acceptodds
Under review as a conference paper at ICLR 2027

BLADE-PDE: Blended-Activation Denoising and LLM-Aided Discovery of Partial Differential Equations from Noisy Data

Abstract

Partial differential equations govern physical systems, and complex systems often require data-driven recovery of their governing equations from observations. Practical observations carry measurement noise, which is challenging because differentiation amplifies the noise and the corrupted derivatives hurt discovery. Implicit network fitting can denoise the observations before discovery, yet different kinds of PDEs prefer different activation representations in such implicit denoising; this motivates a blending activation, which learns a mixture of standard activations adapted to each field. Our denoising method fits the noisy field with an implicit network whose blending activation applies an affine transform to each activation basis, and the affine transform makes a SIREN-like initialization of the sine basis well-defined. LLM-based discovery brings broad prior knowledge and generates free-form candidates without a hand-crafted library, yet the LLM is not specialized in PDEs, and the raw data alone convey limited information about the system. We therefore run an LLM-aided discovery stage on the denoised derivatives that supplies the LLM with two kinds of information; equation skeletons retrieved from a math handbook provide PDE-specific structural knowledge, and mathematical tools extract essential properties of the system that guide the LLM inference and refinement. The LLM receives a property card of invariance and structure probes computed from the data, equation skeletons retrieved from a math handbook, and correlations of a derivative library, and proposes a sparse combination of library terms; mathematical tools such as weak-form, divergence-form, and scaling-invariance filters then score the proposal and return their notes to the LLM for another round of refinement. On the MDBench benchmark the discovery stage recovers 7 of 14 systems from clean data, more than any tested estimator; the full pipeline recovers all 6 systems of the noisy subset at 20 dB where no baseline recovers any; and under a strict leakage-control protocol the method discovers a synthesized equation outside well-known families.

Primary area:
neurosymbolic & hybrid AI systems (physics-informed, logic & formal reasoning, etc.)
Keywords:
PDE discovery, implicit neural denoising, blending activation functions, large language models

Reviews

Among:
Shape:
1234this paper, Bottom 49%
Figure 1. Mean review rating of the reviewed papers at ICLR 2027. This paper averages 2.25 over 4 reviews, above 49% of the others.
Reviewer_bFT3OpenReview ↗
Rating
1
Confidence
3
Informativeness
2
Presentation
3
Soundness
1
Read review
Paper summary

This paper proposes BLADE-PDE, a two-stage framework for discovering PDEs from noisy data. It uses an implicit neural network with blended activations and SIREN-like initialization for denoising, followed by LLM-based equation discovery guided by mathematical tools and handbook information. Experiments on MDBench show improved equation recovery on both clean and noisy data. The authors also provide ablation studies and an additional experiment to evaluate the discovery of a modified PDE.

Critical strengths

The paper addresses the practical challenge of discovering PDEs from noisy data. The method recovers all six equations at 20 dB noise, while the evaluated baselines fail, demonstrating strong empirical performance.

Critical weaknesses
  1. My main concern is that the methodological novelty is not sufficiently established. Previous work has already explored blended activations for PDE modeling (Wang et al., 2023), neural network-based field fitting and derivative estimation for PDE discovery from noisy data (Both et al., 2021), equation discovery guided by mathematical handbooks (Xu et al., 2025), and LLM-assisted PDE discovery with external feedback (Luo et al., 2025).

Although BLADE-PDE introduces a specific SIREN-like initialization and combines these techniques into a unified framework, I believe the authors should more clearly explain the additional technical contributions over these existing methods.

Relevant references (including some already cited in the manuscript):

  1. Although the authors provide component ablations, several important comparisons are missing. For example, how would existing PDE discovery methods perform using the same denoised data? Is the LLM necessary when the same mathematical information and candidate library are available? Without these comparisons, it is difficult to determine whether the improvements mainly come from better denoising or the proposed LLM-based discovery mechanism.

  2. I believe the experiment in Section 4.3 contains a fundamental mathematical issue. The proposed equation is essentially a KdV equation with an additional linear advection term. By introducing a moving coordinate system, y=x−1.5ty = x - 1.5t, followed by simple rescaling, the equation can be transformed into the standard KdV equation. Therefore, this experiment does not demonstrate the discovery of a PDE outside known equation families, as claimed by the authors. This significantly weakens the paper's claim regarding the discovery of previously unknown PDE structures.

Questions for authors

Please refer to the Critical Weaknesses section.

7 comments

Dear Reviewer,

We thank the reviewer for the careful reading and the three concerns, which we answer below, with new experiments where they were needed. In short, BLADE-PDE is library-free: the library is only a guide, and the LLM does free-form discovery, so that it even recovers a PDE with an exponential term eue^{u} in the revised new PDE experiment, a term that is not mentioned anywhere in the prompt. This separates it from the four works the reviewer lists: there the blended activation solves a known PDE, while here it denoises a field whose equation is unknown, with an initialisation whose contribution we measure; our denoiser is tied to no library and improves other estimators; and the LLM is neither restricted to a library or a trained vocabulary nor fed raw data points. The missing comparisons are now run: given our denoised field and derivatives, PDE-FIND (Rudy et al., 2017) and PySR (Cranmer, 2023) each rise from 0/6 PDEs on the raw field to 3/6 (8/18) at 20 dB, against 6/6 (18/18) for BLADE-PDE, and a no-LLM search given the same candidates and math tools recovers less at every noise level from 10 to 40 dB and never the exponential PDE. We agree that the tweaked KdV equation of the submission stays within the KdV family, and the revised experiment replaces it. The revised method enables the LLM to recover all six MDBench PDEs (Ziaei Bideh et al., 2026) at 20 dB even with neither the handbook nor the math tools. We therefore extend the evaluation with open-form PDEs that are in no handbook and no benchmark, and on these the full prompt recovers 3/4 at 20 dB while the prompts without the handbook or without the math tools perform worse.

All numbers come from the revised pipeline. The paper is being updated accordingly, and the tables cited below are in Appendix J of the revised paper. The code of the revised pipeline is in our anonymous repository, https://anonymous.4open.science/r/BLADE-PDE-5D37/, with the complete prompts, reasoning and replies of one example run. The discovery stage no longer receives equation skeletons or a property card; the handbook is a shuffled list of terms with every non-polynomial function left open as f(⋅)f(\cdot); the LLM has no tools and at most three rounds; and the reported equation is a least-squares refit on the terms it chose. The settings of the denoiser and of the discovery stage are in Table 12. Multi-seed counts are written as PDEs (seeds), for example 6/6 (17/18): a PDE counts as recovered when at least two of its three seeds are.

The detailed answers are as follows.

1. We agree that each ingredient has a precedent, and the revised paper states the relation to these four works more clearly (Table 13). ABU-PINN (Wang et al., 2023) blends activations to solve a known PDE with a physics-informed loss. Our denoiser fits a noisy field whose equation is unknown, with no physics term, and its purpose is derivatives up to fourth order; the technical additions are an affine transform on every basis of the blend, with a trainable scale and shift, and the SIREN-like (Sitzmann et al., 2020) compensation on the sine basis alone. The contribution of this initialisation is measured: at 20 dB the mean derivative error is 0.08, against 0.13 with SIREN initialisation on every basis, 0.10 without the compensation and 0.25 for a sine-only SIREN (Table 7). Denoising for PDE discovery is a challenging problem in its own right, because the regression target and most candidate terms are derivatives of the observed field, and differentiation amplifies the noise. DeepMoD (Both et al., 2021) is the classic method: it fits the network together with a sparse regression on a fixed library. Much research since then has aimed at discovery from noisy data, for example PDE-LEARN (Stephany and Earls, 2024), DISCOVER (Du et al., 2024) and EqGPT (Xu et al., 2025). Yet none of the methods we tested handles the strong noise of MDBench, in particular for PDEs with high-order derivatives. We have also run the PINN-based method PDE-LEARN to compare the derivatives: those of its network are less accurate than ours on every system (Table 6), which we attribute to its physics-informed loss, since a network fitted together with the equation being discovered receives wrong physics information when that equation is wrong. Our denoiser uses neither a library nor an equation, so it is a preprocessor that also serves other estimators (Table 5). On the discovery side, EqGPT trains a dedicated generative model on 221 handbook equations and composes equations from a vocabulary of 56 tokens. BLADE-PDE trains no model on equations: a general LLM reads the handbook as a list of terms with every non-polynomial function left open as f(⋅)f(\cdot), together with math tools computed on the data, and it can write a term that is in no vocabulary, such as eue^{u}. LLM4PD (Luo et al., 2025) lets the LLM select terms from a library of seven candidates given raw observation points, and the LLM also proposes the coefficients, which are scored with a PDE solver. BLADE-PDE does none of these: the LLM is not limited to a library, is not given raw data points and does not set the coefficients, and no PDE solver is needed. Instead, the LLM proposes free-form terms, its feedback is math tools computed on the denoised derivatives (correlations, coefficient profiles, fit gains, residual checks and weak-form scores), and the coefficients are a least-squares refit.

2. Both comparisons are now run. For existing discovery methods on the same denoised data: given our denoised field and derivatives, PDE-FIND and PySR each rise from 0/6 to 3/6 at 20 dB and stay there at 30 and 40 dB, against 6/6 for BLADE-PDE (Table 5), so the denoiser transfers to other estimators and the remaining gap belongs to the discovery stage. For the necessity of the LLM: to test whether a script can replace the LLM, we designed scripts that have everything the LLM gets: the same guide and handbook terms as candidates, with open functions filled from a fixed menu that includes exp⁡\exp, and the same math tools, refit and grading. We tested sequential thresholded least squares (Brunton et al., 2016), as in PDE-FIND, and greedy forward–backward selection, each at six thresholds and on two candidate pools (Table 8). The best single setting recovers 23/35 against 33/35 for the LLM stage; no setting recovers all six PDEs of a noisy level, and none recovers the new PDE. With the handbook pool the greedy script does recover the other three open-form equations, whose functions are on its menu, but it then loses most of the clean benchmark. The LLM matters for two reasons. First, the prompt holds many kinds of information, from field statistics and term correlations to coefficient profiles, fit gains, residual checks and weak-form scores, and a hand-crafted script can hardly judge on all of it. Second, the knowledge of the LLM makes the discovery free-form: the new PDE adds an eue^{u} term to a KdV-type equation, no exponential function is mentioned in the prompt, and the LLM still finds it. Moreover, even without the math tools and the handbook the LLM still recovers all six benchmark PDEs at 20 dB (Table 10). The example run in our repository shows how the term is found (new PDE, clean data, seed 1, with the complete prompts, reasoning and replies). The handbook offers only the open form f(u)f(u), and the math tools show how the coefficient of each base symbol varies with uu. In round one the LLM reads that the part of utu_t that depends on uu alone is not constant, tries a straight line and a quadratic in uu, and then finds that this part divided by eue^{u} is constant, between −1.52 and −1.49 across the bins, so it proposes eue^{u} as a single term in place of a polynomial of three; from the same curves it reads uuxuu_x (the coefficient of uxu_x is linear in uu with slope −4) and uxxxu_{xxx} (a flat coefficient). In round two the math tools confirm the proposal: replacing eue^{u} by a constant, uu, u2u^{2} or u3u^{3} lowers R2R^{2} from 0.994 to between 0.72 and 0.85, and no further term adds more than 0.002. The refit gives ut=−1.499eu−3.988uux−0.248uxxxu_t=-1.499e^{u}-3.988uu_x-0.248u_{xxx}.

3. We agree that the tweaked KdV equation of the submission stays within the KdV family: its transport term is removed by a moving frame and its coefficients by rescaling. Our reason for the original experiment was the following. When a PDE is solved, such a transformation yields the solution easily from that of the KdV equation. In PDE discovery, however, the equation is not known, and discovery methods, whether library-based or search-based, do not consider coordinate transformations: searching for a transformation together with a KdV equation is even harder than searching for a KdV equation with an additional term. We were therefore not sensitive enough to the transformation. The claim is revised, and that example is replaced by four open-form equations that are in no handbook and no benchmark (Table 10): the new PDE ut=−4uux−0.25uxxx−1.5euu_t=-4uu_x-0.25u_{xxx}-1.5e^{u}, whose exponential term is not a transport term and is not removed by a change of coordinates, and three equations with the terms sin⁡(u)ux\sin(u) u_x, cos⁡(u)\cos(u) and uxx/(1+u2)u_{xx}/(1+u^{2}). The structure beyond a familiar family is the non-polynomial function itself: the handbook leaves every such function open as f(⋅)f(\cdot), no exponential function is mentioned in the prompt, and the LLM still finds eue^{u}. At 20 dB the full prompt recovers 3/4 of these equations, against 1/4 without the handbook and none without the math tools; the equation with the rational diffusion coefficient (D) is not recovered. On clean data the new PDE is recovered on all three seeds, while the no-LLM scripts never recover it, although exp⁡\exp is on their menu (Table 8).

We thank the reviewer again for the time and the constructive comments.

Reviewer_zptPOpenReview ↗
Rating
4
Confidence
2
Informativeness
2
Presentation
3
Soundness
3
Read review
Paper summary

This paper combines blended-activation neural denoising with an LLM that selects PDE terms using retrieved skeletons and mathematical feedback.

Critical strengths

The denoising results are promising, particularly for PDEs whose governing terms involve higher-order derivatives. The paper addresses a real difficulty in equation discovery: a field approximation that looks accurate can still produce unreliable derivatives. I consider this a potentially useful preprocessing component for PDE identification. The use of mathematical diagnostics and residual feedback to refine an equation proposal is also beneficial.

Critical weaknesses

My main concern is that introducing an LLM does not appear to substantially change the traditional SINDy-style discovery workflow. The proposed method constructs a candidate derivative library, selects a sparse support, and fits coefficients by least squares. The LLM takes on the support-selection role and receives mathematical diagnostics and regression feedback to revise that choice. This may be a useful enhancement, but the underlying process remains close to library-based identification. The manuscript should make clear what new capability this module provides beyond that established framework. I would appreciate if the authors could further provide a controlled comparison demonstrated a substantial recovery benefit from the LLM, beyond the improved derivative estimates and numerical tools.

Questions for authors
  1. With the full BLADE-PDE denoiser held fixed, how does the discovery stage compare with a conventional sparse-regression estimator using the same fields and candidate library? Recovery rate and coefficient error on the existing noisy subset would help determine whether the LLM provides a benefit after the derivative-quality advantage is controlled.

  2. What capability requires the LLM to select the final support rather than guide a sparse-regression module? A comparison or analysis of a simple hybrid using the same structural suggestions would help establish the claimed distinction.

2 comments

Dear Reviewer,

We thank the reviewer for the careful reading and the two concrete questions, which we answer below with new experiments. In short, with our denoiser held fixed, sparse regression on the same fields and the same candidates recovers less than the LLM stage at every noise level from 10 to 40 dB: at 20 dB the best script recovers 4/6 PDEs (14/18 seeds) and PDE-FIND (Rudy et al., 2017) on our derivatives 3/6 (8/18), against 6/6 (18/18). The LLM is needed for the final support because BLADE-PDE is library-free: the library is only a guide, and the LLM does free-form discovery, so that it even recovers a PDE with an exponential term eue^{u} in the revised new PDE experiment, a term that is not mentioned anywhere in the prompt, while no sparse-regression setting recovers this PDE although exp⁡\exp is on its menu. The revised method enables the LLM to recover all six MDBench PDEs (Ziaei Bideh et al., 2026) at 20 dB even with neither the handbook nor the math tools. We therefore extend the evaluation with open-form PDEs that are in no handbook and no benchmark, and on these the full prompt recovers 3/4 at 20 dB while the prompts without the handbook or without the math tools perform worse.

All numbers come from the revised pipeline. The paper is being updated accordingly, and the tables cited below are in Appendix J of the revised paper. The code of the revised pipeline is in our anonymous repository, https://anonymous.4open.science/r/BLADE-PDE-5D37/, with the complete prompts, reasoning and replies of one example run. The discovery stage no longer receives equation skeletons or a property card; the handbook is a shuffled list of terms with every non-polynomial function left open as f(⋅)f(\cdot); the LLM has no tools and at most three rounds; and the reported equation is a least-squares refit on the terms it chose. The settings of the denoiser and of the discovery stage are in Table 12. Multi-seed counts are written as PDEs (seeds), for example 6/6 (17/18): a PDE counts as recovered when at least two of its three seeds are.

The detailed answers are as follows.

1. We ran this comparison. With the denoiser held fixed, we replaced the LLM by sparse regression on the same denoised fields and derivatives, with the guide terms as the candidate library and the same refit and grading (Table 8). We tested sequential thresholded least squares (Brunton et al., 2016), as in PDE-FIND, and greedy forward–backward selection, each at six thresholds. On the recovery rate, sparse regression at its best threshold recovers 3/6, 4/6, 4/6 and 5/6 PDEs at 10, 20, 30 and 40 dB, while the LLM recovers 5/6, 6/6, 6/6 and 6/6, and no threshold recovers all six PDEs at any noise level. On the coefficient error, when a method finds the correct terms its coefficients are the same as those of the LLM stage, because every method fits them by the same least squares on the same derivatives (Table 9). With the same denoiser, the LLM therefore improves which terms are found, not how accurately their coefficients are fitted. We also ran the official MDBench implementations of PDE-FIND and PySR (Cranmer, 2023) on our denoised field and derivatives, and each recovers 3/6 PDEs at 20 dB, again fewer than the LLM (Table 5).

2. The capability is free-form discovery. BLADE-PDE is library-free: the library is only a guide, and the LLM proposes terms in free form, whereas a sparse-regression module can only select among the columns it is given and judges them by coefficient size or fit gain alone. In the revised method regression already serves as the guide: the math tools give the LLM least-squares fits, fit gains and residual correlations, no math tool adds or removes a term, and the reported equation is a least-squares refit on the terms the LLM chose. For the reverse hybrid, in which the suggestions guide a sparse-regression module that selects the final support, we gave the scripts the same structural suggestions that the LLM receives: the guide terms and the handbook terms, with open functions filled from a fixed menu that includes exp⁡\exp, and the same math tools (the "+ handbook" columns of Table 8). The richer suggestions make sparse regression worse: on the noisy subset, at the default thresholds, thresholded least squares falls from 27/72 to 0/72 recovered seeds and greedy selection from 49/72 to 31/72, and no setting recovers the new PDE. With the handbook pool the greedy script does recover the other three open-form equations, whose functions are on its menu, but it then loses most of the clean benchmark. The LLM matters for two reasons. First, the prompt holds many kinds of information, from field statistics and term correlations to coefficient profiles, fit gains, residual checks and weak-form scores, and a hand-crafted script can hardly judge on all of it. Second, the knowledge of the LLM makes the discovery free-form: the new PDE adds an eue^{u} term to a KdV-type equation, no exponential function is mentioned in the prompt, and the LLM still finds it. Moreover, even without the math tools and the handbook the LLM still recovers all six benchmark PDEs at 20 dB (Table 10). The example run in our repository shows how the term is found (new PDE, clean data, seed 1, with the complete prompts, reasoning and replies). The handbook offers only the open form f(u)f(u), and the math tools show how the coefficient of each base symbol varies with uu. In round one the LLM reads that the part of utu_t that depends on uu alone is not constant, tries a straight line and a quadratic in uu, and then finds that this part divided by eue^{u} is constant, between −1.52 and −1.49 across the bins, so it proposes eue^{u} as a single term in place of a polynomial of three; from the same curves it reads uuxuu_x (the coefficient of uxu_x is linear in uu with slope −4) and uxxxu_{xxx} (a flat coefficient). In round two the math tools confirm the proposal: replacing eue^{u} by a constant, uu, u2u^{2} or u3u^{3} lowers R2R^{2} from 0.994 to between 0.72 and 0.85, and no further term adds more than 0.002. The refit gives ut=−1.499eu−3.988uux−0.248uxxxu_t=-1.499e^{u}-3.988uu_x-0.248u_{xxx}.

We thank the reviewer again for the time and the constructive comments.

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

BLADE-PDE first fits noisy field observations with an implicit coordinate network using a learned mixture of five activation functions and per-basis affine transformations. Automatic differentiation supplies derivatives of the fitted field. A subsequent LLM stage receives data-derived property probes, retrieved nameless equation skeletons, and candidate-library correlations, then proposes sparse supports that are filtered and jointly refitted.

The paper reports recovery of seven of fourteen clean MDBench systems and all six systems in its 20 dB noisy subset. Ablations examine denoising, activation initialization, handbook retrieval, and mathematical tools. A further synthetic experiment changes KdV coefficients and adds a linear transport term to test identification under controls on equation names and prompts.

Critical strengths
  1. The paper correctly focuses on a bottleneck upstream of symbolic search. Fitting the observed field accurately is not sufficient if its derivatives are unreliable, and separating a PDE-independent denoiser from the identifier avoids embedding a guessed governing equation into the reconstruction objective. This makes the denoiser potentially useful beyond the particular LLM pipeline.

  2. Providing structured evidence to the LLM is a practical insight. Property probes, derivative correlations, and residual feedback give the model information about a particular field rather than asking it to infer an equation from sparse raw numbers alone. This makes the proposal/refinement loop more interpretable and offers a clearer division between structural selection and numerical coefficient estimation.

  3. The component analyses probe parts of the motivation. The denoiser/initialization ablation indicates that representation choices matter for recovering higher-order systems, while the handbook/tool ablation separates improved search efficiency from eventual recovery. In particular, the finding that mathematical feedback matters more than handbook retrieval is a useful observation about how this pipeline operates.

Critical weaknesses
  1. The denoising insight is plausible but not sufficiently justified. The method is motivated by derivative fidelity, yet the network is trained with field-reconstruction MSE alone. A more flexible activation blend can also fit noise, and a SIREN-like scale matches a branch's initialization without guaranteeing stable high-order derivatives of the full mixture. The existing fixed-tanh and initialization comparisons show that these choices affect recovery, but do not establish whether blending addresses a general derivative-reconstruction problem or mainly supplies useful periodic features for the selected tasks. The missing link is between the stated bottleneck and the particular representation proposed to solve it.

  2. The claimed LLM discovery capability is not clearly distinguished from tool-assisted library selection. Section 3.3 has the LLM choose sparse supports from supplied terms, while Appendix F augments proposals with residual-correlated terms and joint least-squares fits. This is a narrower operation than the free-form discovery motivation suggests. The key question is what reasoning or structural generalization the LLM contributes beyond this numerical identification procedure. A non-LLM selector supplied with the same denoised fields and tool evidence is the most informative missing comparison because it tests that conceptual claim directly.

  3. The example intended to establish discovery beyond familiar equation families does not do so. Equation (2), u_t=-4u u_x-0.25u_xxx-1.5u_x, becomes U_t=-4U U_X-0.25U_XXX under X=x-1.5t and is a rescaled KdV-type equation. Fresh data and changed coefficients test identification, but not a qualitatively new structural capability. Likewise, omitting equation names from prompts or available reasoning text cannot exclude recognition of a familiar mathematical template. This matters because the example is used to support a stronger insight about discovery rather than merely a new coefficient fit.

Questions for authors
  1. What mechanism makes the proposed blend particularly suitable for reconstructing derivatives under noise, rather than only fitting the field? Can existing derivative-error diagnostics or a matched periodic denoiser clarify the distinct contribution of blending and initialization?

  2. What part of the structural identification requires the LLM once candidate terms, property probes, residual correlations, and joint refits are available? Please clarify the implemented search space and the insight a comparison with an equally informed non-LLM selector would test.

  3. Will the claim of discovery outside well-known families be revised given the KdV coordinate transformation? Which existing evidence supports structural generalization beyond changing coefficients or adding a removable transport term within a familiar family?

2 comments

Dear Reviewer,

We thank the reviewer for the careful reading and the three concrete questions, which we answer below, with new experiments where they were needed. In short, BLADE-PDE is library-free: the library is only a guide, and the LLM does free-form discovery, so that it even recovers a PDE with an exponential term eue^{u} in the revised new PDE experiment, a term that is not mentioned anywhere in the prompt. We agree that the tweaked KdV equation of the submission stays within the KdV family, and the revised experiment replaces it. The blend is suited to derivatives and not only to the field: our derivatives are on average more accurate than those of a matched sine-only SIREN (Sitzmann et al., 2020), a Fourier-feature network (Tancik et al., 2020) and Savitzky–Golay smoothing (Savitzky and Golay, 1964), and given our denoised field and derivatives, PDE-FIND (Rudy et al., 2017) and PySR (Cranmer, 2023) each rise from 0/6 PDEs on the raw field to 3/6 (8/18) at 20 dB. A no-LLM search given the same candidates and math tools recovers less at every noise level from 10 to 40 dB and never the exponential PDE. The revised method enables the LLM to recover all six MDBench PDEs (Ziaei Bideh et al., 2026) at 20 dB even with neither the handbook nor the math tools. We therefore extend the evaluation with open-form PDEs that are in no handbook and no benchmark, and on these the full prompt recovers 3/4 at 20 dB while the prompts without the handbook or without the math tools perform worse.

All numbers come from the revised pipeline. The paper is being updated accordingly, and the tables cited below are in Appendix J of the revised paper. The code of the revised pipeline is in our anonymous repository, https://anonymous.4open.science/r/BLADE-PDE-5D37/, with the complete prompts, reasoning and replies of one example run. The discovery stage no longer receives equation skeletons or a property card; the handbook is a shuffled list of terms with every non-polynomial function left open as f(⋅)f(\cdot); the LLM has no tools and at most three rounds; and the reported equation is a least-squares refit on the terms it chose. The settings of the denoiser and of the discovery stage are in Table 12. Multi-seed counts are written as PDEs (seeds), for example 6/6 (17/18): a PDE counts as recovered when at least two of its three seeds are.

The detailed answers are as follows.

1. The derivatives come from automatic differentiation of the fitted network, so they inherit the smoothness of its activations. A sine network is smooth in every derivative, which suits wave-like solutions, but the benchmark also contains steep fronts, chaotic dynamics and slowly decaying diffusive profiles. The matched periodic denoiser is a sine-only SIREN of the same width, depth and training budget (Table 7), and it is uneven across systems: with ω0=3\omega_0=3 it is comparable to ours on Burgers and KS but three times worse on the heat system and more than twenty times worse on advection–diffusion, and with the common ω0=30\omega_0=30 it fits the noise, with an error above 1 on every system. Blending the sine with non-periodic bases makes the fit much more robust, with a mean error of 0.08 against 0.25; this is the contribution of the blend. The contribution of the initialisation is separated by two ABU (Sütfeld et al., 2020) variants that differ only in it: the SIREN-like initialisation of the sine basis improves training, since the mean error rises to 0.13 with SIREN initialisation on every basis and to 0.10 without the compensation, although the per-basis affine map could learn the same scale. As derivative-error diagnostics against other denoisers at 20 dB, ours has the lowest error on every system against Savitzky–Golay smoothing, a Fourier-feature network and the network derivatives of the PINN method PDE-LEARN (Stephany and Earls, 2024) (Table 6). We attribute the larger error of the PINN method to its physics-informed loss: the network is fitted together with the equation being discovered, that equation is wrong on all six systems, and a wrong equation gives the fit wrong physics information, whereas our denoiser uses no physics term. The derivatives also transfer to other estimators: given our denoised field and derivatives, PDE-FIND and PySR each rise from 0/6 to 3/6 at 20 dB and stay there at 30 and 40 dB (Table 5).

2. The part that requires the LLM is the choice of the terms. BLADE-PDE is library-free: the library is only a guide, and the LLM proposes terms in free form. The implemented search space is any expression in the field and its derivatives, with products, quotients, powers and functions; the handbook lists terms with every non-polynomial function left open as f(⋅)f(\cdot), no math tool adds or removes a term, and the reported equation is a least-squares refit on the terms the LLM chose. To test whether a script can replace the LLM, we designed scripts that have everything the LLM gets: the same guide and handbook terms as candidates, with open functions filled from a fixed menu that includes exp⁡\exp, and the same math tools, refit and grading. We tested sequential thresholded least squares (Brunton et al., 2016), as in PDE-FIND, and greedy forward–backward selection, each at six thresholds and on two candidate pools (Table 8). The best single setting recovers 23/35 against 33/35 for the LLM stage; no setting recovers all six PDEs of a noisy level, and none recovers the new PDE. With the handbook pool the greedy script does recover the other three open-form equations, whose functions are on its menu, but it then loses most of the clean benchmark. The LLM matters for two reasons. First, the prompt holds many kinds of information, from field statistics and term correlations to coefficient profiles, fit gains, residual checks and weak-form scores, and a hand-crafted script can hardly judge on all of it. Second, the knowledge of the LLM makes the discovery free-form: the new PDE adds an eue^{u} term to a KdV-type equation, no exponential function is mentioned in the prompt, and the LLM still finds it. Moreover, even without the math tools and the handbook the LLM still recovers all six benchmark PDEs at 20 dB (Table 10). The example run in our repository shows how the term is found (new PDE, clean data, seed 1, with the complete prompts, reasoning and replies). The handbook offers only the open form f(u)f(u), and the math tools show how the coefficient of each base symbol varies with uu. In round one the LLM reads that the part of utu_t that depends on uu alone is not constant, tries a straight line and a quadratic in uu, and then finds that this part divided by eue^{u} is constant, between −1.52 and −1.49 across the bins, so it proposes eue^{u} as a single term in place of a polynomial of three; from the same curves it reads uuxuu_x (the coefficient of uxu_x is linear in uu with slope −4) and uxxxu_{xxx} (a flat coefficient). In round two the math tools confirm the proposal: replacing eue^{u} by a constant, uu, u2u^{2} or u3u^{3} lowers R2R^{2} from 0.994 to between 0.72 and 0.85, and no further term adds more than 0.002. The refit gives ut=−1.499eu−3.988uux−0.248uxxxu_t=-1.499e^{u}-3.988uu_x-0.248u_{xxx}.

3. Yes. We agree that the tweaked KdV equation of the submission stays within the KdV family: its transport term is removed by a moving frame and its coefficients by rescaling. Our reason for the original experiment was the following. When a PDE is solved, such a transformation yields the solution easily from that of the KdV equation. In PDE discovery, however, the equation is not known, and discovery methods, whether library-based or search-based, do not consider coordinate transformations: searching for a transformation together with a KdV equation is even harder than searching for a KdV equation with an additional term. We were therefore not sensitive enough to the transformation. The claim is revised, and that example is replaced by four open-form equations that are in no handbook and no benchmark (Table 10): the new PDE ut=−4uux−0.25uxxx−1.5euu_t=-4uu_x-0.25u_{xxx}-1.5e^{u}, whose exponential term is not a transport term and is not removed by a change of coordinates, and three equations with the terms sin⁡(u)ux\sin(u) u_x, cos⁡(u)\cos(u) and uxx/(1+u2)u_{xx}/(1+u^{2}). The structure beyond a familiar family is the non-polynomial function itself: the handbook leaves every such function open as f(⋅)f(\cdot), no exponential function is mentioned in the prompt, and the LLM still finds eue^{u}. At 20 dB the full prompt recovers 3/4 of these equations, against 1/4 without the handbook and none without the math tools; the equation with the rational diffusion coefficient (D) is not recovered. On clean data the new PDE is recovered on all three seeds, while the no-LLM scripts never recover it, although exp⁡\exp is on their menu (Table 8).

We thank the reviewer again for the time and the constructive comments.

Reviewer_RDcMOpenReview ↗
Rating
2
Confidence
2
Informativeness
2
Presentation
2
Soundness
2
Read review
Paper summary

BLADE-PDE is a two-stage pipeline for discovering PDEs from noisy field data. Stage 1 fits the noisy observations with a coordinate MLP whose hidden activation is an "Adaptive Blending Unit" (ABU): a learned mixture $ \alpha_k $ of five bases $ {e^z-1, \tanh, GELU, softplus, \sin} $ , each wrapped in a trainable affine map $ s_k z + b_k $ . The affine wrap allows a SIREN-like initialization compensation to be applied only to the sine channel ( $ s_{\sin} = \sigma_{SIREN}/\sigma_{Kaiming} $ , equal to $ 1/\omega_0 $ in hidden layers), while the shared linear layers keep a Kaiming initialization. Derivatives are then taken by automatic differentiation. Stage 2 builds a derivative library from the denoised field and asks a general LLM to propose a sparse support and coefficients, given (a) a "property card" computed from the data (parity, conservation, scaling balance, dispersion order, etc.), (b) nameless equation skeletons retrieved from a "math handbook", and (c) library correlations. Mathematical tools (dependency/QR, weak-form, divergence-form, scaling, variational and dispersion filters) score each proposal and send notes back for one more round (two rounds in total); final coefficients come from a joint least-squares refit.

On MDBench, the discovery stage alone recovers 7 of 14 clean systems (the strongest MDBench estimators recover 6; NLS is recovered by none of the tested MDBench methods), with mean coefficient error $ \le $ 1.5% on recovered systems. On the six-system noise subset at 20 dB, the full pipeline recovers 6/6 whereas the baselines that run on noisy fields (PDE-FIND, WSINDy, PySR, Operon, uDSR, plus EqGPT and LLM4ED) recover 0/6. Ablations at 20 dB show 0/6 without the denoiser, 2/6 with a fixed-tanh denoiser, 4/6 for ABU without SIREN-like compensation (missing KdV and KS), and 6/6 for the full model; removing the mathematical tools reduces the LLM stage to 3/6, whereas removing the handbook still gives 6/6 but needs a second round for KdV and KS. A leakage-controlled experiment (opaque IDs, prompt audit, tools disabled, reasoning-transcript scan) on a freshly simulated, modified KdV-type equation ( $ u_t = -4uu_x - 0.25u_{xxx} - 1.5u_x $ ) recovers the support in two rounds with a mean coefficient error of 2.1%.

Critical strengths
  1. The component ablation (Table 2) and LLM-mechanism ablation (Table 3) separate the contribution of the denoiser from that of the LLM stage and trace them to specific failure modes: finite differences on 20 dB data give 0/6; a fixed activation gives 2/6; the blended activation without the sine compensation fails exactly on the two systems with third- and fourth-order terms (KdV, KS); and without the math tools the LLM loses Burgers, Heat 1D and AD, while the handbook mainly saves a round. This is consistent with the claim that derivative quality for high-order terms is the bottleneck, and it is informative for anyone building LLM-based discovery pipelines: the numeric verification tools, not extra prior knowledge, are what makes proposals correct.
  2. All LLM runs use opaque case identifiers, a banned-term audit of prompts, disabled tools, and scanning of reasoning transcripts for known equation names; the LLM only selects the support while coefficients come from least squares; and a freshly simulated equation is used as an additional check. Combined with the per-round trace in the appendix, the output is explicit and inspectable. Many LLM-for-science papers do not address leakage at all, so this protocol is a real strength even though it is not conclusive (see below).
Critical weaknesses
  1. Only one of MDBench's four noise levels is reported (20 dB; 40, 30 and 10 dB are not), and only on the six-system noise subset, which is relatively simple. Each ablation cell is a single binary outcome, with no repeated network trainings, repeated LLM samples, variance or confidence intervals, although both the network fit and the LLM are stochastic. It is therefore hard to judge how robust the result is, or how much would transfer to other noise levels, systems or LLMs.

  2. The noisy comparison is unequal. The baselines receive the noisy data directly, whereas BLADE-PDE adds a trained denoiser, an LLM, up to two refinement rounds and a suite of tools. There is no "baseline + the same denoiser" control, even though the denoiser is described as usable with any downstream identifier. The gain from denoising therefore cannot be separated from the LLM stage. LLM4ED also differentiates the noisy field numerically, which puts it at a disadvantage.

  3. The only denoiser comparisons are finite differences and fixed tanh. There is no comparison against other reasonable denoisers (SIREN or Fourier-feature networks, a standard PINN/DeepMoD-style fit, spline or Savitzky–Golay smoothing). The need for the blended activation, as opposed to a better smoother, is therefore not established. The learned mixture weights $ \alpha_k $ are described as differing across PDE families, but no quantitative report or ablation of the basis set is given.

  4. The method is not truly library-free, which conflicts with the abstract's "free-form candidates without a hand-crafted library" framing. The pipeline proposes sparse combinations over a predefined derivative library with coefficients fixed by least squares, so terms outside the library cannot be found. The leakage test uses a modified KdV whose terms all lie in the library, on a single example. It supports "no verbatim recall" more than "discovery of new expressions". The contents and size of the handbook collection, and whether it contains skeletons of the benchmark equations, are also not described.

  5. The need for the LLM is not shown. The handbook's contribution is small (6/6 without it), the tools carry substantial hand-designed structure, and thresholds such as residual correlation $ > 0.15 $, up to eight additional candidate terms and $ R^2 < 0.95 $ come with no sensitivity analysis. There is no no-LLM control (tools plus sparse regression or greedy search) to show that the LLM is necessary rather than a heuristic search over a small library. On clean data the margin is one system (7/14 vs 6/14), all seven failures are in the "real research problem" group (2D/3D heat, laser, RD, NS), and the noisy experiment uses the simpler subset, so evidence on harder coupled or multi-variable systems under noise is missing.

  6. Reproducibility and cost are insufficiently reported. The coefficient tolerance is relaxed to 15% in the noisy setting and no coefficient errors are reported there. The denoiser hyperparameters (depth, width, $ \omega_0 $, optimizer, epochs) and the baselines' tuning are not given. The LLM used by BLADE-PDE is never named (Grok 4.6 appears only for the LLM4ED baseline). Training time and LLM call/token counts and cost are not reported. Code is to be provided only to reviewers. The contribution is mostly a combination of existing ingredients (SIREN-style initialization, adaptive activation, LLM + verification loop), and the two-round budget limits what can be said about the power of tool-driven refinement.

  7. The method overview figure and the experimental result tables lack clarity and detail.

Questions for authors
  1. What are the results at 10, 30 and 40 dB, and over multiple network seeds and LLM samples (with variance)? Where does the pipeline start to fail?
  2. Do the MDBench baselines (e.g. PDE-FIND, WSINDy, PySR) improve substantially when given the ABU-denoised field and its AD derivatives? How does ABU compare with SIREN, Fourier-feature networks, a plain DeepMoD/PINN fit and spline/Savitzky–Golay smoothing at 20 dB?
  3. Which LLM (name, version, temperature) does BLADE-PDE use, how many calls and tokens per system, and how do results change with a different LLM (including an open-weight one)?
  4. Is there a no-LLM control that uses the same tools with sparse regression or a greedy/beam search over the library? How much of the 6/6 result is attributable to the LLM?
  5. What does the handbook collection contain (size, source), and does it include skeletons that correspond to the benchmark equations (Burgers, KdV, KS, etc.)? Are the leakage-control conclusions robust to removing such entries?
  6. Can you report the hyperparameters of the denoiser, the baseline settings and tuning protocol used for the re-run baselines, and the sensitivity of the tool thresholds?
2 comments

Dear Reviewer,

We thank the reviewer for the careful reading and the six concrete questions, which we answer below, with new experiments where they were needed. In short, BLADE-PDE is library-free: the library is only a guide, and the LLM does free-form discovery, so that it even recovers a PDE with an exponential term eue^{u} in the revised new PDE experiment, a term that is not mentioned anywhere in the prompt. It performs consistently at 20, 30 and 40 dB over three denoiser seeds, recovering all six PDEs on every seed except one seed of the heat system at 30 dB, and it still recovers 5/6 PDEs (14/18 seeds) at 10 dB. The denoiser is useful beyond our pipeline: given our denoised field and derivatives, PDE-FIND (Rudy et al., 2017) and PySR (Cranmer, 2023) each rise from 0/6 PDEs on the raw field to 3/6 (8/18) at 20 dB, and our derivatives are on average more accurate than those of SIREN (Sitzmann et al., 2020), a Fourier-feature network (Tancik et al., 2020) and Savitzky–Golay smoothing (Savitzky and Golay, 1964). The discovery stage does not depend on one LLM: the open-weight DeepSeek V4.1 Flash recovers 5/6 at 20 dB. A no-LLM search given the same candidates and math tools recovers less at every noise level from 10 to 40 dB and never the exponential PDE. The revised method enables the LLM to recover all six MDBench PDEs (Ziaei Bideh et al., 2026) at 20 dB even with neither the handbook nor the math tools. We therefore extend the evaluation with open-form PDEs that are in no handbook and no benchmark, and on these the full prompt recovers 3/4 at 20 dB while the prompts without the handbook or without the math tools perform worse. All hyperparameters are fixed across systems, noise levels and seeds.

All numbers come from the revised pipeline. The paper is being updated accordingly, and the tables cited below are in Appendix J of the revised paper. The code of the revised pipeline is in our anonymous repository, https://anonymous.4open.science/r/BLADE-PDE-5D37/, with the complete prompts, reasoning and replies of one example run. In response to the leakage concern, the discovery stage no longer receives equation skeletons or a property card; the handbook is a shuffled list of terms with every non-polynomial function left open as f(⋅)f(\cdot); the LLM has no tools and at most three rounds; and the reported equation is a least-squares refit on the terms it chose. The settings of the denoiser and of the discovery stage are in Table 12. Multi-seed counts are written as PDEs (seeds), for example 6/6 (17/18): a PDE counts as recovered when at least two of its three seeds are.

The detailed answers are as follows.

  1. Table 4 gives the noise ladder over three denoiser seeds, with one LLM run per denoiser. From 20 to 40 dB all six PDEs are recovered, and the only miss is the heat system on one seed at 30 dB, where the LLM adds small higher-order terms. The pipeline starts to fail at 10 dB, where it recovers 5/6 (14/18), and the cause is mostly the denoiser: for the heat system even a least-squares fit on the true term is more than 15% off. For the variance over LLM samples, we repeated the LLM stage three times on fixed data for the four clean open-form equations, and it recovers the same 3/4 each time.

  2. Yes. Given our denoised field and derivatives, PDE-FIND and PySR each rise from 0/6 to 3/6 at 20 dB and stay there at 30 and 40 dB (Table 5), so the denoiser transfers to other estimators and the remaining gap to BLADE-PDE belongs to the discovery stage. On derivative quality at 20 dB, our denoiser has the lowest error on every system against Savitzky–Golay smoothing, a Fourier-feature network and the network derivatives of the PINN method PDE-LEARN (Stephany and Earls, 2024) (Table 6). We attribute the larger error of the PINN method to its physics-informed loss: the network is fitted together with the equation being discovered, that equation is wrong on all six systems, and a wrong equation gives the fit wrong physics information, whereas our denoiser uses no physics term. Against SIREN (Table 7), a sine network is smooth in every derivative, which suits wave-like solutions, but the benchmark also contains steep fronts, chaotic dynamics and slowly decaying diffusive profiles. A sine-only SIREN is therefore uneven across systems: with ω0=3\omega_0=3 it is comparable to ours on Burgers and KS but three times worse on the heat system and more than twenty times worse on advection–diffusion, and with the common ω0=30\omega_0=30 it fits the noise, with an error above 1 on every system. Blending the sine with non-periodic bases makes the fit much more robust, with a mean error of 0.08 against 0.25. The SIREN-like initialisation of the sine basis also improves training: although the per-basis affine map could learn the same scale, the ABU (Sütfeld et al., 2020) without the compensation has a higher mean error (0.10).

3. BLADE-PDE uses Grok 4.6 through the Cursor SDK, with reasoning effort high and every tool disabled. The interface exposes no sampling temperature, so the provider default applies and runs are not deterministic. The stage makes one call per round for at most three rounds, and Table 11 gives the measured calls and tokens per system. With a different LLM, the open-weight DeepSeek V4.1 Flash in the same setup apart from a longer call limit recovers 5/6 at 20 dB; its miss is the heat system, with the same extra higher-order terms as in reply 1.

4. BLADE-PDE is library-free: the library is only a guide, and the LLM proposes terms in free form. To test whether a script can replace the LLM, we designed scripts that have everything the LLM gets: the same guide and handbook terms as candidates, with open functions filled from a fixed menu that includes exp⁡\exp, and the same math tools, refit and grading. We tested sequential thresholded least squares (Brunton et al., 2016), as in PDE-FIND, and greedy forward–backward selection, each at six thresholds and on two candidate pools (Table 8). The best single setting recovers 23/35 against 33/35 for the LLM stage; no setting recovers all six PDEs of a noisy level, and none recovers the new PDE. With the handbook pool the greedy script does recover the other three open-form equations, whose functions are on its menu, but it then loses most of the clean benchmark. The LLM matters for two reasons. First, the prompt holds many kinds of information, from field statistics and term correlations to coefficient profiles, fit gains, residual checks and weak-form scores, and a hand-crafted script can hardly judge on all of it. Second, the knowledge of the LLM makes the discovery free-form: the new PDE adds an eue^{u} term to a KdV-type equation, no exponential function is mentioned in the prompt, and the LLM still finds it. Moreover, even without the math tools and the handbook the LLM still recovers all six benchmark PDEs at 20 dB (Table 10).

5. The handbook pools the terms of three sources: our original 72 templates of canonical pde classes, the Wikipedia list of nonlinear partial differential equations (Wikipedia, 2026) (30 equations) and the EqWorld digest of Polyanin and Zaitsev (2012) (55 equations). It holds no skeletons, names or coefficients, only a shuffled list of terms with every non-polynomial function left open as f(⋅)f(\cdot). It does contain the terms of the benchmark equations, but the benchmark does not depend on them: it stays at 6/6 with the whole handbook removed (Table 10). The leakage conclusion therefore rests on four open-form equations that are in no handbook and no benchmark, among them the new PDE with an eue^{u} term. On these both components matter: the full prompt recovers 3/4, against 1/4 without the handbook and none without the math tools.

6. Table 12 lists the settings of the denoiser and of the discovery stage; they are the same for every system, noise level and seed. The re-run baselines use the official MDBench adapters unchanged under the benchmark's own protocol (an 80/20 split, the benchmark's hyperparameter grid and fitness, a 12-hour limit per run) with no further tuning. On tool thresholds, in the revised discovery stage the LLM chooses every term itself: no math tool adds or removes a term, and the reported equation is a refit on the terms it chose. The three constants in question only set what is displayed to the LLM after each round: the eight guide terms most correlated with the residual are listed, those with a correlation above 0.15 enter a joint fit that is shown, and a note is added when the hold-out R2R^{2} is below 0.95. Changing the correlation threshold to 0.05 or 0.30, or the R2R^{2} threshold to 0.90 or 0.99, leaves the 20 dB benchmark at 6/6 with the same recovered equations (seed 1, one LLM run per PDE).

We thank the reviewer again for the time and the constructive comments.

Comments on the paper

1 comment
AuthorsOpenReview ↗

Main Additional Results

We have revised the paper to align with the revised method and experiments. For easier reading, we collect here the main tables of the revised experiments, numbered as in Appendix J of the revised paper.

Table 4. Noise ladder on the MDBench noise subset (Ziaei Bideh et al., 2026). The per-system columns count recovered seeds out of three, and "Total" is PDEs (seeds); AD is the two-dimensional advection–diffusion system. "Coeff. error" is the largest relative coefficient error of a recovered equation, as mean and maximum over the recovered seeds.

NoiseAdv.BurgersKdVKSHeat 1DADTotalCoeff. errorTrue terms fit
10 dB3333025/6 (14/18)3.8% / 8.4%5/6 (15/18)
20 dB3333336/6 (18/18)2.5% / 8.2%6/6 (18/18)
30 dB3333236/6 (17/18)1.7% / 8.0%6/6 (18/18)
40 dB3333336/6 (18/18)2.2% / 11.4%6/6 (18/18)

Table 5. MDBench (Ziaei Bideh et al., 2026) estimators, PDE-FIND (Rudy et al., 2017) and PySR (Cranmer, 2023), given our denoised field and its automatic-differentiation derivatives, graded with the same criterion. Three denoiser seeds per noise level; the noise-level columns are PDEs (seeds). The per-system columns count recovered seeds out of the nine at 20, 30 and 40 dB. AD is the two-dimensional advection–diffusion system. "Raw" is the raw noisy field, in single runs per system; "denoised" is our denoised field with its derivatives. The baselines are the official MDBench adapters under the benchmark's protocol. The BLADE-PDE entries are those of Table 4.

MethodInput20 dB30 dB40 dBAdv.BurgersKdVKSHeat 1DAD
PDE-FINDraw0/60/60/6000000
PDE-FINDdenoised3/6 (8/18)3/6 (9/18)3/6 (9/18)809009
PySRraw0/60/60/6000000
PySRdenoised3/6 (8/18)3/6 (10/18)3/6 (10/18)980029
BLADE-PDEdenoised6/6 (18/18)6/6 (17/18)6/6 (18/18)999989

Table 6. Derivative quality at 20 dB, seed 1: relative RMSE against the noiseless field and its derivatives on the grid interior, averaged over the field, the time derivative and the spatial derivatives that appear in each system's equation. The compared methods are Savitzky–Golay smoothing (Savitzky and Golay, 1964), a Fourier-feature network (Tancik et al., 2020) with the same width, depth and training budget as ours, and PDE-LEARN (Stephany and Earls, 2024).

MethodAdv.BurgersKdVKSHeat 1DADMean
BLADE-PDE denoiser0.0550.0800.0400.2060.1110.0290.087
Savitzky–Golay, best single window (41)0.5820.2330.1500.3600.2910.0710.281
Fourier-feature network, frequency scale 10.3150.2130.2641.1470.8051.9260.778
PDE-LEARN (PINN), network derivatives0.6890.1430.2051.2131.1690.1830.600

Table 8. No-LLM control on the same data, candidates, math tools, refit and grading, with three seeds per PDE. Entries are PDEs (seeds). "Default" is the threshold fixed before any run; "best" is the single threshold with the highest total, chosen afterwards. "Thresholded LSQ" is sequential thresholded least squares (Brunton et al., 2016). The "default" and "best" columns use the guide terms as the candidate pool; "+ handbook" adds the handbook terms, with open functions filled from a fixed menu that includes exp⁡\exp, at the default threshold. The open-function equations are B, C and D of Table 10.

SetThresholded LSQ defaultThresholded LSQ bestThresholded LSQ + handbookGreedy selection defaultGreedy selection bestGreedy selection + handbookBLADE-PDE (LLM)
MDBench clean5/7 (14/21)6/7 (16/21)2/7 (6/21)6/7 (17/21)7/7 (21/21)3/7 (9/21)7/7 (21/21)
MDBench 10 dB1/6 (4/18)1/6 (3/18)0/6 (0/18)3/6 (9/18)3/6 (11/18)3/6 (8/18)5/6 (14/18)
MDBench 20 dB2/6 (6/18)2/6 (7/18)0/6 (0/18)4/6 (13/18)4/6 (14/18)1/6 (5/18)6/6 (18/18)
MDBench 30 dB3/6 (8/18)3/6 (8/18)0/6 (0/18)5/6 (15/18)4/6 (13/18)3/6 (9/18)6/6 (17/18)
MDBench 40 dB3/6 (9/18)2/6 (8/18)0/6 (0/18)4/6 (12/18)5/6 (14/18)3/6 (9/18)6/6 (18/18)
New PDE, clean0/1 (0/3)0/1 (0/3)0/1 (0/3)0/1 (0/3)0/1 (0/3)0/1 (0/3)1/1 (3/3)
Open-function, clean0/3 (0/9)0/3 (0/9)2/3 (6/9)0/3 (0/9)0/3 (0/9)3/3 (9/9)2/3 (5/9)
Total14/35 (41/105)14/35 (42/105)4/35 (12/105)22/35 (66/105)23/35 (73/105)16/35 (49/105)33/35 (96/105)

Table 10. Information ablation at 20 dB, seed 1, same denoised fields, one run per PDE, all tools disabled. The four new equations are ut=−4uux−0.25uxxx−1.5euu_t=-4uu_x-0.25u_{xxx}-1.5e^{u} (A), ut=−2sin⁡(u)ux+0.5uxxu_t=-2\sin(u) u_x+0.5u_{xx} (B), ut=−uux+0.5uxx+1.2cos⁡(u)u_t=-uu_x+0.5u_{xx}+1.2\cos(u) (C) and ut=0.5uxx/(1+u2)−uuxu_t=0.5u_{xx}/(1+u^{2})-uu_x (D).

PromptMDBench (6)ABCDNew (4)
Handbook and math tools (full)6✓✓✓✗3
Without handbook6✓✗✗✗1
Without math tools3✗✗✗✗0
Without both6✗✗✗✗0

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

Related papers

Loading the map…

Discussion (0)

Sign in to comment.