acceptodds
Withdrawn from ICLR 2027

ViGa-SR: LLM-based Evolutionary Search for Symbolic Regression with Visual Priors Guidance from Multimodal Representation of Data

Withdrawn from ICLR 2027, so left out of the paper lists and the map.

Abstract

Symbolic regression aims to derive concise formulas that both fit data and reveal the mechanisms underlying it, with evolutionary computation (EC) as the widely adopted approach. However, EC-based SR methods must navigate a vast search space to find suitable equations. Prior knowledge can ease this burden, but existing approaches either rely on single-modal information or apply guidance indiscriminately, neglecting rich multimodal insights that human experts naturally exploit to adjust their priors during evolution. We propose ViGa-SR, a framework that automatically discovers the underlying mathematical structure of data and exploits it as prior knowledge to guide LLM-based evolutionary search. The data is first interpolated and rendered through diverse visualization schemes, from which a multimodal LLM (MLLM) progressively identifies the potential substructures, operators, and coupling relationships among the variables in the data; each identified feature is then numerically verified, with its associated constants calibrated. The identification proceeds hierarchically. After each feature is verified, it is stripped from the data, and the MLLM re-examines the simplified data. This process iterates to expose progressively deeper substructures and terminates when no further regularity can be confirmed. The verified prior knowledge guides the subsequent LLM-based evolutionary search and is re-acquired with altered visualization schemes whenever evolution stagnates or population diversity drops. Compared with purely numerical feature extraction, MLLM-based feature extraction is more robust to noise and sample sparsity, and the use of visualizations as the communication medium keeps the entire process transparent and human-readable. On the Feynman dataset and LLM-SRBench, ViGa-SR matches or exceeds state-of-the-art performance, demonstrating strong feature-extraction capability and noise robustness.

Primary area:
interpretability and explainable AI
Keywords:
Symbolic regression, multimodal large language models, evolutionary computation, scientific discovery

Reviews

Among:
Shape:
1234this paper, Bottom 2%
Figure 1. Mean review rating of the reviewed papers at ICLR 2027. This paper averages 1.33 over 3 reviews, above 2% of the others.
Reviewer_274cOpenReview ↗
Rating
2
Confidence
2
Informativeness
2
Presentation
2
Soundness
2
Read review
Paper summary

This paper proposes ViGa-SR, a multimodal framework for symbolic regression that uses an MLLM to extract structural cues from visualizations of observed data and injects the resulting priors into an LLM-guided evolutionary search. Candidate visual structures are numerically screened before being used, and the extraction process is applied hierarchically by removing identified components and re-rendering the residual signal. The method is evaluated on Feynman, LLM-SRBench, and several additional case studies, where it achieves competitive symbolic recovery performance and shows that visual priors can substantially influence downstream search.

Critical strengths

1.The paper proposes an interesting system-level mechanism for incorporating multimodal information into symbolic regression.

2.Hierarchical residual re-visualization is a promising idea.

Critical weaknesses

1.The experiments do not sufficiently isolate the mechanisms underlying the paper’s central claims. The paper argues that numerical verification makes the visual channel reliable and that hierarchical identification reveals structures missed by one-shot extraction. However, the evaluation does not include the most direct controls needed to establish these claims, and the contributions of these components are therefore not disentangled. The current experiments do not convincingly establish why the method works or which of the proposed mechanisms are necessary.

2.Variable anonymization removes semantic and name-based shortcuts, but it does not prevent the MLLM from recognizing familiar mathematical shapes or function families that may have appeared during pretraining. This distinction is particularly important here because the core contribution is precisely the use of a pretrained multimodal model to infer structure from rendered data. The paper itself acknowledges that anonymization does not eliminate shape recall, and its LSR-Synth experiment is described as a descriptive self-check rather than a measurement of memorization resistance. Therefore, the current evidence does not support the stronger interpretation that ViGa-SR demonstrates general visual structure discovery on equations that are genuinely outside the MLLM’s prior knowledge.

3.Missing closely related work. The related-work discussion appears incomplete with respect to recent structure-aware evolutionary symbolic regression. In particular, Symbolic Regression with Physics-Informed Residual and Structural Sensitivity Pruning (IEEE TEVC, 2026) introduces structural priors and fine-grained sensitivity-based evaluation of symbolic substructures during evolutionary search. While RSSP differs substantially from ViGa-SR in that its priors are physics-derived rather than visually proposed by an MLLM, both works use structured prior information and numerical criteria to guide local structural decisions during symbolic evolution. This work should therefore be discussed explicitly, and the novelty claim should be narrowed from generic “verified/adaptive structure-guided evolutionary search” to the more specific contribution of visually acquired, coordinate-level priors with numerical admission and hierarchical residual re-reading.

4.The paper also has several presentation and writing issues. First, the core research gaps are not mapped one-to-one to the subsequent experiments. Second, the Introduction overgeneralizes the limitations of prior work, which makes the stated research gap appear broader than the actual scope of the contribution. In addition, the Methodology still reads somewhat like a module-by-module system description, and several nontrivial design choices lack sufficient rationale connecting the underlying difficulty to the proposed mechanism.

Questions for authors

1.Could the authors provide a controlled comparison, under the same computational budget, among the following settings: no visual prior, unverified MLLM prior, numerically verified MLLM prior, one-shot extraction, and hierarchical extraction? Without such controls, it is difficult to determine which mechanism is actually responsible for the observed performance gains.

2.How do the authors position ViGa-SR relative to recent structure-aware evolutionary symbolic regression methods such as Symbolic Regression with Physics-Informed Residual and Structural Sensitivity Pruning? Although the two methods differ substantially in the source of structural information, both use fine-grained structural signals and numerical criteria to guide symbolic evolution. A clearer comparison would help determine whether the main novelty lies in visual prior acquisition, numerical admission, hierarchical residual re-reading, or a broader claim of structure-guided search.

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

This paper proposes ViGa-SR, a symbolic regression framework that uses visual information to guide LLM-based evolutionary search. A multimodal LLM identifies candidate structures from data visualizations, which are numerically verified, fitted. The resulting decomposition provides seed expressions and textual guidance for subsequent search. Experiments cover Feynman equations, LLM-SRBench, oscillator examples, and several ablation and robustness studies.

Critical strengths
  1. Broad experimental coverage. The paper evaluates equation recovery, numerical accuracy, feature extraction, and robustness to noise and sampling conditions across several settings.

  2. Acknowledgment of limitations. The paper explicitly discusses restricted visual observability, imperfect feature identification, and limitations in the experimental comparisons.

Critical weaknesses
  1. Converting numerical observations into images can lose information and introduce distortions through interpolation and rendering. Nevertheless, the proposed method bases structural analysis and feature extraction on these transformed images, while omitting direct structural analysis of the original numerical observations. Although numerical fitting and verification are performed afterward, they do not replace direct numerical structure discovery. Visual information may provide useful complementary clues, but the original observations retain details that smoothing and rendering can obscure. The paper does not adequately justify this design choice and lacks controlled comparisons among numerical-only, visual-only, and combined approaches under matched preprocessing and computational budgets.

  2. The overlap with existing visual symbolic regression methods is substantial, and the closest comparisons are missing. Several relevant works already connect visual representations with symbolic discovery:
    [1]“Symbolic Regression with Multimodal Large Language Models and Kolmogorov–Arnold Networks” (LLM-LEx) is cited but not experimentally compared. It also uses plotted functions to inform LLM-generated expressions, followed by numerical parameter fitting and population-based search.
    [2] “Automated Model Discovery via Multi-modal & Multi-step Pipeline” is not cited. It combines multimodal analysis, visual and quantitative evaluation, and iterative model discovery, overlapping with the paper’s broader analysis–verification workflow.
    [3] “Automated Symbolic Law Discovery: A Computer Vision Approach” (DeSTrOI) is not cited. It extracts operator information from visual representations to guide genetic programming, directly anticipating the use of visual structural priors for symbolic search.
    [4] “In-Context Symbolic Regression: Leveraging Large Language Models for Function Discovery” is cited, but its visual extension, ICSR-V, is not directly compared. That extension uses plots of observations and candidate predictions to support iterative expression refinement.
    These methods make visual guidance itself an insufficient novelty claim. The paper needs to distinguish the contribution and design motivation of its verified hierarchical decomposition and demonstrate an advantage over closely related approaches.

  3. The final performance gains are limited relative to the complexity of the pipeline. In Table 2, the solution rate improves only from 31.53% for LLM-SR to 32.02%, while [email protected] improves from 39.64% to 40.23%. ViGa-SR’s NMSE of 0.007 is also worse than LaSR’s 0.001. These results do not establish a consistent advantage across metrics. The framework adds interpolation, multiple visual queries, numerical verification, repeated fitting, hierarchical stripping, and adaptive re-acquisition. Without uncertainty estimates and a complete accounting of runtime, model calls, and numerical optimization costs, the small gains do not convincingly justify this complexity. Taking baseline results from their original reports under different configurations further limits the comparison.

  4. One- and two-dimensional views cannot generally reveal interactions involving several variables. The paper acknowledges examples that its current views cannot represent, but does not establish how reliably the method handles such interactions. These limitations weaken the applicability of the proposed decomposition to complex equations.

Questions for authors

No

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

The paper introduces ViGA-SR, a symbolic regression (SR) approach that uses MLLMs to extract mathematical structures from visualizations of smoothed data and uses these to guide LLM-based evolutionary search. The MLLM identifies structures like poles, zero sets, composite coordinates and multiplication. The approach also removes verified features at each step to provide a simplified dataset which the MLLM then re-examines. They then evaluate on a subset of SRBench and LLM-SRbench and report competitive performance.

Critical strengths
  1. To the best of my knowledge, the paper's proposed combination of MLLM for SR is novel.

  2. Useful and thoughtful groundwork in building up tools for visual representations in SR, such as the classification and identification of various feature types in Table 1.

Critical weaknesses
  1. Clarity of the paper is sometimes poor, with non-standard terms in the field or vague phrases like "name channel", "numerical pre-filter", "unverified structural claim", "crossing-localization AUC".

  2. Missing key citations for highly related mechanisms. While the paper did cite Ai feynman (Udrescu, 2020), it was not cited for the MLP interpolation in their framework in section 2.2. To the best of my understanding, that is one of the early works which introduced this idea, and at least one of these prior SR works which introduces MLP interpolation for SR should be cited in section 2.2 instead of having no citations at all for the section.

  3. Empirical evidence are not reported with sufficient scientific rigour. Error bars from repeated trials are missing, level of significance of the results are not tested and reported as well.

  4. The datasets evaluated on are from synthetic datasets with a known equation/generator, this is a major weakness in terms of evaluation. Are there any results on real-world datasets? For example, SRBench which the paper uses has real-world datasets which the paper can use. These evaluations in the benchmark are missing in the paper.

  5. Additionally, a reader familiar with LLM-SRbench would find that in Table 2, 3 out of 4 of the rows (LLM-SR, LaSR, SGA) are taken from the LLM-SRBench paper directly, and as the paper points out, there are 111 problems, 10 runs each, which corresponds correctly to 350/1110 (31.53%), 70/1110 (6.31%), 110/1110 (9.91%). The presented result of ViGa-SR (Ours) is 32.02%, which corresponds to neither 355/1110 (31.98%) nor 356/1110 (32.07%).

Questions for authors
  1. Please address the weaknesses stated above.

  2. Also, please explain this sentence found in the list of contributions: "Structure always originates in the images; the criteria only judge—refusals, not penalties.". I do not understand the meaning of the last 3 words in relation to the whole sentence.

  3. Please also address the inconsistencies with the benchmarks, particularly how the values in Table 2 were obtained, and also on the missing real-world datasets evaluation.

  4. Can the paper include results and comparison for ViSymRe? ViSymRe is highly similar. Though the paper differentiates by saying that ViSymRe does not use image inputs at inference time, it does use the visualizations internally. ViGa-SR could be said to require an additional overhead. It would be very meaningful to see the comparison with a closely related algorithm that also uses visual representations for SR, to further advance this field of visual representations in SR.

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

Related papers

Discussion (0)

Sign in to comment.