acceptodds
Under review as a conference paper at ICLR 2027

TERM: Enriching Representations in PDE Foundation Models with Term-wise Physical Responses

Abstract

Partial Differential Equation (PDE) foundation models aim to learn transferable solution priors across heterogeneous physical systems, yet existing methods typically represent governing equations as static symbolic structures or combine equations and trajectories only at a global level. Such representations capture which physical terms constitute a PDE, but rarely how each term acts in the current state. Although physical terms have stable mathematical identities, their strength, spatial distribution, and dynamical roles vary along physical trajectories. Motivated by this observation, we propose the Term-wise Enriched Representation Model (TERM), which organizes PDE information at the level of individual physical terms when the governing equation is available. TERM first normalizes the equation into a structured term graph, then computes state-dependent responses for each physical term and explicitly binds them to their symbolic identities. A structure-aware Transformer jointly models these response-enriched term representations and compact trajectory information, while a continuous query decoder enables prediction across equations, resolutions, and continuous coordinates. During autoregressive rollout, the equation structure remains fixed while term responses are recomputed from the evolving predicted state. Experiments across multi-PDE pretraining, equation-specified zero-shot transfer, compositional generalization, cross-resolution prediction, and long-horizon rollout show consistent improvements over the evaluated PDE foundation-model baselines and representation controls. The results suggest that state-dependent, term-level physical representations provide a useful interface for scalable and compositional PDE foundation modeling.

Primary area:
applications to physical sciences (physics, chemistry, biology, etc.)
Keywords:
PDE foundation models, physics-informed representation learning, term-wise physical responses, partial differential equations

Reviews

Among:
Shape:
1234this paper, Top 30%
Figure 1. Mean review rating of the reviewed papers at ICLR 2027. This paper averages 2.67 over 3 reviews, above 70% of the others.
Reviewer_zyZ2OpenReview ↗
Rating
3
Confidence
3
Informativeness
2
Presentation
3
Soundness
3
Read review
Paper summary

This paper proposes a foundation model known as TERM that incorporates state-local governing equation information into its latent representations. The authors argue that global PDE structure is not sufficient for distinguishing different parts of the state space, since mathematical phenomena such as term cancellation may unduly unite very different states. By exploiting a DAG representation of governing equations, TERM informs each token in the context window with its contributions from each term in the governing equations. Results and ablations on standard benchmarks are presented along with some very light theoretical analysis.

Critical strengths

There are many benchmarks and it is clear that the method improves on prior art. The use of generative AI is (mostly) appropriately subtle and I trust that the core contributions come from the authors. Overall I feel that the "pros" outweigh the "cons" and (weakly) recommend acceptance.

Critical weaknesses

The biggest differentiator of this method is also its biggest drawback: it requires exact knowledge of governing equations. This is arguably too restrictive for the majority of situations where foundation models are applied, although I can see a use case as a surrogate for high-fidelity PDE codes. Given this requirement, I think the framing against benchmarks is not 100% transparent; this method is more performant, but also requires strictly more information. The benefit of the analysis in section 4 is also unclear. The first result is a tautology given how the Om\mathcal{O}_m are constructed, and the second result is a straightforward application of the triangle inequality given strong assumptions on the system and observations. There is no experimental verification of either result, so I do not feel that they add anything to the paper. Additionally, TERM relies on Fourier features and is restricted to regular grids, heavily limiting its applicability to realistic computational scenarios.

Questions for authors
  • Why does ℓmt\ell^t_m appear in eqn 3 when no other parts of this equation use it?
  • RmtR^t_m in eqn 4 is not defined.
  • There are various LLM-isms like "binding these views" that are linguistically jarring and obviously machine-generated. I recommend reading the prose carefully and rewriting things in a more human way.
4 comments

Weaknesses: First, we agree that TERM's dependence on a specified equation is a substantive restriction. Its intended setting is an equation-conditioned surrogate: the equation, coefficients and relevant boundary/domain information are available, and the model predicts future fields for different states and problem instances. Unseen equation means absent from the applicable training split but supplied at inference; it does not mean unknown to the system. Cases where the law is unknown, essential variables are unavailable, or the supplied equation or coefficients are inaccurate are outside what the present transfer experiments establish.

Second, TERM computes the numerical term responses from the supplied equation and observed state. A baseline that is not given the equation or these features has less physical conditioning. In the revised manuscript, we have clarified that the main benchmark compares complete methods under their respective input configurations. Its performance gaps cannot by themselves attribute the improvement to architectural design under identical information. The shared-backbone controls provide a more controlled analysis, but the original Unbound Response does not include explicit response-source labels. New source-labeled variants in Sec. 5.4, Table 4 help assess that distinction; their remaining limitations are explained in our reply to Reviewer 8KvM.

Third, for the theoretical analysis in Sec. 4, we agree that the response-sum identity is definitional and that non-injectivity of summation does not by itself prove a prediction gap. The revised Sec. 4 explains the locality of term enrichment and a conditional pooled-response attribution bound. App. A.6 proves the block-diagonal local Jacobian, the attribution and representation-separation bounds, and a filtered linear Fourier-evaluator perturbation bound. These results concern local updates, pooled targets or the stated linear periodic evaluator; they do not prove full-field recovery or stability of the complete model. In particular, the horizon-uniform rollout estimate still requires LTERM<1L_{\mathrm{TERM}} < 1, which the architecture does not enforce. Existing ablations and noise results provide empirical context but do not verify the bounds or their constants.

Finally, considering the Fourier features and regular-grid limitations, TERM consists of three different stages:

  1. Input observations: the reported cross-resolution study starts with history fields sampled on a structured 128×128128\times128 grid. It does not validate the complete pipeline on native unstructured input meshes.
  2. Response evaluation: on the structured input grids, periodic problems use filtered spectral differentiation and nonperiodic problems use boundary-aware finite differences. The choice is a numerical evaluator for the specified equation; it is distinct from coordinate encoding in the decoder.
  3. Output queries: Fourier features encode coordinates supplied to the continuous query decoder. Those output coordinates need not form the input grid. In the existing cross-resolution study, irregular output queries give relative L2L_2 0.0418 for TERM versus 0.0537 for a fixed-grid decoder followed by bilinear interpolation; at 128→256128\to256 the corresponding errors are 0.0395 and 0.0498.

To clarify this, the revised Sec. 5.5 summarizes the irregular-query comparison. These results support off-grid output querying given structured-grid histories, not a demonstrated unstructured-mesh input pipeline. Adapting the input representation and constructing reliable derivatives on irregular geometries require further validation. We now state that boundary in the conclusion and implementation discussion (Secs. 3.2 and 3.4; App. A.4.4).

Question 1 — Role of the log response scale in Equation 3

Q1: The log response scale in Eq. 3 is the logarithm of the term's response-window RMS scale, ℓmt=log⁡amt\ell_m^t=\log a_m^t. The same equation defines the normalized response field and this separate scale feature. The normalized field communicates the spatial pattern while ℓmt\ell_m^t retains amplitude information; the response encoder receives the scale, and the gate in Eq. (4) also uses it. We now call out this subsequent use near the definition.

Question 2 — Definition of the response tokens in Equation 4

Q2: Lowercase ρm,kt\rho_{m,k}^t denotes the numerical response field at frame kk. The response encoder processes the normalized fields, their coordinates, relative times and scale; learned queries then resample the resulting features into a fixed-size token set RmtR_m^t. After projection to a common attention width, RmtR_m^t has shape R×dR\times d. The symbolic term embedding eme_m is the query, while RmtR_m^t supplies keys and values in its term-local cross-attention. We have added the distinction and dimensions in Sec. 3.2.

Question 3 — Unnatural wording and language

Q3: We agree that abstract phrases can obscure the operation and have revised them in the manuscript. For example, the revised conclusion states that TERM enriches symbolic PDE terms with their state-dependent responses before the shared Transformer. The revised method opening also gives a Burgers example before the module-level description. Thank you for pointing out that the presentation should make the operations directly traceable to the equations.

Thanks to the authors for providing additional details and clarifications. Given that I still have concerns about the utility of the theory and fairness of the comparisons, I will maintain my score of "weak accept".

Thank you for reviewing the revision and for clarifying your remaining concerns.

Comparison fairness. TERM and Origo use exactly the same train/test trajectories on all ten benchmarks, the same 10-frame history, a 10-step training rollout, and relative L2L_2 evaluation. They also share 1,000 joint pretraining epochs, a total batch size of 24, and 200 adaptation epochs per task. Section 5.1 and App. A.3.4 state these matched settings and document the different learning rates and schedules. Tables 1 and 3 identify the external values as published references and explain TERM's additional equation and response inputs. Tables 2 and 4 provide the controlled comparisons of response organization.

Use of the theory. We shortened Sec. 4 and tied it to two design choices. With the other response inputs held fixed, term-local fusion gives each response a direct path to its own embedding before the Transformer mixes terms. The pooled-response objective supervises that embedding, and App. A.6 states when reconstruction error and target separation can bound misattribution. The discussion of evaluator sensitivity motivates the TERM-Direct comparison in Sec. 5.5. Those experiments report prediction error under perturbations. The analysis describes the assumptions for the bounds, including contraction of the full predictor for the horizon-uniform result in App. A.2.5.

Reviewer_8KvMOpenReview ↗
Rating
3
Confidence
2
Informativeness
2
Presentation
2
Soundness
2
Read review
Paper summary

TERM is a PDE surrogate solver pretrained jointly on 10 benchmarks, for settings where the governing equation is known. It splits the equation into its additive right-hand-side terms, computes each term numerically on the recent observed frames, and binds each resulting field to that term's symbolic embedding before a shared transformer predicts the next state. It reports gains over existing multi-PDE pretrained models.

Critical strengths

It shows gains compared with existing models.

Critical weaknesses

It is not clear in the unseen equation generalization, what is the limit of TERM – what type of unseen PDE TERM can successfully generalize to?

Questions for authors

Does forcing the "binding" really help? The "Unbound Response" and "Global Response" controls are presented as having the same information as TERM. Do they tell the model which field came from which term? If not, could you rerun them with source-term labels to show TERM's explicit pairing still helps?

1 comment

Weakness 1 — Limits of unseen-equation generalization

W1: We thank the reviewer for the detailed review and feedback. TERM assumes that the target PDE, coefficients, boundary/domain information and the observations needed by the evaluator are available at inference. An unseen equation is absent from the relevant training split but supplied at test time; this is not discovery of an unknown physical law.

For each of five training seeds (0–4), the same frozen ten-task checkpoint is evaluated on both held-out KdV and Allen–Cahn; results are summarized across seeds. Allen–Cahn recombines diffusion and reaction mechanisms represented in pretraining. In KdV, the complete third-order dispersion term is absent during pretraining, but the supplied equation lets the numerical evaluator compute its action on the current state. Thus the model does not discover that operator from trajectories.

The separate strict-composition benchmark trains on the six pairs of advection, diffusion, reaction and dispersion, and holds out four complete three-term combinations in 1D. The 2D study trains on three pairs and holds out A+D+RA+D+R. Coefficient ranges, initial-condition family and solver protocol are shared. TERM's 1D average and 2D errors are 0.0217 and 0.0360, versus 0.0271 and 0.0430 for Unbound. A separate leave-one-viscosity study tests selected FNO-NS coefficient shifts in independently trained models (App. A.4.2).

In the revised manuscript, Sec. 5.4 maps each transfer type to its seen components and held-out structure. These results do not establish success for unknown laws, unavailable variables, inaccurate coefficients or arbitrary distribution shifts. Broader pretraining over combinations, parameters and time scales may expand dynamical coverage; this is a future direction, not a tested guarantee (Sec. 6).

Question 1 — Whether explicit binding helps with source labels

Q1: Regarding term–response binding, the original controls do not attach a source-term label to every response token:

  • TERM uses the equation graph to form a symbolic embedding eme_m for each term and restricts that term's local cross-attention to its own response tokens RmR_m. The association is supplied structurally.
  • Unbound Response retains the symbolic term embeddings and separate response tokens, but each term attends to the union of all response tokens. No explicit source label is attached to each token.
  • Global Response concatenates encoded responses and compresses them into a global bank without preserving an explicit per-token source index.

Therefore, the original statement that Unbound Response and Global Response controls have the same information as TERM was not quite accurate. They start from the same equation and computable response fields, but differ in what source correspondence is made explicit and, in the global case, in how responses are compressed. The original comparison cannot by itself separate the value of source labels from the value of restricting attention.

Source-labeled controls. We add two controls on the strict held-out 1D composition split (Sec. 5.4, Table 4). For each response token Rm,rR_{m,r}, a projection of its concatenation with the source embedding eme_m supplies a label at the same token width. Unbound+Labels preserves global access: every symbolic term can attend to all source-labeled response tokens. Global+Labels globally resamples the labeled tokens. Labels are supplied during both training and testing, and the controls keep their respective fusion structures.

MethodA+D+RA+D+RA+D+SA+D+SA+R+SA+R+SD+R+SD+R+SAverage
Unbound0.02710.02490.02960.02670.0271
Unbound+Labels0.02430.02210.02620.02380.0241
Global+Labels0.02560.02320.02780.02500.0254
TERM0.02190.01970.02380.02150.0217

The values are relative L2L_2 errors and lower is better. Adding labels improves the unbound mean from 0.0271 to 0.0241, about 11.1%; TERM is about 10.0% lower than Unbound+Labels based on these displayed averages. Thus, in this benchmark, providing source information accounts for part of the original gap, while the tested local pairing still performs better than the source-labeled unrestricted-attention control. The Global+Labels comparison additionally changes compression and is not a clean isolation of locality.

Standard deviations are not reported for Unbound+Labels and Global+Labels, so these results do not establish statistical significance. These experiments test the stated architectures and held-out compositions, not every way a model could exploit source labels. Sec. 5.3 points to the labeled comparison; Sec. 5.4 and Table 4 report the results and limitations, and App. A.6 specifies the label projection.

Reviewer_ofC3OpenReview ↗
Rating
2
Confidence
4
Informativeness
2
Presentation
1
Soundness
1
Read review
Paper summary

The authors present TERM, a model for solving PDEs which incorporates information from the governing equations to enrich the operator learning component. The approach has several key components. A graph encoder is used to turn the governing equations into a set of vectors, which remain frozen for all predictions in the rollout. A response evaluator uses a numerical approach, seemingly finite differences or FFT, to extract where each term of the governing equations has a strong effect on the dynamics. TERM has a "binding" operation to tie the governing equation vectors to the response vectors, creating a direct pathway to tie the symbolic and numerical information. The backbone of the model is a transformer, which processes the term vectors and other data tokens, learning the map from input to output functions. A Fourier-based decoder enables continuous querying in the output domain. The model is non-Markovian, taking 10 snapshots as inputs and uses a feedback loop, appending its own predictions to the series of solutions for longer rollouts. The authors train the model and evaluate it on several canonical PDE tasks, comparing with task-specific and foundation models.

Critical strengths
  1. Governing equations contain extremely useful information that have potential to improve the capabilities of operator learning models. It remains unclear amongst the community exactly how this information may be tied in. The idea to couple GE terms explicitly with the numerical phenomena they represent is, at least in my understanding, novel. The overarching goal of this work is also timely and worthwhile.
  2. One of the main questions in neural operator literature is how to enable zero-shot learning. TERM presents some initial results, showing how advection, diffusion, reaction, and dispersion can be learned pair-wise and evaluated zero-shot on a triplet of these phenomena, with better results in comparison to several ablated model setups.
Critical weaknesses
  1. The paper is very difficult to read. Many terms are unclear. The general approach is obfuscated by confusing language and a lack of a clear description. Even after several passes through the paper, I do not feel I completely understand how TERM works.
  2. I reviewed the ablation setups described in the appendix, but it remains unclear to me that what changes in each of the ablated models.
  3. The zero-shot experiments are extremely limited. First, it seems that it is only the vocabulary which is limited — nothing about the physics itself changes. Second, it looks at two models, each trained for their specific tasks. The much more interesting question, in my view, is whether one model can extrapolate and recompose dynamics it observes from training on many PDEs to an unseen governing equation. In the unseen-term KdV experiment, the action of the new term is computed numerically by the response evaluator rather than inferred by the model. As a result, the test seems to rely on having this term available. It is unclear how much of this transfer comes from learned components, as opposed to explicit numerical computation.
  4. Many of TERM’s results are marginally different from other models — ablations and other neural operators. It is difficult to determine whether there is any real advantage to this model, or whether it could be noise in training. No result reports variability across seeds, and gains likely lie within the range of seed-to-seed variability.
  5. The bound in section 4 rests on a condition that, as noted by the authors, is not guaranteed by the architecture. The analysis therefore comes across a fairly weak, and it is unclear what this adds.
  6. The experiments imply many full pretraining runs of the 105M model, but training cost is not reported for TERM or the baselines. This makes it difficult to judge the practical value of this model over much simpler alternatives.
Questions for authors
  1. Is there anything actually different about the physics being learned by the model in the zero-shot experiments?
  2. TERM uses a series of 10 snapshots when predicting future states. Is this consistent across all other models?
  3. How does term perform in the Markovian setting, where only 1 snapshot is given?
  4. What exactly is the pool operation in the response head? Is it a mean pool or max pool? This seems to be undefined. Can the authors provide a clearer explanation of the response evaluator and response head and how exactly they are employed for each problem?
  5. Do gradients flow through the evaluator?
  6. Are the numbers in the tables the result of a single training run or a mean of several seeds?
  7. Were all the baselines trained under the same exact splits and history length? If not, can you retrain the strongest ones under the same conditions?
  8. How does TERM perform when the evaluator is inaccurate or unavailable? For example on coarse grids or with noisy observations?
8 comments

W1: We thank the reviewer for the careful reading and for recognizing the potential of connecting governing-equation terms to their numerical effects. We agree that the original presentation introduced too many components before explaining the complete prediction loop. We have revised Sec. 3 to provide a clearer explanation.

TERM assumes that the target PDE, its coefficients, domain, and boundary conditions are specified. It receives a history Ht=(ut−9,…,ut)\mathcal{H}_t=(u_{t-9},\ldots,u_t) and predicts the next state at requested coordinates. For Burgers' equation ut=−u,ux+νuxxu_t=-u\\,u_x+\nu u_{xx}, we treat −u,ux-u\\,u_x and νuxx\nu u_{xx} as two signed additive terms, including their coefficients. On each of the four most recent states, the numerical evaluator computes the corresponding spatial fields ρadv=−u,ux\rho_{\mathrm{adv}}=-u\\,u_x and ρdiff=νuxx\rho_{\mathrm{diff}}=\nu u_{xx}. These instantaneous term responses indicate the contribution of each supplied term to the time derivative at the observed state, not the complete future solution. Keeping them separate can retain information about large contributions that cancel in their sum.

The PDE graph provides one symbolic embedding eme_m per term, describing its operator structure, coefficient and connections. The numerical fields for each term are normalized by their response-window RMS; the log RMS scale is retained so that normalization does not discard amplitude. A shared response encoder processes normalized fields together with coordinates, relative times and scale, and a learned resampler produces a fixed number of tokens RmR_m for that term. In term-local cross-attention, eme_m reads only RmR_m; a gated residual update produces the enriched term embedding. The shared Transformer then processes all enriched term embeddings together with trajectory summaries, and a coordinate-query decoder predicts the next field.

The newly predicted field replaces the oldest history frame. The evaluator then recomputes responses from the updated history, and the learned model produces the next prediction. The equation graph and symbolic embeddings are reused throughout that rollout; responses and enriched embeddings change with the predicted state. The evaluator supplies numerical features from the equation and visible/predicted states; it does not itself solve the future trajectory, and TERM is not required to learn the supplied differential operator from scratch. The main loss supervises the predicted rollout, with an auxiliary loss on a coarse response summary.

In the revised manuscript, Sec. 3 starts with the inputs, outputs, and a Burgers example. Sec. 3.2 distinguishes numerical response fields and encoded tokens; Sec. 3.4 defines the mean-pooling target, explains the detached evaluator, and gives rollout pseudocode. Appendix A.6 expands that pseudocode and separates term-local enrichment from the shared backbone.

W2: We agree that the original appendix did not make the controls easy to compare. Let SS denote history tokens, eme_m a symbolic term embedding, and RmR_m the encoded response of term mm. TERM updates each eme_m using only its own RmR_m and passes the enriched terms and SS to the backbone. The controlled variants retain the same trajectory encoder, backbone, decoder, data split, optimizer and schedule, while response modules and auxiliary targets change when the corresponding information is removed. Their retained information, modifications, and purposes are summarized in Table 8 of App. A.3.5 and outlined below.

  • Trajectory-only: keeps SS and removes the graph, symbolic tokens, graph bias, response branch and its auxiliary loss; this is the trajectory-only reference.
  • Static Graph: retains SS, symbolic terms and graph bias, but removes the response branch and uses static term embeddings.
  • Response-only: keeps SS and separately computed response features, but removes symbolic term tokens and graph bias from the predictor. The evaluator still uses the specified equation.
  • Global Response: retains the graph and history, and compresses the encoded response tokens into a global bank without explicit source attribution.
  • RHS-Matched: retains the graph and history, but sums signed response fields per state component before encoding, with a matched response-token budget. Global Response instead combines features after encoding.
  • Identity-Free: retains separate response sets, their scales, anonymous term slots and the stated token budget, while withholding symbolic identities and graph relations.
  • Shuffled Binding: retains the local update, but mismatches response–term pairings within each equation in both training and evaluation; the same permutation is used across a response window.
  • Unbound Response: retains symbolic terms and the individual responses, but each term can attend to all response tokens; the original control has no explicit source labels.

Both Response-only and Identity-Free compute separate responses using the specified equation and withhold explicit symbolic identity from the predictor; coefficients still affect response values and scales. Identity-Free explicitly retains anonymous per-term slots, scales and the specified token budget. These controls have overlapping comparison goals, and we do not interpret their results as independent evidence for two distinct mechanisms (App. A.3.5).

The results show that, on KdV, Trajectory-only, Static Graph, Response-only, Global Response, Shuffled Binding, Unbound and TERM obtain relative L2L_2 errors of 0.0169, 0.0142, 0.0135, 0.0119, 0.0124, 0.0114 and 0.0087. RHS-Matched gives 0.0127/0.1010 on KdV/AC versus 0.0087/0.0836 for TERM; Identity-Free gives 0.0122 on KdV. These comparisons support the usefulness of the tested information and organization, but cannot attribute each gap solely to cancellation, pairing, or a single graph feature.

The original Global and Unbound controls lack explicit source labels, so the original comparison changes both source correspondence and fusion. We also report source-labeled controls (Sec. 5.4, Table 4): strict 1D composition averages are 0.0271 for Unbound, 0.0241 for Unbound+Labels and 0.0217 for TERM. This is evidence for the tested configurations, not a universal claim about all fusion methods.

In addition, Frozen Response reuses initial responses during rollout and has higher long-horizon error than TERM. TERM-Direct omits evaluator stabilization; at 5% noise its macro-average error is 0.0597 versus 0.0419 for TERM. TERM-Grid Decoder uses a fixed-grid head and interpolation; at 128→256128\to256 its error is 0.0498 versus 0.0395 for TERM. These test response updating, numerical evaluation and decoding, respectively, and are described separately in App. A.3.5.

The revised Sec. 5.3 links to the appendix configuration table and clarifies the pooling and source-label differences; App. A.3.5 gives the individual constructions, acknowledges the Response-only/Identity-Free overlap, and distinguishes the three additional controls.

W3: The revised Sec. 5.4 and Table 3 clarify that, for each of five independent training seeds (0–4), the same TERM-M checkpoint jointly pretrained on ten source tasks is frozen and evaluated on both held-out KdV and Allen–Cahn. Neither target equation participates in pretraining; Table 3 reports five-seed means, with standard deviations in App. A.5. The target equation, coefficients, boundary information and observed history are provided at inference. Allen–Cahn tests a held-out full equation using known diffusion and reaction mechanisms.

The strict composition benchmark uses separate 1D and 2D training setups (Table 4 and App. A.3.6). With AA, DD, RR and SS denoting advection, diffusion, reaction and dispersion, the 1D model trains on all six pairs and tests on A+D+RA+D+R, A+D+SA+D+S, A+R+SA+R+S and D+R+SD+R+S; the 2D model trains on A+DA+D, A+RA+R and D+RD+R and tests on A+D+RA+D+R. Every term and pair is seen, but each complete test equation is held out, as checked by operator-tree signatures. Coefficient ranges, initial-condition family and solver protocol are shared within each split. Adding R[u]R[u] to A[u]+D[u]A[u]+D[u] changes the complete evolution operator and generally its solution, while reusing known mechanisms. TERM versus Unbound has mean relative L2L_2 0.0217 versus 0.0271 in 1D and 0.0360 versus 0.0430 in 2D. These results concern the tested compositions, not arbitrary unknown physics.

The full KdV term uxxxu_{xxx} is absent from pretraining, but the evaluator computes its action from the supplied equation and observed or predicted state. TERM combines this input with the equation and history to predict future states; responses are recomputed during rollout without future ground truth. The network is not claimed to discover the operator or learn its numerical differentiation from trajectories alone.

The variants below use the same term-response computation, but organize responses differently (relative L2L_2, Table 2):

MethodKdVAllen–Cahn
Global Response0.01190.0964
Unbound Response0.01140.0923
TERM0.00870.0836

TERM's error is 23.7% and 9.4% lower than Unbound on KdV and Allen–Cahn, respectively. These comparisons support an effect of response organization, not numerical evaluation alone. However, original Unbound lacks explicit source labels, so this comparison cannot isolate pairing from source information. The source-labeled controls in Sec. 5.4, Table 4 address part of this confound; the evaluator remains a prerequisite.

Weakness 4 — Seed variability and marginal differences

W4: We agree that reporting a single number makes small differences difficult to interpret. We have added a reference to Table 20 in App. A.5 in Sec. 5.2 (Main Results), immediately after the TERM-200 discussion and before Table 1. Table 20 supplements Table 1 with the mean and standard deviation of relative L2L_2 error over five independent training seeds (0–4) for TERM-M and TERM-200 on all ten benchmarks; TERM-200 uses 200 epochs of task adaptation. For example, TERM-M versus TERM-200 gives 0.0306±0.00170.0306 \pm 0.0017 versus 0.0239±0.00130.0239 \pm 0.0013 on FNO-ν=10−5\nu=10^{-5}, 0.0086±0.000250.0086 \pm 0.00025 versus 0.0033±0.000230.0033 \pm 0.00023 on Burgers, and 0.0798±0.00740.0798 \pm 0.0074 versus 0.0218±0.00180.0218 \pm 0.0018 on NS. App. A.5 also reports five-seed statistics for the controlled ablations. In Sec. 5.6, Fig. 2 now shows five-seed means and standard-deviation error bars for the rollout study. For the 1D composition average, the four equation errors are first averaged within each seed, then the five per-seed averages are used to compute the sample standard deviation. This is not the arithmetic average of the four standard deviations.

Examples (mean ±\pm SD, relative L2L_2):

TaskStatic GraphUnbound ResponseTERM
KdV0.0142±0.000770.0142 \pm 0.000770.0114±0.000580.0114 \pm 0.000580.0087±0.000600.0087 \pm 0.00060
Allen–Cahn0.1115±0.00530.1115 \pm 0.00530.0923±0.00330.0923 \pm 0.00330.0836±0.00650.0836 \pm 0.0065
Burgers0.0101±0.000490.0101 \pm 0.000490.0091±0.000810.0091 \pm 0.000810.0086±0.000250.0086 \pm 0.00025
NS0.0876±0.00280.0876 \pm 0.00280.0821±0.00350.0821 \pm 0.00350.0798±0.00740.0798 \pm 0.0074

For the strict 1D composition average, Unbound gives 0.0271±0.000480.0271 \pm 0.00048 and TERM 0.0217±0.000680.0217 \pm 0.00068. The NS difference between TERM and Unbound is small relative to their reported variability; the summaries alone do not establish statistical significance or a paired-seed confidence interval. We therefore avoid claiming that every numerical advantage is statistically resolved. External baseline values are taken from published reports at their supplied precision; we do not infer standard deviations for those rows.

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

Related papers

Loading the map…

Discussion (0)

Sign in to comment.