acceptodds
Under review as a conference paper at ICLR 2027

NS-WM: Neural-Symbolic World Modeling under Partial Observability

Abstract

World models under partial observability must reuse dynamics when familiar sensor views and operating regimes occur in combinations withheld from training. We study this combination out-of-distribution (combination-OOD) setting with NS-WM, a neuro-symbolic framework that separates view-dependent observation mappings from sparse state-action-regime dynamics. This factorization lets the same transition operate with different sensor mappings, while observation-aware state estimation connects masked histories to the state required for rollout. For controlled systems, an encoder route learns symbolic latent dynamics, and a complementary identification route fits dynamics and sensor maps in reference-sensor coordinates and estimates states by matching visible histories. Each route reuses its current state across alternative action plans; validation-selected mixtures combine their observation predictions. Shared temporal and sensor experts extend the prediction interface to heterogeneous sequences. Across six benchmark suites and an additional controlled population, NS-WM reduces combination-OOD mean squared error (MSE) relative to both evaluated baselines in 13/14 condition-population settings. It improves both combination-OOD forecasting and action-effect skill in all nine controlled settings. Controlled diagnostics recover the active dynamics terms, while showing that accurate equations do not eliminate ambiguity in history-based state estimation.

Primary area:
learning on time series and dynamical systems
Keywords:
Neuro-symbolic world models, Partial observability, System identification, Action-conditioned prediction

Reviews

Among:
Shape:
1234this paper, Top 49%
Figure 1. Mean review rating of the reviewed papers at ICLR 2027. This paper averages 2.33 over 3 reviews, above 51% of the others.
Reviewer_CcotOpenReview ↗
Rating
2
Confidence
4
Informativeness
2
Presentation
2
Soundness
2
Read review
Paper summary

This paper proposes NS-WM, a neural-symbolic framework for world modeling under partial observability, with a focus on compositional generalization across sensor views and operating regimes. The main idea is to separate view-dependent observation mappings from shared state-action-regime dynamics, so that previously seen views and regimes can be recombined at test time.

The framework includes two prediction routes. One identifies sparse dynamics and sensor mappings in reference-sensor coordinates, and estimates the current state by fitting the observed history. The other uses a recurrent encoder to infer a latent state and learns symbolic dynamics in that space. Their predictions are combined using validation-selected weights. The paper also introduces additional forecasting experts for heterogeneous sequence datasets.

Experiments are conducted on six benchmark suites, including a controlled pendulum system and several time-series datasets. The results show improvements over neural world model and direct forecasting baselines on most evaluated settings. The authors also provide diagnostics on recovered dynamics, action effects, and state-estimation failures.

Critical strengths

Separating the dynamics from the observation model seems like a sensible way to approach compositional generalization, especially when the same physical system can be observed through different sensors. I also think the emphasis on state estimation under partial observability is important. A shared dynamics model is not very useful if it cannot be initialized reliably from the available observations, and the paper makes a reasonable effort to address this issue.

The controlled pendulum experiments are interesting, particularly the recovery of interpretable dynamics and the evaluation under alternative action plans. I also appreciate the analysis of failure cases. The state-aliasing example, where different state estimates explain the observed history almost equally well but lead to very different future predictions, is quite informative. It is good that the authors discuss this issue and acknowledge that the fusion strategy does not always help.

Critical weaknesses

My biggest concern is that the full NS-WM framework does not seem to offer a clear advantage over its identification component. In fact, the results in Appendix Table 4 are rather surprising. On the main controlled pendulum setting, the identification-only model achieves much lower MSE than the final mixture on IID, combination-OOD, and parameter-OOD evaluations. The difference is quite large, especially for combination-OOD, where the identification model obtains an MSE of 0.000430 compared to 0.07971 for the final model.

This makes me question how much the neural-symbolic encoder route and the fusion mechanism actually contribute. The paper does acknowledge that identification alone performs better on these test splits, but I think this deserves more attention given that the combination of these components is central to the proposed framework. Why should we prefer the full model over the simpler identification approach? Are there settings where the neural-symbolic route clearly helps, rather than introducing additional prediction errors? I would have liked to see a more systematic comparison of the individual components across the benchmarks.

Related to this, the identification route uses fairly strong structural priors. In the controlled pendulum experiments, the first two reference-sensor channels define the state coordinates, and the prescribed dynamics library already contains the relevant terms needed to represent the underlying system. The method also uses paired multi-view observations during training. These are reasonable assumptions for a structured system identification approach, but they make the problem considerably easier than general world modeling from partial observations. I am not fully convinced that the experiments demonstrate how well the approach works when the physical coordinates are not readily available, or when the true dynamics are not well represented by the predefined library. Since compositional generalization is the main motivation, I think it would be useful to understand how much of the improvement comes from these priors.

I also found the experimental comparisons somewhat difficult to interpret. Although the paper reports results across six benchmark suites, the actual prediction mechanisms differ across them. The controlled systems use the identified dynamics and history-based state estimation, while some heterogeneous sequence tasks rely on additional direct multi-horizon ridge forecasting experts. These are rather different mechanisms, and I am not sure that their results should be interpreted as evidence for the same neural-symbolic world modeling approach. It would help to separate the conclusions supported by the controlled experiments from those supported by the more general forecasting experiments.

The baseline comparison is another concern. The final NS-WM model involves multiple components, additional structural information, model selection, and in some settings an ensemble of symbolic predictors. In contrast, the reported neural and direct forecasting baselines use different architectures and computational budgets. The authors are transparent about this, but it is still difficult to tell which parts of the performance gain are due to the proposed modeling principles. In particular, a stronger comparison against a SINDYc-style model with history-based state estimation, or other factorized state-space models with comparable supervision, would make the results more convincing.

Finally, the statistical evidence seems a little limited given the strength of some of the generalization claims. The main baselines are averaged over three model seeds, while NS-WM is reported as a single selected composite. The additional pendulum population is also used during development, and each held-out pairing has its own trained and selected model. I understand that running the full pipeline repeatedly may be expensive, but some estimate of its variability would be helpful. I would be cautious about drawing strong conclusions from the overall 13/14 result without a more consistent evaluation protocol.

Questions for authors
  1. The identification-only model performs substantially better than the final NS-WM mixture in the main controlled experiments (Appendix Table 4). Could the authors explain why the fusion strategy causes such a large degradation? More importantly, in which settings does the neural-symbolic encoder route provide a clear advantage over identification alone? A component-wise comparison across more benchmarks would be very helpful.

  2. How sensitive is NS-WM to the choice of reference-state coordinates and the predefined dynamics library? The pendulum experiments seem to benefit from a representation that is already close to the true physical state and a library containing the correct functional terms. I would be interested in seeing what happens when these assumptions are relaxed, for example by using a less informative reference view or removing some of the relevant candidate functions.

  3. Could the authors clarify which components of NS-WM are actually used in each benchmark suite? In particular, how much of the performance on the heterogeneous sequence datasets comes from the symbolic dynamics model, and how much comes from the additional direct forecasting experts? A clearer component-level comparison would help establish whether the same underlying mechanism is responsible for the improvements across tasks.

  4. Would it be possible to include a more directly comparable structured baseline, such as sparse system identification combined with history-based state estimation, under similar supervision and computational budgets? I think this would be more informative than comparing only against the neural world model and direct forecaster, especially given how strong the identification component already is.

  5. Could the authors provide some measure of variability for the complete NS-WM pipeline, rather than comparing one selected composite with the average of three baseline seeds? It would also help to clarify how much model selection and validation tuning contribute to the reported results.

22 comments
AuthorsOpenReview ↗

Response to W1 & Q1 (Due to character limits, the reply will be split into multiple messages. Thank you for your understanding.)

We thank the reviewer for raising this central point and agree that it deserves more attention. The analysis is in Appendix F.1 (Tables F.1–F.5).

(1) Why the gap arises. The forecast targets carry sensor noise with variance 4.0e−4 (§4.1), and identification alone reaches 4.2–5.0e−4 on the main test splits, i.e., the noise level; S still beats both K=3 ensembles there (0.0797 vs. 0.258 and 0.103 on combination-OOD; Table F.23). View 2 measures (θ − ω, ω, cos θ) with the first channel hidden, and the visible channels and dynamics are 2π-periodic, so θ and θ + 2π give the same visible history (§4.4, Appendix D). A few aliased validation windows carry 99.8–99.9% of the identification route's validation error and set the static encoder weight on view 2 / channel 0 to 1, while the other eight weights stay at most 0.007 (Eq. 18; Tables F.1–F.2). On combination-OOD splits whose held-out view is not view 2, S equals identification alone.

(2) Settings where the encoder route helps.

  • Aliasing at test time. On (1,1.0) IID, S lowers MSE from 0.2526 to 0.0529 (Table F.1). On the larger test sets of Table F.24, the MSE ratio of S to identification alone on the main combination-OOD split is 0.90, compared with 185 on the Table 4 split (Table F.3). With three new data seeds and all components rebuilt, S is lower on all nine IID splits (16–76% per run; Table F.4), and over the nine controlled splits identification alone and S reach 0.0285 and 0.0460 on the paper's test splits, 0.160 and 0.157 on the Table F.24 episodes, and 0.114 and 0.087 with the new data seeds (Table F.5).
  • No reference coordinates. On the five heterogeneous suites, S combines the symbolic encoder ensemble with the expert: 12/15 endpoints below both K=3 ensembles, against 7/15 for each component alone; on the IID endpoint S is lower than both components and both ensembles in all five suites (e.g., Dysts IID: 0.719 vs. 0.755, 4.28 and 1.08 / 1.07; Table F.20).

(3) Systematic component comparison. The comparison covers Table F.1 (three controlled conditions: identification alone, symbolic encoder ensemble alone, S, validation, skill and fitted weights), Tables F.3–F.5 (larger evaluation sets), Table F.20 (heterogeneous suites) and Table F.7 (route-level ablations: S 12/12 against both ensembles with worst-case ratio 0.815, identification alone 11/12 with 2.713, encoder route alone 4/12 with 1.234). They show that on combination-OOD splits whose held-out view is not view 2, S coincides with identification; where aliasing occurs, S reduces the error on most splits; and without reference coordinates, the combination outperforms each component.

(4) Why prefer the full model. S is best on 24/27 primary endpoints against both the K=3 ensembles and a validation-selected baseline system (Table F.23), while identification alone applies to only 12 of them. Identification alone is a special case of S (w = 0 in Eq. 7), and the paper's selection criterion prefers S ("Validation selection reduces macro-domain MSE by approximately 17% relative to identification alone", §4.4; 21–24% lower validation MSE with new data seeds, Table F.4). S is the only configuration below both baselines on every controlled endpoint (Table F.7).

In short, the Table 4 gap comes from one validation-fitted fusion weight that responds to 2π aliasing in a few validation windows; where such aliasing occurs at test time, S is generally more accurate (14 of the 17 aliased splits with new data seeds, including all nine IID splits), and S is the only configuration that beats both K=3 ensembles on all 12 controlled endpoints (Table F.7).

Revisions in the paper (blue): §4.4 "Component tradeoffs" is rewritten along these lines; §3.4 states that w = 0 recovers the identification route; the Abstract and §5 state when each route is most accurate; the Table 4 caption is revised; new Appendix F.1 (Tables F.1–F.5) and Appendix F.5 (Table F.20).

This question also prompted the cross-benchmark component comparison of Appendix F.1, for which we are grateful.

AuthorsOpenReview ↗

Tables for Response to W1 & Q1 (1/3) (Due to character limits, the reply will be split into multiple messages. Thank you for your understanding.)

Tables for Response to W1 & Q1 (1/3)

Table F.1 (Appendix F) — Component MSE in the three primary controlled conditions (extension of Table 4). Identification = identification route alone; Symbolic ensemble = K=3 encoder-route ensemble alone; Final mixture = the validation-selected S of Table 2. "Validation" is validation macro-domain MSE, the selection quantity. Skill = combination-OOD action-effect skill. Bold = best component in each row (values within 1% are all bold). W_enc lists the fitted encoder weights (rows: views 0–2; columns: channels 0–2). Forecast targets carry sensor noise with standard deviation 0.02 (§4.1), i.e., a noise variance of 4.0e−4.

ConditionSplitIdentificationSymbolic ensembleFinal mixture
(2,1.0)Validation0.22001.1140.1823
IID0.0004980.138330.04035
Comb-OOD0.0004300.093020.07971
Param-OOD0.0004190.184150.16849
Comb-OOD skill0.95140.92190.9432
W_enc[[0, .0067, 0], [0, 0, 0], [1, .0069, 0]]
(0,0.7)Validation0.44810.40630.1692
IID0.0004300.08250.0354
Comb-OOD0.0004100.02180.000410
Param-OOD0.0004170.08840.0331
Comb-OOD skill0.95140.89250.9514
W_enc[[0, .0004, 0], [.0004, 0, 0], [1, 0, 0]]
(1,1.0)Validation0.43972.6310.2737
IID0.25260.09660.0529
Comb-OOD0.0004530.02700.000453
Param-OOD0.0004190.02510.0036
Comb-OOD skill0.96220.94690.9622
W_enc[[0, .0031, 0], [0, 0, 0], [1, .0016, 0]]

Table F.2 (Appendix F) — Periodic state aliasing and the fitted weight on view 2 / channel 0. Aliased windows are windows whose history-based state estimate is offset by ±2π (squared error ≈ (2π)² ≈ 39.5), detected post hoc by checking whether a 2π shift of the estimate reduces the forecast error; this diagnostic uses future targets and plays no role in inference or selection. "Share" = fraction of the identification route's squared error from aliased windows. The last column refits the weight of Eq. (16) without the aliased validation windows (diagnostic only). In test splits that include view 2, view 2 / channel 0 accounts for 89.7–99.8% of the squared error of S (98.8–99.8% in the main condition).

ConditionSplitAliased windows (episodes)Share of identification squared errorView-2/ch-0 MSE on aliased samples: identification → encoderFitted w(view 2, ch 0)Refit without aliased windows
(2,1.0)Validation24 (1)99.8%39.5 → 31.51.00.0023
IID / Comb-OOD / Param-OOD0 / 0 / 0——
(0,0.7)Validation34 (3)99.9%39.5 → 12.61.00.0008
IID / Comb-OOD / Param-OOD0 / 0 / 0——
(1,1.0)Validation40 (2)99.9%39.5 → 22.91.00.0025
IID23 (2)99.8%39.5 → 7.97
Comb-OOD / Param-OOD0 / 0——

Table F.3 (Appendix F) — Trained models on the test episodes of Table F.24 (30 episodes per regime; IID spans two regimes). Nothing is refit or selected, and S reproduces Table F.24. The last column gives the same ratio on the primary test splits (Table F.1). Aliased windows appear in all seven splits that include view 2. S is below both K=3 ensembles in 9/9 cells; identification alone is below both in 8/9. Bold = lower value (both if within 1%).

ConditionSplit (episodes)Aliased episodesIdentificationSS / IdentificationSame ratio on the paper's split
(2,1.0)IID (60)10.02060.04492.281
Comb-OOD (30)61.0950.9900.90185
Param-OOD (30)10.02480.09353.8402
(0,0.7)IID (60)40.12110.11890.9882
Comb-OOD (30)04.12e−44.12e−41.001.00
Param-OOD (30)10.02480.04261.779
(1,1.0)IID (60)50.12650.10550.830.21
Comb-OOD (30)05.87e−45.87e−41.001.00
Param-OOD (30)10.02480.01450.588.6
AuthorsOpenReview ↗

Tables for Response to W1 & Q1 (2/3) (Due to character limits, the reply will be split into multiple messages. Thank you for your understanding.)

Tables for Response to W1 & Q1 (2/3)

Table F.4 (Appendix F) — New data seeds with larger validation and test sets. Three new data seeds × three conditions = 9 runs; all NS-WM components rebuilt with the paper's per-condition training settings (identification route refit, three encoder models retrained, fusion weights refit on validation); 14 training episodes per regime as in the paper, 20 validation and test episodes per regime; selection rule unchanged. Entries are seed means (validation macro-domain MSE in "Validation" rows, test MSE otherwise); "Change" is relative to identification alone. Over the 9 runs: validation contains aliased windows in 9/9, and the fitted w(view 2, ch. 0) ranges from 0.25 to 1.0 (all other weights ≤ 0.011); 17 of 27 test splits contain aliasing (IID 9/9), and S is lower on 14 of them, including all nine IID splits (16–76% per run). Conditions sharing a data seed reuse evaluation episodes where their cells overlap, so counts over runs are paired. N and D are not part of this experiment (see Table F.24). Absolute errors exceed those in Table 4 because aliased episodes dominate the squared error.

ConditionSplitAliased episodes (per seed)IdentificationSChange
(2,1.0)Validation2 / 3 / 2 of 400.06060.0478−21.2%
IID6 / 5 / 4 of 400.17410.1317−24.3%
Comb-OOD3 / 0 / 2 of 200.34760.3038−12.6%
Param-OOD1 / 1 / 0 of 200.026320.02645+0.5%
(0,0.7)Validation2 / 5 / 6 of 400.22420.1736−22.6%
IID5 / 4 / 8 of 400.21150.1391−34.2%
Comb-OOD0 / 0 / 0 of 204.10e−44.19e−4+2.0%
Param-OOD1 / 1 / 0 of 200.026310.02667+1.4%
(1,1.0)Validation6 / 4 / 5 of 400.12690.0963−24.1%
IID7 / 7 / 4 of 400.20970.1414−32.6%
Comb-OOD0 / 0 / 0 of 206.17e−45.79e−4−6.3%
Param-OOD1 / 1 / 0 of 200.026320.01371−47.9%

Table F.5 (Appendix F) — Mean test MSE over the nine controlled test splits. Three conditions × {IID, combination-OOD, parameter-OOD}, equal weights. For the new data seeds, the nine cells are the seed means of Table F.4; the last column gives the change for each seed.

Evaluation setTest episodes per splitIdentificationSChangePer seed
Primary test splits (Table F.1)5–100.02850.0460+61.8%—
Test episodes of Table F.24 (Table F.3)30–600.1600.157−1.9%—
New data seeds (Table F.4)20–400.1140.087−23.4%−23.3% / −33.6% / −17.4%

Table F.7 (Appendix F) — Route-level ablations over the 12 controlled endpoints. Worst-case ratio = maximum of MSE / min(N-ens, D-ens) over the nine MSE endpoints; a value below 1 means that no MSE endpoint is worse than the stronger K=3 baseline. Dense library refits the identification route without sequential thresholding. All variants are refit on training data; no test-based selection is used.

ConfigurationEndpoints beating both K=3 ensemblesWorst-case MSE ratio
Full NS-WM (validation-selected S)12/120.815
Identification route only11/122.713
Encoder route only4/121.234
Identification route, dense library5/1252.19
Identification route, no history fitting3/1268.08

Table F.20 (Appendix F) — Component decomposition on the heterogeneous suites (15 endpoints, inference only). Entries count endpoints with lower test MSE than both K=3 ensembles. The identification route does not apply to these suites (§3.5). The three endpoints where S does not win are those already reported in Table 2. On the IID endpoint, S is lower than the symbolic ensemble alone, the expert alone and both K=3 ensembles in all five suites (e.g., Dysts IID: S 0.719; symbolic ensemble alone 0.755; expert alone 4.28; N-ens / D-ens 1.08 / 1.07).

Suite (endpoints)Symbolic ensemble aloneExpert aloneS (combination)
Application (3)323
Dysts (3)101
TEP (3)223
V2-all (3)023
V2-synthetic (3)112
Total7/157/1512/15
Reviewer_efowOpenReview ↗
Rating
2
Confidence
4
Informativeness
2
Presentation
2
Soundness
1
Read review
Paper summary

This paper proposes NS-WM, a neural-symbolic world modeling framework for partially observable environments. The method separates view-dependent observation mappings from shared state-action-regime dynamics to enable compositional generalization to unseen combinations of sensor views and operating regimes.

The proposed framework combines an encoder-based symbolic dynamics model with a reference-coordinate system identification approach. The two components independently estimate latent states from partially observed histories and generate action-conditioned forecasts. Their predictions are combined using validation-selected fusion weights. The paper evaluates the framework on controlled dynamical systems and several heterogeneous sequence datasets.

While the problem is interesting and the reported results show improvements over the selected neural baselines, I have several concerns regarding the necessity of the proposed components, the fairness of the experimental comparisons, and the strength of the empirical evidence.

Critical strengths

The paper addresses a meaningful problem: compositional generalization in world models under partial observability.

Separating sensor-specific observation mappings from shared dynamics is a reasonable inductive bias for generalizing to unseen view-regime combinations.

The paper evaluates both absolute forecasting accuracy and action-effect prediction, and includes useful analyses of symbolic equation recovery and state-estimation ambiguity.

Critical weaknesses
  1. The full NS-WM framework performs substantially worse than its identification-only component.

This is my primary concern.

Appendix C, Table 4 shows that the identification-only component consistently achieves much lower prediction error than the final validation-selected mixture in the main controlled experiments.

Specifically, the reported MSEs are:

IID: 0.000498 for identification alone versus 0.04035 for the final mixture.

Combination-OOD: 0.000430 versus 0.07971.

Parameter-OOD: 0.000419 versus 0.16849.

Thus, on the main combination-OOD test, the full model has approximately 185 times the error of its identification-only component.

This is a substantial discrepancy. Although the authors acknowledge that the fusion is not consistently beneficial, the results raise a more fundamental question about whether the proposed full framework is necessary.

The strong performance relative to neural baselines appears largely attributable to the identification component rather than the complete neural-symbolic fusion architecture.

I would expect a convincing demonstration that the additional encoder-based route and fusion mechanism provide consistent benefits, or a clearer justification for why the complete model should be preferred over the substantially simpler identification-only alternative.

  1. The experimental comparisons do not sufficiently control for structural priors and computational budgets.

The proposed identification component makes use of relatively strong structural assumptions.

In particular, the controlled experiments define a two-dimensional state using reference-sensor coordinates and employ a prescribed library of polynomial, trigonometric, action, regime, and interaction features.

This library is well aligned with the underlying pendulum dynamics.

In contrast, the neural and direct forecasting baselines do not appear to receive equivalent structural priors. Moreover, the authors explicitly acknowledge that model capacity, ensemble size, identification priors, and search budgets differ across the compared systems.

Consequently, it is difficult to determine whether the reported improvements arise from the proposed architecture or from the substantially stronger inductive bias and additional optimization procedures available to the symbolic identification route.

A more controlled comparison with existing sparse system identification approaches, structured neural state-space models, and appropriately matched computational budgets would strengthen the empirical claims.

  1. The heterogeneous benchmarks do not consistently evaluate the same proposed mechanism.

The paper reports results across six benchmark suites, but the actual modeling procedures differ substantially between the controlled systems and the heterogeneous sequence datasets.

In the controlled experiments, NS-WM uses explicitly identified symbolic dynamics and observation-aware state initialization.

For several heterogeneous datasets, however, the method additionally relies on ridge-regression forecasting experts, direct multi-horizon predictions, and dataset-specific validation mixtures.

These components do not necessarily share the same reusable symbolic transition structure that motivates the main framework.

Furthermore, some heterogeneous datasets do not provide genuine action-conditioned interventions. For instance, the TEP action descriptors encode scheduled manipulation metadata rather than measured actuator controls.

Therefore, the experiments demonstrate the performance of a collection of related prediction pipelines, but provide limited evidence that one unified neural-symbolic world modeling mechanism consistently generalizes across these settings.

  1. Limited statistical reliability of the main comparisons.

The experimental reporting also raises concerns.

The primary neural and direct baselines are averaged over three model seeds, whereas the final NS-WM predictor is a single validation-selected composite. As acknowledged in Appendix B, no independent full-pipeline uncertainty estimate is available.

The main controlled evaluation contains only five physical episodes, with numerous overlapping history windows and multiple action branches derived from the same histories.

These observations are not independent samples, making the large number of evaluated histories or prediction endpoints insufficient evidence of independent experimental replication.

Moreover, the additional pendulum population was examined during development and is explicitly not presented as a prospectively blinded evaluation.

The reported improvements are encouraging, but stronger independent replications and matched uncertainty estimates are needed to establish their robustness.

  1. The claimed generalization remains limited to relatively restricted settings.

The most convincing equation-recovery and action-conditioned experiments are performed on a low-dimensional controlled pendulum system with an appropriately chosen symbolic feature library.

It remains unclear how well the method would scale to more complex nonlinear dynamics, high-dimensional partially observed systems, or environments where the correct symbolic basis is not known in advance.

The paper also reports state-aliasing failures and adverse parameter-OOD outcomes, showing that accurate symbolic dynamics do not necessarily ensure reliable state inference or extrapolation.

Finally, downstream closed-loop control performance is not evaluated, limiting the evidence for the method's practical utility as a world model.

Questions for authors

Given that identification alone substantially outperforms the final NS-WM mixture in Table 4, what is the empirical justification for retaining the encoder-based route and fusion mechanism?

How would the results change when comparing against stronger system identification and structured state-space baselines with comparable prior knowledge and computational budgets?

Which experimental results directly demonstrate the benefit of reusable symbolic dynamics on heterogeneous sequence datasets, independently of the additional ridge-regression forecasting experts?

Can the authors provide full-pipeline results across multiple independent seeds or data populations, with appropriate uncertainty estimates?

How does NS-WM perform when the prescribed symbolic library does not closely match the underlying physical system?

20 comments
AuthorsOpenReview ↗

Response to W1 & Q1 (Due to character limits, the reply will be split into multiple messages. Thank you for your understanding.)

We thank the reviewer for raising this central point. The analysis is in Appendix F.1 (Tables F.1–F.5).

On the ratio of 185. The forecast targets carry sensor noise with standard deviation 0.02 (§4.1), i.e., variance 4.0e−4, and identification alone reaches 4.2–5.0e−4 on the three main test splits. The 0.0797 of S on combination-OOD remains below both K=3 ensembles (0.258 and 0.103; Table F.23). Averaged over the nine controlled splits, identification alone and S reach 0.0285 and 0.0460 on the paper's test splits, 0.160 and 0.157 on the larger test sets of Table F.24, and 0.114 and 0.087 with new data seeds and larger evaluation sets (Table F.5).

Where the gap comes from. View 2 measures (θ − ω, ω, cos θ) with the first channel hidden; the visible channels and the dynamics are 2π-periodic, so θ and θ + 2π give the same visible history and differ only in the hidden channel (§4.4, Appendix D). In 24, 34 and 40 validation windows of the three conditions, the history fit converges to the 2π alias (squared error ≈ (2π)² ≈ 39.5). These windows carry 99.8–99.9% of the identification route's validation error, so least squares sets the encoder weight on view 2 / channel 0 to 1, while the other eight weights are at most 0.007 (Eq. 18; Tables F.1–F.2). The main test splits contain no aliased windows, so there S inherits the encoder error on this single channel; on combination-OOD splits whose held-out view is not view 2, S equals identification alone (Table F.1).

On whether the gains come from the identification component. In the controlled suite, the identification route is an integral part of S and implements the paper's first two contributions (shared sparse state–action–regime dynamics with history-fitting state estimation, §3.2). S adds what identification alone lacks: it is the only configuration that beats both K=3 ensembles on all 12 controlled endpoints (Table F.7), and it reduces the error when the history fit converges to a 2π-offset state (§5). The identification route exists only for the controlled pendulum; four of the 13 settings in the 13/14 result come from heterogeneous suites (Table F.25), where S combines the encoder route with the expert and reaches 12/15 endpoints, against 7/15 for each component alone (Table F.20).

Evidence for the encoder route and fusion.

EvidenceSettingIdentification aloneSSource
Below both K=3 ensembles12 controlled endpoints (worst-case ratio)11/12 (2.713)12/12 (0.815)Table F.7
9 cells, test episodes of Table F.248/99/9Table F.3
27 primary endpoints, best entrydefined on 12 only24/27Table F.23
Without reference coordinatesfive heterogeneous suitesnot applicable12/15 (each component 7/15)Table F.20
Lower error under aliasing(1,1.0) IID, paper's test split0.25260.0529Table F.1
main comb-OOD, Table F.24 episodes1.0950.990Table F.3
new data seeds: IID splits where S is lower—9/9Table F.4
Mean test MSE, nine splitspaper's test splits / Table F.24 episodes / new data seeds0.0285 / 0.160 / 0.1140.0460 / 0.157 / 0.087Table F.5
Validation (selection criterion)three primary conditions0.2200 / 0.4481 / 0.43970.1823 / 0.1692 / 0.2737Table F.1

Why S rather than identification alone. Identification alone is a special case of S (w = 0 in Eq. 7), and S has the lower validation error in all three conditions, so it is the selected predictor. Table F.1 is read together with Tables F.2 and F.4 (combined in the first table below): identification alone is lower only on primary splits without aliased windows, while S is below both baselines on every primary split and lower on all nine IID splits and 14 of the 17 aliased splits with new data seeds.

Revisions in the paper (blue): §4.4 "Component tradeoffs" is rewritten along these lines; §3.4 states that w = 0 recovers the identification route; the Abstract and §5 state when each route is most accurate; the Table 4 caption is revised; new Appendix F.1 (Tables F.1–F.5).

We thank the reviewer again for pressing on this point, which led to the component analysis of Appendix F.1.

AuthorsOpenReview ↗

Tables for Response to W1 & Q1 (1/4) (Due to character limits, the reply will be split into multiple messages. Thank you for your understanding.)

Tables for Response to W1 & Q1 (1/4)

Combined view of Tables F.1, F.2, F.4 and F.23 (all values from the revised PDF). Left: the primary data (data seed 2027), comparing S with the K=3 baselines (Table F.23); identification alone is listed for reference with its number of 2π-aliased test windows (Table F.2). Right: three new data seeds with all NS-WM components rebuilt and larger validation and test sets (Table F.4, seed means), with aliased episodes per seed. Validation rows give validation macro-domain MSE, the selection criterion; other rows give test MSE. Bold: left, lowest of S, N-ens and D-ens (validation rows: lower of S and identification); right, lower of identification and S (both if within 1%).

ConditionSplitSN-ensD-ensIdent. (aliased windows)New seeds: aliased eps.New seeds: Ident.New seeds: SChange
(2,1.0)Validation0.1823––0.2200 (24)2 / 3 / 2 of 400.06060.0478−21.2%
IID0.040350.13570.20880.000498 (0)6 / 5 / 4 of 400.17410.1317−24.3%
Comb-OOD0.079710.25800.10330.000430 (0)3 / 0 / 2 of 200.34760.3038−12.6%
Param-OOD0.168490.20750.20670.000419 (0)1 / 1 / 0 of 200.026320.02645+0.5%
(0,0.7)Validation0.1692––0.4481 (34)2 / 5 / 6 of 400.22420.1736−22.6%
IID0.03540.067890.097630.000430 (0)5 / 4 / 8 of 400.21150.1391−34.2%
Comb-OOD0.0004100.020550.039190.000410 (0)0 / 0 / 0 of 204.10e−44.19e−4+2.0%
Param-OOD0.03310.071680.074330.000417 (0)1 / 1 / 0 of 200.026310.02667+1.4%
(1,1.0)Validation0.2737––0.4397 (40)6 / 4 / 5 of 400.12690.0963−24.1%
IID0.05290.10750.093100.2526 (23)7 / 7 / 4 of 400.20970.1414−32.6%
Comb-OOD0.0004530.024130.047020.000453 (0)0 / 0 / 0 of 206.17e−45.79e−4−6.3%
Param-OOD0.00360.022720.030340.000419 (0)1 / 1 / 0 of 200.026320.01371−47.9%
Mean of nine test splits0.04600.02850.1140.087−23%

Reading: S has the lower validation error in every condition on both data sets and is therefore the selected predictor; on the primary data, S is below both K=3 baselines on every test split, while identification alone is lower only on splits with no aliased windows and exceeds both baselines on the one split that has them ((1,1.0) IID); with all components rebuilt on new data seeds, S is lower on all nine IID splits and on 14 of the 17 aliased splits, and its mean test error is 23% lower (Tables F.4–F.5).

Table F.7 (Appendix F) — Route-level ablations over the 12 controlled endpoints. Worst-case ratio = maximum of MSE / min(N-ens, D-ens) over the nine MSE endpoints; a value below 1 means that no MSE endpoint is worse than the stronger K=3 baseline. Dense library refits the identification route without sequential thresholding. All variants are refit on training data; no test-based selection is used.

ConfigurationEndpoints beating both K=3 ensemblesWorst-case MSE ratio
Full NS-WM (validation-selected S)12/120.815
Identification route only11/122.713
Encoder route only4/121.234
Identification route, dense library5/1252.19
Identification route, no history fitting3/1268.08
AuthorsOpenReview ↗

Tables for Response to W1 & Q1 (2/4) (Due to character limits, the reply will be split into multiple messages. Thank you for your understanding.)

Tables for Response to W1 & Q1 (2/4)

Table F.24 (Appendix F) — Fresh test episodes (controlled pendulum). New episode seeds, disjoint from every seed used in development (including data seed 2028); simulator and splits identical to the paper's, 30 episodes per regime (IID: 60). Trained checkpoints are reused unchanged, with no refitting or selection. Cells show MSE / action-effect skill; bold = best in each row. Ep. wins = episodes on which S has lower MSE than both N and D; p = one-sided exact sign test. S is best in all nine cells on both metrics, and its episode wins are significant at p < 0.003 in every cell.

ConditionSplitEp.SNDN-ensD-ensEp. winsp
(2,1.0)IID600.0449 / 0.9702.846 / 0.7731.945 / 0.6972.801 / 0.7871.900 / 0.75558/601.6e−15
(2,1.0)Comb-OOD300.990 / 0.9431.309 / 0.8431.082 / 0.7721.228 / 0.8831.016 / 0.85323/302.6e−3
(2,1.0)Param-OOD300.0935 / 0.9560.1908 / 0.9020.1354 / 0.8820.1486 / 0.9250.1071 / 0.90725/301.6e−4
(0,0.7)IID600.1189 / 0.9640.4451 / 0.8880.6193 / 0.5640.3705 / 0.9180.5662 / 0.61851/601.5e−8
(0,0.7)Comb-OOD304.12e−4 / 0.9581.920 / 0.6401.710 / 0.5091.524 / 0.7251.501 / 0.60430/309.3e−10
(0,0.7)Param-OOD300.0426 / 0.9520.1333 / 0.8640.1460 / 0.5610.1050 / 0.9070.1154 / 0.63328/304.3e−7
(1,1.0)IID600.1055 / 0.9560.2021 / 0.9100.2126 / 0.8000.1885 / 0.9230.1879 / 0.83259/605.3e−17
(1,1.0)Comb-OOD305.87e−4 / 0.991107.8 / 0.530101.7 / 0.320107.6 / 0.569101.5 / 0.34330/309.3e−10
(1,1.0)Param-OOD300.0145 / 0.9580.1180 / 0.8960.1104 / 0.7720.1065 / 0.9110.0916 / 0.80229/302.9e−8

Table F.23 (Appendix F) — Matched comparisons, including a validation-selected baseline system (primary scope, 27 endpoints). Inference only. N, D and the S backbones share the paired-view, all-view and action-effect loss weights (0.2 each), the optimizer, the training budget (160 epochs, patience 35) and the checkpoint criterion. N-ens / D-ens average seeds 11/22/33 (K=3, matching S). N⊕D is the Eq. (16) per-view/channel validation fusion of N-ens and D-ens. B* is the baseline system selected by validation macro-domain MSE from {N₁₁, N₂₂, N₃₃, D₁₁, D₂₂, D₃₃, N-ens, D-ens, N⊕D} using the validation selection rule of S. No test metric enters any selection. S is best in 24/27 rows against both the K=3 ensembles and B*.

Suite / conditionSplit / metricSN-ensD-ensN⊕DB*
ApplicationIID0.10230.12880.13100.12710.1317
ApplicationComb-OOD5.55e−80.19870.0012050.14420.00784
ApplicationDataset-OOD37.6246.2747.6946.2145.50
Pend. (2,1.0)IID0.040350.13570.20880.20810.4207
Pend. (2,1.0)Comb-OOD0.079710.25800.10330.10310.2011
Pend. (2,1.0)Comb. skill0.94320.90180.88680.89820.8771
Pend. (2,1.0)Param-OOD0.16850.20750.20670.20880.2364
Pend. (0,0.7)IID0.035370.067890.097630.068790.06879
Pend. (0,0.7)Comb-OOD4.10e−40.020550.039190.020840.02084
Pend. (0,0.7)Comb. skill0.95140.92600.74990.90880.9088
Pend. (0,0.7)Param-OOD0.033110.071680.074330.047770.04777
Pend. (1,1.0)IID0.052880.10750.093100.095540.1034
Pend. (1,1.0)Comb-OOD4.53e−40.024130.047020.024130.02337
Pend. (1,1.0)Comb. skill0.96220.93370.87520.93370.9339
Pend. (1,1.0)Param-OOD0.0036010.022720.030340.023240.02084
DystsIID0.71901.0801.0700.97270.9727
DystsComb-OOD0.011610.013070.010390.010350.01035
DystsSystem-OOD0.0038890.0039720.0027040.0029440.002944
TEPIID0.19530.73360.35330.38350.3835
TEPComb-OOD0.16740.17500.18610.17560.1756
TEPTarget-OOD2.1012.6382.4022.4382.438
V2-allIID0.086670.090440.093320.087440.08744
V2-allComb-OOD4.12e−77.42e−54.62e−53.36e−53.36e−5
V2-allDataset-OOD275.5295.6312.8297.3297.3
V2-syntheticIID0.20910.22460.22750.22080.2208
V2-syntheticComb-OOD0.37170.58510.53820.58090.5809
V2-syntheticDataset-OOD32.6630.2631.1130.7730.77
Reviewer_T3KEOpenReview ↗
Rating
3
Confidence
2
Informativeness
2
Presentation
3
Soundness
2
Read review
Paper summary

This paper proposes NS-WM, a neuro-symbolic framework for combination-OOD prediction under partial observability. It separates view-dependent observation mappings from shared sparse state-action-regime dynamics, and combines an encoder-based latent predictor with a reference-coordinate identification and history-fitting route. Experiments on controlled systems and heterogeneous sequence suites show promising performance on withheld factor combinations and alternative action prediction. The paper also provides useful diagnostics of recovered dynamics, state ambiguity, and prediction fusion.

Critical strengths

S1. Well-motivated prediction setting. The paper considers an interesting form of compositional generalization where individual sensor views and operating regimes are observed during training but their specific pairings are withheld. The factorization of sensing and dynamics is an intuitive and potentially useful inductive bias.

S2. Promising empirical results. NS-WM outperforms both reported neural baselines in 13 of 14 combination-OOD condition-population settings and improves action-effect skill in all nine controlled settings.

S3. Useful diagnostic analyses. The recovered pendulum dynamics closely match the reference equations in the prescribed coordinate system and library. The analyses of periodic state aliasing and fusion failures are also informative.

S4. Transparent reporting of limitations. The authors acknowledge unequal computational budgets, dependence between overlapping evaluation windows, lack of full-pipeline uncertainty estimates, and cases where fusion or distributional extrapolation does not improve performance.

Critical weaknesses

W1. The necessity and effectiveness of the complete neuro-symbolic architecture are not established.

The most significant concern is that the identification-only component substantially outperforms the final NS-WM predictor on all three main controlled test splits. For example, Table 4 (p.16) reports combination-OOD MSEs of 0.000430 for identification alone and 0.07971 for the final mixture. Similar large gaps occur for IID and parameter-OOD prediction.

Although the authors appropriately acknowledge this observation, it significantly weakens the motivation for the proposed two-route fusion framework. The reported gains over N/D do not establish that the additional encoder and fusion components are beneficial. A clearer comparison and analysis of the conditions under which each route contributes are needed.

W2. Baseline comparisons do not sufficiently isolate the claimed methodological contribution.

The main comparisons use a neural world model and a direct forecaster. However, the identification route uses a strong reference-coordinate prior, a prescribed symbolic feature library, paired full-channel training observations, and numerical history fitting (Section 3.2 and Appendix A, pp.3–4 and 12–14).

Moreover, the compared systems differ in architecture, supervision objectives, ensemble size, and search budgets (Section 4.1, p.6). It is therefore unclear whether the improvements arise specifically from the proposed compositional factorization or from these additional structural priors and supervision.

A structured SINDYc-style baseline with the same observation mapping and state-fitting protocol, as well as matched neural ablations, would substantially strengthen the empirical evidence.

W3. The current component analyses do not fully explain why the approach works.

Table 5 removes individual terms from a frozen identified model without retraining or re-estimating the states. This demonstrates dependence on particular dynamics terms but does not isolate the effects of the full training components.

In particular, the contributions of symbolic constraints, history-based state estimation, paired-view supervision, action-effect supervision, and prediction fusion have not been independently assessed through controlled end-to-end ablations (Sections 3.2–3.4 and Appendix C–D).

Questions for authors

Q1 — Most important. Given that the identification-only predictor outperforms the final NS-WM mixture by a large margin in Table 4, what empirical evidence supports the necessity of the encoder route and fusion? Are there controlled settings in which the identification route fails but the full model provides a consistent improvement?

Q2 — Important. Can the authors provide matched comparisons that control for reference-coordinate access, paired-view supervision, action-effect losses, ensemble size, and model-selection budget? In particular, how does a SINDYc-style model with the same sensor mappings and history-fitting procedure compare?

Q3 — Important. Could the authors report independent full-pipeline seed variation, or uncertainty estimates aggregated at the independent episode level, for the main combination-OOD experiments?

18 comments
AuthorsOpenReview ↗

Response to W1 & Q1(Due to character limits, the reply will be split into multiple messages. Thank you for your understanding.)

We thank the reviewer for this important point. All new results are in Appendix F.1 (Tables F.1–F.5).

(1) Why S is the reported predictor.

  • Identification alone is a special case of S (w = 0 in Eq. 7); fusion decides on validation data, per view and channel, whether to depart from it.
  • The paper's selection criterion prefers S: "Validation selection reduces macro-domain MSE by approximately 17% relative to identification alone" (§4.4), with lower validation MSE in all three conditions (Table F.1) and 21–24% lower validation MSE with the new data seeds (Table F.4).
  • S beats both K=3 ensembles on 12/12 controlled endpoints (worst-case ratio 0.815, vs. 11/12 and 2.713 for identification alone; Table F.7) and is best on 24/27 primary endpoints (Table F.23). Identification alone exists on only 12 of these 27 endpoints.

(2) Where S is more accurate.

  • Paper's test splits (Table F.1). On (1,1.0) IID, 2 of 10 test episodes are aliased, and S lowers MSE from 0.2526 to 0.0529.
  • Reading Tables F.1, F.2 and F.4 together (combined in the first table below). The primary test splits contain no aliased windows except (1,1.0) IID (Table F.2); where aliasing occurs, on that split and across the new data seeds (Table F.4), S is the more accurate predictor, and on every primary split S is below both K=3 baselines.
  • New data seeds, all components rebuilt (Table F.4). With 20 validation and test episodes per regime, aliasing appears in every validation set and every IID test split; S is lower than identification alone on all nine IID splits (16–76% per run) and on 14 of the 17 aliased test splits.
  • Mean error (Table F.5). Over the nine controlled splits, identification alone and S reach 0.0285 and 0.0460 on the paper's test splits, 0.160 and 0.157 on the test sets of Table F.24, and 0.114 and 0.087 with the new data seeds (lower for each of the three seeds).

(3) What Table 4 shows. The forecast targets carry sensor noise with standard deviation 0.02 (§4.1), i.e., variance 4.0e−4, and identification alone reaches 4.2–5.0e−4 on the main test splits, so the ratios in Table 4 are measured against the noise floor. View 2 measures (θ − ω, ω, cos θ) with the first channel hidden; the visible channels and the dynamics are 2π-periodic, so θ and θ + 2π give the same visible history and differ only in the hidden channel. This is the ambiguity diagnosed in §4.4 and Appendix D. In 24, 34 and 40 validation windows of the three conditions, the history fit converges to the 2π alias (squared error ≈ (2π)² ≈ 39.5). These windows carry 99.8–99.9% of the identification route's validation error, so least squares sets the encoder weight on view 2 / channel 0 to 1, while the other eight weights stay at most 0.007 (Eq. 18; Tables F.1–F.2). The encoder route sees the same history but is not tied to one history-fitted branch, and on aliased windows its error on this channel is below the full-period error (7.97–31.5 vs. 39.5). The main test splits contain no aliased windows, so there S inherits the encoder error on this single channel; on combination-OOD splits whose held-out view is not view 2, S equals identification alone (Table F.1).

In short, S is the configuration selected on validation and the only one that beats both K=3 ensembles on all 12 controlled endpoints (Table F.7); identification alone is its w = 0 special case and reaches the sensor-noise floor on the alias-free test splits of Table 4, while S is generally more accurate where periodic aliasing occurs (14 of the 17 aliased splits).

Revisions in the paper (blue): §4.4 "Component tradeoffs" is rewritten along these lines; §3.4 states that w = 0 recovers the identification route; the Abstract and §5 state when each route is most accurate; the Table 4 caption now states that its test splits contain no aliased windows and that validation prefers S; Appendix D ("Validation mixture") points to the new Appendix F.1 (Tables F.1–F.5).

We thank the reviewer again for this question; the resulting analysis now makes explicit when each route contributes.

AuthorsOpenReview ↗

Tables for Response to W1 & Q1 (1/4) (Due to character limits, the reply will be split into multiple messages. Thank you for your understanding.)

Tables for Response to W1 & Q1 (1/4)

Combined view of Tables F.1, F.2, F.4 and F.23 (all values from the revised PDF). Left: the primary data (data seed 2027), comparing S with the K=3 baselines (Table F.23); identification alone is listed for reference with its number of 2π-aliased test windows (Table F.2). Right: three new data seeds with all NS-WM components rebuilt and larger validation and test sets (Table F.4, seed means), with aliased episodes per seed. Validation rows give validation macro-domain MSE, the selection criterion; other rows give test MSE. Bold: left, lowest of S, N-ens and D-ens (validation rows: lower of S and identification); right, lower of identification and S (both if within 1%).

ConditionSplitSN-ensD-ensIdent. (aliased windows)New seeds: aliased eps.New seeds: Ident.New seeds: SChange
(2,1.0)Validation0.1823––0.2200 (24)2 / 3 / 2 of 400.06060.0478−21.2%
IID0.040350.13570.20880.000498 (0)6 / 5 / 4 of 400.17410.1317−24.3%
Comb-OOD0.079710.25800.10330.000430 (0)3 / 0 / 2 of 200.34760.3038−12.6%
Param-OOD0.168490.20750.20670.000419 (0)1 / 1 / 0 of 200.026320.02645+0.5%
(0,0.7)Validation0.1692––0.4481 (34)2 / 5 / 6 of 400.22420.1736−22.6%
IID0.03540.067890.097630.000430 (0)5 / 4 / 8 of 400.21150.1391−34.2%
Comb-OOD0.0004100.020550.039190.000410 (0)0 / 0 / 0 of 204.10e−44.19e−4+2.0%
Param-OOD0.03310.071680.074330.000417 (0)1 / 1 / 0 of 200.026310.02667+1.4%
(1,1.0)Validation0.2737––0.4397 (40)6 / 4 / 5 of 400.12690.0963−24.1%
IID0.05290.10750.093100.2526 (23)7 / 7 / 4 of 400.20970.1414−32.6%
Comb-OOD0.0004530.024130.047020.000453 (0)0 / 0 / 0 of 206.17e−45.79e−4−6.3%
Param-OOD0.00360.022720.030340.000419 (0)1 / 1 / 0 of 200.026320.01371−47.9%
Mean of nine test splits0.04600.02850.1140.087−23%

Reading: S has the lower validation error in every condition on both data sets and is therefore the selected predictor; on the primary data, S is below both K=3 baselines on every test split, while identification alone is lower only on splits with no aliased windows and exceeds both baselines on the one split that has them ((1,1.0) IID); with all components rebuilt on new data seeds, S is lower on all nine IID splits and on 14 of the 17 aliased splits, and its mean test error is 23% lower (Tables F.4–F.5).

Table F.7 (Appendix F) — Route-level ablations over the 12 controlled endpoints. Worst-case ratio = maximum of MSE / min(N-ens, D-ens) over the nine MSE endpoints; a value below 1 means that no MSE endpoint is worse than the stronger K=3 baseline. Dense library refits the identification route without sequential thresholding. All variants are refit on training data; no test-based selection is used.

ConfigurationEndpoints beating both K=3 ensemblesWorst-case MSE ratio
Full NS-WM (validation-selected S)12/120.815
Identification route only11/122.713
Encoder route only4/121.234
Identification route, dense library5/1252.19
Identification route, no history fitting3/1268.08
AuthorsOpenReview ↗

Tables for Response to W1 & Q1 (2/4)(Due to character limits, the reply will be split into multiple messages. Thank you for your understanding.)

Tables for Response to W1 & Q1 (2/4)

Table F.24 (Appendix F) — Fresh test episodes (controlled pendulum). New episode seeds, disjoint from every seed used in development (including data seed 2028); simulator and splits identical to the paper's, 30 episodes per regime (IID: 60). Trained checkpoints are reused unchanged, with no refitting or selection. Cells show MSE / action-effect skill; bold = best in each row. Ep. wins = episodes on which S has lower MSE than both N and D; p = one-sided exact sign test. S is best in all nine cells on both metrics, and its episode wins are significant at p < 0.003 in every cell.

ConditionSplitEp.SNDN-ensD-ensEp. winsp
(2,1.0)IID600.0449 / 0.9702.846 / 0.7731.945 / 0.6972.801 / 0.7871.900 / 0.75558/601.6e−15
(2,1.0)Comb-OOD300.990 / 0.9431.309 / 0.8431.082 / 0.7721.228 / 0.8831.016 / 0.85323/302.6e−3
(2,1.0)Param-OOD300.0935 / 0.9560.1908 / 0.9020.1354 / 0.8820.1486 / 0.9250.1071 / 0.90725/301.6e−4
(0,0.7)IID600.1189 / 0.9640.4451 / 0.8880.6193 / 0.5640.3705 / 0.9180.5662 / 0.61851/601.5e−8
(0,0.7)Comb-OOD304.12e−4 / 0.9581.920 / 0.6401.710 / 0.5091.524 / 0.7251.501 / 0.60430/309.3e−10
(0,0.7)Param-OOD300.0426 / 0.9520.1333 / 0.8640.1460 / 0.5610.1050 / 0.9070.1154 / 0.63328/304.3e−7
(1,1.0)IID600.1055 / 0.9560.2021 / 0.9100.2126 / 0.8000.1885 / 0.9230.1879 / 0.83259/605.3e−17
(1,1.0)Comb-OOD305.87e−4 / 0.991107.8 / 0.530101.7 / 0.320107.6 / 0.569101.5 / 0.34330/309.3e−10
(1,1.0)Param-OOD300.0145 / 0.9580.1180 / 0.8960.1104 / 0.7720.1065 / 0.9110.0916 / 0.80229/302.9e−8

Table F.23 (Appendix F) — Matched comparisons, including a validation-selected baseline system (primary scope, 27 endpoints). Inference only. N, D and the S backbones share the paired-view, all-view and action-effect loss weights (0.2 each), the optimizer, the training budget (160 epochs, patience 35) and the checkpoint criterion. N-ens / D-ens average seeds 11/22/33 (K=3, matching S). N⊕D is the Eq. (16) per-view/channel validation fusion of N-ens and D-ens. B* is the baseline system selected by validation macro-domain MSE from {N₁₁, N₂₂, N₃₃, D₁₁, D₂₂, D₃₃, N-ens, D-ens, N⊕D} using the validation selection rule of S. No test metric enters any selection. S is best in 24/27 rows against both the K=3 ensembles and B*.

Suite / conditionSplit / metricSN-ensD-ensN⊕DB*
ApplicationIID0.10230.12880.13100.12710.1317
ApplicationComb-OOD5.55e−80.19870.0012050.14420.00784
ApplicationDataset-OOD37.6246.2747.6946.2145.50
Pend. (2,1.0)IID0.040350.13570.20880.20810.4207
Pend. (2,1.0)Comb-OOD0.079710.25800.10330.10310.2011
Pend. (2,1.0)Comb. skill0.94320.90180.88680.89820.8771
Pend. (2,1.0)Param-OOD0.16850.20750.20670.20880.2364
Pend. (0,0.7)IID0.035370.067890.097630.068790.06879
Pend. (0,0.7)Comb-OOD4.10e−40.020550.039190.020840.02084
Pend. (0,0.7)Comb. skill0.95140.92600.74990.90880.9088
Pend. (0,0.7)Param-OOD0.033110.071680.074330.047770.04777
Pend. (1,1.0)IID0.052880.10750.093100.095540.1034
Pend. (1,1.0)Comb-OOD4.53e−40.024130.047020.024130.02337
Pend. (1,1.0)Comb. skill0.96220.93370.87520.93370.9339
Pend. (1,1.0)Param-OOD0.0036010.022720.030340.023240.02084
DystsIID0.71901.0801.0700.97270.9727
DystsComb-OOD0.011610.013070.010390.010350.01035
DystsSystem-OOD0.0038890.0039720.0027040.0029440.002944
TEPIID0.19530.73360.35330.38350.3835
TEPComb-OOD0.16740.17500.18610.17560.1756
TEPTarget-OOD2.1012.6382.4022.4382.438
V2-allIID0.086670.090440.093320.087440.08744
V2-allComb-OOD4.12e−77.42e−54.62e−53.36e−53.36e−5
V2-allDataset-OOD275.5295.6312.8297.3297.3
V2-syntheticIID0.20910.22460.22750.22080.2208
V2-syntheticComb-OOD0.37170.58510.53820.58090.5809
V2-syntheticDataset-OOD32.6630.2631.1130.7730.77

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

Related papers

Loading the map…

Discussion (0)

Sign in to comment.