including models released by other groups, and the direct comparison with simulator ground truth is particularly useful. The paper is well written and clear about its limitations.

Critical weaknesses
  1. The matched probe relies on finding similar states with different actions. Its accuracy drops as states become less similar, and it is unclear how well it works with less diverse datasets. Atari uses a random policy, which makes matching easier.
  2. The inverse model direction check performs similarly on LIBERO and TD-MPC2 and can also evaluate actions outside the dataset. The advantage of the matched probe is mainly on Atari, where the inverse model performs poorly.
  3. The corrupted model experiments are convincing, but the results are somewhat expected by construction. On original models, the matched probe is not significantly better than retrieval overall. Its clearest advantage comes from a smaller LIBERO comparison.
  4. The connection to planning is still weak. No probe significantly predicts planning performance among the original models, and improving the matched probe during training does not improve control.
Questions for authors
  1. How fast does the probe degrade as action diversity decreases?
  2. With a strong inverse dynamics model available, is there a reason to use the matched probe?
  3. Can it predict planning performance among original models on harder tasks?
  4. Could the lack of control gains from the loss reflect planning tasks that are too easy?
4 comments
AuthorsOpenReview ↗

Response to reviewer 4Wkj (1/2)

We thank the reviewer for the supportive review and the specific questions. We ran new experiments, including 152 new LIBERO models trained on image features and a released robot video world model (A2World).

Note: For each model, the probe produces one score while the simulator gives its true direction on the logged swaps (D_logged in the paper). Unless stated otherwise the numbers below are the Spearman correlation between the two across models. A value of 1 means the probe and the simulator put the models in the same order. Intervals are 95% bootstrap intervals over training objectives.

W1 (Regarding similar states and less diverse data)

The matched probe degrades slowly when the pairs get worse. We ran new checks on our proprioceptive LIBERO models ("30 models" are the 6 objectives x 5 seeds of Table 9).

  • Less similar states: we cut the 1800 pairs that the paper also ran in the simulator (the 5 nearest partners of each of the 360 positions) into 5 groups by the distance between the two start states. The estimate's cosine with the simulator drops from 0.90 in the closest group to 0.70 in the farthest. But computed on the farthest group alone the matched probe still ranks the 30 models at 0.84 and the 92 paper models at 0.98.

  • Less diverse data: with only 8, 4 or 2 other demos per task to match against (normally about 15) it ranks the 30 models at 0.89, 0.85 and 0.80. Retrieval gets 0.14 with all the data.

These are LIBERO's human teleoperated demos (50 per task) so the results do not depend on a random logging policy as in Atari. For less diverse actions please see our response to Q1.

W2 (Regarding the inverse check)

Both direction probes are ours. On our proprioceptive LIBERO models they are about equal (0.89 and 0.92, Table 9) and for actions outside the data we recommend the inverse check. New test: at the 360 held out positions we ran the simulator under 3 translation-only action sequences each, sampled the way our planner samples its candidates. Computed on these sequences the inverse check predicts the true direction at 0.68 across 65 of our LIBERO models (the 30 of Table 9 plus the 35 models of the Table 13 training study that use no contrastive loss). Retrieval gets 0.38.

Regarding the advantage of the matched probe mainly on Atari, we trained new LIBERO models on image features. They use the same transformer as in the paper but the state is 1536 DINOv2 ViT-S/14 features of LIBERO's two cameras (we rendered all demos and all simulator branches ourselves). There the inverse model has a held-out R² of 0.67 but the inverse check only gets 0.09, 0.02 and 0.07 at the three model sizes. The matched probe gets 0.90, 0.87 and 0.86 (Table R1).

W3 (Regarding the transform test and the advantage on original models)

The true direction under each transform is known by construction and that is what makes it a test. Whether a probe follows it is not known in advance and the existing probes do not. Retrieval even rates the reflected and rotated responses as more faithful than the originals on our LIBERO models, the released TD-MPC2 models and the released Atari models (Tables 1, 4 and 6). We found the same on A2World [1], a released robot video world model with 2.5B parameters. At 90 held out positions retrieval ranks the logged action chunk first at all 90 for the reflected copy and for the rotated copy (69 of 90 for the released model).

[1] Learning Transferable Dynamics Priors from Action to World Modeling, ECCV 2026

We agree that the transform test is necessary but not sufficient (Section 4 of the paper). So we compared the probes on many more original models, including the image feature models at three sizes.

Table R1. Spearman correlation with D_logged on models as trained (no transformed copies). Higher is better. "30" are the 6 objectives x 5 seeds of Table 9. "90" add the 60 models of our training study (Table 13 plus three new contrastive settings) and "65" are the 90 without the 25 contrastive models. The last column is matched probe minus retrieval with its 95% interval.

modelsmatched probe (ours)inverse check (ours)retrievallead over retrieval
proprioceptive 0.81M, 300.920.890.140.78 [0.02, 1.49]
proprioceptive 0.81M, 650.920.910.660.26 [0.06, 0.66]
proprioceptive 0.81M, 900.970.95−0.201.17 [0.56, 1.67]
image feature 1.2M, 300.900.09−0.070.96 [0.11, 1.88]
image feature 5.5M, 300.870.020.120.75 [0.04, 1.52]
image feature 26.8M, 300.860.070.040.82 [0.17, 1.61]
image feature 1.2M, 900.970.09−0.471.44 [0.86, 1.84]

The matched probe leads retrieval significantly in all seven sets (the first row is Table 9 of the paper), so the advantage on original models is not limited to one small comparison. On the released TD-MPC2 and Atari models as trained the two probes remain tied as the paper reports.

AuthorsOpenReview ↗

Response to reviewer 4Wkj (2/2)

W4 (Regarding the connection to planning)

Errors of direction are what break planning and the existing probes miss them. In Table 1 a reflected response makes model predictive control miss a goal 4 cm away by 32.0 cm and a rotated one by 24.9 cm, while the trained models miss by 0.7 cm. Retrieval rates both as more faithful than the trained models. On a new harder task the matched probe predicts planning success across 90 original models and retrieval does not (please see Q3). We agree that matched swap alignment does not improve control (Q4).

Q1 (how fast the probe degrades as action diversity decreases)

Slowly. We sorted the same 1800 pairs by how much the two 8 step action sequences differ (in normalized units) and cut them into 5 groups of 360.

Table R2. Matched probe on our proprioceptive LIBERO models by action difference between the two segments of a pair. "Estimate" is the mean cosine between a pair's logged estimate and what the simulator says. The last two columns are the Spearman correlation between the matched probe and the simulator's direction, both computed on that group only. Higher is better in the last three columns.

groupmedian action differenceestimate30 models92 paper models
1 (most similar actions)1.90.700.760.97
23.20.740.890.99
34.40.830.970.99
45.80.820.880.98
5 (most different actions)8.20.830.690.97

With the most similar actions the estimate is less accurate (0.70 against 0.83) but the probe still ranks the 30 models at 0.76 and the 92 models at 0.97. The 30 model score moves between 0.69 and 0.97 with no downward trend as the actions get more similar.

Q2 (is there a reason to use the matched probe when a strong inverse model is available?)

For actions in the data and with an accurate inverse model there is little reason. On our proprioceptive LIBERO models the two probes are about equal. The reason to prefer the matched probe is that a decent fit of the inverse model does not guarantee that the inverse check works. On the image feature models the fit is R² 0.67 but the check is at 0.09, 0.02 and 0.07 (W2). The matched probe needs no inverse model.

Q3 (planning performance among original models on harder tasks)

New test: we ran model predictive control with 90 of our LIBERO models as trained (no transforms) on a harder task. The goal is 10 cm away instead of 4, an episode has 40 steps and the planner executes 4 actions before it replans (in the paper it replans after every action). Every model starts from the same 24 states (2 in each of 12 tasks) which gives 2160 episodes.

Table R3. Spearman correlation across models between each measure and the success rate on the harder task, with 95% intervals. Success means the gripper ends within 2 cm of the goal. Higher is better. "90 models" are the 30 of Table 9 plus the 60 models of our training study. "55 models" leave out the objectives that lower D_logged by more than 0.1 (retrieval's contrastive loss, inverse dynamics at weight 1, shuffled targets).

measure90 models55 models
true direction on logged swaps (D_logged, needs the simulator)0.83 [0.57, 0.90]0.34 [−0.03, 0.55]
matched probe (ours)0.76 [0.44, 0.88]0.07 [−0.32, 0.38]
validation error (lower error counted as better)0.90 [0.70, 0.94]0.64 [0.12, 0.87]
retrieval−0.37 [−0.78, 0.17]−0.06 [−0.61, 0.47]

Across the 90 models the matched probe predicts success and retrieval does not. Among the 55 models even the simulator's true direction does not predict success significantly, so a direction probe cannot be expected to. There validation error predicts best. We therefore recommend reporting validation error together with a direction probe. Validation error cannot see a response that is reflected around the logged action (Proposition 4) but a direction probe can.

Q4 (could the lack of control gains reflect planning tasks that are too easy?)

We tested this with the harder task of Q3 and it did not show a clear gain from matched swap alignment either. At the weight we fixed in advance (0.01) neither task shows a significant gain. At weight 0.001 the success rate on the harder task goes from 0.68 to 0.78 (uncorrected p = 0.007) but the final distance to the goal does not change, and the gain does not survive a Holm correction over our 12 control comparisons of matched swap alignment. So we do not claim a control benefit.

We hope this addresses your concerns.

Thank you so much for the work you have done.

The analysis on degradation answer my question about the matching. Also the image feature results answer my questions about the inverse check. The comparison on model is much broader within LIBERO.

My concerns have been addressed, I increased the soundness score and kept my positive review score.

Dear Reviewer 4Wkj,

Thank you for your supportive review and for raising the soundness score. We are glad our responses addressed your concerns.

Best regards, The authors

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

The paper asks whether simulator-free "action probes" can detect when a world model's response to an action change points in the wrong direction. Using simulators that can restore a saved state and branch under a different action (LIBERO, DeepMind Control, Atari), the authors measure the true effect of action changes for 235 world models (92 trained by the authors on LIBERO proprioception, 117 released TD-MPC2 models, 26 released DIAMOND Atari models) and build control models whose faithfulness is known by reflecting, rotating, rescaling or removing a trained model's response while leaving its predictions at the logged actions unchanged. They prove that probes reading only response size cannot detect direction, show that logged-action retrieval rewards response size and scores reflected models higher, and that training on retrieval's criterion lowers true direction. They propose the matched probe, which scores the cosine between the model's response to swapping the actions of two nearby logged segments and the difference of their recorded outcomes, together with a direction-scored inverse-model check; both track true direction across the three model groups. The same pairs are turned into a training loss (matched swap alignment), which improves direction on the authors' LIBERO models but not control, and hurts released TD-MPC2 models. All propositions are checked in Lean 4 and code is provided.

Critical strengths

1. A principled, reusable methodology for validating action probes, with honest self-calibration. The response-transform test (reflect / rotate / rescale / remove while keeping predictions at logged actions fixed) is a clean sanity check in the spirit of control tasks and randomization checks, and the paper calibrates it with a "reference, no data or model" row that exposes how much of a probe's score on transformed copies comes from symmetry alone. The experimental engineering (controller-state reset in MuJoCo with a measured noise floor, loader checks for released checkpoints, bootstrap intervals with the correct resampling unit, every number regenerable from checkpoints) is careful.

2. An actionable negative result about a probe the community uses. The demonstration that logged-action retrieval rewards response size regardless of direction, prefers reflected and rotated copies, and (when used as a training criterion) makes models less faithful under both random and logged action changes, is a concrete warning for anyone using retrieval-style controllability scores for model selection or as a training target. The paper also reports its own negative results (matched swap alignment does not improve control and harms released models; no probe predicts control among as-trained models) in the main text.

Critical weaknesses

1. The matched probe answers a different question from the one planning needs, and its estimate shares its data with the model under test. The probe can only be asked about action swaps that occur in the logged data. The paper's own Corollary 5 shows that logged data cannot certify response direction away from the logged actions, and its own measurements show that on-support and off-support direction differ substantially. But a planner including the paper's own MPC experiment queries exactly the off-support actions. The title question is therefore answered positively only for the case that matters least for planning, while the abstract and introduction frame the result more broadly. Second, the logged estimate and the model are built from the same data: the authors observe that the probe is closer to the truth than their symmetric-noise analysis predicts, and that it overstates the gain of training on its own criterion, which indicates that it partly rewards agreement with the data rather than with the environment. Third, the direction-scored inverse-model check (which needs no matched pairs, can query off-support actions and is at least as accurate wherever an inverse model can be fit) is presented as secondary, so the paper's recommendation is not clearly the strongest tool it has.

2. The comparative advantage over retrieval is established only in a toy setting, and the setting where the question matters most is absent. The evidence that the matched probe ranks models better than retrieval comes from the 30 LIBERO models the authors trained themselves: small proprioception-only transformers differing in six objectives and seeds. On the 143 released models the difference between the two probes is not significant, and within TD-MPC2 tasks retrieval is numerically ahead. Moreover, the TD-MPC2 ground truth is measured in each model's own latent space, so cross-model correlations there largely reflect task difficulty. No robot video world model is evaluated, but these are the models for which action-following is currently in question and the setting of the concurrent work the paper positions itself against. The only pixel models are Atari diffusion models under a uniformly random logging policy, the most favorable case for a matched-pair probe. The Goodhart result is confounded: the contrastive loss required an extreme temperature to train at all and worsens held-out prediction error substantially, so the paper cannot separate "direction gets worse" from "the model gets worse overall". The practical prerequisite of the probe (i.e., similar states with different actions in the logged data) is not quantified on real robot datasets with expert demonstrations. Finally, the training contribution is in effect negative: matched swap alignment does not improve control, lowers direction when fine-tuning released models, and a random-partner ablation retains most of its gain, so the "matched" element is not what drives it.

Questions for authors

Q1 scope of the comparative claim. The advantage of the matched probe over retrieval is significant only on the 30 LIBERO models the authors trained; on released models the differences are not significant and within TD-MPC2 tasks retrieval is numerically ahead. Do the authors consider the recommendation to prefer the matched probe over retrieval supported beyond their own models? Could authors provide evidence closer to practical use for example a robot video world model, or pixel-based LIBERO models?

Q2 on-support vs. off-support. Planners query actions outside the logged data, where the paper shows direction is much lower and where Corollary 5 says logged data cannot certify it. How to demonstrate the practical value of the matched probe for planning-oriented evaluation, and in that setting would the direction-scored inverse check be their recommended tool rather than the matched probe?

Q3 shared data between estimate and model. The estimate and the model under test are derived from the same logged data, and the authors observe that the probe rewards agreement with the data and overstates the gain of training on its criterion. Is there any mitigations, and how would a practitioner without a simulator detect that the estimate is less accurate than the model being evaluated? (the paper mentions held-out error as a warning, please elaborate on how reliable that is).

Q4 the Goodhart result and retrieval variants. The contrastive loss needed a temperature of 10⁻³ to train and worsens validation MSE. Can the conclusion that training on retrieval's criterion lowers true direction be reproduced at a weight that does not hurt prediction accuracy? Does a size-normalized variant of retrieval, which Proposition 3 would suggest, still fail the transform test and the 30-model ranking?

Q5 applicability. The matched probe requires similar states with different actions. How available are such pairs in typical expert-demonstration robot datasets, what is the distribution of action differences among the pairs used on LIBERO, and what are the main obstacles to applying the probe to video world models?

7 comments
AuthorsOpenReview ↗

Response to reviewer VWBx (1/4)

We thank the reviewer for the detailed and technical review. We ran new experiments for all five questions, including 152 new LIBERO models trained on image features and a released robot video world model (A2World).

Note: For each model, the probe produces one score while the simulator gives its true direction on the logged swaps (D_logged in the paper). Unless stated otherwise the numbers below are the Spearman correlation between the two across models. A value of 1 means the probe and the simulator put the models in the same order. Intervals are 95% bootstrap intervals over training objectives.

W1 (The matched probe answers a different question from the one planning needs, and its estimate shares its data with the model under test.)

The matched probe only checks actions that are in the data and the paper states this (Section 7: "Yes, but only for actions supported by the logged data"). This is not specific to our probe. Checks built from logged data alone cannot verify a model's response to off-support actions without extra assumptions (Corollary 5). The existing methods in the literature also work within this limit. Off policy evaluation assumes the logged data cover the evaluated actions or adds smoothness assumptions to go beyond them [1]. In recent world model evaluation, WorldGym compares generated videos with real robot videos under the recorded actions of a validation split and for action sequences outside its training data it checks the generated videos only qualitatively [2]. NVIDIA scores action conditioned Cosmos-Predict2.5 on 100 Bridge test episodes against the ground truth videos, so also under the recorded actions [3]. We do not agree that the on-support case matters least for planning (please see our response to Q2).

[1] Off-Policy Evaluation Beyond Overlap: Sharp Partial Identification Under Smoothness, ICML 2024

[2] WorldGym: World Model as An Environment for Policy Evaluation, arXiv 2025

[3] World Simulation with Video Foundation Models for Physical AI (Cosmos-Predict2.5), arXiv 2025

Regarding shared data, the probe takes its pairs only from validation and test demos which the models never train on (Appendix E). The link that remains does not change the ranking of the models (please see our response to Q3).

Regarding the inverse check, both direction probes are ours and for off-support actions we do recommend the inverse check (Q2). But it is not at least as accurate wherever an inverse model can be fit. On our new image feature models the inverse model has a held-out R² of 0.67 but the inverse check only gets 0.09, 0.02 and 0.07 at the three model sizes. The matched probe gets 0.90, 0.87 and 0.86 (Table R1). The paper shows the same on Atari (0.74 against 0.94, Appendix C.3).

W2 (The comparative advantage over retrieval is established only in a toy setting, and the setting where the question matters most is absent.)

As suggested, we have added results for released robot video world model (A2World) and image-based LIBERO models at three sizes. The results are in Q1. We agree that the TD-MPC2 truth partly reflects task difficulty. But task difficulty cannot explain the transform test because it compares each model with copies of itself. Retrieval ranks the reflected TD-MPC2 models above the originals (mean rank 0.86 against 1.72, Table 4). Atari's random policy does make matching easier so our main evidence is from LIBERO's human demos. However, our additional results (seen in Q1) show that we achieve a higher Spearman correlation compared to retrieval.

The Goodhart result holds at matched accuracy and with a softer temperature (check our response to Q4).

On matched swap alignment, we agree it does not improve control and lowers direction on the released TD-MPC2 models. It is a secondary contribution. But the matched partner is what drives its gain. On logged swaps (what the loss targets) a random partner keeps only 0.020 of the 0.046 gain and shuffled targets give −0.164 (Table 3). Under random perturbations its gain is about 2.5 times that of delta prediction, the best existing objective we tried (+0.124 against +0.050, Table 13).

AuthorsOpenReview ↗

Response to reviewer VWBx (2/4)

Q1 (scope of the comparative claim)

As suggested, we now have evidence beyond the 30 models from the following settings: Image-based LIBERO, Robot video world model (A2World).

Image-based LIBERO: the new models use the same transformer as in the paper but the state is 1536 DINOv2 ViT-S/14 features of LIBERO's two cameras. We rendered all demos and all simulator branches ourselves with one renderer, which avoids the image mismatch of Appendix E. We trained 152 models with 1.2M parameters and the 30 models of Table 9 again with 5.5M and 26.8M parameters. They predict image features and not pixels.

Table R1. Spearman correlation with D_logged on models as trained (no transformed copies). Higher is better. "30" are the 6 objectives x 5 seeds of Table 9. "90" add the 60 models of our training study (Table 13 plus three new contrastive settings) and "65" are the 90 without the 25 contrastive models. For validation error a lower error is counted as better. The last column is matched probe minus retrieval with its 95% interval.

modelsmatched probe (ours)inverse check (ours)retrievalvalidation errorlead over retrieval
proprioceptive 0.81M, 300.920.890.140.610.78 [0.02, 1.49]
proprioceptive 0.81M, 650.920.910.660.690.26 [0.06, 0.66]
proprioceptive 0.81M, 900.970.95−0.200.871.17 [0.56, 1.67]
image feature 1.2M, 300.900.09−0.070.060.96 [0.11, 1.88]
image feature 5.5M, 300.870.020.12−0.150.75 [0.04, 1.52]
image feature 26.8M, 300.860.070.04−0.100.82 [0.17, 1.61]
image feature 1.2M, 900.970.09−0.470.291.44 [0.86, 1.84]

The matched probe is the best measure in every row and its lead over retrieval is significant in every row (the first row is Table 9 of the paper). On the 90 models it is also significantly above validation error (lead 0.10 [0.03, 0.27] on proprioception and 0.68 [0.15, 1.26] on image features).

Robot video world model: we tested A2World [4] (2.5B parameters, fine tuned by its authors on LIBERO from Cosmos-Predict2). At 90 held out positions we gave it the logged 20 step action chunk, 3 Gaussian perturbations of it, the matched partner's chunk and the null action. We rendered the true outcome of every chunk except the null action in the simulator and then transformed the model's responses as in Table 1.

[4] Learning Transferable Dynamics Priors from Action to World Modeling, ECCV 2026

Table R2. Transform test on A2World, means over the 90 positions, measured in pixels. "True direction" is D_logged. The size probe is zero action divergence. "Retrieval rank" is the mean rank of the logged chunk among 4 candidates (the logged chunk and its 3 perturbations), 0 means ranked first and chance is 1.5.

copy of A2Worldtrue directionsize proberetrieval rankmatched probe
as released+0.0250.0230.32+0.009
reflected response−0.0250.0230.00−0.009
rotated response−0.0010.0230.00+0.000
response x4+0.0250.3720.00+0.009

The size probe does not change under reflection or rotation. Retrieval ranks the logged chunk first at all 90 positions for every transformed copy (69 of 90 as released). The matched probe flips sign under reflection and goes to zero under rotation like the true direction. This is one model so it only tests these failure modes and not the ranking.

Q2 (on-support vs. off-support)

We conducted new experiments to show that direction on logged swaps tells us about the actions which a planner tries and inverse check can be used for off-support actions. At the 360 held out positions we ran the simulator under 3 translation-only action sequences each, sampled the way our planner samples its candidates, and measured each model's true direction on them.

Table R3. Spearman correlation between each measure and the true direction on planner-style actions, across 65 of our LIBERO models (the 30 of Table 9 plus the 35 models of the Table 13 training study that use no contrastive loss). Only the inverse check is computed on the planner's sequences. The last two rows need the simulator and are shown for reference.

measurecomputed onneeds simulatorcorrelation
inverse check (ours)the planner's sequencesno0.68
matched probe (ours)logged datano0.63
validation error (lower error counted as better)logged datano0.70
retrievallogged datano0.38
true direction on logged swaps (D_logged)logged swapsyes0.77
true direction under random perturbations (D_random)random perturbationsyes0.34

Continued further...

AuthorsOpenReview ↗

Response to reviewer VWBx (3/4)

Q2 (continued)

D_logged predicts the direction on the planner's actions much better than D_random does (0.77 against 0.34, difference 0.43 [0.12, 0.79]). So the on-support case is useful for planning-oriented evaluation. The matched probe is also above retrieval (lead 0.25 [0.03, 0.61]). For off-support actions we recommend the inverse check since it is the only one that can be computed on the planner's own sequences. Its inverse model is still fit to logged transitions so there it extrapolates.

Q3 (shared data between estimate and model)

The pairs come only from held out demos but a link remains and the paper reports it (the probe reads 0.756 where noise independent of the model would give about 0.69, Appendix C.3). This link does not change the ranking of the models.

  • New experiments: we computed the probe on each position's nearest partner and the true direction on its 2nd to 5th nearest partners, so on different swaps. The probe still ranks the 30 models at 0.94.

  • On the image feature models the probe reads 0.81 to 0.86 times the true direction for all 18 training objectives, so the order of the models is kept.

Mitigations: take the pairs from held out demos (as we do) and do not judge a model that was trained on the logged estimate by the matched probe alone. For matched swap alignment the probe shows a gain of +0.073 where the true gain is +0.046 (Appendix C.4).

To detect an estimate that is less accurate than the model, held-out prediction error is sensitive but not specific. Across the 117 fine tuning runs of released TD-MPC2 models (39 tasks x 3 weights) it went up in 97 of the 99 runs where direction fell but also in 11 of the 18 runs where direction rose.

Q4 (the Goodhart result and retrieval variants.)

Yes. It holds at matched accuracy and the extreme temperature is not the cause.

Table R4. Adding retrieval's contrastive loss (Appendix K) to next state MSE. 5 seeds per row, each compared with the MSE only model of the same seed. "Val ratio" is validation MSE with the loss divided by without (1.00 means same accuracy). "Retrieval rank" is the mean rank of the logged action among 16 candidates, without → with the loss (lower means retrieval likes the model more, chance is 7.5). "Change in D_logged" is negative when true direction gets worse and it dropped in 5 of 5 seeds in every row. p is from a paired t test over seeds.

modelsweighttemperatureval ratioretrieval rankchange in D_loggedpfrom
proprioceptive, 0.81M0.110⁻³4.230.90 → 0.15−0.291< 0.001paper, Table 3
proprioceptive, 0.81M0.0110⁻³2.330.90 → 0.47−0.239< 0.001paper, Table 3
proprioceptive, 0.81M0.00110⁻³1.610.90 → 0.31−0.127< 0.001new
proprioceptive, 0.81M0.0110⁻²3.920.90 → 0.13−0.265< 0.001new
proprioceptive, 0.81M0.1none1.020.90 → 0.77−0.0010.017paper models, new numbers
image feature, 5.5M0.110⁻³1.024.90 → 0.14−0.033< 0.001new
image feature, 26.8M0.110⁻³1.004.24 → 0.42−0.0210.002new

In the last row the val ratio stays within 0.992 to 1.004 in every seed and direction still drops in all 5 seeds. So the drop is not explained by a less accurate model. Row 4 uses a temperature 10 times softer and direction still drops. Row 5 is the no temperature setting that Appendix K calls "never trained". Its loss stays at log 5 but the gradient is not zero so the model still moves a little.

For size normalized retrieval we rescale each model's responses to the same mean size before ranking the candidates.

Table R5. Size normalized retrieval on our proprioceptive LIBERO models. The first five columns are mean ranks of the logged action for the 15 trained models of Table 1 and their transformed copies (lower means rated more faithful, chance is 7.5). The last column is the Spearman correlation with D_logged on the 30 models.

probetrainedreflectedrotatedx0.25x430 models
retrieval (Table 1)0.920.790.414.140.020.14
size normalized retrieval0.810.690.360.810.810.63

It ranks the 30 models better than plain retrieval (0.63 against 0.14) and the gap to the matched probe's 0.92 is not significant (0.28 [−0.11, 1.18]). But it still fails the transform test since it rates the reflected and rotated copies as more faithful than the trained models. With size fixed, the remaining term 2eᵀr of Proposition 3 depends on the model's own error e and not on the true response.

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

This paper investigates simulator free evaluation probes for world model action faithfulness, referring to a world model’s ability to follow correct action directions that match the real environment changes triggered by agent actions. This work empirically examines five existing probes and proposes two direction probes, i.e., the matched probe and direction form inverse model check. Experiments are conducted across 235 world models, including 92 self-trained LIBERO robot manipulation models, 117 TD-MPC2 models for control tasks, and 26 DIAMOND Atari frame prediction models.

Critical strengths
  1. The paper provides mathematical proofs for why existing simulator-free probes fail and strengthens these claims by formally verifying theoretical propositions using the Lean proofs. This contribution has methodological impact for the world model research community.
  2. The testing across 235 models spanning three distinct domains, providing supportive empirical evidence. It reports uncertainty, distinguishes transformed from ordinary models, and explicitly acknowledges unsuccessful training transfer and absent control improvements. This transparency makes the empirical claims more trustworthy.
Critical weaknesses
  1. The proposed matched probe only assesses action faithfulness for the actions that lie within the support of logged training data, meaning that it cannot evaluate OOD action perturbations. And the matched probe also requires logged data containing different actions starting from sufficiently similar states. This data dependency makes it hard in real-world settings where data diversity is limited or finding reliable nearest-neighbor pairs becomes difficult, which can potentially introduce estimation noise or bias.
  2. The matched swap alignment training benefit is narrow and is inconsistent across models. Specifically, it has limited improvement when it is applied to fine-tuning pre-trained TD-MPC2 models. This is because the logged data estimate is sometimes less accurate than the model's own predictions, leading to degraded performance. However, this weakens its practical value as a plug-and-play alignment tool for existing foundational world models.
Questions for authors
  1. Please refer to Weaknesses.
  2. When scaling to high-dimensional state spaces, where raw Euclidean distance over the observation space fails to capture semantic proximity, how sensitive is the proposed matched probe’s performance to the selection of nearest-neighbor distance metrics and thresholds?
  3. In real-world robot datasets where precise state matching is hard to acquire, and the observation noise is commonly large, could you explain how the performance of matched probe degrades if increasing the maximum distance between state pairs that beyond the thresholds in the LIBERO experiments?
  4. Does the harmful effect of retrieval training persist when the prediction accuracy is held comparable to the baseline? Table 3 shows contrastive retrieval training also increases validation MSE. Could you report direction faithfulness under smaller retrieval loss weights where the main prediction MSE stays matched to the baseline?
11 comments
AuthorsOpenReview ↗

Response to reviewer iusp (1/3)

We thank the reviewer for the careful read and the concrete questions. We ran new experiments for all three questions. One note on the numbers. A probe gives every model one score, and the simulator gives every model its true direction on logged swaps (D logged in the paper). Unless we say otherwise, the numbers below are the Spearman correlation between the two across models. Here “1” means that sorting the models by the probe's score gives exactly the same order as sorting them by the simulator's true direction.

W1 (Regarding need for actions in logged data)

The matched probe only checks actions that are in the data. But this is not specific to our probe. Checks built from logged data alone cannot verify a model's response to OOD action perturbations without extra assumptions, because the data contain no outcome for an action nobody took. The existing methods in the literature also work within this limit. Off policy evaluation assumes the logged data cover the evaluated actions, or adds smoothness assumptions to go beyond them [1]. Offline model based RL does not trust the model away from the data, and penalizes state action pairs where the model is uncertain or that the data do not cover (MOPO [2], MOReL [3]). Recent world model evaluations score predictions against recorded videos under the recorded actions. NVIDIA scores action conditioned Cosmos-Predict2.5 on 100 Bridge test episodes against the ground truth videos [4]. WorldGym compares generated videos with real robot videos under the recorded actions of a validation split. For action sequences that are not in its training data it checks the generated videos qualitatively [5]. The recent measures of response direction that we know of all compare the response with an executed one [6, 7, 8] (Section 6 of the paper).

[1] Off-Policy Evaluation Beyond Overlap: Sharp Partial Identification Under Smoothness, ICML 2024

[2] MOPO: Model-based Offline Policy Optimization, NeurIPS 2020

[3] MOReL: Model-Based Offline Reinforcement Learning, NeurIPS 2020

[4] World Simulation with Video Foundation Models for Physical AI (Cosmos-Predict2.5), arXiv 2025

[5] WorldGym: World Model as An Environment for Policy Evaluation, arXiv 2025

[6] WorldSimProbe: Diagnosing Simulator Faithfulness in Action-Conditioned World Models for Embodied Manipulation, arXiv 2026

[7] Do Robotic World Models Really Follow Actions? Diagnosing and Aligning Action-Conditioned Generation for Policy Learning, arXiv 2026

[8] The Intervention Gap in Latent World Models, arXiv 2026

For actions outside the data we would use our inverse check (an inverse model reads the action back from the predicted states, and we score the direction of the change). It needs no simulator, but it does assume the inverse model still holds for those actions. New test: at the 360 held out positions we ran the simulator under 3 translation-only action sequences each, sampled the way our planner samples its candidates, and measured each model's true direction on them. Table R1 shows how well each measure predicts this true direction across 65 of our LIBERO models (the 30 of Table 9 plus the 35 models of the Table 13 training study that use no contrastive loss).

Table R1. Spearman correlation across 65 models between each measure and the true direction on planner-style actions. Higher is better. Only the inverse check is computed on the planner's sequences. The other probes are computed on the logged data, as usual. The last two rows need the simulator and are shown for reference.

measurecomputed onneeds simulatorcorrelation
inverse check (ours)the planner's sequencesno0.68
matched probe (ours)logged datano0.63
validation error (lower error counted as better)logged datano0.70
retrievallogged datano0.38
true direction on logged swaps (D logged)logged swapsyes0.77
true direction under random perturbations (D random)random perturbationsyes0.34

Both of our probes are well above retrieval. Direction on logged swaps, which is what the matched probe estimates, also says more about the planner's actions than direction under random perturbations does (0.77 against 0.34).

On the data requirement, the probe needs similar states with different actions. LIBERO has 50 human teleoperated demos per task, recorded with no matching in mind, and that was enough. Q2 and Q3 show what happens when the pairs get worse.

AuthorsOpenReview ↗

Response to reviewer iusp (2/3)

W2 (Regarding narrow matched swap alignment)

We agree it is a secondary contribution and not plug and play for released models (we did not make this claim in the paper). But we do not think it is inconsistent. In our two settings it helps where the logged estimate is at least as accurate as the model, and hurts where it is less accurate.

  • On our LIBERO models it raises direction under random perturbations at all three weights (Table 13). At weight 0.01 the gain is +0.124 (about 2.5 times the gain of delta prediction, the best existing objective we tried), with +0.046 on logged swaps and −0.152 in counterfactual error, all significant after Holm correction (Table 3).
  • The gain comes from the matched estimate. A random partner gives only +0.020 on logged swaps, and shuffled targets give −0.164 (Table 3).
  • On the released TD-MPC2 models it lowers direction (Table 14), because there the logged estimate agrees with the simulator at only 0.61 while the models are already at 0.87 (Sec. 5.4).

Additional experimental results: Across the 117 fine tuning runs (39 tasks x 3 weights), held-out error went up in 97 of the 99 runs where direction fell. So it works as a warning sign without a simulator.

Q2 (distance metric, high dimensional states)

Regarding the selection of nearest-neighbor distance metrics and thresholds, there is no distance threshold, the probe just takes the nearest segment from another held out demo of the same task. In Table R2 we show additional ablation results.

Table R2. Matched probe on our proprioceptive LIBERO models (the 0.81M parameter transformers of the paper) when one thing about its pairs changes. Higher is better. "30 models" are the 6 objectives x 5 seeds of Table 9, and "65 models" add the 35 from Table 13. Noise is Gaussian, in standard deviations of each state dimension, added to every recorded state the probe uses. Noise and demo rows are averages over 5 draws.

change30 models65 models
none (as in the paper: standardized Euclidean distance over the 21 state dims)0.920.92
Euclidean distance, not standardized0.940.89
end effector position only0.910.85
joint angles only0.910.89
cosine distance0.950.90
noise of 0.1 std0.760.87
noise of 0.3 std0.650.81
only 8 other demos per task to match against (normally about 15)0.890.90
only 4 other demos per task0.850.88
only 2 other demos per task0.800.86
retrieval with the full clean data, for comparison0.140.66

So the choice of metric moves the 30 model score between 0.91 and 0.95, and every row stays above retrieval (last row).

For high dimensions we trained new models on image features. They use the same transformer, but the state is 1536 DINOv2 ViT-S/14 features of LIBERO's two cameras (we rendered all demos and all simulator branches ourselves, with one renderer). Matching in this feature space picks a different partner than matching by robot state at 334 of the 360 positions. Still, both kinds of pairs rank the 92 image feature models (the paper's 92 configurations) at 0.88. On the 30 ordinary image feature models the matched probe gets 0.90 and retrieval −0.07.

Q3 (larger distance between pairs, observation noise)

Each pair's estimate gets less accurate with distance, but the probe's ranking of the models (how closely its order of the models matches the simulator's order) does not get worse. We took the 1800 pairs that the paper also ran in the simulator (the 5 nearest partners of each of the 360 positions), sorted them by the the standardized distance between the two start states, and cut them into 5 groups of 360. See table R3.

Table R3. Matched probe by distance between the two states of a pair. "Estimate" is the mean cosine between a pair's logged estimate and what the simulator says. The last two columns are the matched probe computed on that group only. Higher is better in the last three columns.

groupmedian distanceestimate30 models92 paper models
1 (closest)0.350.900.640.98
20.500.820.910.98
30.670.760.890.98
40.890.730.970.99
5 (farthest)1.330.700.840.98

The estimate drops from 0.90 to 0.70. But computed on the farthest group alone, the matched probe still ranks the 30 models at 0.84 and the 92 models at 0.98. The 0.64 in the closest group is there because the models differ least on those pairs (the std of true direction across the 25 models other than inverse dynamics w=1 is 0.005, against 0.008 to 0.018 in the other groups), so their order is hardest to recover.

For observation noise, see the two noise rows of Table R2. With noise of 0.1 and 0.3 std added to every recorded state the probe uses, the matched probe gets 0.76 and 0.65 on the 30 models (0.87 and 0.81 on the 65), against 0.14 and 0.66 for retrieval on clean data.

AuthorsOpenReview ↗

Response to reviewer iusp (3/3)

Q4 (retrieval training at matched accuracy)

Yes, the harm persists. Table R4 has every weight we trained, including the smaller ones you asked for.

Table R4. Adding retrieval's contrastive loss (Appendix K) to next state MSE. 5 seeds per row, each compared with the MSE only model of the same seed. "Val ratio" is validation MSE with the loss divided by without (1.00 means same accuracy). "Retrieval rank" is the mean rank of the logged action among 16 candidates, without → with the loss (lower means retrieval likes the model more, chance is 7.5). "Change in D logged" is negative when true direction gets worse, and it dropped in 5 of 5 seeds in every row. p is from a paired t test over seeds.

modelsweighttemperatureval ratioretrieval rankchange in D loggedpfrom
proprioceptive, 0.81M0.110⁻³4.230.90 → 0.15−0.291< 0.001paper, Table 3
proprioceptive, 0.81M0.0110⁻³2.330.90 → 0.47−0.239< 0.001paper, Table 3
proprioceptive, 0.81M0.00310⁻³1.780.90 → 0.27−0.206< 0.001new
proprioceptive, 0.81M0.00110⁻³1.610.90 → 0.31−0.127< 0.001new
proprioceptive, 0.81M0.1none1.020.90 → 0.77−0.0010.017paper models, new numbers
image feature, 1.2M0.110⁻³1.184.87 → 0.61−0.056< 0.001new
image feature, 5.5M0.110⁻³1.024.90 → 0.14−0.033< 0.001new
image feature, 26.8M0.110⁻³1.004.24 → 0.42−0.0210.002new

Rows 3 and 4 are the smaller weights. They hurt direction less (−0.127 at weight 0.001 against −0.291 at 0.1) but still in every seed, and validation error still goes up by at least 49% in every seed. So with the temperature we could not match accuracy on these small models.

The last row does match it. With 26.8M parameters the val ratio stays within 0.992 to 1.004 in every seed, retrieval's rank still improves from 4.24 to 0.42, and direction still drops in all 5 seeds. So the drop is not explained by a less accurate model.

Row 5 is the no temperature setting that Appendix K calls "never trained" (its loss stays at log 5). The gradient is not zero though, so the model still moves a little. Direction changes by less than 0.001, as Appendix K says.

We hope this addresses your concerns.

Comments on the paper

1 comment
AuthorsOpenReview ↗

General response and revised paper

We thank the three reviewers for their careful reviews. We have answered every weakness and question in the three threads and we have now uploaded a revised paper. New revised text is in blue.

The following major additional results are included in the revised manuscript:

  1. We added image based LIBERO models at three sizes. In all seven sets of LIBERO models the matched probe ranks the models better than retrieval, the inverse check and validation error (lower part of Table 2, Appendix T).
  2. We tested a released robot video world model (A2World). Retrieval rates its reflected, rotated and enlarged copies as more faithful than the released model itself (Appendix V).
  3. We repeated the training on retrieval's criterion at equal accuracy and tested size normalized retrieval (Appendix U).
  4. We tested the matched probe with worse pairs: other distance measures, noise, fewer demos, more distant states and more similar actions (Appendix X).
  5. We added planner style actions and a harder control task with 90 models (Appendix Y).

The abstract now states the scope of the probe and the conclusion is rewritten with the new results. The supplementary code contains the scripts for the new experiments.

We would be grateful if the reviewers could read our responses and the revised paper and let us know whether they address your concerns. We are happy to answer any further question before the discussion closes on 18 November (AoE).

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

Related papers

Loading the map…

Discussion (0)

Sign in to comment.