LLMs as Post-hoc Auditors of Physiological Plausibility in Symbolic Regression: A Clinician-Evaluated Case Study
Jorge López-Varela, J. Ignacio Hidalgo, José-Manuel Muñoz, Omar Costilla-Reyes, Esther Maqueda, Jesus Moreno-Fernandez, Tomás González-Vidal, J. Manuel Velasco, et al.
cs.AI
Quick verdict
Case study: LLMs help rank symbolic regression models under clinician oversight, but flawed interpretations and a 3-rater co-author panel limit the evidence.
Summary
This case study extends the authors' prior work on body-fat-percentage symbolic regression from NHANES 2017-2018 (2,403 adults). Four grammar-guided GP models — two DSGE, two CFG-GP, with test R² 0.837–0.849 (Table 2) — were submitted to three LLMs: Gemma3-27B (deterministic baseline, temperature 0, seed 42), DeepSeek-R1, and GPT-5.1 Thinking, each asked to interpret expressions term-by-term and rank the models on interpretability, physiological plausibility, simplicity, and practical use, over three runs apiece. Three clinicians (all co-authors) rated outputs on 1–5 Likert statements. The central result: comparative rankings received consistently more favorable clinician assessments than isolated interpretations across all three LLMs. Yet the LLMs delivered physiologically false explanations fluently — Gemma3 repeatedly asserted men have higher body fat than women — and the reasoning models' mid-rank orderings drifted across runs while Gemma3 remained invariant; clinician reception of the three LLMs was nevertheless similar. The authors conclude LLMs suit comparative auditing under expert oversight, not autonomous validation. Limitations are substantial: three raters preclude statistical inference; evaluators are unblinded co-authors; no inter-rater or human–LLM agreement metrics are computed; the prompt's sign heuristic — acknowledged invalid for nonlinear terms — structurally biases the interpretation task; and no code release is mentioned.
Review complete
Assessment at a glance
5
risk flags identified
● high 1
● medium 2
● low 2
5 strengths · 5 weaknesses · 5 author questions
AI text signal = Binoculars zero-shot detector score over 20 sampled passages (Fast-DetectGPT co-reported) — absence does not prove human authorship, and a flag does not imply misconduct.
Levels reflect verification complexity — not a quality score. Evolve never rates your paper.
Key claims under review
- Across Gemma3, DeepSeek-R1, and GPT-5.1 Thinking, clinicians rated the comparative ranking of four SR models more favorably than any isolated term-level interpretation.
- LLMs produce physiologically incorrect statements (e.g., Gemma3's repeated claim that men have higher body fat than women) with the same fluency as correct ones, so clinician oversight is indispensable.
- Reasoning models (DeepSeek-R1, GPT-5.1 Thinking) show run-to-run instability in middle ranks while the deterministic Gemma3 baseline reproduces identical orderings, yet clinician reception of all three LLMs was largely indistinguishable.
- LLMs are therefore suited to comparative model selection as an expert common-sense filter under human oversight, not as autonomous arbiters of biomedical validity.
Strengths
Genuine, well-positioned gap
Section 2 maps three LLM-in-EC paradigms (adaptive operators, scientific priors, co-evolution, citing EvoPrompt/LLM-SR/LASR) and identifies post-hoc plausibility auditing of already-interpretable SR models as unexplored; the research question is qualitative and cleanly scoped.
Expert validation loop that demonstrably caught errors
Clinicians flagged Gemma3's repeated false claim that men have higher body fat than women (Section 5.1, Tables 4-5; Section 8), and noted hip circumference signals peripheral rather than visceral fat — concretely demonstrating why oversight matters rather than asserting it.
Accuracy-controlled model set and transparent protocol
Self-critical interpretation of its own evidence
Informed LLM deployment choices
Weaknesses
Evaluator circularity
No quantitative agreement or reliability analysis
Prompt design biases the flagship contrast
Unexamined LLM claims reported uncritically
Reporting gaps and internal inconsistencies
Full Review Report
$1.99 · one-timeUnlock the complete preliminary analysis Evolve produced for this paper.
Report preview
Scope & Claims
This case study (arXiv 2609.11431, an extended version of a journal submission) asks whether LLMs can audit symbolic regression (SR) models for interpretability and physiological plausibility. It reuses four grammar-guided GP models from the authors' prior EvoStar work [21] — DSGE-M21/M18 and CFG-M9/M3, all at test R² 0.837-0.849 on NHANES 2017-2018 body-fat data (2,403 adults, nine features → DXA fat %, Table 2) — and submits them to Gemma3-27B (deterministic, T=0, seed 42), DeepSeek-R1, and GPT-5.1 Thinking. Two tasks: term-level interpretation (analyzed for Gemma3 and DeepSeek-R1 only) and comparative ranking on four criteria, three runs per LLM, the first run serving as the evaluated reference. Three clinician co-authors rate all outputs on 1-5 Likert statements (Tables 4-14).
Central claims: (1) clinicians rate comparative rankings more favorably than isolated interpretations across all LLMs; (2) LLMs deliver physiologically false statements fluently, so expert oversight is mandatory; (3) LLMs fit comparative auditing under oversight, not autonomous validation. All three are directionally credible, but the evidence base is tiny and methodologically self-referential.
The remaining sections cover Methodology Assessment, Statistical Evidence, Results & Reproducibility, Limitations & Risks, and Verdict & Recommendations.
- All 5 strengths & 5 weaknesses, analyzed in detail
- 5 risk flags with severity, section & reviewer questions
- Methods & statistical evidence assessment
- 5 questions a peer reviewer would ask
- Full six-dimension evaluation report & recommendations
One-time purchase · this paper only
One-time purchase, this paper only. AI-generated preliminary analysis — not peer review.
Frequently asked questions
Can LLMs reliably judge whether a symbolic regression model is medically plausible?
According to this study, not on their own. Clinicians found LLM comparative rankings of candidate models useful, but the LLMs also produced physiologically false explanations delivered with confident fluency — for example, Gemma3 repeatedly asserted that men have higher body fat percentage than women, which is wrong. The recommended role is a comparative screening and selection aid that always operates under expert human review, not an autonomous validator.
Which LLM performed best in this study?
No consistent winner. The three clinicians rated all three LLMs' comparative rankings similarly. GPT-5.1 Thinking was praised for making its evaluation criteria explicit, DeepSeek-R1 for finer-grained physiological reasoning, and Gemma3-27B (run at temperature 0) was the only fully reproducible baseline — but it also made the most flagrant repeated error (claiming men have more body fat than women).
What data and models were used?
The four symbolic regression models estimate total body fat percentage from NHANES 2017-2018 data (2,403 US adults; nine anthropometric/demographic features; DXA-measured fat percentage as target). The models come from the authors' earlier study and were built with grammar-guided genetic programming variants (GE/DSGE and CFG-GP), all reaching test R² of roughly 0.84. Gemma3-27B, DeepSeek-R1, and GPT-5.1 Thinking then interpreted and ranked them.
How big was the clinician panel and is the evidence statistically strong?
Three clinicians evaluated the LLM outputs on 1-5 agreement statements. All three are co-authors of the paper, no blinding is described, and the authors themselves acknowledge that with three raters genuine agreement cannot be separated from insensitivity of the evaluation instrument. The findings should be read as directional, anecdote-grade evidence rather than statistically established results.
About this review
This is an AI-generated preliminary review by Evolve.Pub, produced from the openly available text of the paper. It is not peer review and does not represent the authors' views. The canonical record of the paper is at the publisher link above.
- Figures 2-27 are images; the LLM output texts are additionally quoted in the body, so no substantive content is lost, but figure-embedded details beyond the quoted text cannot be verified.
- The Acknowledgments section is empty in the source text (the heading is immediately followed by References); funding, ethics, and conflict-of-interest statements, if any, are unavailable.
Go deeper on this paper
Wrote a paper? Stress-test it before reviewers do.
Upload your manuscript for the same structured scrutiny — reasoning, statistics, evidence, figures, and reproducibility.