acceptodds
Under review as a conference paper at ICLR 2027

Does LLM Dream of Differential Equation Discovery?

Abstract

Discovering a governing equation from data means choosing its terms and estimating their coefficients. We ask whether a general-purpose multimodal model, given no feedback, can do both itself, with no library of candidate terms and no optimiser fitting the coefficients. Four such models are shown ten dynamical systems in twelve ways, each run returning five candidate equations, and a run discovers the equation when a screened candidate has exactly the true terms and every coefficient within 10% of its value. Each model is also run with the data withheld, so that what it reads from the data can be told from what it recalls. In its strongest configuration GLM discovers ten laws, Qwen and DeepSeek eight and Gemma four, including laws none of the models writes without data, with coefficients they estimate themselves, and their reasoning shows how: they read values where competing terms vanish, and GLM tests what it reads. Under added noise the models seldom say that the data are noisy, yet on several systems they keep writing the true law. The hardest laws are missed because one of their less common terms is never written.

Primary area:
neurosymbolic & hybrid AI systems (physics-informed, logic & formal reasoning, etc.)
Keywords:
equation discovery, differential equation, LLM

Reviews

Among:
Shape:
1234this paper, Bottom 10%
Figure 1. Mean review rating of the reviewed papers at ICLR 2027. This paper averages 1.67 over 3 reviews, above 10% of the others.
Reviewer_G4ieOpenReview ↗
Rating
2
Confidence
2
Informativeness
2
Presentation
3
Soundness
2
Read review
Paper summary

The paper evaluates whether four general-purpose multimodal LLMs can propose differential-equation structures and numerical coefficients without feedback or external coefficient correction. It studies ten systems across twelve configurations, with data-withheld, prompt, noise, and regression comparisons. Five candidates per run are Pareto-screened, and success means a retained candidate matches the true terms and coefficients within 10%.

Critical strengths
  1. A substantial empirical sweep separates structure recovery from coefficient accuracy across four models, ten systems, and twelve configurations, with data-withheld, prompt, and noise controls.

  2. Detailed appendices examine failure stages, threshold and screening sensitivity, statistical uncertainty, regression retuning, and non-identifiability. The paper openly reports important limitations rather than hiding them.

Critical weaknesses
  1. The motivation and incremental contribution need sharper justification. Isolating a model’s unaided capability is a legitimate research question, but this setting alone does not establish substantial novelty or practical value. LLM-SR already studies full-equation generation and removal of iterative refinement in separate ablations; the paper should clarify what new understanding its combined protocol adds. Likewise, retaining correct candidates with two existing screening criteria is a useful auxiliary result, but does not establish reliable final selection. The contribution therefore rests primarily on the empirical insights.

  2. Those insights remain insufficiently deep and generalizable. Data helping, missing terms blocking recovery, and candidate generation differing from selection are largely intuitive; detailed counts are useful, but the study does not sufficiently test their causes or derive validated guidance. Written-reasoning categories do not establish the actual computation, and no-data failures do not exclude data-cued recognition. Randomized non-round coefficients and independent solutions would strengthen the inference claim; missing-term hints, aligned derivative inputs, and same-model fitting/feedback comparisons would test the proposed bottlenecks. Without such evidence, the results mainly characterize behavior on the chosen instances rather than explain discovery capability more broadly.

Questions for authors
  1. What important uncertainty does the unaided setting resolve, and what precisely extends LLM-SR’s related ablations? Which new understanding, rather than the protocol combination itself, is the main contribution?

  2. Can randomized non-round coefficients, transformed equations, and independent solutions distinguish data-based inference from recognition of familiar laws?

  3. Which findings yield validated practical guidance? Could missing-term hints, pointwise-aligned derivatives, and matched fitting/feedback arms test the explanations for failure and input effects? Please also distinguish candidate retention from final selection.

1 comment
AuthorsOpenReview ↗

What models contribute before fitting and how to use their proposals

Thank you for the careful reading. We revised the Introduction to explain why measuring a model’s own coefficient estimates matters even when a researcher ultimately plans to fit the equation numerically.

About relation to LLM-SR:

LLM-SR tested the effects of removing iterative refinement and external coefficient fitting in separate experiments, measuring prediction error. We removed both together and checked the terms and coefficients the model wrote. On fresh observations, a model could propose the correct terms and adjust its coefficient estimates before an optimiser or feedback could influence its answer (Section 4.3). LLM-SR, Section 3.5

About what the fresh trajectories showed:

We kept the equation families familiar and generated new observations with changed coefficients. For example, DeepSeek identified the correct terms on both the wave and Kuramoto–Sivashinsky systems. Its estimates followed the new values for the wave equation but stayed near the published values for Kuramoto–Sivashinsky. Recognising the equation form could therefore coexist with estimating a coefficient. The written explanations leave the model’s internal calculation unresolved. The experiment used fixed coefficient shifts and new initial conditions. Continuous sampling, transformed equations and unfamiliar structures remain open (Figure 3).

About using a proposed equations:

A researcher can ask for equations complete with numerical coefficients, then check whether they still fits observations from a different initial condition. On one system, a second trajectory distinguished two equations that fit the first equally well. Presentation also mattered. Across the tested systems, models recovered more equations on average from numerical samples than from plots of the same data (Sections 4.1 and 4.5).

About choosing a single equation:

A run counted as successful if the retained set contained the correct terms with accurate coefficients. The screen judged each proposal by its fit with the written coefficients and by the stability of coefficients fitted on different parts of the trajectory. A proposal with the correct terms could therefore lose to an alternative that fit better and had equally stable coefficients. Choosing just the equation with the best fit gave 13 to 66 correct choices per 100 runs across models, within two successes of the whole set for each model in its strongest configuration. Retaining alternatives exposed cases where one trajectory was consistent with several equations (Section 4.5, Tables 56–57).

About hints, derivative inputs and feedback:

Naming a missing term from the known equation would give the model part of the answer. We identified missing terms after the unaided runs by comparing the proposals with the known equations. We tested whether supplying derivative values helped. For data that varied across space, the observations and derivatives were sampled separately. Explicitly pairing them at the same locations remains untested. We also fitted coefficients to the terms the models proposed, on fresh solutions selected for accurate fitting. The fits improved recovery, but the effect of returning that feedback to the model for another attempt remains open (Table 1).

Reviewer_uE7NOpenReview ↗
Rating
1
Confidence
2
Informativeness
2
Presentation
2
Soundness
2
Read review
Paper summary

This paper evaluates differential equation recovery by LLMs. Unlike previous work, which has used specialized models and formats to recover differential equations from data, this paper evaluates how well standard multi-modal models do at this task and manipulates a variety of factors such as how the data are presented. The results show variation across models in their ability to recover the underlying laws for given domains.

Critical strengths

The strength of this paper is extending previous analyses of differential equation recovery into a new setting where little scaffolding is applied to models. The evaluations show meaningful variation across models in this task.

Critical weaknesses

The move to the unscaffolded setting is relatively incremental over previous work that has explored differential equation recovery, as it is likely that any user interested in this application is likely to use some kind of scaffold.

The key results are focused on the number of laws recovered by the models, and the numbers involved are small because there are only twelve base configurations. In general, I was concerned by the absence of statistical significance tests or error bars in the reported results. This makes it difficult to evaluate claims about one model being better than another. For example, Figure 4 is entirely focused on very small sample differences, and the standard errors in Table 1 are very large. I would like to see all qualitative statements in the results being supported by the relevant statistical tests, with error bars included in Figures 2 and 3.

The AI usage statement indicates liberal use of AI and this is reflected in the text, which read like Claude in many places.

Questions for authors

None

1 comment
AuthorsOpenReview ↗

Unaided coefficient estimates and uncertainty in model comparisons

Thank you for the detailed review. We added uncertainty intervals and revised the Results to distinguish differences supported across the ten systems from rankings that remained uncertain.

About the model’s contribution before fitting:

A fitting algorithm can improve a coefficient regardless of how the model chose its original value. To measure the model’s contribution, we changed the coefficients used to generate fresh observations and scored the numbers the model wrote. For example, DeepSeek’s estimates followed changes in the coefficient governing wave speed. On another system, it found the correct terms but kept its coefficients near their published values. Figure 3 shows the written estimates. Table 1 shows what an external fit added on solutions selected to support accurate fitting.

About comparing models:

The twelve configurations were different ways to present the data and prompt the model. Repeated calls measured variability on the same observations. For comparisons between models, we paired their results on each of the ten systems and corrected for multiple comparisons. In their strongest configurations, GLM and DeepSeek each recovered more equations than Gemma (adjusted p-value of 0.047 for each). The evidence left GLM’s lead over Qwen or DeepSeek uncertain. Section 4.1 reports the supported comparisons, and Appendix E.4 gives the tests and intervals.

About noise and coefficient errors:

Recovery declined as noise increased for GLM and DeepSeek. The trends for Qwen and Gemma remained inconclusive. Table 2 replaces the submitted noise plot (Figure 4), and Table 35 gives the trend tests. In the submitted Table 1, parentheses showed the standard deviation of the models’ coefficient errors. The revised Table 61 gives the sample sizes and intervals for the mean error where the estimates varied.

About error bars in the submitted figures:

Figure 2 now shows 95% intervals for how often each model recovered the equation on repeated calls with the same data. The submitted Figure 3, now Figure 5, illustrated a single run in each example. Each point represented one proposed equation, so there was no repeated measurement behind that point. Table 56 gives the comparison across runs.

About writing and AI use:

We shortened the Results and moved detailed counts to the appendix. The revised disclosure identifies where language models assisted with code, checks and writing. The authors reviewed the text, code and results and take responsibility for them.

Reviewer_W3BzOpenReview ↗
Rating
2
Confidence
2
Informativeness
1
Presentation
2
Soundness
2
Read review
Paper summary

The paper tests whether general-purpose multimodal LLMs can propose both the structure and numerical coefficients of differential equations from sampled trajectories or plots, without tools, coefficient-fitting feedback, or iterative refinement. Across ten systems, it finds substantial but uneven success, compares against data-withheld prompts and sparse regression, and attributes failures mainly to missing terms, inaccurate coefficients, and truncated responses.

Critical strengths
  1. Requiring models to produce coefficients as well as symbolic structure distinguishes this evaluation from pipelines where an external optimizer supplies the numerical answer.
  2. Experimental sweep is extensive. Four models, twelve input configurations, and repeated runs allow useful comparisons of numerical inputs, plots, derivative fields, and one- versus two-call prompting.
  3. Useful controls and transparent reporting. The paper includes data-withheld prompts, prompt ablations, noise tests, and detailed failure analyses. It openly acknowledges round coefficients, nonidentifiability, configuration selection on the evaluation runs, and provider-dependent truncation.
  4. Separates proposal from screening. Evaluating candidate equations using data-based objectives is sensible, and examining whether a true term was ever proposed helps locate a genuine bottleneck.
Critical weaknesses

A model can fail to produce an equation without data but recognize a familiar trajectory once shown it. The data may therefore trigger retrieval of a known law and its conventional coefficients, rather than support independent numerical inference.

This is particularly relevant because nine fields come from earlier studies, several are unchanged published arrays, and every coefficient is round. The control establishes that data affect the answer; it does not establish that the answer was derived rather than retrieved. The conclusion that coefficients for laws not produced in the control “come from the data” is stronger than the experiment supports.

Questions for authors

Can you distinguish data-triggered recall from coefficient inference? Please test fresh trajectories with continuously randomized coefficients, changed units, and unfamiliar combinations of terms, and measure whether estimates track the generating parameters. What is the accuracy of one equation selected without ground truth? Report first-candidate accuracy, screened-set size, incorrect-candidate retention, and a deployable selection rule alongside candidate-set success. Do recovered equations generalize across initial conditions? Can you evaluate on held-out trajectories and distinguish genuinely unidentified laws from errors by the model?

1 comment
AuthorsOpenReview ↗

Coefficient adaptation, candidate selection and held-out validation

Thank you for the careful review. We added an experiment in which the equation form stayed the same but its coefficients changed. We also revised Section 4.2 to acknowledge that a correct answer given only when observations were present could still reflect recognition of a familiar system.

About coefficient estimates on fresh trajectories:

We generated fresh trajectories from eight familiar equation families using new initial conditions and three fixed coefficient shifts. For the wave equation, DeepSeek recovered the correct terms and estimated the coefficient governing wave speed within 10% of each new value. On another system, the same model recovered the terms but kept the coefficients near their published values. Figure 3 compares the estimates when the correct terms survived screening. We varied the domain size in part of the experiment. Continuous coefficient sampling, changes of physical units and unfamiliar term combinations remain open.

About choosing one equation:

One rule selected whichever of the five proposals best fit the observations with the coefficients the model wrote. In each model’s strongest configuration, the rule selected the correct equation in 13 to 66 of 100 runs. For GLM, the rule succeeded in 66 runs, the first proposal was correct in 74, and the screened set contained a correct equation in 67. The screen retained one to two candidates on average per answered run. Depending on the model, 35% to 75% of the retained candidates had incorrect terms (Tables 56–57).

About why screening could discard the generating equation:

The screen measured how well each proposal fit the observed trajectory with its written coefficients and how much its coefficients varied when fitted separately to different parts of that trajectory. A proposal with the correct terms could fit poorly if a written coefficient was inaccurate. Even an accurate proposal could lose on Burgers, where a different equation also satisfied the original trajectory. Small numerical differences sometimes gave the alternative a better score on one measure and an equal or better score on the other. A second trajectory separated the two equations. Across all systems, checking the unchanged proposals on a new trajectory preserved every previously successful run and added successes only on Burgers. This check assessed agreement with new observations. Forward simulation remains untested (Section 4.5, Table 59).

Comments on the paper

1 comment
AuthorsOpenReview ↗

Fresh coefficients, uncertainty estimates and equation selection

Thank you for the careful reviews. We revised the abstract and Introduction to clarify the aim of the study. We asked whether a model given observations of a system could propose both the terms of its governing equation and their numerical coefficients before receiving feedback or help from a fitting algorithm.

We added an experiment that kept familiar equation forms but changed the coefficients used to generate new observations. Repeating the published coefficients would therefore miss the new values. When the correct terms survived screening, some of the models’ estimates followed the changes and others stayed near the published values. This gives evidence of coefficient estimation even when the model might have recognised the equation family (Section 4.3).

We also added uncertainty estimates to the model comparisons. Because each model could propose several equations per run, Section 4.5 now reports both how often the correct equation appeared among the retained proposals and how often a rule using only the observations selected it.

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

Related papers

Loading the map…

Discussion (0)

Sign in to comment.