acceptodds
Under review as a conference paper at ICLR 2027

Action Learning for Scientific Simulations

Abstract

Understanding and controlling how systems evolve is central to science and engineering, where many problems are defined on trajectories rather than individual states. Existing learned methods often predict local dynamics or generate trajectories, while representative path inference, mechanism steering and constrained trajectory optimization require an objective on complete paths. We introduce LEAP (Learned Energy on Action Paths), which treats the stochastic action itself as the learning target and recovers it directly from trajectory samples. LEAP decomposes the unknown action to an evaluable reference process and learns the remaining path space correction from complete trajectories, yielding a differentiable objective on paths. Once learned, the same action can be reused for path ranking, representative path inference, mechanism tilting, constraints and sampling. Across stochastic transition, orbital and molecular systems, LEAP recovers physical drift from terminally selected trajectories, represents a nonadditive action correction induced by a persistent latent regime and reuses a fixed molecular action for completion, basin preferences, angular constraints and temperature transfer. Code is available https://anonymous.4open.science/r/leap-iclr-7E4D/README.mdhere.

Primary area:
probabilistic methods (Bayesian methods, variational inference, sampling, UQ, etc.)
Keywords:
scientific ML, stochastic dynamical systems, trajectory modeling, generative modeling, action learning

Reviews

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

This paper proposes LEAP (Learned Energy on Action Paths), a framework for learning differentiable action functionals directly from complete trajectories. By estimating a path-level density-ratio correction relative to a tractable reference process, LEAP enables trajectory optimization, sampling, and constraint-based inference without retraining. Theoretical analyses establish action recovery under suitable assumptions and characterize nonadditive trajectory dependencies. Experiments on stochastic transitions, orbital dynamics, and molecular simulations demonstrate accurate action recovery and reuse across different inference objectives.

Critical strengths
  1. The paper introduces a useful perspective by learning reusable objectives over complete trajectories rather than focusing on local dynamics or trajectory generation. This formulation naturally supports multiple downstream inference tasks.

  2. The theoretical analysis provides a principled foundation for action recovery and nonadditive trajectory dependencies. Experiments across several physical systems demonstrate the effectiveness and versatility of the proposed framework.

Critical weaknesses
  1. The core learning objective relies on contrastive density-ratio estimation, a well-established technique. Although applying this approach to complete trajectories is interesting, the paper does not sufficiently clarify its methodological advantages over existing energy-based or density-ratio learning approaches with comparable path-level representations.

  2. LEAP requires sufficient overlap between the reference and target trajectory distributions, which may be difficult to achieve in complex or rare-event systems. Furthermore, the molecular experiments operate on reduced two-dimensional coordinates, leaving the scalability and effectiveness of the framework on more complex molecular configurations insufficiently

The paper presents an interesting and well-motivated framework for learning reusable trajectory-level objectives, supported by theoretical analysis and promising experimental results. However, the methodological novelty beyond existing density-ratio estimation remains insufficiently established, and the dependence on reference distributions raises concerns about generalizability.

Questions for authors

What are the fundamental methodological differences between LEAP and conventional density-ratio estimation applied to complete trajectories?

How does LEAP perform when the reference distribution has poor overlap with the observed trajectories, particularly for rare-event transitions?

Can the authors provide evidence that LEAP remains effective for higher-dimensional molecular representations beyond the two-dimensional collective coordinates?

5 comments

We thank the reviewer for their thoughtful feedback and for recognizing the value of learning reusable objectives over complete trajectories, and the theoretical motivation and experimental diversity. We have clarified the method contributions and extended the experiments on reference overlap, rare events and higher dimensional molecular representations in the updated pdf.

W1/Q1. What distinguishes LEAP from conventional density-ratio estimation on complete trajectories?

The paper does not sufficiently clarify its methodological advantages over existing energy-based or density-ratio learning approaches.

LEAP builds on density ratio estimation to learn the stochastic action as a reusable objective for trajectory inference. Our reference-based construction has two important properties: (i) conditioning the reference on the event selecting the data preserves relative action differences (Proposition A.3), and (ii) under shared diffusion, the structured correction identifies the physical drift difference (Eq. 9), allowing the learned action to be reused under new boundary conditions. Proposition 3.1 further connects the contrastive gradient to conditional maximum likelihood without evaluating the event probability. For persistent latent regimes, Proposition A.5 characterizes the nonadditive interactions captured by global corrections.

We clarified these contributions in Sections 1.1 and 3 and added comparisons with generic path energies on the double well, using the same selected trajectories, reference and objective. With 5,000 training trajectories, we have the following results (Section C.4, Tables 13-14).

CorrectionTraining event NRMSE ↓Unseen events NRMSE ↓
LEAP, structured0.100.10-0.14
Sum pooled encoder0.170.16-0.18
Transformer0.680.77-1.00

We also evaluate on unseen terminal events without retraining, conditioning the reference on each event while keeping the learned correction fixed. This shows that the structured correction improves action recovery over generic path energies and supports reuse under new terminal conditions.

W2/Q2. How does reference overlap affect LEAP, particularly for rare events?

LEAP requires sufficient overlap between the reference and target trajectory distributions, which may be difficult to achieve in complex or rare-event systems.

We agree that reference overlap is important for accurate density ratio estimation. We extended our experiments to study terminal event conditioning, reference overlap and computational cost (Sections C.1 and D).

We evaluate increasingly rare terminal events, with probabilities down to 9e-6. LEAP achieves drift accuracy comparable to conditional maximum likelihood across these events, while directly conditioned Brownian and linear Gaussian references avoid rejection sampling. At the rarest event, conditional MLE with rejection sampled trajectories requires 7,909 times as many simulated path steps (Section C.1). We also study overlap across trajectory lengths and dimensions. At 64 dimensions, a Brownian reference gives path KL 30 nats and drift MSE 0.056. Using a reference with half the true drift decreases these to 7.4 nats and 1.3e-4. With a fitted linear reference that does not require the true drift, increasing K from 8 to 128 decreases drift MSE from 0.024 to 0.0046 (Section D). These results quantify the effect of poor overlap and show how reference choice and additional samples improve action recovery. Constructing efficient conditioned references for more complex nonlinear systems is an interesting direction for future work.

W2/Q3. Does LEAP extend beyond two-dimensional molecular coordinates?

Can the authors provide evidence that LEAP remains effective for higher-dimensional molecular representations beyond the two-dimensional collective coordinates?

We extended the molecular experiments by training separate actions using 2, 4 and 8 torsions and 48 internal coordinates (Section F.2).

RepresentationLEAP AUROC ↑LEAP completion ↓Reference completion ↓
2 torsions0.9850.1160.132
4 torsions0.9800.1100.122
8 torsions0.9750.1100.124
48 internal coordinates0.9930.1140.130

LEAP achieves high held out trajectory discrimination accuracy across all representations and consistently reduces sparse completion error compared with the reference action. With 4 and 8 torsions, LEAP has completion error 0.110, vs 0.134-0.164 for conditional DDPM and flow matching. Action learning and downstream inference therefore extend beyond the original two backbone angles.

We thank the reviewer again for the helpful suggestions and hope these additions address the concerns. We would be happy to answer/clarify any remaining questions during the discussion period.

Thank you,

Authors

Thank you for the detailed rebuttal and additional experiments. Your responses have addressed many of my concerns and strengthened the paper. I have brief follow-up questions.

  • Were the baseline models compared under similar model capacities and training settings?
  • How can a suitable reference process be selected when the underlying dynamics are unknown?
  • Does the performance remain stable as the molecular representation becomes more complex and higher-dimensional?

We thank the reviewer for their follow up questions and are glad that the additions have helped address their concerns and strengthen the paper.

Were the baseline models compared under similar model capacities and training settings?

Yes, we use the same training data, similar model capacities and training budgets for each comparison. We clarify the following in greater detail:

  • Path EBMs (additional comparison): We additionally trained the same 3 path energy architectures using contrastive divergence with Langevin sampling and a replay buffer. We kept the networks, 5000 selected trajectories, training updates and checkpoint selection fixed. At the same cost per update, LEAP's contrastive training achieves correction NRMSE of 0.10, 0.17 and 0.68 for structured, sum pooled and Transformer energies vs 0.15, 0.27 and 1.0 for conventional EBM training. This shows that LEAP's contrastive training improves recovery even with the same path energy architectures.
  • Conditional dynamics (App. C.1): Conditional MLE uses the same potential network, optimiser and number of updates as LEAP, or the same cubic drift family. With 128 trajectories, LEAP decreases drift MSE from 0.34 to 0.12. While both achieve similar accuracy with 5000 trajectories, LEAP does not evaluate the event probability.
  • Molecular models (App. F): LEAP, binary NRE and velocity regression share the same correction family, with each tuned over 4 learning rates and 2 widths (39k-205k parameters). Conditional DDPM and flow matching have approximately 200k parameters, with approximately 4 times as many optimizer updates as LEAP.
  • History models (App. E.3): GRU and neural CDE mixture density models use the same accumulated orbital features, and we tune them over 3 widths and 3 learning rates (50k-1.1M parameters). Increasing training from 100 to 300 epochs does not result in improved recovery. LEAP-Global achieves higher path ordering accuracy on each run as well.

We therefore show that LEAP improves action recovery compared to conventional EBM training, achieves better drift recovery with limited selected data and gives more accurate trajectory ordering than history-conditioned dynamics models. The learned action can also be reused across inference tasks without retraining.

How can a suitable reference process be selected when the underlying dynamics are unknown?

We can construct references with information available from the observed trajectories. For example, on the double well, we use a Brownian reference with diffusion estimated from the observed quadratic variation; in this setting, LEAP recovers the drift as accurately as conditional MLE and also recovers the potential barrier (App. C.1). For alanine, we fit the reference drift to the training increments by least squares and estimate the diagonal diffusion from the residuals. We also condition this reference using bridge proposals, with acceptance rates of 9-22% across the torsion and internal coordinate representations.

We also evaluate the reference fit before training by measuring the likelihood of observed trajectories under the reference. The reference fits the data better in 48 internal coordinates (reference likelihood gap of 1.6 nats per transition) than in aligned Cartesian coordinates (30-38 nats per transition), which helps guide the choice of representation. We also find that increasing the number of reference samples improves recovery when there is limited overlap. Using a fitted linear reference in 64 dimensions, increasing K from 8 to 128 decreases drift MSE from 0.024 to 0.0046 (see App. D).

Does performance remain stable as molecular representations become more complex and higher dimensional?

Yes, across the representations tested. At 600 K, the held-out AUROC is 0.985, 0.980 and 0.975 for 2, 4 and 8 torsions, and 0.993 for 48 internal coordinates (App. F.2). LEAP decreases sparse completion error by approximately 10-12% compared to the reference action across all representations. With 4 and 8 torsions, LEAP gets a completion error of 0.110 vs 0.134-0.164 for conditional DDPM and flow matching.

These results show that action learning is effective in higher-dimensional molecular representations (i.e., as the representations become more complex), and the same learned action also supports different inference objectives without having to retrain.

We thank the reviewer once again for their constructive questions, and we will update the PDF to reflect these additional results. We would be happy to answer any further questions. We hope our responses and revisions resolve your concerns; if so, we would greatly appreciate it if you would consider updating your assessment of our paper.

Thank you,

Authors

Reviewer_BDkgOpenReview ↗
Rating
3(+1)
Confidence
4
Informativeness
2
Presentation
3
Soundness
2
Read review
Paper summary

This paper proposes LEAP (Learned Energy on Action Paths), a framework for learning stochastic action functionals directly from trajectory data. The key idea is to decompose the unknown path action into an analytically tractable reference action and a learned correction, where the latter is estimated through contrastive learning of path-space density ratios.

The authors introduce local, global, and hybrid parameterizations of the action correction. The local formulation is closely related to Girsanov likelihood ratios, while the global formulation can capture nonadditive dependencies across an entire trajectory. Once learned, the action can be reused for trajectory optimization, sampling, and inference under additional preferences or constraints.

The method is evaluated on a double-well stochastic system, an orbital dynamics system with hidden parameters, and alanine dipeptide molecular simulations. The experiments demonstrate drift recovery from terminally selected trajectories, representation of nonadditive path corrections, and reuse of learned actions for trajectory completion and constrained inference.

Critical strengths
  1. Learning a reusable action functional is an interesting perspective. Once the action is available, different trajectory inference tasks, including path optimization, mechanism preferences, and constraints, can be formulated without retraining the model.

  2. The global action parameterization can capture dependencies induced by persistent latent variables, which cannot generally be represented by additive Markovian transition costs. The orbital dynamics experiment provides a useful illustration of this capability.

Critical weaknesses
  1. The central construction of LEAP is to estimate the path-space density ratio between the data and a reference process using a contrastive objective, and then combine this estimate with the known reference action. This is closely related to existing noise-contrastive density-ratio estimation and Girsanov reweighting.

More importantly, the paper does not clearly establish what fundamental advantage is gained by learning the action directly rather than learning the underlying dynamics and constructing the corresponding path action. For Markovian systems, the local LEAP formulation is essentially equivalent, in terms of representation, to learning the drift or transition density. For non-Markovian systems, a general history-dependent dynamics model can also represent the complete path distribution. Therefore, the conceptual distinction between action learning and dynamics learning is not sufficiently supported by either methodological advances or empirical evidence.

  1. The double-well experiment compares LEAP primarily with Kramers–Moyal estimation and local velocity regression. However, these baselines are applied directly to terminally selected trajectories, whose local conditional dynamics differ from the original dynamics. LEAP explicitly accounts for terminal selection through the reference path distribution, giving it an important advantage in this comparison. A more meaningful evaluation should include dynamics-learning approaches that correctly account for terminal conditioning, such as conditional maximum likelihood or Doob-transform-based approaches.

Similarly, the orbital experiment demonstrates the limitations of additive transition models, but does not establish an advantage over general history-dependent or latent-variable dynamics models. Without these comparisons, it is difficult to determine whether LEAP offers a genuine advantage over existing dynamics-learning approaches.

  1. LEAP requires a tractable reference process with sufficient overlap with the data path distribution. Constructing such a reference may itself be challenging for high-dimensional systems and rare transitions. In particular, sampling reference trajectories satisfying terminal constraints can require nontrivial conditional or rare-event sampling. Poor overlap may also make path-space density-ratio estimation increasingly difficult as trajectory length or dimensionality grows.

The paper does not demonstrate how an effective reference process can be constructed when the underlying dynamics are largely unknown, or whether the resulting computational cost offers an advantage over conventional approaches.

  1. The main demonstration of nonadditive action learning uses a specially constructed orbital system with a persistent hidden mass perturbation. Furthermore, the reported performance depends strongly on accumulated trajectory features derived from the nominal physical dynamics. Without these features, the global model performs close to the constant-prediction baseline.

This raises concerns about whether the global action model can learn general non-Markovian path dependencies without substantial problem-specific feature engineering. More convincing evidence on less structured systems would strengthen this contribution.

5.A more fundamental concern arises when the learned action is used to identify minimum-action or most probable paths. LEAP learns the negative log-density of discretized trajectories, but maximizing the probability density of a discrete trajectory is generally not equivalent to identifying the most probable path of the underlying continuous-time stochastic process.

For example, for an SDE with additive noise, the Onsager–Machlup functional generally contains a drift-divergence correction in addition to the quadratic residual term appearing in the Euler–Maruyama transition likelihood. This distinction matters when defining most probable paths through small-tube probabilities in continuous path space.

Consequently, even if LEAP accurately recovers the discrete trajectory distribution, minimizing its learned action does not necessarily yield the most probable continuous-time path in the Onsager–Machlup sense. The paper should clarify which notion of a most probable path is being considered and whether its path-optimization results remain valid or consistent under temporal refinement. This issue is particularly important because trajectory optimization is presented as a central downstream application of the learned action.

Questions for authors
  1. What concrete advantages does LEAP provide over learning a probabilistic dynamics model and constructing its path action? In particular, can the authors demonstrate improvements in statistical efficiency, computational cost, or downstream trajectory inference under comparable model capacity and training data?

  2. Can the authors compare LEAP with dynamics-learning methods that explicitly account for terminal conditioning in the double-well experiment, and with history-dependent or latent-variable likelihood models in the orbital experiment? These comparisons would be important for assessing whether the reported advantages are specific to the action-learning formulation.

  3. How sensitive is LEAP to the overlap between the reference and data path distributions, particularly as trajectory length and system dimensionality increase? What is the practical cost of constructing and sampling a suitable reference process for rare-event trajectories?

  4. To what extent can LEAP-Global learn nonadditive action corrections without the accumulated, physics-informed trajectory features used in the orbital experiment? Can the authors provide evidence of this capability on a less specially structured system?

  5. The learned action is based on discrete-time trajectory likelihoods, whereas the Onsager–Machlup functional governing most probable continuous-time paths generally includes additional terms, such as the drift-divergence correction for additive-noise SDEs. How do the authors reconcile these two formulations? In particular, under what conditions can minimizing the learned discrete-time action be interpreted as identifying a most probable continuous-time path, and are the resulting paths consistent under temporal refinement?

2 comments

We thank the reviewer for their detailed feedback and for recognizing the importance and usefulness of learning reusable stochastic actions and modeling dependencies induced by latent variables. We have extended our experiments to address the concerns about conditional dynamics models, overlap of the reference and continuous time path inference. We have also updated the pdf to reflect the revisions (see changes in blue).

W1/Q1. What are the advantages of learning an action directly?

What concrete advantages does LEAP provide over learning a probabilistic dynamics model and constructing its path action?

LEAP learns a reusable stochastic action by correcting an evaluable reference process. Under shared diffusion, the structured correction recovers the physical drift (Eq. 9), while Proposition 3.1 connects contrastive learning from terminally selected trajectories to conditional maximum likelihood without evaluating the event probability. Compared to conditional MLE using the same potential network, LEAP decreases drift MSE from 0.34 to 0.12 with 128 selected trajectories (Table 7), which shows an advantage in limited data regime.

We also added comparisons with generic path energies on the double well, using the same selected trajectories, reference and training objective (Section C.4). With 5,000 training trajectories, we get

CorrectionTraining event NRMSE ↓Unseen events NRMSE ↓
LEAP, structured0.100.10-0.14
Sum pooled encoder0.170.16-0.18
Transformer0.680.77-1.00

We additionally evaluate three unseen terminal events without retraining, conditioning the reference on each event while keeping the learned correction fixed (Tables 13-14). These results show how the structured correction supports accurate action recovery and reuse across different inference conditions.

W2/Q2. How does LEAP compare with conditional and history-dependent dynamics models?

A more meaningful evaluation should include dynamics-learning approaches that correctly account for terminal conditioning.

We thank the reviewer for this suggestion. We added conditional MLE and Doob h-transform regression baselines that account for terminal conditioning (Section C.1, Table 7). LEAP improves recovery with limited selected data and gets accuracy that is comparable to conditional MLE with 5,000 trajectories, while Doob regression has higher drift MSE (0.0037).

We also added history conditioned GRU and neural CDE mixture density models trained by maximum likelihood using the same accumulated orbital features (Section E.3).

ModelCorrection NRMSE ↓Pairwise ordering accuracy ↑
LEAP-Global (K=64)0.1100.985
GRU mixture density model0.1310.932
Neural CDE mixture density model0.1370.933

LEAP-Global achieves more accurate action recovery and trajectory ordering, extending our comparison to expressive history conditioned dynamics models.

W3/Q3. How does reference overlap affect recovery and computational cost?

Poor overlap may also make path-space density-ratio estimation increasingly difficult as trajectory length or dimensionality grows.

We agree that reference overlap affects the accuracy of density ratio estimation, especially as trajectory length and dimensionality increase. We have extended our experiments to evaluate these effects and see how reference choice can improve recovery (Sections C.1 and D).

Conditioning the reference on the same terminal event decreases drift MSE from 1.26 to 0.00101 (Table 8). Across event probabilities from 0.020 to 9e-6, LEAP achieves accuracy comparable to conditional MLE while avoiding rejection sampling through directly conditioned references. At the rarest event, conditional MLE with rejection sampled paths needed 7,909x more simulated path steps.

We also vary trajectory length and dimensionality. At 64 dimensions, using half the true drift in the reference decreases path KL from 30 to 7.4 nats and drift MSE from 0.056 to 1.3e-4. We also evaluate a fitted linear reference that does not require the true drift, where increasing the reference budget from K=8 to K=128 reduces drift MSE from 0.024 to 0.0046 (Section D). These results show how reference overlap and reference sampling affect action recovery.

W4/Q4. Can LEAP-Global learn without physics-informed features?

More convincing evidence on less structured systems would strengthen this contribution.

We thank the reviewer for their suggestion and have added a ring potential with persistent hidden rotation (Section E.3). Using only raw trajectory coordinates, LEAP-Global gets correction NRMSE 0.27 vs 0.97 for LEAP-Local. This shows nonadditive action learning without the accumulated physical features used in the orbital benchmark.

We have also motivated Proposition A.5 in Section 3 (Eq. 20), which shows how persistent latent variables induce interactions across transitions.

W5/Q5. How does the learned discrete action relate to continuous-time path inference?

Minimizing its learned action does not necessarily yield the most probable continuous-time path in the Onsager-Machlup sense.

We have clarified the distinction between discrete maximum density paths and continuous time Onsager-Machlup paths (Section 2.4). Under temporal refinement, the discrete minimizer approaches the Freidlin-Wentzell minimizer, with path error decreasing from 0.015 to 9.3e-4, while the distance to the Onsager-Machlup minimizer stays near 0.89 (Table 11).

We also show that LEAP's structured correction recovers an explicit drift, allowing us to construct and optimize the corresponding Onsager-Machlup functional. We get mean action gaps of 0.010-0.025 nats for observation intervals up to 0.02 (Table 12), demonstrating that the recovered drift supports continuous time path inference.

We thank the reviewer again for the helpful suggestions and hope that these additions address the concerns. We would be happy to answer and clarify any further questions during the discussion period.

Thank you,

Authors

Reviewer_eA2LOpenReview ↗
Rating
3
Confidence
3
Informativeness
2
Presentation
2
Soundness
2
Read review
Paper summary

The authors reframe learning for scientific simulations as predicting how probably an entire trajectory is, rather than learning a given state transition. An example given is that while NSDEs learn a drift governing a system's local evolution, the author's method, LEAP, learns a scalar objective assigning a cost to a whole trajectory. On several experiments, LEAP shows strong performance; notably on the double-well potential.

Critical strengths

The experimental evaluation is thorough and covers a broad range of scientific tasks. Several of these are directly relevant to my own research, and I expect many other researchers will feel similarly. This breadth makes it immediately clear where LEAP fits into the current landscape of scientific modelling methods, and where it could be useful.

The theoretical analysis is strong, particularly Prop A.5 which shows why persistent latent variables induce actions that cannot be represented by memoryless transition-additive models. As a disclaimer, I have not carefully checked the proof. To me this seems like a more interesting theoretical result than the paper gives credit for, perhaps it should be included in the body with some motivation? apologies if I missed this.

Critical weaknesses

The relationship to energy-based models (EBMs) is not explored in much detail, beyond a brief discussion of transition-level EBMs in Section 1.1. Formally, LEAP is essentially a conditional EBM on path space, with pθ(Φ∣c)∝exp⁡[−Sθ(Φ;c)]p_\theta(\Phi\mid c)\propto \exp[-S_\theta(\Phi;c)], trained using standard contrastive density-ratio estimation. I would like to see a more substantial comparison with existing EBMs, as it is unclear what methodological novelty LEAP introduces beyond defining the energy over complete trajectories.

The empirical performance of LEAP is not especially convincing:

  • Table 9, binary NRE outperforms LEAP handily in NRMSE
  • Table 13, binary NRE and simple velocity regression outperform LEAP on completing MD trajectories
  • Table 3, physical features appear necessary for LEAPs performance

I appreciate that the authors report these negative results so openly. I interpret them less as evidence that LEAP is a poor method, and more as evidence that the authors have not convincingly demonstrated its advantages over existing EBMs, density-ratio estimators, and other established approaches. My concern is therefore not with LEAP's performance in isolation, but rather whether it offers a meaningful improvement over existing methods.

Questions for authors

What distinguishes LEAP from applying standard noise-contrastive density-ratio estimation to complete trajectories?

Why does binary NRE outperform LEAP in Tables 9 and 13? What advantage does the proposed categorical comparison objective provide? I would like to see some interpretation of this wrt where to, and where not to, use LEAP.

In Table 3, the global model fails without accumulated physical features. How well does LEAP generalize when such features are unavailable, and are they usually available in practice? It seems to me they may be, but I would appreciate if the authors could comment on how this affects applicability of their method.

The authors argue in Section 4.2 that LEAP captures nonadditive path dependencies that transition-level models cannot. Could they compare against a history-conditioned model, such as a neural CDE with a probabilistic output head, which can also represent these dependencies?

5 comments

We thank the reviewer for the thoughtful feedback and for recognizing the experimental breadth and theoretical analysis. We also appreciate the suggestion to emphasize Proposition A.5, which we now discuss further in Section 3 (Eq. 20), to better motivate why we need these global actions. We have also updated the pdf to reflect additions and changes (as seen in blue).

W1/Q1. What distinguishes LEAP from conventional path EBMs and density ratio estimation?

It is unclear what methodological novelty LEAP introduces beyond defining the energy over complete trajectories.

LEAP builds on density ratio estimation to learn a reusable stochastic action for trajectory inference. We combine an evaluable reference process with a structured correction that identifies the physical drift under shared diffusion (Eq. 9), allowing the action to be reused under new inference conditions. For terminally selected trajectories, Proposition 3.1 shows that contrastive gradient approaches the conditional maximum likelihood without evaluating event probability, which provides a way to learn physical dynamics from selected trajectories. We clarified these contributions in Sections 1.1 and 3 and added comparisons with generic path energies on the double well using the same selected trajectories, reference and objective. With 5,000 training trajectories, we get:

CorrectionTraining event NRMSE ↓Unseen events NRMSE ↓
LEAP, structured0.100.10-0.14
Sum pooled encoder0.170.16-0.18
Transformer0.680.77-1.00
We evaluate three unseen terminal events without retraining (Section C.4, Tables 13-14). This shows that the structured correction improves action recovery over generic path energies and transfers to unseen terminal events without retraining.

W2/Q2. Why does binary NRE outperform LEAP in some experiments?

Why does binary NRE outperform LEAP in Tables 9 and 13? What advantage does the proposed categorical comparison objective provide?

We thank the reviewer for this question. Both objectives estimate the path log density ratio, but the categorical objective compares trajectories within the same context, thus allowing context dependent normalization constants to cancel. Proposition 3.1 also shows that the gradient approaches the conditional maximum likelihood gradient as the number of reference samples increases, connecting it to learning from event-conditioned trajectories.

We extended the comparisons across reference budgets and molecular simulations (Sections C.4 and F.2). On the double well, both objectives get drift MSE near 0.001 across K=1-32, while binary NRE achieves lower orbital error at K=4 (0.175 vs 0.228). Binary NRE also gets slightly lower completion error in the original molecular experiment (0.0897 vs 0.0931). On additional molecular simulations, LEAP, binary NRE and velocity regression have similar completion errors (0.1145, 0.1157 and 0.1151, respectively). We prefer categorical comparisons for event-conditioned action learning and their conditional likelihood interpretation, while binary NRE can achieve lower error at smaller reference budgets. The key contribution of LEAP is learning a structured action that can be reused across inference tasks.

W3/Q3. How dependent is LEAP-Global on accumulated physical features?

In Table 3, the global model fails without accumulated physical features.

We note that the accumulated orbital features summarize how the nominal gravitational dynamics respond to the uncertain solar mass, so they provide information across transitions and can be constructed when a nominal simulator is available. We have clarified this in Section 4.2 and added another ring potential with a persistent hidden rotation. We use only raw trajectory coordinates, and see that LEAP-Global has correction NRMSE 0.27 vs 0.97 for LEAP-Local (Section E.3), showing that global corrections can capture nonadditive dependencies without the features used in the orbital experiments.

W4/Q4. Can history-conditioned dynamics models capture the same dependencies?

Could they compare against a history-conditioned model, such as a neural CDE with a probabilistic output head?

We thank the reviewer for this suggestion. We added GRU and neural CDE mixture density models trained by maximum likelihood using the same accumulated orbital features (Section E.3).

ModelCorrection NRMSE ↓Pairwise ordering accuracy ↑
LEAP-Global (K=64)0.1100.985
GRU mixture density model0.1310.932
Neural CDE mixture density model0.1370.933

These results extend our comparison to more expressive history-conditioned dynamics models and show the benefit of directly learning the complete path action for accurate action recovery and trajectory ordering.

We thank the reviewer again for the helpful suggestions and hope these additions address the concerns. We would be happy to clarify/answer any remaining questions during the discussion period.

Thank you,

Authors

Thanks for the rebuttal. The clarified contributions and new experiments, particularly the GRU and neural CDE baselines, are appreciated. My remaining concern is that the path-energy comparison ablates LEAP’s own architecture rather than comparing against existing EBMs. I’m open to raising my score after seeing the other reviews.

Comments on the paper

2 comments

We thank the reviewer for their response following the experiments and revisions. We want to further clarify the concern about advantages of learning a stochastic action directly over learning the underlying dynamics.

While a sufficiently expressive history-conditioned dynamics model can represent the same path distribution, this type of representation capacity does not necessarily provide the same action recovery or downstream inference performance. Our experiments have tested these following differences:

  • Physical recovery: LEAP achieves lower drift MSE than conditional MLE with the same potential network and 128 terminally selected trajectories. Our structured correction (Eq 9) also enables reuse under unseen terminal events without retraining by conditioning the reference on each new event.
  • Non-additive actions: With the same features, LEAP-Global achieves a higher pairwise trajectories order accuracy than neural CDE models and history-conditioned GRU.
  • Existing EBMs: LEAP achieves lower correction NRMSE across all architectures at about the same matched cost per update with the same path energy architecture and training data and updates.

We have also addressed the concerns about reference construction, nonadditive action learning without features, and the relationship between discrete actions and continuous time Onsager-Machlup inference. These results show the advantages of learning a reusable action directly from trajectories, including physical recovery from selected data and inference under new conditions.

We believe the additional theoretical and empirical evaluation substantially address the original concerns. We hope this provides grounds for a stronger assessment; if there are any concerns remaining, we would welcome further clarification and are happy to address them, and incorporating additional clarifications in the final version of the paper.

Thank you,

Authors

We thank the reviewers for their constructive feedback. We appreciate that RVEs recognized the value of learning reusable objectives, eA2L highlighted the breadth of scientific applications and analysis and BDkg emphasized action reuse and nonadditive path modeling.

We would like to clarify our contribution and summarize how our revisions address the main concerns.

Contributions

The primary novelty is to introduce action learning as a trajectory-level objective: recovering an evaluable stochastic action directly from complete trajectories as a reusable objective for trajectory inference. Existing learned methods typically predict local dynamics or generate trajectories (Fig. 1). We learn the stochastic action itself, enabling path ranking, representative path inference, mechanism tilting, constraints and sampling without retraining for a range of downstream tasks.

We provide two key capabilities:

  • Physical recovery from selected trajectories: We address the problem that fitting terminally selected transitions can recover conditioned dynamics instead of the underlying physical dynamics. By matching reference conditioning (Proposition A.3) and learning structured action corrections (Eq. 9), we recover the underlying drift and reuse the learned correction under unseen terminal events. Our contrastive gradient also approaches conditional maximum likelihood without having to evaluate event probabilities (Proposition 3.1).

  • Nonadditive action learning: We characterize how persistent latent variables induce dependencies across transitions in which additive Markovian actions cannot represent (Proposition A.5). We then introduce global path corrections to represent these dependencies from complete trajectories.

We also added controlled comparisons to show advantages over existing approaches. With the same path energy architectures and data, LEAP achieves lower action recovery error than existing EBM training. We also compare against conditional MLE and Doob h-transform regression, showing improved drift recovery over conditional MLE with limited data, and history-conditioned GRU and neural CDE models, demonstrating higher trajectory ordering accuracy. For molecular trajectory completion, we further evaluate against DDPM and flow matching.

This shows the benefit of action learning for trajectory-level reasoning, enabling physical recovery from selected trajectories and reuse of the learned stochastic action for inference under new conditions.

Reference construction and generalizability

We can construct references from observed trajectories without knowing the underlying physical dynamics or requiring an oracle reference. We estimate diffusion for stochastic transitions and fit reference drift and diffusion for molecular simulations. We also evaluate reference overlap and show how conditioning and additional reference samples improve recovery.

For generalizability, we conducted experiments that span different scientific systems:

  • Stochastic transitions, where we recover physical drift from terminally selected trajectories, including rare events with probabilities down to 9e-6, and reuse the learned correction under unseen terminal conditions.
  • Nonadditive dynamics, where we show global action learning from raw coordinates without physics-informed features on a ring potential with persistent hidden rotation.
  • Orbital dynamics, where we evaluate 31- and 42-body systems with persistent latent mass perturbations, demonstrating nonadditive action learning on large trajectories.
  • Molecular dynamics, where we extend alanine dipeptide experiments to 2, 4 and 8 torsions and 48 internal coordinates and show improved sparse completion across these representations and high trajectory discrimination accuracy.

We also clarify the relationship between discrete maximum density paths and continuous time Onsager-Machlup (OM) paths. We show how the recovered drift supports construction and optimization of the OM functional (Section 2.4, Tables 11-12).

We believe our theoretical results, comparisons and empirical evaluations have addressed the reviewers' concerns and established the contribution of action learning. More broadly, our formulation provides a reusable trajectory objective for inference under changing preferences and constraints, with potential applications in robotics and many sequential decision-making settings.

We thank the Area Chair and reviewers for their consideration. We welcome any remaining questions and are happy to address them, including incorporating additional feedback into the final version of the paper.

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

Related papers

Loading the map…

Discussion (0)

Sign in to comment.