acceptodds
Under review as a conference paper at ICLR 2027

Localize, Edit, and Learn: Structure-aware Symbolic Regression for Network Dynamics

Abstract

Describing network dynamics through the discovery of symbolic equations is critical for understanding complex systems. However, as a network's dynamics are jointly governed by its nodes' local processes and mutual interactions, where both encapsulate diverse symbolic components, existing symbolic regression methods struggle to precisely localize erroneous components and translate diagnostic evidence into controlled structural edits. Furthermore, despite the rise of large language models (LLMs) in evolutionary symbolic regression for network dynamics, a dedicated policy update paradigm is needed to let LLMs learn from historical edits and internalize such capability, instead of merely reusing experience through prompts. We introduce Network Symbolic Patch Evolution (NetSPE), a framework that connects localization, editing, and policy learning for equation generation. Guided by network structural priors, NetSPE uses residual contributions to localize a target component to be updated, and then instructs an LLM to generate constrained local replacements. We further propose Structural Patch Preference Optimization (SPPO), which adapts the editing policy using preferences among alternative edits generated under the same editing state and evaluated against their parent. Experiments on two real-world and one synthetic benchmarks show NetSPE's efficacy in its prediction accuracy, equation quality, and generalizability.

Primary area:
applications to physical sciences (physics, chemistry, biology, etc.)
Keywords:
Symbolic Regression, Network Dynamics, Equation Discovery

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_N25KOpenReview ↗
Rating
2
Confidence
3
Informativeness
2
Presentation
2
Soundness
2
Read review
Paper summary

This paper introduces NetSPE, a method for discovering dynamical equations from network observation data. It leverages prior knowledge of the network structure to organize mathematical expressions, employs residual analysis to identify branches requiring modification, and determines specific editing locations through subtree pruning and coefficient refitting. An LLM generates local replacements at these identified locations, while an SPPO (Structured Policy Optimization) mechanism updates the generation strategy based on the performance of different replacements under the same editing state. Finally, the optimal formula is selected via cross-validation. The approach is evaluated across 25 tasks spanning three benchmarks: synthetic network dynamics, bacterial growth, and epidemic spread.

Critical strengths
  1. The research problem is clearly defined, and the methodology is well-aligned with the research objectives. The paper addresses how to translate aggregate prediction errors into specific structural modifications, enabling the model to accumulate experience from these adjustments. The process—comprising "branch localization, subtree localization, local editing, and strategy update"—follows a clear logical progression, while type constraints and fixed aggregation structures establish well-defined boundaries for editing operations. As a framework for formula discovery in the context of network dynamics, this design holds significant research value.
  2. The experiments demonstrate meaningful system-level gains and include analysis at the structural level. Using two different LLM backbones, the authors report results from primary experiments, modular ablation studies, structural recovery tasks, and an analysis of search trajectories. Compared to the baseline method ND², the Llama-based version achieved a reduction in average Normalized Mean Squared Error (NMSE)—from 0.0437, 1.0672, and 0.3819 to 0.0084, 0.8256, and 0.2395, respectively—across synthetic, bacterial, and epidemic-related tasks. Furthermore, the authors present examples of both successful and unsuccessful structural recovery, allowing readers to examine the nuances between model fitting and the underlying structural representation.
Critical weaknesses
  1. The paper defines structural accuracy by allowing numerical constants in a returned expression to be reassigned before checking algebraic equivalence. This may help separate operator-structure recovery from coefficient-estimation error, but the manuscript does not clearly explain why this criterion is the most appropriate measure for the claimed scientific objective. The authors should clarify what capability this metric is intended to capture, how it differs from exact recovery of the fitted equation, and how the results should be interpreted when structural accuracy and predictive performance disagree.

  2. The contribution of the network-aware representation is not sufficiently isolated from the imposed structural constraints. NetSPE has access to richer network information, including neighboring states, edge attributes, and aggregation operations, while the general LLM baselines primarily operate on local scalar features. Moreover, the proposed method imposes a fixed network-aware expression structure. Consequently, the reported performance gains may partly arise from additional information and a more constrained hypothesis space rather than the proposed search and editing mechanism itself. A fair comparison using matched network information and comparable expression spaces is needed to establish the independent contribution of NetSPE.

Questions for authors
  1. Can the five general-purpose LLM baselines access neighbor states, edge attributes, and aggregation operators? Are there results under a shared graph-aware execution environment?

  2. Since Cs−Cz=Vs−VzC_s-C_z=V_s-V_z, what additional benefit does the joint term provide in branch routing? Have you compared it with marginal-score routing under the same history and exploration settings?

  3. How is neutral subtree replacement defined for different operator contexts? How are invalid or non-finite ablations scored?

  4. Does SPPO remain effective after controlling for patch validity and duplication? How does it compare with DPO, GRPO, or supervised learning on the same patch data?

  5. How many candidates and successful edits come from LLM patches, atomic edits, and deterministic templates? Are ablation refitting and final cross-validation included in the reported budget?

  6. Can you report how many returned equations are actually equivalent to the ground truth using their fitted coefficients, rather than becoming equivalent only after reassigning constants? Given NetSPE’s 17/3017/30 and 15/3015/30 recoveries versus ND²’s 16/3016/30, what structural-recovery advantage is supported?

I am open to reconsidering my rating if the rebuttal convincingly addresses the concerns above.

5 comments

We sincerely appreciate your time and thoughtful review of our work. The following is our response to your remaining concerns.

Response to Weakness 1 and Question 6

Our SA metric evaluates whether the returned expression can exactly represent the ground-truth equation through finite reassignment of its numerical constants, thereby assessing structural search capability separately from coefficient-estimation error. It does not imply that the fitted coefficients are already correct. Since coefficients of redundant terms may be set to zero, it also does not guarantee a minimal expression.

We conducted an additional verification of the original 30 runs for each method:

MethodOriginal SANumerical Agreement ≤ 10⁻⁶ (Fitted Coefficients)
NetSPE–Llama17 / 3017 / 30
NetSPE–Qwen15 / 3015 / 30
ND²16 / 3013 / 30

The results show that all NetSPE instances counted as successes under the original SA metric also pass the numerical check above with their actual fitted coefficients. Therefore, these successes do not rely solely on constant reassignment during evaluation.

Response to Weakness 2 and Question 1

In the original experiments, the five LLM-based baselines primarily received local scalar features and did not share NetSPE’s full graph-based execution environment. Thus, those comparisons alone do not isolate the effects of network information, structural constraints, and the proposed search mechanism.

To address this concern, we added comparisons with three strong baselines FunctionEvolve, IGSR, and RESTART augmented with neighbor information and structural priors (denoted “w/ NP”). While preserving each baseline’s original search procedure, we provided information about all neighbor nodes and incorporated NetSPE’s GLV and epidemic readout priors into FunctionEvolve’s generation and evolution contexts, IGSR’s expression-generation prompts, and RESTART’s candidate proposal, refinement, and analysis prompts.

We report task-averaged NMSE and ACC_10 for each benchmark as NMSE / ACC_10. Llama-3.1-8B

MethodSyntheticBacteriaEpidemic
FunctionEvolve w/ NP0.0855 / 24.045.1096 / 22.320.7053 / 24.64
IGSR w/ NP0.2771 / 13.251.0007 / 1.690.3372 / 15.68
RESTART w/ NP0.2345 / 15.681.1090 / 2.600.4164 / 10.64
NetSPE0.0084 / 83.690.8256 / 29.240.2395 / 25.56

Qwen3-4B

MethodSyntheticBacteriaEpidemic
FunctionEvolve w/ NP0.0828 / 24.355.2872 / 19.470.3424 / 23.34
IGSR w/ NP0.2793 / 13.071.0245 / 0.810.3506 / 13.06
RESTART w/ NP0.3459 / 11.081.0293 / 8.820.5406 / 8.29
NetSPE0.0092 / 80.370.8759 / 28.920.2269 / 24.28

In addition to these three augmented baselines, we included SymBOL (TPAMI) as an existing graph-based and LLM-based baseline, using Llama-3.1-8B as its backbone. This provides a complementary comparison with a method explicitly designed to incorporate graph structure. This additional comparison demonstrates that, under the current evaluation setting, NetSPE retains a performance advantage over a method that also incorporates LLMs and network interaction modeling.

MethodSyntheticBacteriaEpidemic
SymBOL0.0494 / 55.791.2375 / 1.510.3625 / 13.88
NetSPE0.0084 / 83.690.8256 / 29.240.2395 / 25.56

Response to Question 2

The joint contribution decomposition accounts for each channel’s standalone explanatory power and its incremental explanatory power given the other channel, while ensuring that Cs+Cz=VszC_s+C_z=V_{sz}. Directly normalizing the marginal scores depends only on Vs,VzV_s,V_z and therefore cannot distinguish cases with identical marginal explanatory power but different joint explanatory power.

For example, when Vs=0.2V_s=0.2, Vz=0.1V_z=0.1, and Vsz=0.8V_{sz}=0.8, directly normalizing the marginal scores gives (0.667,0.333)(0.667, 0.333), whereas normalizing the residual contributions gives (0.5625,0.4375)(0.5625, 0.4375). The latter reflects the additional explanatory information provided jointly by the two channels, allocating this information symmetrically and giving the branches more balanced editing opportunities. Direct marginal-score normalization depends only on VsV_s and VzV_z and therefore cannot reflect differences in joint explanatory value when these marginal scores remain unchanged.

As described in Appendix D.2 and Algorithm 1, normalized residual contributions are combined with historical editing utility and uniform exploration to determine branch-selection probabilities. The probabilities above illustrate the residual-contribution component before this combination. Therefore, although the joint term does not affect the initial selection based on the largest contribution, it can affect subsequent branch-selection probabilities. We will clarify this distinction in the manuscript, the joint term incorporates joint explanatory information into the allocation of editing opportunities, rather than changing the branch ranking.

Response to Question 3

The same replacement rule applies across parent operators. For example, setting uu to zero in a+ua+u yields aa, whereas doing so in a×ua\times u yields zero. Therefore, this score measures the impact on fitting performance of disabling the subtree’s output while allowing the remaining coefficients to re-adjust. It does not measure the subtree’s independent contribution while holding the outputs of other operations unchanged.

For ablations that fail structural validation or coefficient fitting, including those rejected because of non-finite predictions, we mark their contribution as unavailable and exclude them from the current target selection. We do not assign them a zero or arbitrarily large contribution. This distinguishes an unsuccessful ablation evaluation from evidence about subtree importance, where a numerical failure alone does not justify prioritizing a subtree for editing. Target selection therefore uses the available contribution scores from successfully evaluated ablations.

We will clarify the zero-valued replacement rules, the interpretation of the contribution score, and the handling of failed ablations in the revised manuscript.

Response to Question 4

To distinguish the contribution of policy learning from that of candidate sampling and selection, we added three ablation variants under matched total sampling and evaluation budgets: Frozen policy, Frozen+best-of-G, and Frozen+random. Frozen policy generates one candidate at a time without updating the generation policy. Frozen+best-of-G generates G candidates per group and selects the best, whereas Frozen+random randomly selects one candidate from each such group.

In particular, Frozen+best-of-G matches SPPO in group size, candidate evaluation, and selection rules, while keeping the generation policy fixed. As shown below, SPPO achieves better overall search performance, supporting an additional benefit from policy learning beyond multi-candidate sampling and selection. (NMSE / ACC_10)

BenchmarkFrozen+RandomFrozen policyFrozen+best-of-GSPPO
Synthetic0.0631 / 60.980.0381 / 68.470.0607 / 67.420.0084 / 83.68
Bacteria0.8509 / 27.800.8482 / 28.090.8468 / 28.390.8256 / 29.24
Epidemic0.2679 / 21.950.2683 / 25.760.2683 / 24.010.2395 / 25.56

We additionally compared SPPO with DPO and GRPO as alternative policy-update methods within NetSPE. (NMSE / ACC_10)

DatasetNetSPE (SPPO)DPOGRPO
Synthetic0.008411 / 0.83690.007128 / 0.74340.024217 / 0.7519
Bacteria0.825636 / 0.29240.898579 / 0.29230.878835 / 0.2897
Epidemic0.239465 / 0.25560.276176 / 0.23850.276634 / 0.2349

SPPO achieves higher ACC_10 on all three benchmarks and lower NMSE on Bacteria and Epidemic than both alternatives. These results support the effectiveness of SPPO for patch-policy learning in our search framework.

Regarding SFT, our motivation is to learn from the relative utility of candidate patches generated for the same search state. Standard SFT on successful patches instead learns to imitate those patches without explicitly contrasting their relative utility. This motivates our preference-based formulation.

Response to Question 5

We additionally counted the valid, deduplicated candidates and successful edits produced by each source across 25 tasks spanning the three benchmarks.

ModelLLM patchesAtomic editsDeterministic templatesLLM share of successful edits
Llama9,835 / 1,4342,039 / 65412,124 / 38158.1%
Qwen4,551 / 1,1341,916 / 59311,617 / 36954.1%

Each source column reports valid, deduplicated candidates / successful edits. The last column reports the proportion of all successful edits contributed by LLM patches.

All three sources produced improvements. LLM patches contributed more than half of the successful edits with both models. Atomic edits generated fewer candidates but achieved a higher rate of local improvement, while deterministic templates provided additional opportunities for improvement.

Regarding computational budgets, we distinguish proposal counts from runtime. The proposal budget counts attempts from LLM patches and atomic edits, including failed and duplicate attempts. Deterministic templates, subtree ablations and their associated coefficient refitting, final cross-validation, and final coefficient refitting are excluded from this count.

Thank you for the detailed response and the additional experiments. The rebuttal meaningfully addresses most of my concerns, particularly through the graph-aware baseline comparisons, the frozen-policy controls, and the clarification of the implementation and evaluation procedures.

However, my main concern about the evidence for equation discovery remains. For a paper aimed at discovering governing equations, I place substantial weight on recovering the underlying symbolic relationships. The proposed SA criterion establishes that a returned expression can represent the ground-truth equation after reassigning its constants, potentially including setting coefficients of redundant terms to zero. For example, the same returned expression f^(x)=ax+bx2+csin⁡x\hat f(x)=ax+bx^2+c\sin x can be counted as structurally correct for three different ground-truth laws—xx, x2x^2, and sin⁡x\sin x—simply by reassigning its coefficients. This remains possible even when the actual fitted expression retains spurious terms. Consequently, SA can award full structural credit to an expression containing several alternative relationships without establishing that the method has identified which relationship actually governs the data. The additional check using fitted coefficients is helpful, yet numerical agreement at a specified tolerance is still insufficient to establish symbolic recovery. Flexible approximations, including polynomial expansions, can achieve very high numerical accuracy without identifying the underlying functional form.

I will maintain my original rating of 2 (weak rejection). If the other reviewers favor acceptance and the AC considers the existing evaluation sufficient for the paper’s intended scope and contributions, I would not oppose acceptance.

Reviewer_pgeUOpenReview ↗
Rating
3
Confidence
2
Informativeness
2
Presentation
1
Soundness
2
Read review
Paper summary

This paper introduces NetSPE, a framework for symbolic regression (SR) on network dynamics that combines structural priors, large language model (LLM)–based local editing, and preference optimization. Unlike prior LLM-based SR methods that generate entire equations in one shot or rely on generic tree mutations, NetSPE explicitly exploits the two-branch structure of network systems: a self-dynamics branch and a neighbor-interaction branch. The framework operates in an evolutionary loop: (1) a hierarchical localization step uses residual diagnostics and subtree ablation to identify which branch and which abstract syntax tree (AST) subtree within it should be edited; (2) an LLM is prompted with the current equation, diagnostic context, and editing intent to propose a constrained local replacement (a “patch”); (3) Structural Patch Preference Optimization (SPPO) updates the LLM’s policy by comparing alternative patches generated under the same editing state, favoring those that improve over the parent equation. Experiments on one synthetic and two real-world network benchmarks show that NetSPE outperforms existing LLM-based and non-LLM SR methods in prediction accuracy and, on synthetic data, structural recovery. The authors also provide ablation studies and case analyses to justify each component.

Critical strengths
  1. Clear problem-solution fit.​ The paper correctly identifies that generic LLM-based SR struggles with network dynamics because it ignores the relational structure (self vs. interaction). NetSPE’s two-branch prior and typed AST edits directly address this, making the search space more structured and interpretable.

  2. Novel integration of localization, editing, and learning.​ While individual components (structural priors, LLM editing, preference tuning) exist in isolation, NetSPE is the first to close the loop: diagnostics guide where to edit, the LLM proposes local patches, and SPPO internalizes effective edits. This is a coherent pipeline that moves beyond simple prompt-based memory.

  3. SPPO design is well‑motivated.​ Grouping patches by the same editing state and using pairwise preferences relative to the rollout policy is a sensible adaptation of GRPO/DPO to equation evolution. Length normalization and KL regularization further improve stability.

  4. Comprehensive experimental evaluation.​ The paper tests on both synthetic (with ground truth) and real-world datasets, uses two different LLM backbones (Llama‑3.1‑8B, Qwen3‑4B), and includes ablations for each proposed component (hierarchical localization, SPPO, final selection). Case studies illustrate the search trajectory qualitatively.

  5. Reproducibility.​ Code is anonymously released, which is commendable.

Critical weaknesses
  1. Writing and clarity significantly hinder readability.​ Section 4 introduces formalism (editing state ξ, SPPO objectives) without a concrete, step‑by‑step example. Key concepts like the four edit intents (grow, replace, simplify, rewrite) are never illustrated. Figure 1 is referenced but its caption does not compensate for the lack of narrative. As a result, the core contributions—especially how localization feeds into editing—are difficult to grasp on first reading. This is not merely a presentation issue; it may cause reviewers to undervalue the method.

  2. Insufficient evidence that SPPO truly learns.​ The paper reports final metrics but omits critical diagnostics: (a) What percentage of editing states yield informative groups (both positive and negative patches)? If this fraction is low, SPPO rarely updates. (b) How does performance change if SPPO is replaced with a simple selector that picks the best‑of‑G patches without policy update? (c) Is there a measurable shift in the LLM’s output distribution before/after training? Without this, SPPO could be perceived as a sophisticated sampling filter rather than a learning mechanism.

  3. Experimental protocol has gaps.​ (a) Only 3 independent runs per benchmark are reported; given LLM stochasticity, this is borderline low for statistical significance. (b) Real‑world benchmarks (Bacteria Growth, Epidemic Transmission) lack structural accuracy (SA) scores or even qualitative equation comparisons; only NMSE and ACC0.1 are given. Since the paper claims “structure‑aware” discovery, some measure of equation quality beyond prediction error is needed. (c) Computational cost (GPU hours, wall‑clock time per problem, ablation overhead) is not reported, making it hard to assess practicality. (d) The baseline comparison in Table 1 mixes non‑LLM methods (ND², LLC) with LLM methods, but LLM baselines are not augmented with the same two‑branch prior. An ablation in Appendix F.2 suggests this matters, yet the main table does not reflect it, potentially overstating NetSPE’s gains.

  4. Limited theoretical novelty.​ The individual pieces—network SR priors, LLM‑based AST editing, preference optimization—are each drawn from recent literature. The main contribution is their integration into a domain‑specific loop. While valuable, this may be viewed as incremental by reviewers expecting a more fundamental advance.

  5. Potential backbone dependence not fully explored.​ Although NetSPE works with both Llama and Qwen, the paper does not test smaller or larger models, nor does it analyze how backbone capability interacts with localization quality. This leaves open whether the method generalizes beyond the chosen 4–8B models.

Questions for authors
  1. Can you report (i) the proportion of editing states that produce informative groups, (ii) the average number of positive/negative patches per group, and (iii) a control experiment where SPPO is replaced by random selection among the G sampled patches? This would clarify whether the policy update is essential or merely a side effect of diverse sampling.

  2. Please provide total GPU hours, approximate time per iteration, and the overhead of subtree ablation (which requires refitting coefficients for each candidate). This is crucial for assessing scalability, especially as equation complexity grows.

  3. For Bacteria Growth and Epidemic Transmission, could you include (a) representative discovered equations (similar to Table 2) and (b) a human‑evaluation or approximate structural similarity score? Without this, it is unclear whether NetSPE finds meaningfully interpretable dynamics or just fits well.

  4. In the main table, would it be possible to add a version of the strongest LLM baseline (e.g., FunctionEvolve or RESTART) augmented with the same self/interaction branch prior used by NetSPE? This would isolate the benefit of the localization+SPPO loop from the benefit of structural constraints.

  5. Consider adding a concrete, worked example (e.g., one step of the FHN case from initialization to final edit) directly in Section 4 before the formalism. This would greatly lower the entry barrier for readers and highlight the intuition behind hierarchical localization and typed patches.

  6. Given the stochasticity of LLM sampling, could you increase to at least 5 runs for the main results, or provide confidence intervals beyond standard deviation? This would strengthen the statistical claims.

  7. The four edit intents (grow, replace, simplify, rewrite) are mentioned but never analyzed. Which intent contributes most? Is there a schedule or adaptive selection? A brief analysis would enrich the methodology.

3 comments

Thank you for your time and valuable feedback. Below, we provide our responses to the remaining concerns you have raised.

Response to Weakness 1, Question 5 and Question 7

In each editing step, NetSPE first identifies which branch to modify the local dynamics branch or the neighbor interaction branch and then selects a target subtree within that branch through subtree ablation and coefficient refitting. The edit intent specifies the direction of the local modification, and the LLM generates multiple candidate patches conditioned on the parent equation, editing location, and structural constraints. Each patch is applied to the parent equation, and the resulting equation undergoes coefficient fitting before being evaluated relative to the parent. Candidates generated under the same editing state are then used to construct preference feedback for updating the editing policy. Therefore, localization determines where to edit, the edit intent guides how to edit, and SPPO uses feedback from candidate edits to improve subsequent patch generation.

To make this process more concrete, we will add a worked example of a complete editing step to the case study section. The example will walk through the parent equation, branch and subtree selection, edit intent, candidate patches, and the resulting equation and evaluation results. We will also clarify the definitions of the four edit intents and provide corresponding examples of local expression changes. Finally, we will improve the caption and accompanying description of Figure 1 to make the correspondence between the illustrated stages and the concrete operations clearer.

Response to Weakness 2 and Question 1

We provide additional candidate-group statistics and ablation experiments with three SPPO variants to examine the availability of preference signals and the contribution of policy updates, respectively.

First, we analyze candidate-group composition using the Llama backbone. Across the three benchmarks, each informative group contains approximately three candidates contributing positive or negative preference signals on average.

BenchmarkInformative Group RatioAvg. Positive Patches per Informative GroupAvg. Negative Patches per Informative Group
Synthetic20.11%1.281.79
Bacteria8.30%1.221.72
Epidemic8.40%1.221.95

Second, to distinguish the contribution of policy learning from that of candidate sampling and selection, we added three ablation variants under matched total sampling and evaluation budgets: Frozen policy, Frozen+best-of-G, and Frozen+random. Frozen policy generates one candidate at a time without updating the generation policy. Frozen+best-of-G generates G candidates per group and selects the best, whereas Frozen+random randomly selects one candidate from each such group. SPPO achieves better overall search performance, supporting an additional benefit from policy learning beyond multi-candidate sampling and selection. (NMSE / ACC_10)

BenchmarkFrozen+RandomFrozen policyFrozen+best-of-GSPPO
Synthetic0.0631 / 60.980.0381 / 68.470.0607 / 67.420.0084 / 83.68
Bacteria0.8509 / 27.800.8482 / 28.090.8468 / 28.390.8256 / 29.24
Epidemic0.2679 / 21.950.2683 / 25.760.2683 / 24.010.2395 / 25.56

Response to Weakness 3(a) and Question 6

Our evaluation covers three benchmarks comprising 25 subtasks. Each NetSPE run involves candidate generation, equation evaluation, coefficient fitting, and online policy updates, making repeated full experiments computationally expensive. To balance evaluation coverage and computational resources, we conduct three independent runs per subtask. Tables 5 and 6 in the appendix of the original manuscript already report the mean and standard deviation over these three runs for each subtask, showing task-specific performance and variability across runs. We will make these results more explicitly referenced in the main text so that readers can assess stability beyond the aggregate results.

Response to Weakness 3(b) and Question 3

Since the ground-truth governing equations are unknown for Bacteria Growth and Epidemic Transmission, we cannot directly apply the structural accuracy (SA) metric used for the synthetic tasks. To provide evidence beyond predictive metrics, we will include the discovered equations for both datasets in the revised manuscript, following the format of Tables 9 and 10, together with brief interpretations of the self-dynamics and neighbor-interaction terms in representative equations.

Response to Weakness 3(c) and Question 2

We will report the total GPU hours, average time per iteration, and additional overhead of subtree ablation and the associated coefficient refitting in the revised manuscript.

Response to Weakness 3(d) and Question 4

We added comparisons with three representative strong LLM baselines: FunctionEvolve (Xia et al., 2026b), IGSR (Saveliev et al., 2026), and RESTART (Li & Pan, 2026). While preserving each baseline’s original design, we provided all three with information about all neighboring nodes so that they could use the corresponding network context. We also aligned the network structural priors with NetSPE as closely as possible. Specifically, we incorporated the GLV and epidemic readout structural priors used by NetSPE into FunctionEvolve’s generation and evolution contexts, IGSR’s expression-generation prompts, and RESTART’s candidate proposal, refinement, and analysis prompts. We report the avg NMSE and ACC_10 across tasks as (NMSE/ACC_10), w/ NP means the baseline with neighbor information and structural priors.

Llama-3.1-8B

MethodSyntheticBacteriaEpidemic
FunctionEvolve w/ NP0.0855 / 24.045.1096 / 22.320.7053 / 24.64
IGSR w/ NP0.2771/13.251.0007/1.690.3372 / 15.68
RESTART w/ NP0.2345 / 15.681.1090 / 2.600.4164 / 10.64
NetSPE0.0084 / 83.690.8256 / 29.240.2395 / 25.56

Qwen3-4B

MethodSyntheticBacteriaEpidemic
FunctionEvolve w/ NP0.0828 / 24.355.2872 / 19.470.3424 / 23.34
IGSR w/ NP0.2793 / 13.071.0245 / 0.810.3506 / 13.06
RESTART w/ NP0.3459 / 11.081.0293 / 8.820.5406 / 8.29
NetSPE0.0092 / 80.370.8759 / 28.920.2269 / 24.28

In addition to these three augmented baselines, we included SymBOL as an existing graph-based and LLM-based baseline, using Llama-3.1-8B as its backbone. This provides a complementary comparison with a method explicitly designed to incorporate graph structure. This additional comparison demonstrates that, under the current evaluation setting, NetSPE retains a performance advantage over a method that also incorporates LLMs and network interaction modeling.

MethodSyntheticBacteriaEpidemic
SymBOL0.0494 / 55.791.2375 / 1.510.3625 / 13.88
NetSPE0.0084 / 83.690.8256 / 29.240.2395 / 25.56

Response to Weakness 4

Our contribution is not a new general-purpose preference optimization objective, but a method for addressing a specific challenge in discovering network dynamics equations: how to translate equation-level prediction errors into actionable local structural edits and learn from their outcomes.

Addressing this challenge requires three connected designs. First, hierarchical localization combines branch diagnostics with subtree ablation to translate overall error feedback into a specific editing location. Second, type-constrained local patches restrict modifications to the selected location, allowing local structural changes while preserving the rest of the equation and its network aggregation constraints. Finally, SPPO compares candidate patches under the same parent equation, editing location, and edit intent. This makes preference feedback reflect candidate quality within a shared editing context, reducing confounding from differences across editing contexts.

The methodological contribution therefore lies in establishing an explicit connection between error diagnosis, local structural editing, and editing-policy learning. Structural priors specify which equations can be represented, but do not themselves determine where to edit, how to propose local modifications, or how to use editing feedback to improve subsequent search.

Response to Weakness 5

We accounted for potential backbone dependence in our experimental design by evaluating NetSPE with Llama-3.1-8B and Qwen3-4B, which differ in both model family and parameter count (in Tab 1). NetSPE achieves competitive results with both backbones, providing empirical evidence of its applicability across different backbone settings. Evaluating two backbones is also consistent with the experimental setups of prior LLM-based symbolic regression methods, such as LLM-SR and Deliberate Evolution, with the latter using the same two models as our work.

Furthermore, we compare NetSPE with other LLM-based symbolic regression methods under the same backbone settings. Controlling for backbone choice allows the comparison to more directly reflect the contributions of different equation search and editing mechanisms. Thus, our experiments both assess NetSPE’s performance across backbones and provide support for the effectiveness of the method itself.

I thank the authors for their rebuttal, which has addressed most of my concerns. However, given my limited expertise in this area, I do not feel sufficiently confident to make a further judgment. Therefore, I will maintain my original score.

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

This paper proposes Network Symbolic Patch Evolution (NetSPE), a symbolic regression framework for network dynamics. It represents governing equations with task-specific structural priors and typed expression-tree slots. A hierarchical localization mechanism first uses residual-related signals to select an editable branch, then measures subtree importance through removal, coefficient refitting, and validation error. An LLM produces constrained structural patches, supplemented by atomic actions and deterministic motifs. Structural Patch Preference Optimization (SPPO) then updates the patch-generation policy using positive and negative sibling edits sampled under the same editing state. A final grouped cross-validation stage chooses the output equation. The authors evaluate NetSPE on synthetic network dynamics and two real-world datasets, Bacteria Growth and Epidemic Transmission, using predictive errors and synthetic structural recovery metrics, together with ablations, out-of-distribution evaluations, and synthetic equation-recovery case studies. The results are promising, especially on synthetic tasks, but several attribution, robustness, and interpretability questions remain.

Critical strengths

This paper proposes a network symbolic equation editing scheme. The work integrates network-aware equation representation, local abstract syntax tree editing, and feedback-based learning into a coherent search pipeline. Restricting edit operations to structurally valid subtrees is the method for preserving relational substructures while optimizing candidate equations.

Critical weaknesses

1.The method has limited generality. NetSPE predefines the overall expression structure of network dynamics. However, for higher-order network interactions, nonlinear aggregation functions, and other more general and broader scenarios, it seems unable to handle them, and the paper does not fully discuss experiments or performance in these scenarios. By contrast, the compared method ND^2 remains effective in more general scenarios and is discussed thoroughly.

2.The entire NetSPE framework (neighborhood aggregation, interaction branches, and hierarchical residual localization) is built on the premise that the input network topology is completely accurate. However, in many real-world network dynamics scenarios, the network topology itself is inferred from data and may contain topological noise such as false-positive edges and missing edges; in some scenarios, the topology may even be completely unknown. All experiments in this paper use fully accurate topology for inference and do not evaluate the impact of topological perturbations. When the topology contains errors, aggregating neighbor features is inherently biased, which contaminates residual diagnostics and subtree ablation contributions and directly causes hierarchical edit localization to fail. The method design appears not to consider these key issues.

3.The role of SPPO is unclear: whether the performance improvement actually stems from the multiple candidate solutions generated each time or from SPPO itself. In addition, SPPO's constraints are very strong: guidance can be activated only when the parent equation, branch, target subtree, and edit intent are all the same.

4.The performance improvements on Bacteria Growth and Epidemic Transmission are limited, and in real-world dataset scenarios it is impossible to evaluate the structural interpretability of the discovered equations. It cannot be determined whether the method merely achieves good numerical fitting to the data or actually discovers the underlying mechanisms of the real-world scenarios.

5.Baselines are insufficient. In the related work, the paper emphasizes that "Despite these advances, most existing LLM-based SR methods target tabular data or low-dimensional dynamical systems without explicitly accounting for the node–edge relational structure of complex network dynamics in their symbolic representations," but there are existing related LLM-based works that can serve as general methods for discovery in both low-dimensional and high-dimensional network scenarios, such as SymBOL (TPAMI). Baseline coverage in low-dimensional symbolic regression scenarios is also insufficient, e.g., LASR (NIPS) and Sr-scientist (ICLR). In addition, some non-LLM classical symbolic regression methods, such as advanced baselines based on genetic programming, Transformer, and Monte Carlo search, are not covered.

Questions for authors

1.Could the proposed method be experimentally evaluated and described in detail in more general and broadly applicable scenarios, such as higher-order network interactions and nonlinear aggregation functions, with reference to ND^2?

2.Could the authors explain in detail and verify the impact of network topology on the performance of the proposed method, including its performance under unknown network topology, topological noise, and state noise?

3.Further analysis is needed regarding the role of SPPO.

4.Could the authors further explain, from the perspective of equation mechanisms, the performance of the discovered real-world equations rather than merely their numerical performance?

5.Could more baseline methods be added?

3 comments

Thank you so much for your time and valuable feedback. Your comments have greatly helped us improve our work.

Response to Weakness 1 and Question 1

NetSPE’s structural priors do not restrict all tasks to a fixed additive output form. Specifically, the epidemic task uses the editable readout in Appendix B.3, Eq. (12), to combine local node states with aggregated neighbor information. Therefore, fixing the base aggregation operation does not imply that the aggregated information can influence predictions only linearly. In the revised manuscript, we will clarify the structural priors and editable components for each task, as well as the support for nonlinear transformations after aggregation. The benchmarks used in this study do not include evaluation tasks with explicit hyperedge structures and corresponding higher-order interactions, and our existing experiments have not established NetSPE’s effectiveness in these settings. NetSPE’s core contribution is to improve equation discovery under given structural priors through hierarchical localization, local symbolic editing, and policy learning. Extending the representation to more general higher-order interactions and validating its effectiveness remain directions for future research.

Response to Weakness 2 and Question 2

We added a topology perturbation experiment by introducing 5% additional edges into the graphs of the synthetic tasks and comparing NetSPE with ND2 under the same perturbations. We chose synthetic data because the network topology used to generate the data is known and matches the original input topology, allowing us to clearly define and control the introduced topological errors. In contrast, Bacteria and Epidemic are real-world datasets whose networks cannot be assumed to fully and accurately represent the true interaction structures as they don’t have the ground-truth equations. The results show that topology perturbations increase the prediction errors of both methods, but NetSPE exhibits a smaller absolute increase in NMSE. Notably, NetSPE’s NMSE under perturbed topology remains close to that of ND2 under the original topology. Under the same topology perturbations, NetSPE also achieves lower NMSE and higher ACC_10 than ND2.

MethodSettingNMSEACC_10
ND2Clean0.0437269.25
NetSPEClean0.0084183.69
ND25% edge addition0.1057340.27
NetSPE5% edge addition0.0487851.58

Response to Weakness 3 and Question 3

We added three ablation variants: Frozen policy, Frozen+best-of-G, and Frozen+random, and compared them under matched sampling and evaluation budgets. Frozen policy keeps the generation policy fixed and generates one candidate at a time. Frozen+best-of-G generates G candidates per group using the frozen policy and selects the best candidate, whereas Frozen+random randomly selects one of the G candidates generated in the same way. In particular, Frozen+best-of-G uses the same number of candidates per group and the same evaluation and selection rules as SPPO, with only the generation policy frozen. SPPO achieves better overall performance than the frozen-policy variants. This indicates that SPPO’s gains extend beyond candidate sampling and selection, with policy learning providing an additional contribution to final search performance. (NMSE / ACC_10)

BenchmarkFrozen+RandomFrozen policyFrozen+best-of-GSPPO
Synthetic0.0631 / 60.980.0381 / 68.470.0607 / 67.420.0084 / 83.68
Bacteria0.8509 / 27.800.8482 / 28.090.8468 / 28.390.8256 / 29.24
Epidemic0.2679 / 21.950.2683 / 25.760.2683 / 24.010.2395 / 25.56

The same-edit-state constraint is intended to ensure that candidate patches used for learning are comparable. When the parent equation, branch, target subtree, and edit intent are identical, performance differences between candidates more directly reflect the quality of the patches themselves, rather than differences in edit locations or contexts. This constraint applies to the editing context used for within-group comparisons, it does not require the entire search to remain in the same edit state.

Response to Weakness 4 and Question 4

Since Bacteria Growth and Epidemic Transmission lack known ground-truth equations, we analyze the relationships expressed by the model through two illustrative examples of the final equations and regard them as mechanistic hypotheses that can be further tested.

For example, for Bacteria Low-12, the final equation can be written as

=x_i\left[ r_i-\beta x_i+x_i^2 -\frac{\alpha}{K+x_i+\epsilon} \sum_{j\in\mathcal N_i}x_j \right],$$ where $r_i$ is the fitted node-specific growth term, $\beta=3.2783$, $\alpha=0.1521$, $K=0.0904$, and $\epsilon=10^{-6}$. For nonnegative abundances, the neighbor interaction term is negative. With total neighbor abundance held fixed, its inhibitory coefficient $\alpha x_i/(K+x_i+\epsilon)$ approaches saturation as the focal node’s abundance increases. Thus, this equation proposes an abundance-dependent neighbor inhibition hypothesis. For Epidemic CHI, the final equation is $$\hat y_i =\tanh(a x_i)-b h_i+cS_{x,i}-dS_{h,i},$$ $$S_{x,i}=\sum_{j\in\mathcal N_i}\mathrm{flow}_{ji}x_j$$ $$S_{h,i}=\sum_{j\in\mathcal N_i}\mathrm{flow}_{ji}h_j$$ Here, $x_i$ denotes the normalized current case count, $h_i$ denotes the normalized sum of the current and previous day’s case counts, and the fitted coefficients $a,b,c,d$ are all positive. This equation combines a saturating local case response, recent case history, and flow-weighted neighboring case information, providing explicit local and network predictive relationships that can be examined individually. These equations not only provide numerical predictions but also express state dependence and network interactions as inspectable symbolic relationships. The bacterial equation reveals the abundance-dependent inhibitory relationship captured by the model, while the epidemic equation explicitly combines local case history with flow-weighted neighbor information. These examples demonstrate that NetSPE can generate interpretable dynamical hypotheses that can be further tested. ## Response to Weakness 5 and Question 5 We added SymBOL (TPAMI), as suggested by the reviewer, as an additional baseline. We selected SymBOL primarily because its high-dimensional network dynamics setting combines LLM-driven symbolic search with network interaction modeling, making it directly relevant to our task and enabling a more targeted comparison. we evaluated SymBOL based on Llama-3.1-8B on Synthetic, Bacteria Growth, and Epidemic Transmission with the same data splits and evaluation metrics as the other methods. NetSPE achieves lower NMSE and higher ACC_10 across all three benchmarks. This additional comparison demonstrates that, under the current evaluation setting, NetSPE retains a performance advantage over a method that also incorporates LLMs and network interaction modeling. (NMSE / ACC_10) |Method|Synthetic|Bacteria|Epidemic| |---|---|---|---| | SymBOL | 0.0494 / 55.79 | 1.2375 / 1.51 | 0.3625 / 13.88 | |NetSPE|0.0084 / 83.69|0.8256 / 29.24|0.2395 / 25.56|

We believe the additional information we provided during the rebuttal, including new experimental results, additional details, and clarification on misunderstandings, is helpful for answering some of your key questions. Please let us know whether or not our response has been able to address your previous concerns, and we are more than happy to discuss further on any unclear points.

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

This manuscript studies the problem of symbolic regression for complex networks, that is, the process of generating closed form mathematical equations that characterize the dynamics of a given network, wherein the network is defined as a time parameterized dynamic graph with changing node and edge signals. Recently, LLM-based methods have been developed to generate and/or iteratively refine such equations. However, the manuscript argues that most of these methods do not explicitly train a policy to improve context-dependent local structural edits. To address this, the manuscript introduces an implicit bias for LLM-driven symbolic regression wherein equation refinement is separated into three operations (identifying where to apply a refinement/edit, applying a refinement, learning how to best refine based on verifiable outcomes of prior edits).

This is implemented as network symbolic patch evolution (NetSPE). Under this framework, equations are represented as ASTs within graph-aware templates and a plethora of heuristics are used to select where to apply a refinement under this representation. Then, an LLM is conditioned on a given state from the last step and generates a refinement. Finally, a novel SPPO strategy is used to sample several alternative refinements for the given state, whose outcome is used to update the LLM using a regularized pairwise preference objective, and finally a refinement/edit is selected.

Empirical results show improvements in predictive accuracy and symbolic equation recovery on different real world tasks.

Critical strengths
  1. the main insight and motivation is fundamentally sound -- e.g. the network aware ADT representation and the localization mechanism built on this data structure efficiently applies a graph-based inductive bias that prior works seem to lack. Analytic discussion and careful ablations moreover provide interpretability and clearly isolate the contributions of the work.
  2. The empirical results are strong -- This is especially true in the synthetic benchmark, where both NMSE and structural accuracy improves substantially over baselines. Additionally, they test their method with two different backbones which shows some support for the method being architecture agnostic. Ablations moreover demonstrate interpretability in the performance gains of the important components of the work.
Critical weaknesses
  1. The experiments do not properly isolate the effect of the proposed method as the experiments with the proposed method also seem to include domain specific information. Appendix B seems to indicate that NetSPE includes application specific structural constraints to each run (e.g. bacterial dynamics use a partially fixed GLV-style structure while epidemic dynamics uses predefined readout structures). Yet, it seems the reported general LLM baselines are either given access to task agnostic generic node information, or are given task specific information via ICL (prompting). If true, this is not a fair comparison as such experiments do not fully isolate the proposed localization or policy-learning mechanisms as they are operating on different hypothesis spaces. More precise comparisons would instead be necessary.
  2. The isolated contributions of of the policy learning are not properly established. Although the ablation clearly shows that removing SPPO degrades performance, the manuscript doesn't show how the learned patch generation policy improves across training updates. There does not seem to be any information on important aspects sa the number of accepted updates, changes in patch generation success rates, or performance on validation editing states. This is particularly problematic as learning to edit is a key insight and claimed contribution.
Questions for authors
  1. Can the author's clarify the effective graph specific information provided the NetSPE experiments versus the baselines? Differences in search-space restrictions? Total compute budget given to each baseline versus NetSPE?
  2. Could the authors provide more evidence that SPPO improves the patch-generation policy? For example, the authors could evaluate the initial and updated policies on a held-out editing state set and report metrics sa the rate at which the parent equation improves or average objective improvement under matched sampling and evaluation budgets.
3 comments

We sincerely thank you for your time and for providing thoughtful and constructive feedback on our work. The comments have been invaluable in helping us identify areas for improvement.

Response to Weakness 1 and Question 1

We added comparisons with three representative strong LLM baselines: FunctionEvolve (Xia et al., 2026b), IGSR (Saveliev et al., 2026), and RESTART (Li & Pan, 2026). While preserving each baseline’s original design, we provided all three with information about all neighboring nodes so that they could use the corresponding network context. We also aligned the network structural priors with NetSPE as closely as possible. Specifically, we incorporated the GLV and epidemic readout structural priors used by NetSPE into FunctionEvolve’s generation and evolution contexts, IGSR’s expression-generation prompts, and RESTART’s candidate proposal, refinement, and analysis prompts. We report the avg NMSE and ACC_10 across tasks as (NMSE/ACC_10), w/ NP means the baseline with neighbor information and structural priors.

Llama-3.1-8B

MethodSyntheticBacteriaEpidemic
FunctionEvolve w/ NP0.0855 / 24.045.1096 / 22.320.7053 / 24.64
IGSR w/ NP0.2771 / 13.251.0007 / 1.690.3372 / 15.68
RESTART w/ NP0.2345 / 15.681.1090 / 2.600.4164 / 10.64
NetSPE0.0084 / 83.690.8256 / 29.240.2395 / 25.56

Qwen3-4B

MethodSyntheticBacteriaEpidemic
FunctionEvolve w/ NP0.0828 / 24.355.2872 / 19.470.3424 / 23.34
IGSR w/ NP0.2793 / 13.071.0245 / 0.810.3506 / 13.06
RESTART w/ NP0.3459 / 11.081.0293 / 8.820.5406 / 8.29
NetSPE0.0092 / 80.370.8759 / 28.920.2269 / 24.28

Additionally, to assess NetSPE against an existing approach that combines graph structure with LLMs, we included SymBOL (TPAMI) as a Graph-based+LLM-based baseline using Llama-3.1-8B as the backbone. This additional comparison demonstrates that, under the current evaluation setting, NetSPE retains a performance advantage over a method that also incorporates LLMs and network interaction modeling.

MethodSyntheticBacteriaEpidemic
SymBOL0.0494 / 55.791.2375 / 1.510.3625 / 13.88
NetSPE0.0084 / 83.690.8256 / 29.240.2395 / 25.56

Regarding the search sampling budget, NetSPE uses 1,000 samples, as do LLM-SR and RESTART, while PiT-PO uses 2,500. FunctionEvolve’s configuration starts with 20 initial expressions and runs for up to 30 evolutionary rounds. Each round selects 5 parent expressions, with the LLM proposing 20 offspring expressions per parent. So, the nominal generation budget is approximately 3,020 samples.

Response to Weakness 2 and Question 2

To provide additional evidence on changes in patch-generation success rates following policy updates, we define a successful patch as one that yields a child equation with a better validation objective value than its parent. Across the 25 subtasks within the three benchmarks, the success rate improved in 20 cases, remained unchanged in 4, and declined in 1. Among the 20 subtasks that saw an improvement, the smallest gain was 4.55 percentage points, with 10 of these subtasks achieving an increase of 10 percentage points or more. Here we present the overall gain from three benchmarks.

BenchmarkBefore UpdateAfter UpdateGain
Synthetic49.43%55.68%+6.25
Bacteria47.42%52.58%+5.15
Epidemic37.50%48.61%+11.11
Overall46.38%53.33%+6.96

We believe the additional information we provided during the rebuttal, including new experimental results, additional details, and clarification on misunderstandings, is helpful for answering some of your key questions. Please let us know whether or not our response has been able to address your previous concerns, and we are more than happy to discuss further on any unclear points.

Thanks for your rebuttal and additional experiments. Unfortunately, the additional comparisons to LLM baselines do not establish the equivalence of the effective search spaces available to each method, especially given the large distinction between test time prompted priors and explicitly enforced structural constraints (as outlined in my original review). Additionally, without confirming that the reported patch success rates are evaluated on identical independent editing states, it is difficult to isolate the contribution of policy learning. I thus maintain my score.

Comments on the paper

1 comment
Area_Chair_RZ96OpenReview ↗

Please join the author-reviewer discussion

Dear all reviewers,

Our author and reviewer discussion will end soon in 18 November (AOE). If you have not yet joined the discussion, now it is the time to start it. We highly appreciate your time on discussing and exchanging thoughts to keep our ICLR standards. Many thanks!

Bests,
Your AC

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

Related papers

Loading the map…

Discussion (0)

Sign in to comment.