acceptodds
Under review as a conference paper at ICLR 2027

Dynamical System Discovery from Specifications and Partial Observations

Abstract

Dynamical system discovery is often framed as fitting ordinary differential equations to observed trajectories. In many applications, however, the desired model is defined not only by prediction error but also by a *task specification*: a natural-language description of task-critical mechanisms that the model must represent. In postprandial glucose modeling, for example, a meal response simulator must represent how ingested carbohydrates enter the bloodstream over time; a post-meal decision support model must distinguish meal absorption from delayed insulin action; and a physiological model must expose hepatic regulation. These task-critical mechanisms are often not measured, creating a mismatch between what the data observe and what the model must represent. We study this setting as *task-specified dynamical system discovery* under *specification-induced partial observability*. We propose a framework that treats the state specification itself as an object of discovery. Given a task description and partially observed time series, the framework searches over observed and latent states, analytic dynamics, observation mappings, and causal initialization rules. A proposer large language model (LLM) constructs and revises candidate models through a graph meta model that represents dependencies. A deterministic verifier checks model validity and executable task requirements, while a separate LLM critic supplies scientific feedback for revision. Continuous parameters are fitted on training trajectories using collocation initialization and rollout refinement. Validation performance guides candidate selection and acceptance of pruning. Experiments demonstrate that compared to baselines, our method more reliably recovers task-critical mechanisms while achieving better prediction in unseen, held-out interventions.

Primary area:
applications to physical sciences (physics, chemistry, biology, etc.)
Keywords:
Dynamical system discovery, partial observability, large language models

Reviews

Among:
Shape:
1234this paper, Bottom 27%
Figure 1. Mean review rating of the reviewed papers at ICLR 2027. This paper averages 2.00 over 3 reviews, above 27% of the others.
Reviewer_pp8WOpenReview ↗
Rating
2(+1)
Confidence
3
Informativeness
2
Presentation
3
Soundness
3
Read review
Paper summary

This paper proposes a framework for discovering dynamical systems from natural-language task specifications and partially observed time series. The main contribution of the paper is to jointly discover state variables including latent states and governing equations, guided by task-specific mechanistic requirements. The proposed framework combines LLM-based candidate generation and critique with deterministic verification and numerical parameter fitting. Experiments across different systems, including glucose-insulin dynamics, a chemical reactor, and a synthetic system, show improved predictive accuracy and mechanism compliance over several baselines.

Overall, the paper has a promising methodological idea, but I feel that the current evaluation does not yet establish robustness to limited observations or scalability to complex dynamical systems.

Critical strengths
  1. The paper has an interesting new problem formulation. Jointly discovering the state representation and governing equations is a meaningful extension beyond equation discovery with a predefined state space.

  2. The proposed solution combines LLM-based proposals with deterministic verification and numerical fitting. This is more reliable and interpretable than relying on LLM generation alone.

  3. Its evaluation has an emphasis on mechanism compliance and prediction under unseen interventions. This is valuable, as trajectory fitting alone does not establish mechanistic validity.

Critical weaknesses

I am quite concerned about the evaluation and whether the propose framework works for more general and complex ddynamical systems.

  1. The experiments cover only a small number of structured systems. It is unclear to me whether the method can handle high-dimensional, strongly coupled, or multiscale dynamics. No systematic complexity-scaling or computational-cost analysis is provided in the paper.

  2. Although the paper focuses on partial observations, it does not systematically investigate sensitivity to observation sparsity, measurement noise, or the choice of observed variables.

  3. The framework assumes that the supplied specification is correct and sufficiently informative. It is unclear how robust the method is to incomplete, ambiguous, or overly restrictive specifications.

Questions for authors

The questions are related to the identified weaknesses:

  1. How does performance and computational cost scale with the number of state variables, candidate interactions, and latent states? Could the authors provide a controlled complexity-scaling experiment?

  2. How sensitive is the method to the amount and quality of observed data? In particular, how does performance change under increased measurement noise, sparser sampling, or fewer observed state variables?

  3. How does the framework behave when task specifications are incomplete or partially incorrect? Can it detect inconsistencies between the specification and the observed data?

5 comments
AuthorsOpenReview ↗

Response 1/3: Scaling of Performance and Computational Cost

Scaling

W1: The experiments cover only a small number of structured systems. It is unclear to me whether the method can handle high-dimensional, strongly coupled, or multiscale dynamics. No systematic complexity-scaling or computational-cost analysis is provided in the paper. Q1: How does performance and computational cost scale with the number of state variables, candidate interactions, and latent states? Could the authors provide a controlled complexity-scaling experiment?

We thank the reviewer for this insightful question. We conduct experiments in a setting similar to the alien device benchmark, which allows us to control dimensionality, coupling, and scales of dynamics without worrying about the physical meanings.

Setup

Given a one-dimensional input uu, the states zz are generated by z˙i=−dizi+∑j≠iSijzj+fi(z)+bitanh⁡(αu),\dot z_i=-d_i z_i+\sum_{j \neq i} S_{ij}z_j+f_i(z)+b_i\tanh(\alpha u), where did_i is the decay, representing time scales, SijS_{ij} models coupling between states, and fif_i is a randomly generated function. A one-dimensional output is generated by y˙=−λy+g(z).\dot y=-\lambda y+g(z).

We increase the number of states from 4 to 8 and to 12 (high-dimensional), double the coupling (strongly coupled), and increase the spread of decay rates (multiscale dynamics).

GroupCharacteristicsNumber of systems
4-state base systems3 hidden states2
8-state base systems7 hidden states2
12-state base systems11 hidden states2
8-state, stronger couplingcoupling Sij→2SijS_{ij} \rightarrow 2 S_{ij}2
8-state, wider timescaledecay spread max⁡i,jdi/dj=2→5,6\max_{i,j} d_i/d_j=2 \rightarrow 5,62

Results

We evaluate the performance under unseen intervention with more detailed metrics of response relative squared errors (RSE, evaluating response magnitudes), directional accuracy (evaluting response directions), and shape error (evaluating temporal behaviors). The definitions of these metrics are in the first response to Reviewer Ewor. We run three seeds for each experiment.

Ground-truth systemRetainedResponse RSE ↓Directional accuracy ↑Shape error ↓Total time (min)Proposer tokens (k)
4 states (3 hidden)4/60.54 [1.04, 0.23]0.85 [0.51, 1.00]0.46 [0.99, 0.04]46 [56, 4]178 [342, 138]
8 states (7 hidden)6/61.02 [4.91, 0.24]0.87 [0.35, 0.99]0.43 [1.00, 0.19]16 [85, 7]171 [245, 143]
12 states (11 hidden)6/60.46 [1.00, 0.08]0.90 [0.48, 1.00]0.25 [1.00, 0.08]13 [69, 3]158 [220, 136]
8 states, 2× linear coupling6/60.58 [4.58, 0.18]0.84 [0.38, 0.98]0.43 [0.91, 0.16]14 [66, 9]175 [217, 137]
8 states, wider decay spread6/62.67 [11.88, 1.00]0.84 [0.05, 0.99]0.55 [1.00, 0.33]14 [51, 5]158 [343, 145]

Note: in 2 out of the 6 runs for the 4-state systems, the search did not produce an accepted initial model, therefore skipping the remaining 14 iterations. Thus, 4/6 means four runs produced a retained model and two failed at initial construction.

Conclusion

We can see that our method has stable performance, time complexity, and token usage across most of the scenarios. The two exceptions are the much larger response RSE under wider decay spread, and the two failed runs in the 4-state systems.

AuthorsOpenReview ↗

Response 2/3: Sensitivity to Observation Sparsity, Measurement Noise, and Choice of Observed variables

Sensitivity to Observation Sparsity, Measurement Noise, and Choice of Observed variables

W2: Although the paper focuses on partial observations, it does not systematically investigate sensitivity to observation sparsity, measurement noise, or the choice of observed variables. Q2: How sensitive is the method to the amount and quality of observed data? In particular, how does performance change under increased measurement noise, sparser sampling, or fewer observed state variables?

We thank the reviewer for this important question.

Robustness to Measurement Noise

We vary the level of measurement noises on T2-hard of the glucose-insulin benchmark.

NoiseRetainedTest NMSE ↓Response RSE ↓Directional accuracy ↑Shape error ↓
0%3/31.21 [1.21, 1.17]0.91 [1.00, 0.87]0.66 [0.59, 0.71]0.66 [0.70, 0.64]
1%3/31.24 [1.39, 0.64]0.86 [0.93, 0.84]0.59 [0.59, 0.78]0.68 [0.70, 0.64]
5%2/30.90 [0.90, 0.90]0.87 [0.89, 0.84]0.69 [0.59, 0.79]0.67 [0.69, 0.65]
10%3/30.82 [1.35, 0.35]0.82 [0.85, 0.37]0.80 [0.57, 0.99]0.67 [0.70, 0.33]

Robustness to Observation Sparsity

We vary the sampling intervals on T2-hard of the glucose-insulin benchmark.

Sampling intervalRetainedTest NMSE ↓Response RSE ↓Directional accuracy ↑Shape error ↓
1 min3/31.21 [1.21, 1.17]0.91 [1.00, 0.87]0.66 [0.59, 0.71]0.66 [0.70, 0.64]
2 min3/31.23 [1.41, 0.98]0.96 [1.00, 0.95]0.59 [0.58, 0.80]0.69 [0.71, 0.69]
4 min3/30.94 [0.95, 0.93]0.85 [1.00, 0.84]0.74 [0.59, 0.74]0.71 [0.71, 0.66]

Robustness to Observed Variables

Tasks T1 and T2, each of which has easy and hard tiers, were designed to address the robustness of observed variables.

BenchmarkPrediction targetsObserved StatesRatio of observed to predicted
T1 easyGpG_pEGP,Uii,E,GtEGP,U_{ii},E,G_t4:1
T1 hardGpG_pGtG_t1:1
T2 easyGp,I,UG_p,I,UEGP,Uii,E,GtEGP,U_{ii},E,G_t4:3
T2 hardGp,IG_p,IUiiU_{ii}1:2
BenchmarkRetainedTest NMSE ↓Response RSE ↓Directional accuracy ↑Shape error ↓
T1 easy3/30.20 [0.21, 0.19]0.16 [0.17, 0.15]0.99 [0.99, 1.00]0.13 [0.13, 0.12]
T1 hard3/30.12 [0.19, 0.05]0.15 [0.17, 0.13]0.99 [0.98, 0.99]0.11 [0.13, 0.08]
T2 easy2/30.10 [0.10, 0.10]1.00 [1.00, 1.00]0.41 [0.41, 0.41]0.91 [0.91, 0.91]
T2 hard3/31.21 [1.21, 1.17]0.91 [1.00, 0.87]0.66 [0.59, 0.71]0.66 [0.70, 0.64]

Conclusion

The proposed method is robust to measurement noise and observation sparsity. It is robust to choices of observed variables except T2-hard, which is particularly challenging due to the 1:2 observed to prediction ratio.

AuthorsOpenReview ↗

Response 3/3: Incomplete or Incorrect Task Specifications

Incomplete or Incorrect Task Specifications

W3: The framework assumes that the supplied specification is correct and sufficiently informative. It is unclear how robust the method is to incomplete, ambiguous, or overly restrictive specifications. Q3: How does the framework behave when task specifications are incomplete or partially incorrect? Can it detect inconsistencies between the specification and the observed data?

Rationale behind assuming correct specifications

As we discussed in Sec. 6 of the paper, we assumed that the task specification is correct. We believe that this is a reasonable assumption, because the specification merely requires the existence of certain pathways to fulfill the task, but does not specify the forms of the dynamics. This happens in scenarios where the user has a goal in mind (e.g., insulin therapy design). They can ask for a model including the components to achieve their goal (e.g., insulin pathway), but do not know how the components should look like in an unknown system.

Robustness to incomplete or incorrect specifications

But we do agree that it is meaningful to test how robust the method is to incomplete, over-restrictive, and incorrect specifications. On CSTR-hard, we ran the search under five task specifications while keeping everything else the same:

  • Complete: the original specification (Full).
  • Over-restrictive: adds a requirement of exactly one latent state.
  • Incomplete: removes the jacket heat-exchange requirement.
  • Incorrect: requires, falsely, that the reactor temperature not depend on the jacket feed temperature.

We evaluate the performance under unseen intervention with more detailed metrics of response relative squared errors (RSE, evaluating response magnitudes), directional accuracy (evaluting response directions), and shape error (evaluating temporal behaviors). We run three seeds for each experiment.

SpecificationRetainedTest NMSE ↓Response RSE ↓Directional accuracy ↑Shape error ↓
Complete (Full)3/30.37 [0.75, 0.17]0.14 [0.16, 0.05]0.99 [0.99, 0.99]0.04 [0.05, 0.04]
Over-restrictive2/30.33 [0.43, 0.24]0.16 [0.17, 0.15]0.99 [0.99, 0.99]0.05 [0.05, 0.05]
Incomplete3/30.84 [0.88, 0.43]0.44 [0.44, 0.17]0.72 [0.72, 0.99]0.32 [0.32, 0.05]
Incorrect3/34.79 [4.94, 0.34]0.93 [1.00, 0.16]0.62 [0.01, 0.99]0.86 [1.00, 0.05]

Conclusions.

  • Incorrect specification resulted in the worst performance. This is because under our assumption of correct specification, we have deterministic validity check and LLM critics to ensure that the specification is met.
  • Results on over-restrictive and incomplete specifications give practical guidance in using the system.
    • When unsure, we should provide a relaxed, incomplete specification, rather than a incorrect specification.
    • Specification on the number of latent states (e.g., for model complexity considerations) has less impact as specification on mechanisms.
Reviewer_EworOpenReview ↗
Rating
3(+1)
Confidence
3
Informativeness
2
Presentation
2
Soundness
3(+1)
Read review
Paper summary

This paper investigates task-specified dynamical system discovery under partial observability, aiming to construct interpretable dynamical models that satisfy both observed data and task-specific mechanistic requirements. The authors propose an LLM-guided framework that iteratively constructs and refines candidate models, including latent state representations, their dependencies, analytical governing equations, and initialization rules. The framework incorporates a deterministic verifier to check model validity and task requirements, an LLM-based scientific critic to provide feedback, and numerical optimization for parameter fitting. Candidate models are further refined through pruning and selected based on validation performance. Experiments across glucose–insulin dynamics, a chemical reactor, and a synthetic dynamical system compare the proposed approach with several symbolic regression and LLM-based baselines. The reported results show improvements in prediction accuracy and mechanism compliance, while an additional intervention case study illustrates the potential benefits of task-guided model construction.

Critical strengths
  1. I find the idea of including task-critical latent states and their governing dynamics in the model search space quite interesting. Rather than relying on a predefined state representation, the proposed framework allows an LLM to iteratively construct and refine dynamical models based on fitting feedback and scientific constraints. I think this is a promising direction for automatically discovering interpretable dynamical models from partial observations.

  2. I also appreciate that the authors provide some initial experimental evidence across different dynamical systems, including comparisons with existing baselines and ablation studies. The reported improvements in prediction accuracy and mechanism compliance, together with the intervention case study, suggest that incorporating task-specific requirements into model discovery can be beneficial.

Critical weaknesses
  1. I think the experimental evaluation is still insufficient to fully support the paper's central contribution. Although the proposed framework achieves promising results over the selected baselines, it remains unclear how much of the improvement comes specifically from automatically discovering latent-state representations, rather than incorporating task-specific mechanistic priors. Comparisons with conventional latent-state models or manually specified mechanistic models would help clarify this distinction. Additionally, the paper provides limited analysis of the actual discovered states and equations, including their consistency across different random seeds. Since automatic state-space discovery is a central contribution, I think more direct evidence of its benefits and reliability is needed.

  2. I am not fully convinced that the current experiments sufficiently validate the correctness of the discovered mechanisms. The mechanism compliance metric primarily evaluates whether the generated models satisfy certain predefined functional requirements, but this does not necessarily establish that the underlying latent dynamics are correctly recovered. Although the intervention case study provides some encouraging evidence, it is limited to a particular setting and does not directly establish the validity of the discovered latent mechanisms across tasks. I think this distinction is important given the paper's claim of reliably recovering task-critical mechanisms.

Questions for authors
  1. Could the authors provide a concrete example illustrating how the LLM iteratively constructs and refines a dynamical model? Although the paper describes the overall workflow in detail, I found it difficult to understand how the model structure actually evolves during the search. For example, it would be helpful to show the initially proposed latent states and equations, what feedback motivated subsequent revisions, which states or interactions were added, removed, or modified, and how these changes affected the fitted model. Such an example would help clarify what the LLM actually discovers and optimizes throughout the process.

  2. Since the experiments involve three random seeds per case, I wonder how consistent the discovered model structures are across different runs. Do the LLM-generated models tend to identify similar latent states, dependencies, and governing equations, or do they produce substantially different structures? If the structures differ, how does this variation affect predictive performance and mechanism compliance? I think a comparison of the discovered models across seeds would provide useful evidence about the stability and reliability of the proposed approach.

  3. Could the authors further investigate how much of the reported improvement comes from introducing additional latent states? For example, how would the proposed framework compare with a model that does not introduce new latent states, or one that uses a fixed number of latent states without automatically searching for their representations? I would be particularly interested in whether the improvements arise from the additional modeling capacity itself or from discovering task-relevant state representations.

  4. How can we assess whether the mechanisms generated by the LLM are actually correct or reliable? I understand that the proposed mechanism compliance metric checks whether the fitted equations satisfy certain functional requirements. However, different dynamical mechanisms may satisfy the same requirements while exhibiting different behaviors. Since the experiments use simulated dynamical systems with known reference mechanisms, could the authors provide a more direct evaluation of the discovered mechanisms, for example by comparing their response directions, magnitudes, and temporal behaviors under unseen interventions? I think this would help establish whether the method discovers meaningful dynamics beyond satisfying predefined structural constraints.

  5. Related to the previous question, I wonder how sensitive the final model performance is to the correctness of the discovered mechanism. For example, if two candidate models achieve similar training accuracy but contain different latent-state structures or governing equations, how differently do they perform under unseen conditions? Could the authors compare models with appropriate and misspecified mechanisms, ideally under comparable model complexity and fitting conditions? This could help clarify whether discovering the correct mechanism, rather than simply introducing additional states or satisfying basic structural requirements, is responsible for the improvement.

  6. Finally, could the authors clarify the advantages of LLM-based automatic mechanism discovery compared with manually specifying a reasonable mechanistic model? For example, given the same public task requirements, one could manually introduce latent states representing delayed insulin action and then fit their governing equations using the available data. How close can the automatically discovered models get to such manually designed models in terms of predictive accuracy, mechanism fidelity, and complexity? Since the simulated benchmarks also have known reference dynamics, an additional comparison with reference-based models could provide a useful upper-bound perspective, although I understand that these models would have access to information unavailable to the proposed method. I think such comparisons would help demonstrate the practical value of automatically constructing task-specific dynamical models.

9 comments
AuthorsOpenReview ↗

Correctness of Discovered Mechanisms (W2, Q4) - Part 1/2

Thank you for the insightful review and for recognizing the novelty of the problem formulation. We first address your questions on evaluating correctness of discovered mechanisms. For in-depth discussion, we present a detailed case study on the CSTR-hard benchmark here; statistics on the full benchmark suite will follow once those experiments finish.

Correctness of Discovered Mechanisms (W2, Q4)

W2: The mechanism compliance metric primarily evaluates whether the generated models satisfy certain predefined functional requirements, but this does not necessarily establish that the underlying latent dynamics are correctly recovered.

Q4: Could the authors provide a more direct evaluation of the discovered mechanisms, for example by comparing their response directions, magnitudes, and temporal behaviors under unseen interventions?

Thank you for emphasizing the distinction between satisfying a mechanistic requirement and recovering the correct dynamics. We agree that mechanism compliance alone does not establish correctness. We added a direct evaluation of discovered mechanisms, in terms of their response directions and magnitudes under unseen interventions. As shown in the CSTR-hard benchmark below, our method achieves the highest overall correctness among an expanded set of external baselines.

Background on CSTR-hard

In CSTR, a reactor mixes the reactant while a surrounding jacket exchanges heat with the reactor.

  • Control inputs: concentration CfC_f of reactant feed, and temperature TjfT_{jf} of jacket feed (the cooling/warming liquid entering the jacket);
  • Prediction target: the reactor temperature TT, whose dynamics is governed by T˙=α(Tf−T)+βe−ER/TC−γ(T−Tj),\dot{T}=\alpha(T_f-T)+\beta e^{-E_R/T} C - \gamma (T-T_j), where the reactant concentration inside the reactor CC and the jacket temperature TjT_j become unobserved states in CSTR-hard.
  • Three mechanisms to recover: the three terms in the temperature dynamics: feed transport, reaction heat, and heat transfer to jacket.

Metrics for Response Directions, Magnitudes, and Temporal Behaviors

Unseen interventions. We applied 24 interventions, including extended input ranges (e.g., feed temperature and concentration increasing and decreasing 2-3 times beyond the training range), different input shapes (e.g., pulses of unseen widths, sinusoids of unseen periods), different operating points (e.g., unseen initial temperatures), and combinations of the above. Under these interventions, the prediction target (i.e., reactor temperature) goes well beyond the range in the training trajectories.

Metrics directly evaluating the three mechanisms. For a mechanism m(t)m(t), namely a term of the temperature dynamics equation, we define its response Δm(t)\Delta m(t) as its value in the intervention run minus its value in the matched control run: Δm(t)=mintervention(t)−mcontrol(t).\Delta m(t)=m_{\mathrm{intervention}}(t)-m_{\mathrm{control}}(t).

  • Response directions are measured by directional accuracy, the standard measure of whether predictions get the direction of change right (Pesaran and Timmermann, 1992), with each time point weighted by the size of the realized change (Blaskowitz and Herwartz, 2011): A=∑t∣Δmtrue(t)∣ 1[sign⁡Δm(t)=sign⁡Δmtrue(t)]∑t∣Δmtrue(t)∣.A=\frac{\sum_t |\Delta m_{\mathrm{true}}(t)|\ \mathbb{1}\big[\operatorname{sign} \Delta m(t)=\operatorname{sign}\Delta m_{\mathrm{true}}(t)\big]}{\sum_t |\Delta m_{\mathrm{true}}(t)|}. Higher directional accuracy is better.
  • Response magnitudes are measured by relative squared errors (RSE), defined as (Eqn. (42) in Appendix F) RSE=∑t∥Δm(t)−Δmtrue(t)∥2∑t∥Δmtrue(t)∥2.\mathrm{RSE}=\frac{\sum_t \|\Delta m(t)-\Delta m_{\mathrm{true}}(t)\|^2}{\sum_t \|\Delta m_{\mathrm{true}}(t)\|^2}. Lower RSE is better.
  • Temporal behaviors are measured by shape errors, defined as the uncentered (anomaly) correlation between Δm(t)\Delta m(t) and Δmtrue(t)\Delta m_{\mathrm{true}(t)} (Murphy, 1988; Murphy and Epstein, 1989; Gupta et al., 2009) S=1−max⁡(0,ρ)2,ρ=∑tΔm(t) mtrue(t)∑tΔm(t)2 ∑tΔmtrue(t)2.S=1-\max(0,\rho)^2,\qquad \rho=\frac{\sum_t \Delta m(t)\,m_{\mathrm{true}}(t)}{\sqrt{\sum_t \Delta m(t)^2\,\sum_t \Delta m_{\mathrm{true}}(t)^2}}. Lower shape error is better.
AuthorsOpenReview ↗

Correctness of Discovered Mechanisms (W2, Q4) - Part 2/2

Correctness of Mechanisms: Comparison with Ablated and External Baselines

We compare our method against

  • ablations: latent states not allowd ("No latent") and task specification removed from the prompt ("No specification"); and
  • external baselines: SINDy, PySR, D3, and four new baselines
    • SINDy with a LLM-designed library informed by the specification ("spec-SINDy"), as suggeseted by Reviewer 7pHK,
    • LLM-SR (Shojaee, 2025),
    • LLM-ODE (Bideh and Gryak, 2026), and
    • HDTwinGen (Holt et al., 2024b).

All LLM-based baselines receive the same prompt as our method.

We run three seeds for every method. For each seed, we calculate the means of directional accuracy and shape error and medians of RSEs (because RSEs are not bounded) over 24 interventions. Then we report "median [worst, best]" over three seeds.

MethodFeed transport: directions↑Feed transport: RSE↓Feed transport: shape error↓Reaction heat: directions↑Reaction heat: RSE↓Reaction heat: shape error↓Jacket heat: directions↑Jacket heat: RSE↓Jacket heat: shape error↓
Full0.98 [0.97, 0.98]0.26 [0.34, 0.14]0.14 [0.17, 0.022]0.95 [0.56, 0.96]0.69 [1.0, 0.22]0.16 [0.23, 0.080]0.88 [0.87, 0.90]0.31 [0.45, 0.29]0.23 [0.33, 0.20]
No latent0.97 [0.97, 0.98]0.46 [0.61, 0.14]0.17 [0.20, 0.022]0.95 [0.42, 0.96]0.71 [1.0, 0.22]0.24 [0.25, 0.16]0.88 [0.88, 0.90]0.53 [0.56, 0.31]0.25 [0.33, 0.25]
No specification0.60 [0.22, 0.98] (2 seeds)0.70 [1.0, 0.40]0.21 [0.27, 0.15]0.42 [0.42, 0.42] (2 seeds)1.0 [1.0, 1.0]0.63 [1.0, 0.25]0.22 [0.22, 0.22]1.0 [1.2, 1.0]0.54 [1.0, 0.54]
SINDy0.21 [0.21, 0.21]24 [24, 24]0.40 [0.40, 0.40]0.40 [0.40, 0.40]130 [130, 130]1.0 [1.0, 1.0]0.35 [0.35, 0.35]620 [620, 620]0.83 [0.83, 0.83]
PySR———0.42 (1 seed)1.00.250.32 [0.29, 0.35] (2 seeds)1.0 [1.1, 1.0]0.33 [0.35, 0.30]
spec-SINDy0.19 [0.15, 0.21]30 [31,000, 9.1]0.41 [0.91, 0.39]0.40 [0.02, 0.63]92 [290,000, 26]1.0 [1.0, 0.50]0.40 [0.32, 0.53]510 [1,000,000, 420]0.74 [0.87, 0.68]
LLM-SR0.75 [0.54, 0.88]0.30 [0.34, 0.087]0.026 [0.042, 0.026]0.60 [0.52, 0.72]0.76 [230,000, 0.35]0.48 [0.50, 0.45]0.73 [0.52, 0.73]1.4 [750,000, 0.32]0.24 [0.32, 0.24]
LLM-ODE0.22 [0.22, 0.43]1.0 [1.0, 1.0]0.27 [0.27, 0.27]0.84 [0.42, 0.95]0.35 [1.0, 0.33]0.34 [0.45, 0.25]0.22 [0.21, 0.22]1.2 [1.9, 1.2]0.54 [0.54, 0.54]
D30.98 [0.98, 0.98]0.27 [0.32, 0.25]0.061 [0.075, 0.056]0.94 [0.87, 0.96]0.94 [1.0, 0.44]0.25 [0.25, 0.21]0.85 [0.83, 0.89]0.57 [0.60, 0.50]0.27 [0.28, 0.25]
HDTwinGen0.91 [0.90, 0.98]0.74 [0.93, 0.48]0.085 [0.23, 0.013]0.86 [0.12, 0.89]0.86 [1.0, 0.79]0.60 [1.0, 0.35]0.61 [0.31, 0.87]1.0 [1.2, 0.80]0.31 [0.54, 0.24]

PySR does not have an identifiable feed transport term, and has an identifiable reaction heat term in one seed and jacket heat in two seeds. SINDy and spec-SINDy have large RSEs in reaction heat and jacket heat because these two terms have large magnitudes but largely cancel each other in the equation.

Summary. Our method comes on top in 7 out of 9 metrics for mechanism correctness. LLM-SR and LLM-ODE each claims one metric.

References for the definition of directional accuracy:

  • Pesaran, M. H., and Timmermann, A. (1992). A simple nonparametric test of predictive performance. Journal of Business & Economic Statistics, 10(4), 561--565.
  • Blaskowitz, O., and Herwartz, H. (2011). On economic evaluation of directional forecasts. International Journal of Forecasting, 27(4), 1058--1065.

References for the definition of shaping errors:

  • Murphy, A. H. (1988). Skill scores based on the mean square error and their relationships to the correlation coefficient. Monthly Weather Review, 116(12), 2417--2424.
  • Murphy, A. H., and Epstein, E. S. (1989). Skill scores and correlation coefficients in model verification. Monthly Weather Review, 117(3), 572--581.
  • Gupta, H. V., Kling, H., Yilmaz, K. K., and Martinez, G. F. (2009). Decomposition of the mean squared error and NSE performance criteria: Implications for improving hydrological modelling. Journal of Hydrology, 377(1–2), 80--91.

References for new baselines:

  • Shojaee, P., et al. (2025). LLM-SR: Scientific equation discovery via programming with large language models. ICLR.
  • Bideh, A.Z. and Gryak, J. (2026). LLM-ODE: Data-driven discovery of dynamical systems with large language models. arXiv:2603.20910.
  • Holt, S., et al. (2024b). Automatically learning hybrid digital twins of dynamical systems. NeurIPS.
AuthorsOpenReview ↗

Where Does The Improvement Come From? (W1, Q3)

Where Does The Improvement Come From? (W1, Q3)

W1: Although the proposed framework achieves promising results over the selected baselines, it remains unclear how much of the improvement comes specifically from automatically discovering latent-state representations, rather than incorporating task-specific mechanistic priors. Comparisons with conventional latent-state models or manually specified mechanistic models would help clarify this distinction.

Q3: Could the authors further investigate how much of the reported improvement comes from introducing additional latent states?

Thank you for asking us to distinguish the contributions of introducing latent states and incororating task specifications. Simply put, task specification helps to reduce the search space of dynamical models: no-specification ablations see larger performance degradation than no-latent ablations. No-latent ablations can perform reasonably in terms of target NMSEs, but suffer in response correctness under interventions that depends on the missing latent states.

The new results on correctness of discovered mechanisms already supported the conclusion above (see the "No latent" and "No specification" rows of the table in the response). Below is more detailed analysis.

Contribution from Incorporating Task Specification

We first analyze the no-specification ablation on CSTR-hard, where we remove the task specification from the prompt while leaving latent-state construction available.

SeedTraining NMSETest NMSEComments
Full (0-2)0.15 [0.036, 0.18]0.37 [0.17, 0.75]
No-spec 00.160.34heat exchange not modeled -> heat transfer mechanism has directional accuracy 22%
No-spec 10.814.8T˙\dot{T} does not depend on TT (no feedback)
No-spec 21.044.9T˙\dot{T} close to zero

We observed two difficulties under no specification.

  • Difficulty in finding a structure that fits. This happened in seed 1 and 2, whose equations either have no temperature feedback or do not change at all. High training NMSEs indicate that the equation structures are incorrect, resulting in high test NMSEs.
  • Difficulty in discovering correct mechanisms. This happened in seed 0, which fit and predicted the temperature almost as well as Full. But the jacket feed temperature TjfT_{jf} entered the equation through a generic relaxation rather than as heat exchange with the reactor. So the heat exchange mechanism moves in the wrong direction for most of the response (directional accuracy 0.22 vs 0.87).

Contribution from Introducing Latent States

Next, we analyze the no-latent ablation, which retains task specification but restricts the dynamics to observed states.

We also extend the analysis beyond no-latent ablations to five external LLM-based baselines (spec-SINDy, LLM-SR, LLM-ODE, D3, HDTwinGen). They receive the same prompt with task specification and latent states allowed. However, their methods have no internal mechanisms for proposing additional latent states. In our experiment, these 5 external baselines produced 15 models (3 seeds each) with no latent states.

Introducing latent states has less impact on target NMSEs than incorporating task specification.

FullNo-latentNo-specification
Test NMSE0.37 [0.17, 0.75]0.74 [0.17, 0.80]4.8 [0.34, 4.9]

There are two reasons.

  1. In theory, latent states can be constructed by delayed copies of observed states (Taken's Theorem) if observation noises are "sufficiently small".
  2. In CSTR, the reactor's unobserved states settle within 0.2–0.4 time units, much faster than the training inputs change. So the temperature balance can be written with the inputs in their place.

Latent states have higher impact on mechanism correctness than target NMSEs. For more in-depth analysis, we group the held-out interventions by the hidden state they depend on:

  • Related interventions: an intervention that depends on a hidden state when the input driving that state changes within about one time unit.
  • Unrelated interventions: an intervention that depends on factors other than the latent state.
ModelLatent state modeledRelated InterventionsRelated interventions: no-latent ablationsRelated interventions: external baselinesInterventions that depend on other factors
Full, seed 0delayed feed concentration0.028[0.12, 0.26][0.074, 1.0]0.23
Full, seed 1delayed jacket-feed temperature0.018[0.094, 0.24][0.071, 0.67]0.16

We can see that for both ablations and external baselines, latent states are important in producing task-required mechanisms, especially under interventions that probe that particular latent state.

Reviewer_7pHKOpenReview ↗
Rating
1
Confidence
2
Informativeness
1
Presentation
2
Soundness
1
Read review
Paper summary

Problem: The authors want to find equations of dynamical systems under a certain type of constraint. This constraint they call a “task-specification”. This “task-specification” is a natural language description of what the identified dynamical systems/equations are supposed to be used for. The example task specification they give is “Design an insulin replacement therapy. Require meal absorption and delayed insulin action”. However, the data that is available to fit the dynamics models on does not include measurements of the specified mechanisms (“meal absorption” and “delayed insulin action”). Meaning these are latent variables or latent dynamics that need to be uncovered.

Solution: The authors propose to solve this problem by using an LLM-augmented/driven inference pipeline. They define an iterative process of

  • an LLM proposing a parametrized “candidate ODE” formulated as a graph model
  • a deterministic verifier checking the candidate for validity and for the specified requirements
  • fitting the proposed candidate ODE to the data
  • another LLM acts as a judge and provides a “scientific critic” (advisory only) that the proposer LLM uses to revise the candidate ODE.

Validation: They apply their proposed model to a custom benchmark of synthetically generated data. There are three types of models (ground truth latent dynamics) used to generate these synthetic datasets:

  • glucose-insulin model from Dalla Man et al 2007
  • continuous stirred-tank reactor “CSTR” model from Bequette 1998
  • custom randomly generated model based on a “predefined analytic equation grammar”.

For the glucose-insulin model there are 7 cases across two tasks (T1: meal response, T2: absorption–action disentanglement) and two observability tiers (easy/hard). For T1-easy, there are 4 variations: the original model, an obfuscated variant (names of variables are changed and the glucose-insulin context is removed from the prompt), a perturbed variant (the functional form of two exchange terms in the model is changed), and a perturbed & obfuscated variant. T1-hard, T2-easy and T2-hard use only the original, named model.

The authors report the average performance on all of these datasets. As baselines, they compare against: SINDy, PySR, D3, and a GPT-5.6 agent.

Critical strengths

Using LLMs to constrain dynamical systems identification (identifying governing equations) sounds like a reasonable research direction worth exploring in more detail.

Critical weaknesses

Motivation / Problem Statement: The “gap” that is identified and the “problem” that is presented hardly make sense. The way the problem and the solution are framed is highly circular.

Claim: the task specification defines what mechanisms the model needs to have, and the method then “recovers” them. But to write a specification like “a delayed insulin-action pathway through which plasma insulin regulates the insulin-dependent contribution to total glucose disposal”, you already need to know the mechanism. Whoever can write this sentence can also just directly write down the dynamics model they have in mind and fit it to the data. So what’s the point in being this imprecise and going through natural language first?

On top of that: the deterministic verifier enforces the specified pathways during the search, and the “mechanism compliance” metric then checks more or less the same pathway properties. According to App. D.6, this metric was even developed after inspecting preliminary results. So the mechanisms are imposed and then counted as “recovered”. The authors also assume the task specification is always correct (Limitations). So the framework can never tell you that your hypothesis is wrong, which is exactly what you would want from a scientific modelling tool. To be constructive: the authors would either need to show a concrete setting where the natural-language specification contains information that an expert could not write down as a model, or reframe this as a search within an expert’s hypothesis space and compare against that.

Important related work not considered The main claim of this paper is recovering unobserved latent states/dynamics from partial observations. That’s an identifiability question. The paper proposes a method on identifying latent dynamical systems but doesn't discuss identifiability beyond noting that the state specification is not uniquely determined (C2). I'd recommend discussing identifiability in more depth:

  • identifiable latent variable models for time series: TCL/PCL (Hyvärinen & Morioka 2016, 2017), Hyvärinen et al. 2019, iVAE (Khemakhem et al. 2020), ∆-SNICA (Hälvä et al. 2021)
  • Equation discovery beyond vanilla SINDy:
    • SINDy autoencoders (Champion et al. 2019): the most direct predecessor for jointly learning latent coordinates and sparse dynamics.
    • The follow-up for partial measurements (Bakarji et al. 2023) is cited but only dismissed in one sentence and not compared against.

Without any identifiability argument, the “discovered” latent states are just variables the LLM put a name on. The authors don't compare their method to the ground-truth latents, even though the simulator provides them.

Benchmark datasets are not well designed What’s the point in picking these example models?

  • 7 of the 9 cases come from the same Dalla Man simulator. The obfuscated/perturbed variants only differ by renaming or by a multiplier on two exchange terms. In T1-easy most terms of the target equation (EGP, Uii, E, Gt) are given as auxiliary inputs. So the task boils down to learning a meal absorption filter plus a few coefficients.
  • The “alien device” is a single random draw from a fixed tanh-RNN-like template. Coefficients and selection criteria aren’t reported, and the observed channels were selected by a simulator-side sensitivity analysis, i.e. using privileged access to the ground truth. Its only requirement (input-driven memory) appears to be satisfied by any model in which the input enters the ODE of the target; the authors should confirm.
  • For CSTR, App. D.3 reports a bug in the data export, and it’s not stated whether the results use corrected data. (Sec. 5.1 also says the CSTR case is obfuscated, App. D says it’s named.)

Existing results are not even properly reported Tables 2 and 3 average over all nine cases. There are no per-case results anywhere, not even in the appendix. So it’s impossible to tell how the method does on the obfuscated/perturbed variants or the hard tiers. Q4 (memorization) is posed in Sec. 5 but never answered. Compliance is 88.9% in Table 2 but 85.0% in the text. And 88.9% with SD 33.3 is exactly 8/9 cases passing with CSTR at zero, so this metric is basically binary per case. The “no specification” ablation, which is the central claim of the paper, only shows up in a single-seed post-hoc case study (Fig. 4). And in that case study the perturbed quantity Gt is itself a supplied input, so it’s really just testing the coefficients on supplied auxiliary channels, not a latent mechanism.

The baselines are pretty weak SINDy/PySR can’t introduce latent states, so their ~0% compliance is largely by construction (as the authors acknowledge in Sec. 5.2). The obvious baselines for this problem are missing:

  • an expert-specified model (e.g. the Bergman minimal model, or the standard CSTR energy balance);
  • SINDy with a library informed by the specification (could even be LLM-designed);
  • latent equation discovery (e.g. SINDy autoencoders);
  • HDTwinGen (Holt et al. 2024b), the closest prior method that actually uses natural-language modelling requirements. It’s cited but not compared against.

For the GPT-5.6 agent, neither setup nor budget is described.

Misspecification of D3 in Appendix D.5 is deeply concerning The description of this baseline does not match the method they cite, so it is unclear which implementation and configuration were actually run.

  • App. D.5 lists D3 among “symbolic regression approaches” and describes it as “neural network-guided search with confidence interval-based model selection” with “automatic complexity regularization”. None of this matches Holt et al. (2024a). D3 is an LLM-agent framework that writes PyTorch models and simply picks the lowest validation loss. The confidence intervals in that paper are only used for reporting results.
  • App. A attributes LLM-ODE’s symbolic evolution to Holt et al. (2024a) as well.
  • App. D.6 evaluates D3 as an increment map not scaled by the sampling interval. But D3 defines its models as state derivatives trained with a Δt-scaled one-step loss. If the same convention was used for the rollouts, this alone could explain D3’s test NMSE above 2 under free rollout.
  • The LLM backbone and iteration budget for D3 are not reported.
Questions for authors

See weaknesses.

REFERNCES:

  1. Hyvarinen, A. & Morioka, H. Unsupervised Feature Extraction by Time-Contrastive Learning and Nonlinear ICA. in Advances in Neural Information Processing Systems vol. 29 (Curran Associates, Inc., 2016).
  2. Hyvarinen, A. & Morioka, H. Nonlinear ICA of Temporally Dependent Stationary Sources. in Proceedings of the 20th International Conference on Artificial Intelligence and Statistics 460–469 (PMLR, 2017).
  3. Champion, K., Lusch, B., Kutz, J. N. & Brunton, S. L. Data-driven discovery of coordinates and governing equations. Proc. Natl. Acad. Sci. U.S.A. 116, 22445–22451 (2019).
  4. Hyvarinen, A., Sasaki, H. & Turner, R. E. Nonlinear ICA Using Auxiliary Variables and Generalized Contrastive Learning. Journal of Machine Learning Research (2019).
  5. Khemakhem, I., Kingma, D., Monti, R. & Hyvarinen, A. Variational Autoencoders and Nonlinear ICA: A Unifying Framework. in Proceedings of the Twenty Third International Conference on Artificial Intelligence and Statistics 2207–2217 (PMLR, 2020).
  6. Hälvä, H. et al. Disentangling Identifiable Features from Noisy Data with Structured Nonlinear ICA. Preprint at https://doi.org/10.48550/arXiv.2106.09620 (2021).
4 comments
AuthorsOpenReview ↗

Response 1/3 - Task Specification and Mechanism Correctness

Task specification does not give the answer

The way the problem and the solution are framed is highly circular. But to write a specification like “a delayed insulin-action pathway through which plasma insulin regulates the insulin-dependent contribution to total glucose disposal”, you already need to know the mechanism. Whoever can write this sentence can also just directly write down the dynamics model they have in mind and fit it to the data. So what’s the point in being this imprecise and going through natural language first?

Thank you for this comment. We would like to clarify that a specification does not reveal the dynamics model. Our envisioned use case is when a user has a goal/task in mind (e.g., insulin therapy design) and asks for a model with task-critical components (e.g., an insulin pathway). But the system is unknown, and the user does not know how the components should look like.

A specification like “a delayed insulin-action pathway through which plasma insulin regulates the insulin-dependent contribution to total glucose disposal” only requires that the glucose dynamics G˙\dot{G} has a term dependent on the insulin. It does not specify the functional form or the representation of insulin.


Deterministic Verifier vs Mechanism Compliance

On top of that: the deterministic verifier enforces the specified pathways during the search, and the “mechanism compliance” metric then checks more or less the same pathway properties. According to App. D.6, this metric was even developed after inspecting preliminary results. So the mechanisms are imposed and then counted as “recovered”.

We thank the reviewer for pointing this out. The correctness of discovered mechanisms was also raised by Reviewer Ewor. Please see our first two responses to Reviewer Ewor Correctness of Discovered Mechanisms (W2, Q4).

We have developed a new set of metrics that evaluate the response directions, magnitudes, and temporal behaviors of the discovered mechanisms, and demonstrated that our method has higher accuracy, comparing to an extended set of external baselines (including four new baselines of spec-SINDy, LLM-SR, LLM-ODE, and HDTwinGen).


Identifiability and Correctness of Latent States/Dynamics

The main claim of this paper is recovering unobserved latent states/dynamics from partial observations. That’s an identifiability question. The paper proposes a method on identifying latent dynamical systems but doesn't discuss identifiability beyond noting that the state specification is not uniquely determined (C2). Without any identifiability argument, the “discovered” latent states are just variables the LLM put a name on. The authors don't compare their method to the ground-truth latents, even though the simulator provides them.

We thank the reviewer for this important question, which is shared by Reviewer Ewor. This is also related to your previous question on mechanism compliance.

In short, we directly evaluated the correctness of latent dynamics in terms of response directions, magnitudes, and temporal behaviors. We demonstrted that our method has higher accuracy, comparing to an extended set of external baselines (including four new baselines of spec-SINDy, LLM-SR, LLM-ODE, and HDTwinGen).


Assumption of Correct Task Specification

The authors also assume the task specification is always correct (Limitations). So the framework can never tell you that your hypothesis is wrong, which is exactly what you would want from a scientific modelling tool. To be constructive: the authors would either need to show a concrete setting where the natural-language specification contains information that an expert could not write down as a model, or reframe this as a search within an expert’s hypothesis space and compare against that.

We agree with the reviewer’s proposed framing: automated model construction within a hypothesis space constrained by expert knowledge. Such a constrained hypothesis space can still be large. For example, the domain expert may not have the mathematical knowledge to write down exact equations, and a modeling expert is expensive. Another perhaps more interesting scenario is when the system is unknown.

Therefore, the utility of our framework is assessed against specification-informed alternatives. We have added these comparisons against four new LLM-based external baselines.

We do not claim that natural language supplies knowledge an expert could never encode mathematically; the intended benefit is automating the otherwise manual or unknown choices needed to turn an incomplete qualitative specification into a model. Automatic specification revision remains outside the present scope.

We evaluated the robustness of our method to incomplete, over-restrictive, and incorrect specifications. Please see our response 3/3 to Reviewer pp8W.

AuthorsOpenReview ↗

Response 2/3 - Benchmark Design and New External Baselines

Benchmark Design Rationale

Benchmark datasets are not well designed What’s the point in picking these example models?

Thank you for asking us to clarify the benchmark rationale. Our intended questions are: How does the modeling task change the required representation? How does reducing available measurements affect discovery? How do semantic cues and departures from familiar dynamics affect the proposed models?

Dalla Man has more tasks than the other two, because it provides controlled comparisons. Using one reference system lets us vary task requirements, observability, and semantic information while retaining a common physical foundation. Named and obfuscated pairs preserve the numerical trajectories, isolating the contribution of semantic cues. The perturbed cases introduce state-dependent nonlinear multipliers into the exchange laws, rather than merely changing two constant coefficients. They test adaptation away from familiar equations. These controls assess reliance on prior knowledge (i.e., memorization).

T1-easy is deliberately an assisted problem, and learning a causal meal-absorption representation is therefore the intended task. Supplying EGP, Uii, E, and tissue glucose isolates the construction of meal-response dynamics from the challenge of reconstructing the entire physiological system. T1-hard removes several supplied channels, while T2 additionally requires insulin dynamics and delayed insulin-dependent disposal; T2-hard further reduces available measurements.

The alien device tests modeling without a named physical reference. It is one frozen system selected from 40 generated candidates, using predefined stability and behavioral-richness criteria independently of method performance. The telemetry selection does use simulator-side sensitivity analysis on the training design, so that the resulting channels supplied equally to all methods. We also confirm that the easy requirement permits the output state itself to carry input history: an input-driven output ODE can satisfy this limited requirement without additional latent states. But the models are still evaluated on numerical evidence.

For CSTR, the submitted case was named and easy; the “obfuscated” description is incorrect. The submitted results used the historical export affected by the documented integration defect. Specifically, adaptive integration skipped some forcing windows, producing trajectories inconsistent with the recorded inputs. The corrected generator integrates across explicit forcing boundaries and is checked against independent solvers. All the results in the rebuttal and in the revised manuscript are obtained under the corrected dataset.


New External Baselines

The baselines are pretty weak SINDy/PySR can’t introduce latent states, so their ~0% compliance is largely by construction (as the authors acknowledge in Sec. 5.2).

We thank the reviewer for this comment on external baselines. We have added four new LLM-based baselines:

  • SINDy with a LLM-designed library informed by the specification ("spec-SINDy"), as suggested,
  • LLM-SR (Shojaee, 2025),
  • LLM-ODE (Bideh and Gryak, 2026), and
  • HDTwinGen (Holt et al., 2024b).

Results against these new benchmarks on CSTR-hard benchmark:

MethodFeed transport: directions↑Feed transport: RSE↓Feed transport: shape error↓Reaction heat: directions↑Reaction heat: RSE↓Reaction heat: shape error↓Jacket heat: directions↑Jacket heat: RSE↓Jacket heat: shape error↓
Ours0.98 [0.97, 0.98]0.26 [0.34, 0.14]0.14 [0.17, 0.022]0.95 [0.56, 0.96]0.69 [1.0, 0.22]0.16 [0.23, 0.080]0.88 [0.87, 0.90]0.31 [0.45, 0.29]0.23 [0.33, 0.20]
spec-SINDy0.19 [0.15, 0.21]30 [31,000, 9.1]0.41 [0.91, 0.39]0.40 [0.02, 0.63]92 [290,000, 26]1.0 [1.0, 0.50]0.40 [0.32, 0.53]510 [1,000,000, 420]0.74 [0.87, 0.68]
LLM-SR0.75 [0.54, 0.88]0.30 [0.34, 0.087]0.026 [0.042, 0.026]0.60 [0.52, 0.72]0.76 [230,000, 0.35]0.48 [0.50, 0.45]0.73 [0.52, 0.73]1.4 [750,000, 0.32]0.24 [0.32, 0.24]
LLM-ODE0.22 [0.22, 0.43]1.0 [1.0, 1.0]0.27 [0.27, 0.27]0.84 [0.42, 0.95]0.35 [1.0, 0.33]0.34 [0.45, 0.25]0.22 [0.21, 0.22]1.2 [1.9, 1.2]0.54 [0.54, 0.54]
HDTwinGen0.91 [0.90, 0.98]0.74 [0.93, 0.48]0.085 [0.23, 0.013]0.86 [0.12, 0.89]0.86 [1.0, 0.79]0.60 [1.0, 0.35]0.61 [0.31, 0.87]1.0 [1.2, 0.80]0.31 [0.54, 0.24]

We also implemented SINDy autoencoders as suggested. Note that SINDy autoencoders uses a neural network and therefore is different from all baselines. It achieves superior numerical performance in target NMSEs (one order of magnitude lower). But we cannot interpret how the task-required mechanisms are modeled and cannot evaluate the mechanism correctness.

AuthorsOpenReview ↗

Response 3/3 - Implementation Details of GPT-5.6 Agent and D3

Implementation Details of GPT-5.6 Agent and D3

For the GPT-5.6 agent, neither setup nor budget is described. Misspecification of D3 in Appendix D.5 is deeply concerning The description of this baseline does not match the method they cite, so it is unclear which implementation and configuration were actually run.

Thank you for highlighting these reproducibility concerns.

GPT-5.6 Agent

We used gpt-5.6-sol with high reasoning effort and a hosted Python code interpreter. The configured budget was 12 tool calls, 30,000 output tokens, and a 20-minute timeout per provider request, with at most two provider attempts following failures.

D3

We acknowledge that Appendix D.5 inaccurately describes D3. D3 is an LLM-driven model-generation framework, with parameter fitting and validation-based model selection. We will remove the incorrect statements about confidence-interval-based selection and automatic complexity regularization.

However, the suggested Δt error does not describe our implementation. Although the paper uses derivative notation, the authors’ released code at commit ee86212 computes pred_state = states[:, t] + dx_dt, without an explicit Δt multiplier, in both training and validation. Our adapter consistently uses this convention during fitting and recursive rollout: xk+1=xk+Fθ(xk,uk).x_{k+1}=x_k+F_\theta(x_k,u_k). We therefore did not train with a Δt-scaled derivative loss and subsequently omit Δt during evaluation. Adding a multiplier only during rollout would change the fitted update map. The reported free-rollout error is not a result of the alleged scaling mistake.

The configuration is kept the same as the paper as possible: GPT-OSS-120B, five generations, one model-generation call per generation, and an 8,192-token output limit, selecting the lowest validation NMSE.

Dear Reviewer 7pHK,

Thank you for the time and effort you have devoted to reviewing our work and giving constructive feedback.

As the rebuttal discussion period is nearing its end, we would be grateful if you could review our responses and let us know whether our clarifications have addressed your concerns. To summarize, we have

If any questions remain, we would welcome the opportunity to discuss them and provide further explanations while the discussion is still open.

Thank you again for your time and consideration!

Comments on the paper

1 comment
AuthorsOpenReview ↗

Summary of Rebuttal

Dear area chair and reviewers,

Thank you all for the time and effort in the review process. We would like to briefly summarize the rebuttal discussion.

We thank all three reviewer for acknowledging the novelty and importance of our problem formulation in task-specified dynamical system discovery.

The reviewers' insightful questions and comments have helped us greatly improved our paper. In particular, we have

We hope this summary assists your assessment of our paper. We sincerely appreciate the AC's coordination and all reviewers’ constructive contributions to improving the paper.

Best regards,

The authors

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

Related papers

Loading the map…

Discussion (0)

Sign in to comment.