acceptodds
Under review as a conference paper at ICLR 2027

What Do Latent Predictive Vehicle Representations Retain? Measuring State, Geometry, and Local Response

Abstract

Models of vehicle dynamics learned from logged states and commands complement physics-based models, and latent world models, which predict in a learned representation, are used to plan and train controllers in other domains. Vehicle controllers are usually specified in physical terms: costs, limits, and references depend on position, yaw angle, speed, and yaw rate, and the optimizer compares or differentiates predicted outcomes across nearby commands. A latent model placed in such a controller must therefore let these quantities be recovered and must change its predictions with commands as the vehicle does, and prediction error on its own latent targets measures neither. We contribute a measurement protocol for action-conditioned latent predictors with a physical readout that separately tests retention, physical-neighborhood organization, forecasting, and local response to command perturbations, using an untrained-encoder reference and three matched response paths that locate errors in the representation or the predictor. In a case study of a temporal joint-embedding predictive model trained on signals logged in IPG CarMaker, the representations retain the measured planar outputs, though an untrained encoder of the same architecture retains them slightly better; future-command input improves one-second forecasts with retention nearly unchanged; and responses to small command pulses diverge from the simulator already in latent coordinates, raising regret when choosing among nearby commands in all comparisons. Updating the predictor on responses corrects them locally at a cost in forecast accuracy. Measuring retention, forecasting, and local response separately is thus what qualifies a predictive latent as a candidate model for control, and the protocol provides the basis for its closed-loop evaluation.

Primary area:
applications to robotics, autonomy, planning
Keywords:
representation learning, latent world models, action-conditioned prediction, vehicle dynamics, self-supervised learning, state representation, representation geometry, local sensitivity analysis, Jacobians

Reviews

Among:
Shape:
1234this paper, Bottom 27%
Figure 1. Mean review rating of the reviewed papers at ICLR 2027. This paper averages 2.00 over 4 reviews, above 27% of the others.
Reviewer_8G7pOpenReview ↗
Rating
2
Confidence
3
Informativeness
2
Presentation
2
Soundness
2
Read review
Paper summary

The paper investigates whether action conditioned latent predictive representations retain physically meaningful vehicle states, preserve physical neighborhood structure and accurately capture local responses to control inputs. The authors propose a measurement protocol that separately evaluates state retention through frozen readout probes, physical organization through nearest neighbor overlap and collision metrics, predictive accuracy against persistence and supervised baselines, and local control response through matched perturbations of the simulator, target representation, and predicted representation.

The main contribution is a systematic evaluation framework demonstrating that state retention, forecasting accuracy, and local control fidelity are distinct properties of predictive latent models.

The authors validate the protocol using a temporal joint-embedding predictive architecture with a controlled neural ODE predictor trained on IPG CarMaker simulation data.The results suggest that low prediction error alone is insufficient to establish a latent model's suitability for model-predictive control. However, the validation is limited to noise-free simulation on a single route with one automated driver and six vehicle configurations; real-world vehicle transfer and closed-loop control performance remain unverified.

Critical strengths

1.A systematic evaluation framework that disentangles distinct properties of latent predictive representations. The paper evaluates state retention, physical neighborhood organization, forecasting accuracy, and local control-response fidelity using complementary metrics and controlled comparisons.

  1. A consequential empirical finding supported by controlled experiments. The results show that reasonable forecasting accuracy does not guarantee accurate local responses to control perturbations: predicted responses can diverge substantially from simulator responses and lead to worse rankings of nearby candidate commands.
Critical weaknesses

The validation is limited to noise-free simulation on a single route with one automated driver and six vehicle configurations; real-world vehicle transfer and closed-loop control performance remain unverified.

Questions for authors

How sensitive are these results to the magnitude and duration of the perturbations, the choice of operating points, and the normalization used to calculate response error and regret? Could the authors provide evidence that the observed discrepancies are robust across these choices rather than specific to the evaluation setup?

The untrained encoder performs better than the trained encoder on state recovery and neighborhood preservation. How do the authors interpret this result, and what evidence supports attributing the observed representation properties to predictive training rather than architectural inductive biases or properties of the input data?

1 comment
AuthorsOpenReview ↗

Specific Response to Reviewer 8G7P

We thank the reviewer for the careful comments. We wrote a general response summarizing the improvements to the reviewed paper, and below are the specific responses to your comments:

One route and driver, noise-free simulation, no closed loop.

The study is repeated on a second road and with a second driver (§3; Table 3; Appendix G.8). §6 states the remaining limits: simulation only, no closed-loop evaluation, no independent command excitation because the interface cannot add a signal to a driver command, milder regimes in the replications, and intervals over six vehicles.

Sensitivity to magnitude, duration, operating point and normalization.

  • Magnitude. Each accepted direction is estimated at three pulse sizes that must agree within 15%, and the conclusions hold for thresholds of 10–30% (§4; §5.2).
  • Duration. Both pulse families were repeated at 0.25 s and 0.1 s. At every duration the predictors under-respond to future pulses (gains 0.19–0.47) and over-respond to history pulses (errors 5.0–9.4), and shorter pulses bring no systematic improvement (Appendix G.10).
  • Operating points. Two anchors in each of three conditions (Tables 3, 15 and 42), and forecast skill by speed, yaw-rate and braking regime (Figure 8).
  • Normalization of the response error. Three channel weightings give predictor errors of 0.995–1.043 for future pulses and 4.99–8.23 for history pulses in arms B and E (Appendix G.2).
  • Normalization of regret. Regret is reported with absolute and offset-removed costs and in physical units (Tables 22 and 36). On the second road the candidates differ little in cost, so normalized regret is unstable there, which Table 3 notes.

Why does the untrained encoder do better, and what is attributable to training?

The reference outputs are among the encoder inputs, so retaining them is a property of the inputs and the architecture. An MLP on the raw history reaches R² 0.999, and larger readouts narrow the gap between trained and untrained encoders from 0.015 to 0.005 (§5.1; Table 17). A retention score is therefore informative only next to an untrained reference (§6). Training does change what the other lenses measure. With data and architecture fixed, masking target commands raises neighbor overlap by 0.052 [0.037, 0.064] in 6/6 vehicles, close to the untrained level, and improves forecasting and history response. These effects recur with the second driver (+0.051) and are not resolved on the second road (Table 35), so this attribution depends on the data.

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

The paper proposes a protocol for assessing action-conditioned latent models for vehicle control. It separately evaluates retention of physical quantities, latent organization, forecasting, and responses to small command changes. The case study uses simulated driving data from IPG CarMaker.

Critical strengths
  • The proposed protocol is pertinent. It evaluates each property separately.
Critical weaknesses
  • The paper is hard to read. The abstract has very long sentences, making the ideas difficult to follow.
  • The results are also hard to understand. The values reported in Tables 1 and 2 are difficult to match with the numbers in the text.
  • Given the model used in the paper, it is unclear how broadly the findings translate to JEPA world models.
  • Figure 1 was not informative to me.
Questions for authors
  • How was the latent dimension of 64 chosen? Do the results change with smaller or larger latent dimensions?
  • Could the authors justify why they chose VICReg.style regularization for their model?
  • Since both models receive x,y positions as inputs, does the supervised forecaster perform better because it predicts physical output directly, or because it uses a different architecture?
2 comments
AuthorsOpenReview ↗

Specific Response to Reviewer HHEX

We thank the reviewer for the careful comments. We wrote a general response summarizing the improvements to the reviewed paper, and below are the specific responses to your comments:

Readability of the abstract and Figure 1.

The abstract is rewritten so that each sentence follows from the previous one, with the same question and contributions and the results of the revision. Figure 1 is redrawn, with a legend for its three colours. Panel (a) shows training with one block per operation, the future commands entering the predictor, the moving-average target encoder, and the command mask of each arm. Panel (b) shows the three response paths of §5.2 as three rows that share the same blocks, ending in the response vectors Y, T and P.

Numbers in Tables 1 and 2 are hard to match with the text.

The text reports paired effects with bootstrap intervals over vehicles, and the tables report medians per arm. These are different estimands, as §5 states, and Table 10 places them side by side for every contrast in the text. For example, masking target commands changes the one-second forecast by +0.042 [0.010, 0.081] as a paired effect and by 0.018 as a difference of medians. Table 2 now shows future and history pulses side by side, and Table 35 gives the contrasts of the replicated conditions.

How broadly do the findings translate to JEPA world models?

The model is an experimental instrument and the protocol is the contribution (§1). The revision varies the transition (Table 4), the latent width and learning rate (Appendix G.7), the target space (Table 20), and the route and driver (Table 3). The encoder family, the moving-average target and the VICReg-style terms are fixed (§6). We describe the findings as evidence about this action-conditioned latent predictive model in CarMaker.

Latent dimension 64, and other widths.

The width follows the implementation of Ulmen et al. (2025) on which the predictor is based and was not selected by search (§3). In the search, widths 32, 64 and 128 give validation errors of 0.118, 0.111 and 0.112 for the original transition, so width barely matters there. Wider latents help the other two transitions, and the best configuration overall still misses the reference scales (§5.2; Appendix G.7).

Why VICReg-style regularization?

It comes from the same implementation, with the default VICReg weights (§3). The terms discourage collapse and redundant coordinates. They impose neither physical-neighborhood preservation nor correct command sensitivity, which the protocol measures separately (Appendix F).

Does the supervised forecaster do better because of physical outputs or architecture?

Table 20 matches the architecture, inputs and budget, and changes only the regression target.

MLP targetR² 1 sFuture-pulse errorHistory-pulse error
Physical outputs0.9830.9263.531
Latent, arm B0.9632.7794.470
Latent, arm E0.9641.9964.387

Under this matched budget, the physical target forecasts and responds better, so the architecture alone does not explain the difference. Two limits apply. Output dimension and readout path still differ, and most latent-target fits were still improving at the epoch cap, so the comparison holds at this budget and does not show an irreducible disadvantage of latent targets (§5.2; Appendix G.11).

I thank the authors for the response and the revision. The revised abstract and Figure 1 are now much clearer. However, the main claims of the paper have been toned down, and the main contribution is now the protocol. In my opinion, the paper in its current form belongs to a benchmark track rather than the main conference track, and I therefore maintain my score of Weak Reject.

I encourage the authors to continue in this direction and apply the protocol to other objectives (e.g., LeJEPA SIGReg objective or I-JEPA-style masked prediction) and to pretrained latent world models, as I find this line of work promising.

Reviewer_w5gXOpenReview ↗
Rating
2
Confidence
4
Informativeness
1
Presentation
2
Soundness
1
Read review
Paper summary

The paper proposes a diagnostic protocol for vehicle latent predictive models, separately evaluating physical information retention, neighborhood organization, predictive capability, and local command responsiveness. The authors validate the protocol on a self-generated CarMaker dataset using a JEPA architecture, five training conditions, six vehicle configurations, and three random seeds. The results show that the recoverability of physical quantities does not necessarily imply accurate responsiveness, and that distorted responses may also lead to poorer candidate command selection. The paper’s main contributions lie in its diagnostic framework and empirical findings.

Critical strengths

Clearly distinguishes representational recoverability from control-relevant responsiveness and examines its consequences for decision-making.

Critical weaknesses

The empirical coverage supporting the central findings is limited. All five training conditions are based on the same latent model architecture, and the data are generated using the same route and driving controller. Moreover, the main responsiveness experiments include only one scenario per vehicle. Although the six vehicle configurations and multiple random seeds improve the repeatability of the results, they are insufficient to determine whether the observed failures depend on the specific architecture or data generation process. This limits the evidence supporting the generalizability of the diagnostic protocol.

The added value and practical benefits of the diagnostic protocol have not been sufficiently established. The paper’s main contribution is the integration of representation probing, neighborhood analysis, and local response measurements. However, it does not clearly demonstrate what substantive new capabilities this integration provides compared with existing representation probing and dynamics diagnostics tools. Although responsiveness fine-tuning improves some response metrics, it generally degrades the original predictive performance, and its ability to improve closed-loop control has not been evaluated. Therefore, while the paper provides reasonably convincing evidence that such models can fail, the evidence remains limited regarding how the proposed protocol can guide reliable model selection or improvement. This limits my assessment of the significance of its contributions.

Questions for authors

Could the authors provide additional data and experiments under more diverse conditions, such as different routes, driving controllers, or latent model architectures, to evaluate the robustness of the central findings and further clarify the scope of their applicability?

Could the authors also elaborate on the unique value of the proposed diagnostic protocol compared with existing evaluation tools, and explain how it can guide model selection or improvement? Demonstrating an approach that improves response consistency while preserving predictive accuracy would help strengthen the paper’s methodological contribution and practical significance.

2 comments
AuthorsOpenReview ↗

Specific Response to Reviewer W5GX

We thank the reviewer for the careful comments. We wrote a general response summarizing the improvements to the reviewed paper, and below are the specific responses to your comments:

Limited coverage: one architecture, one route and driver, one scenario per vehicle (Q1).

Routes and drivers. The full design is repeated on a second road with the same driver and on the primary road with a second driver that has different acceleration limits and drives at a median of 60 instead of 109 km/h (§3; Table 3; Appendix G.4, G.8).

  • Replicated in both: the untrained encoder retains the outputs better (90/90 fits each), the predictors under-respond to future pulses and over-respond to history pulses already in latent coordinates, and calibrated regret exceeds that of keeping the template.
  • Not replicated: the forecast gain from future-command input (+0.015 [−0.001, 0.034] and −0.004 [−0.019, 0.011], against +0.067 in the primary condition). The abstract no longer claims it. The target-masking effects recur with the second driver only.

Architectures. Three transitions under one training regime (Table 4), a search over transition, width and learning rate (Appendix G.7), a matched MLP with latent or physical targets (Table 20), and a supervised forecaster. The encoder family and objective are shared, which §6 states.

Scenarios. Each condition has two response anchors, and the primary pulses are repeated at two shorter durations (Appendix G.10). No closed-loop controller is evaluated, and the paper makes no closed-loop claim.

What does the protocol add, and how does it guide model selection (Q2)?

In this study, forecast scores and probes alone lead to conclusions that the other lenses reverse.

  1. Model selection. Among the fixed-encoder transitions, forecast accuracy selects the GRU, response error the ODE with spectral normalization, and calibrated regret the ODE without it. Selecting by forecast picks the model with the largest response error, 5.909 (Table 4; §6).
  2. Retention. Probes report high retention (R² 0.968–0.974), yet an untrained encoder scores higher in 89/90 fits, and in 90/90 in each replication (§5.1).
  3. Pose. An R² of 0.887 coexists with 78.6 m and 91.6° RMSE (§5.1).
  4. Supervised forecasting. R² of 0.98–0.995 coexists with response errors near or above those of a zero response (§5.2).
  5. Where to act. Pose fails already in the decoded realized target, so the readout must change. Response error arises in the predictive path before the readout (§5.2). §6 names the next step for each.

Fine-tuning degrades forecasting. Can response improve while accuracy is preserved?

Not with what we tested, and the paper says so. Predictor updates lower response error in 12/12 fits but in only about half of the accepted cases (52/104 and 50/104), and they raise forecast error in 12/12 fits. A forecast-preserving variant still raises it in 11/12 (§5.2; Appendix G). None of the transitions or searched configurations reaches the response scale. The protocol is the instrument that would show such an improvement.

Thank you for your response and clarifications. I have noted the updates and will maintain my original recommendation.

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

This paper studies action-conditioned latent predictive models for vehicle control and proposes a measurement protocol that separately evaluates physical-state retention, physical-neighborhood organization, forecasting, and local command response. Its central diagnostic compares simulator outputs, decoded embeddings of realized future windows, and decoded model predictions to distinguish response discrepancies in the representation/readout path from those introduced by prediction. The CarMaker study includes six vehicle configurations, five training conditions, and three random seeds, totaling 90 model fits. The results show that observed physical quantities remain recoverable, although untrained encoders generally support better recovery; improvements over latent persistence do not guarantee accurate physical forecasts or command responses. A finite-candidate ranking experiment further shows that response errors can lead to suboptimal local action choices. The main contribution is a diagnostic protocol and a vehicle-simulation case study, rather than a new architecture or control algorithm.

Critical strengths

The paper introduces a control-relevant diagnostic perspective that conventional evaluations can miss. It explicitly distinguishes physical-state accessibility, forecasting accuracy, and command sensitivity. Comparisons against untrained encoders, physical persistence, and supervised forecasting help avoid interpreting strong probe scores or low latent prediction losses as evidence of control competence. This distinction is useful for assessing whether latent world models are suitable for action optimization.

The local-response experiments and discrepancy-localization design are carefully constructed. Matched command pulses from common initial conditions, scale checks across perturbation amplitudes, and comparisons among the simulator, target/readout, and predictor/readout paths go beyond purely correlational probing. Decoder-free response comparisons, a second anchor, and action-ranking analyses with absolute prediction offsets removed further support the presence of response mismatch within the predictive path and its potential relevance to local decisions.

Critical weaknesses

1.The protocol’s broader diagnostic value is not sufficiently validated across diverse models with strong forecasting capability. The experiments use a single encoder and controlled neural ODE combination. In Table 1, even restricting attention to the action-conditioned arms A, B, and E, only 5/18, 7/18, and 11/18 fits, respectively, outperform physical persistence at one second, while a supervised forecaster using the same inputs is substantially more accurate. I appreciate the analyses showing response mismatch on subsets that outperform local persistence. However, exceeding persistence alone does not establish strong dynamics prediction. It therefore remains unclear whether the protocol reveals an independent control-relevant deficiency in competitive latent models or predominantly reflects the limited capability of this particular predictive pipeline. This limits the general methodological contribution, rather than merely raising a concern about achieving state-of-the-art performance.

2.Important alternative explanations for the response mismatch remain unresolved. Training uses closed-loop logs from one route and one driver, whereas response evaluation uses affine open-loop command templates and local perturbations. This permits intervention-response measurement but does not disentangle insufficient independent action excitation, limitations of the particular model or objective, and distribution shift. Additionally, Appendix F states that training data were generated with a custom CarMaker build, while response experiments used the stock binary, without establishing equivalence of their dynamics or local responses. This does not demonstrate that the builds differ, but the possibility should be excluded before attributing mismatch to the learned predictive path. The three-path comparison localizes computational discrepancies; it does not, by itself, identify their causes.

Questions for authors

1.Simulator-build clarification: What modifications distinguish the custom build from the stock binary? Can the authors provide output and pulse-response comparisons under identical initial conditions and commands to establish that the training environment and response-evaluation reference have equivalent dynamics?

2.Limited additional validation: Does the central response mismatch persist for a latent predictor with substantially stronger physical forecasting performance, obtained through reasonable tuning or one alternative predictor with matched inputs? The existing subset analysis against persistence is helpful, but does not fully establish persistence of the failure in a strong latent forecasting model.

3.Attribution and data-support clarification: To what extent are the affine templates and perturbed history–command combinations supported by the training distribution? Can the available evidence distinguish insufficient action excitation in closed-loop logs from effects of the representation objective or predictor? If not, the paper should clearly separate the demonstrated response mismatch from its hypothesized mechanisms.

4.Action-ranking clarification: What absolute cost gaps and physical-unit state errors correspond to the normalized regret values, and are these differences practically meaningful for control? Please also reconcile the Limitations statement that ranking uses seed-17 fits with the Figure 11 caption stating that each point combines three seeds.

2 comments
AuthorsOpenReview ↗

Specific Response to Reviewer JFAD

We thank the reviewer for the careful comments. We wrote a general response summarizing the improvements to the reviewed paper, and below follow the specific responses to your comments:

W1 / Q2. Does the mismatch persist for a latent predictor with stronger forecasts?

Yes, within the models we could obtain, and the strongest forecaster responds worst. With the arm E encoders and readout frozen, three transitions are fitted to the same targets under one stopping rule (Table 4; Appendix G.6):

TransitionR² 1 sBeats persistenceFuture-pulse error / gainCalibrated regret
ODE with spectral normalization0.91718/180.981 / 0.780.214
ODE without it0.93418/181.693 / 1.630.150
GRU0.95318/185.909 / 5.840.168

Forecast accuracy and response error order the three transitions in opposite directions. A search over jointly trained models (three transitions, widths 32/64/128, three learning rates, selected on validation episodes) gives a model with better directional agreement (cosine 0.571) that over-responds (gain 1.23, error 1.163) (Appendix G.7). A supervised forecaster with R² of 0.981–0.995 has future-pulse errors of 0.93–1.24 at the primary anchors (§5.2; Appendix G.5).

We also state the limit plainly. None of these latent models is an accurate absolute forecaster: arm E has one-second RMSE of 78.6 m and 91.6°, against 1.03 m and 2.69° for constant-turn-rate extrapolation, which starts from the measured current pose. Decoding the realized target already gives 46.5 m, so absolute pose is limited by the readout, which must recover a pose accumulated since the episode start (§5.1; Table 23). No latent model meets the pose, velocity or principal response scales (§5.2; Table 43). The separation of forecast and response is therefore shown at this level of latent competence and for an accurate supervised forecaster (§6, Limitations), and the conclusion names the next step for each located cause.

W2 / Q3. Competing explanations and support of the templates.

They are not fully separated, and the paper now separates what is tested from what is hypothesized (§6).

  • Predictor constraint: Removing spectral normalization under identical targets raises the future-pulse gain from 0.78 to 1.63 and the command sensitivity of the vector field about eightfold (Table 31; Appendix G.2). This is consistent with the constraint limiting response size. Training length also differs, and the response without it is still wrong.
  • Target content: Masking target commands lowers the history-pulse error in the primary condition and with the second driver, and not on the second road (Table 35).
  • Independent action excitation: The simulator interface can overwrite a driver command but cannot add an exogenous signal to it (§6). As descriptors only, the history linearly predicts one-second steering with R² of 0.82–0.93 (Table 37), and the two collections realize nearly the same speeds (Table 24).
  • Template support: Restricting to pulses within the 95th percentile of validation distances leaves the errors in place: 0.996 for future pulses and 25.2 for history pulses. These distances describe proximity to logged windows and do not establish independent action support (Appendix G.2).

What is demonstrated is the mismatch itself, already in latent coordinates, in three conditions (Tables 3 and 13). Its mechanism remains hypothesized.

Q1. Custom and stock builds.

The custom build adds logging and an optional controller bridge that is disabled here. Stock CarMaker 15.1 and two executables of the custom build were compared on all six vehicles with identical initial conditions and steering, gas and brake pulses. Across 72 comparisons the largest relative difference between response vectors is 1.87e-8, and the custom executables reproduce all 22 training channels exactly on sampled training prefixes (Appendices F and G.2). These are checks of sampled conditions.

Q4. Physical meaning of the regret, and the seed statement.

At one second, the lateral difference between the selected and the best candidate is a median 1.4 cm (90th percentile 6.3–7.3 cm, maximum 19.5 cm), and the candidates span a median 2.8 cm laterally and up to 3.7° in yaw (Tables 22 and 36). The paper therefore presents the ranking as a local ordering diagnostic and not as evidence of a planning benefit (§5.2). Appendix G.9 gives the exact identity that decides the ranking. The ranking uses all three seeds, as the Figure 11 caption says, and the configuration search, the larger readouts and the predictor updates use one seed. §6 now states both and lists further seeds among the next steps.

Thanks for the authors' additional experiments. The clarification on simulator consistency checks and random seed settings has addressed some of my concerns, so I will raise the Soundness score from 1 to 2. However, in the action ranking experiment, the median horizontal span of candidates is only 2.8 cm. I believe the results support a local ranking diagnosis, but the support for broader practical control implications remains limited. Furthermore, there is still insufficient validation of the applicability of the relevant conclusions to potential models with strong physical prediction capabilities. Overall, I maintain my original rating.

Comments on the paper

1 comment
AuthorsOpenReview ↗

General Response to Reviewers

We thank the reviewers. The reviews share four concerns: whether the forecast–response mismatch only reflects a weak pipeline, what the errors are in physical units, how far the findings hold beyond one route and one driver, and what is demonstrated as opposed to hypothesized. The revision keeps the question, the contributions and the structure, and adds the following.

  • Two more route and driver conditions. The full design (five arms, six vehicles, three seeds, all lenses, two response anchors) is repeated on a second road and with a second driver (§3; Table 3; Appendix G.4, G.8). Both are milder regimes than the primary one, which the paper states.
  • Physical magnitudes and reference scales. Errors are reported in physical units next to persistence and kinematic references (§5.1; Table 23). They are read against application-motivated scales for pose, velocity and response error, presented as analogies and not as certification (§4; Appendix G.1).
  • A controlled transition comparison. Three transitions are fitted to the same frozen targets under one stopping rule (Table 4; Appendix G.6). It replaces the earlier comparison, which mixed architecture and training length.
  • A search for a better latent model over transition, latent width and learning rate (Appendix G.7). No evaluated latent model reaches the reference scales, and the paper says so in the abstract, §5.2 and §6.
  • Robustness: pulse durations (Appendix G.10), the structure and physical size of the ranking test (Appendix G.9), and a supervised forecaster evaluated on all seeds in the three conditions (Appendix G.5).

Three claims of the submission changed:

  • The forecast gain from future-command input appears only in the primary varied collection, so the abstract no longer states it (§5.1).
  • The effects of masking target commands recur with the second driver and are not resolved on the second road (§5.1; §6).
  • The supervised forecaster reaches high forecast accuracy in every condition and still responds inaccurately, with a kind of error that depends on the condition (§5.2).

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

Related papers

Loading the map…

Discussion (0)

Sign in to comment.