- Rating
- 2
- Confidence
- 3
- Informativeness
- 2
- Presentation
- 2
- Soundness
- 3
Read review
Paper summary
This paper introduces EvoMO‑SR, an LLM‑driven symbolic regression (SR) framework. The method uses a Large Language Model (Llama‑3.1‑8B‑Instruct) to generate equation skeletons, which are then fitted by an external optimizer (L‑BFGS‑B). The framework incorporates a multi‑objective survival selection mechanism to balance accuracy (NMSE) and complexity (AST node count), and a substructure guidance module that extracts additive subtrees from evolved expressions and reuses them in subsequent mutations. Experiments on the LSR‑Synth benchmark and two physics‑based datasets (oscillator1/2) demonstrate that EvoMO‑SR outperforms existing LLM‑based and traditional SR baselines in seven out of eight domain‑split comparisons.
Critical strengths
-
The paper tackles an important and timely problem: leveraging LLMs for interpretable scientific discovery via symbolic regression.
-
The proposed substructure guidance idea is intuitive and could potentially help mitigate bloating and improve search efficiency.
-
Empirical results show promising numerical accuracy gains, especially on out‑of‑distribution (OOD) splits, using a relatively small LLM backbone.
Critical weaknesses
-
The manuscript suffers from overly long paragraphs with unclear transitions between ideas, making the narrative difficult to follow. For instance, the Related Work section is presented as a single, unbroken block of text, which severely hampers readability and fails to group related methods logically.
-
The paper positions itself as a novel combination of LLM generation, GP‑style evolution, and substructure extraction. However, the distinction from prior LLM‑based SR works—particularly LLM‑SR (Shojaee et al., 2025a) and LaSR (Grayeli et al., 2025)—is not clearly articulated. It is unclear what fundamental limitation of existing approaches is addressed, or what theoretical benefit the added modules bring. The authors state that EvoMO‑SR “enhances GP search with LLM‑generated concepts,” but LaSR already does this; the essential difference between the two methods is never explained. The current presentation gives the impression of a straightforward fusion of existing techniques rather than a principled advancement.
-
The first sentence of the Related Work states: “SR is a supervised machine learning method that aims to retrieve the mathematical expression that best fits a given dataset.” This characterization is debatable. While deep‑learning‑based SR methods may use synthetic expressions as supervision, many search‑based algorithms (e.g., GP, evolutionary strategies) rely solely on a fitness function (e.g., MSE) without requiring ground‑truth expression labels. Labelling SR universally as “supervised” is therefore inaccurate and potentially misleading.
-
Table 3 reports median NMSE and symbolic accuracy (SA) values across runs. In symbolic regression, the primary goal is to discover at least one expression that is both sufficiently accurate and parsimonious. Averaging over all runs via the median obscures whether the method successfully found such an expression in individual trials. Reporting best‑run performance or success rates (e.g., fraction of runs recovering a ground‑truth subtree) would be more informative.
-
The ablation experiment in Table 4 only compares performance with/without substructure guidance on two specific datasets (oscillator1 and oscillator2). This is inadequate to establish the general utility of the proposed components. A thorough ablation should: (1) Isolate the contribution of each major module: (a) LLM‑only prediction (no evolution), (b) GP‑only search (no LLM), (c) full model without substructure guidance. (2) Evaluate on a substantially larger and more diverse set of expressions (beyond just two cases) to guard against dataset‑specific overfitting. Without such experiments, it is difficult to attribute performance gains to the proposed fusion rather than chance.
Questions for authors
-
Novelty Clarification: Could you explicitly delineate the core differences between EvoMO‑SR and LaSR/LLM‑SR? What specific weakness of those methods does your framework overcome, and how does the multi‑objective survival selection fundamentally improve upon their sampling‑based or single‑objective strategies?
-
Theoretical Justification: Beyond empirical results on two datasets, is there any theoretical analysis or intuition suggesting that substructure guidance should accelerate convergence or improve symbolic recovery? How are substructures scored independently of the full expression’s fitness?
-
SR as Supervised Learning: Do you agree that many SR algorithms (especially evolutionary ones) do not use expression labels and therefore should not be classified as supervised? Would you consider revising this statement to avoid confusion?
-
Evaluation Protocol: Why was the median chosen as the summary metric in Table 3? Would it be possible to report the number of successful recoveries (e.g., SA > threshold) or the best‑of‑run performance to better reflect SR’s goal of finding a single high‑quality expression?
-
Expanded Ablation: Could you provide additional ablation results that remove the LLM or the GP loop entirely, and test on a wider range of problems (e.g., from LSR‑Synth or randomly generated functions) to demonstrate that the improvements are not limited to the two oscillator datasets?
We clarify novelty-related concerns; we show additional experiments for substructures; we clarify SR definition; we clarify the evaluation protocol and provide best expressions for randomly generated functions; we provide an additional ablation study.
We thank the reviewer for the constructive feedback and for recognizing the importance of the problem.
Novelty Clarification: More detailed qualitative comparisons are available in Appendix E. EvoMO-SR introduces an explicit population-based evolutionary loop. While LaSR performs SR with an evolutionary loop as well, its search engine is managed by PySR. LLM-SR uses some evolutionary-related components, but its loop mainly evolves expressions based on a multi-island sampling approach. The overall framework includes the possibility to customize other evolutionary-related parameters (as described in Appendix A.1). Concerning multi-objective optimization, EvoMO-SR exploits numerical accuracy and expression complexity to determine population survival during the evolutionary loop. The ablations in Appendix C.3 (Tables 5 and 6) show that this consistently reduces expression complexity. Concerning the objection raised by the reviewer, we would like to clarify that the sentence "enhances GP search with LLM‑generated concepts" actually refers to LaSR in our manuscript (see Section 2). In fact, our method does not use LLM-generated concepts to enhance the search. This is what distinguishes us from LaSR in what we use to guide LLM-based search.
Theoretical Justification: The intuition in integrating substructure guidance is that mathematical pieces discovered by the LLM during the evolution may remain useful when modifying other candidates. By retaining such substructures in an archive could be a way in which the search can reuse potentially valuable symbolic information. For the concerns raised, we carried out additional experiments on oscillator2, CRK34 problem (LSR-Synth), and the P3 problem (randomly generated function), each for 5 runs. We investigated whether reused substructures were beneficial by removing the added substructure for each incorporation case. Results shown that removing the substructure increased ID NMSE in 72.7% of cases and OOD NMSE in 52.1% of cases (higher is worse). Additionally, for each substructure present in the ground truth, we recorded their improvement rate (on parent expressions) and their presence in the final expression. For example for oscillator2, sin(t) improved parents in 72.2% of cases and was present in 5/5 final expressions; for CRK34 A2 the improvement was in 57.6% of parents, presence 5/5; for P3 x02 the improvement was in 70% of parents, presence 4/5
Based on the results obtained, we aim to show the value of introducing the substructure guidance mechanism, without, however, claiming that it is, a method that consistently improves performance.
SR as Supervised Learning: we agree that many SR algorithms are tested over datasets that do not have ground truth expressions. However, there is a distinction with the actual label of the SR datasets. A ground truth is not a label associated with an input. Dataset’s input-output pairs in SR datasets are numerical only, since SR is a regression algorithm. In fact, SR methods can solve standard regression problems (as seen in SRBench - La Cava, William, et al., 2021). Since the SR outputs a symbolic expression, it is surely compared against the ground truth expression when available, but it does not constitute the label. When the ground truth is not available, the SR output (i.e., the symbolic expression) is evaluated by computing the predicted y, and compared against the real y (the target value in the dataset). For these reasons, the SR is recognized in literature as a supervised algorithm.
Evaluation Protocol: taking the median into account ensures that problems where these methods fail do not have too large an impact on overall performance. For P1, the best expression (based on the lowest NMSE) is obtained in 3 out of 5 runs, where the retrieved expression is the same and is equal to the ground truth. For P2, the best expression found was exactly equal to the ground truth expression. For P3, the best expression has SA_Term = 0.33 and is 1.51–1.95x02+0.33x12+2.36x22+1.37cos(x1)+6.17×10−7x0. For P4, the best expression has SA_Term = 0.4 and is −2.10–2.16sin(x0)–1.08cos(x0)+1.9x12+2.38x1x2−2.23cos(x2)+0.0001x22−0.0002x1+0.000004x2+0.0002sin(x1). All the substructures in the ground truth are present in the final expression, with some spurious terms with low coefficients.
Expanded Ablation: We acknowledge that performing more ablation studies could be beneficial. However, we would like to clarify that in Appendix C.3.1 (Figure 13) we compared results obtained by using GP-only mutation and LLM-guided mutation. Following your suggestion, we carried out ablations on randomly generated functions (P1 and P4) with (wg) and without guidance (wog).