Skip to main content
preprint

LLMs as Post-hoc Auditors of Physiological Plausibility in Symbolic Regression: A Clinician-Evaluated Case Study

arXiv preprint arXiv:2609.11431, 2026

Research summary

A 2026 preprint tests whether large language models can help clinicians compare symbolic-regression equations for body-fat estimation. A panel of clinicians rated the outputs.

Research focus and contribution

The preprint, posted to arXiv on September 10, 2026, extends the authors’ earlier work on estimating total body fat with symbolic regression and evolutionary algorithms, and the arXiv listing notes that it is an extended version of a paper submitted to a journal. Symbolic regression produces explicit mathematical formulas rather than opaque model weights, but a formula that fits data well may still combine variables in ways that make no physiological sense. The study asks whether large language models (LLMs) can help clinicians review candidate formulas for that kind of plausibility. Dr. Costilla Reyes, listed with Equ Healthcare, is one of the coauthors, led by Jorge López-Varela and J. Ignacio Hidalgo of the Universidad Complutense de Madrid, in a collaboration with clinicians.

How the study was run

The candidate formulas came from the earlier study, which evolved them with grammar-guided genetic programming on the 2017–2018 NHANES survey, a United States national health survey that includes body fat measured by dual-energy X-ray absorptiometry in adults. Three LLMs, Gemma 3, DeepSeek-R1, and GPT-5.1 Thinking, were each asked, over repeated runs, to explain the terms of each formula and to rank the formulas on criteria including interpretability and plausibility. A panel of clinicians from different hospitals, who are coauthors, rated the outputs.

Findings and connections

Across the models, clinicians rated the comparative rankings more favorably than the explanations of single equations. Setting candidate formulas side by side against criteria such as interpretability and plausibility gives a reviewer a structured starting point, and it matches how clinicians weigh competing models in practice.

The authors conclude that the results support comparative auditing under expert oversight: the language model acts as a review aid that brings fluent, organized commentary to the table, while the clinician’s judgment decides what a formula means physiologically. This is a practical route to combining the transparency of symbolic regression with the breadth of language models, so that clinical knowledge guides which equations are worth trusting. The paper belongs to Dr. Costilla Reyes’s broader line of research on interpretable models for healthcare, which includes symbolic-regression methods developed at MIT, such as LaSR and InceptionSR, and the body-fat estimation study published in 2025 with several of the same collaborators.

Publication details

Authors
Jorge López-Varela, J. Ignacio Hidalgo, José-Manuel Muñoz, Omar Costilla-Reyes, Esther Maqueda, Jesus Moreno-Fernandez, Tomás González-Vidal, J. Manuel Velasco, Oscar Garnica
Venue
arXiv preprint arXiv:2609.11431
Year
2026

BibTeX citation

@misc{costilla2026llmsaspost,
  title = {LLMs as Post-hoc Auditors of Physiological Plausibility in Symbolic Regression: A Clinician-Evaluated Case Study},
  author = {Jorge López-Varela and J. Ignacio Hidalgo and José-Manuel Muñoz and Omar Costilla-Reyes and Esther Maqueda and Jesus Moreno-Fernandez and Tomás González-Vidal and J. Manuel Velasco and Oscar Garnica},
  year = {2026},
  
  archivePrefix = {arXiv},
  eprint = {2609.11431},
  doi = {10.48550/arXiv.2609.11431},
}

Topics