Who Judges the Judge? Evaluating LLM-as-a-Judge for French Medical open-ended QA

Ikram Belmadani; Oumaima El Khettari; Pacôme Constant dit Beaufils; Richard Dufour; Benoit Favre

cs.CLMar 4, 2026

Who Judges the Judge? Evaluating LLM-as-a-Judge for French Medical open-ended QA

Authors: Ikram Belmadani, Oumaima El Khettari, Pacôme Constant dit Beaufils, Richard Dufour, Benoit Favre

Organizations: Aix-Marseille Univ., CNRS, LIS UMR 7020, 13000 Marseille, France · Nantes Univ., École Centrale Nantes, CNRS, LS2N, UMR 6004, 44000 Nantes, France · Nantes Université, CHU Nantes, PHU 11: Santé Publique, Clinique des données, INSERM, CIC 1413, 44000 Nantes, France · Nantes Université, CNRS, INSERM, L’institut du thorax, 44000 Nantes, France ·

Abstract

Automatic evaluation of medical open-ended question answering (OEQA) remains challenging due to the need for expert annotations. We evaluate whether large language models (LLMs) can act as judges of semantic equivalence in French medical OEQA, comparing closed-access, general-purpose, and biomedical domain-adapted models. Our results show that LLM-based judgments are strongly influenced by the model that generated the answer, with agreement varying substantially across generators. Domain-adapted and large general-purpose models achieve the highest alignment with expert annotations. We further show that lightweight adaptation of a compact model using supervised fine-tuning (SFT) and Group Relative Policy Optimization (GRPO) substantially improves performance and reduces generator sensitivity, even with limited data. Overall, our findings highlight the need for generator-aware evaluation and suggest that carefully adapted small models can support scalable evaluation in low-resource medical settings.

Figures & tables

Figure 1: Overview of the evaluation and alignment pipeline. French medical OEQA questions are answered by multiple LLM generators and annotated by a clinician for binary semantic equivalence. LLM evaluators judge the same instances using an identical prompt. A lightweight evaluator (Phi-3.5-mini) is further aligned via SFT and GRPO, and improvements are validated through paired significance testing.

Judge	P	R	F1	Accuracy	Pearson $r$
MedGemma-27B	52.17	72.00	60.50	71.80	40.67
Qwen-80B	64.62	56.00	60.00	77.60	44.77
Gemini-2.5-Pro	65.35	44.00	52.59	76.20	38.81
GPT-5.1	76.19	32.00	45.07	76.60	38.27
Phi-3.5-mini	35.94	98.00*	52.59	47.00	27.49
SFT-Phi-3.5-mini	38.81	91.33	54.47	54.20	29.79

Table 1: Agreement between LLM judges and the medical expert on binary equivalence judgments. Best scores are shown in bold. The symbol * indicates the best-performing variant among the base, SFT-aligned, and GRPO-aligned models.

Metric	Pearson $r$
ROUGE-L	25.40
BLEU	17.16
BERTScore-F1	14.88

Table 2: Pearson correlation between automatic similarity metrics and expert equivalence judgments.

Figure 2: Heatmap of F1 scores for each judge model across answer-generating models.

Figure 3: Comparison of F1 scores of Phi models on the text generated by multiple LLMs

Comparison (A vs. B)	Acc (A)	Acc (B)	$\Delta$ Acc (Acc (B) - (Acc (A))	McNemar p
Base vs. SFT	47.00	54.20	7.20	1.25E-06
Base vs. GRPO	47.00	71.40	24.40	1.44E-16
SFT vs. GRPO	54.20	71.40	17.20	2.64E-11

Table 3: McNemar test results for accuracy differences between the base, SFT-only, and GRPO-aligned evaluators on the test set.

Appendix figures & tables7 assets

Supplementary material from the paper’s appendix.

Appendix

Hyperparameter	SFT	GRPO
Training set size	166	184
Validation set size	18	–
Number of epochs	5	2
Learning rate	$5\times 10^{-6}$	$1\times 10^{-5}$
Batch size	1	1
Gradient accumulation steps	4	4

Appendix

Table 4: Hyperparameters used for SFT and GRPO alignment of the Phi-3.5 evaluator.

Comparison (A vs. B)	Pearson r(A)	Pearson r(B)	$\Delta$ Prearson r (r(B) - r(A))	Bootstrap p	Permutation p
Base vs. SFT	27.49	29.79	2.3	0.5533	0.4255
Base vs. GRPO	27.49	36.33	8.84	0.4985	0.1106
SFT vs. GRPO	29.79	36.33	6.54	0.5138	0.1899

Appendix

Table 5: Paired bootstrap and permutation test results for correlation differences between the base, SFT-only, and GRPO-aligned evaluators on the test set.

Comparison (A vs. B)	F1 (A)	F1 (B)	$\Delta$ F1 (F1 (B) - F1 (A))	Bootstrap p	Permutation p
Base vs. SFT	52.59	54.47	1.88	0.5049	0.092
Base vs. GRPO	52.59	57.06	4.47	0.5002	0.1278
SFT vs. GRPO	54.47	57.06	2.59	0.5338	0.3389

Appendix

Table 6: Paired bootstrap and permutation test results for F1-score differences between the base, SFT-only, and GRPO-aligned evaluators on the test set.

Model Variant	TP	FP	TN	FN	F1
PHI-3.5 (Base)	147	262	88	3	0.526
PHI-3.5 + GRPO (after SFT-60 steps)	142	246	104	8	0.528
PHI-3.5 + GRPO (after SFT to convergence)	95	88	262	55	0.571
PHI-3.5 + GRPO (from scratch)	139	219	131	11	0.547

Appendix

Table 7: Ablation of alignment schedules for Phi-3.5-mini. We compare the base model, GRPO applied from scratch, GRPO after a short SFT warm-up, and GRPO applied after SFT convergence. The strongest performance is obtained when GRPO is applied after SFT loss stabilization.

Question

Citez par quels moyens réaliser une laryngoscopie indirecte ?

Clinical Context

Un homme de 60 ans consulte pour une dysphagie aux solides s’aggravant depuis 2 mois. Il a perdu 6 kg en 6 mois. L’examen ORL par laryngoscopie indirecte retrouve une lésion d’allure tumorale du sinus piriforme droit.

Ground Truth (Expert Reference)

– Au miroir (miroir laryngé avec lumière frontale, de Clar)

– Au nasofibroscope

Appendix

Table 8: Illustrative example of semantic equivalence judgments in French medical OEQA. Answers with widely different formulations, levels of detail, and terminology can still be judged equivalent by the clinician as long as the essential medical content is preserved, while concise but incorrect answers are rejected.

Judge	Generator	Accuracy	P	R	F1	Spearman $\rho$
GEMINI-2.5-PRO	Gemma-3-4B-IT	75.00	69.23	51.43	59.02	42.54
GEMINI-2.5-PRO	SFT-LLaMA-13B-Chat	78.00	100.00	24.14	38.89	42.93
GEMINI-2.5-PRO	MedGemma-4B-IT	70.00	58.33	41.18	48.28	28.87
GEMINI-2.5-PRO	Qwen-4B-IT	70.00	60.61	54.05	57.14	34.31
GEMINI-2.5-PRO	SFT-Qwen-4B	88.00	63.64	46.67	53.85	47.89
GPT-5.1	Gemma-3-4B-IT	75.00	100.00	28.57	44.44	45.43

Appendix

Table 9: Detailed agreement metrics between LLM judges and the expert annotations, broken down by answer-generating model.

Generating Model	Mean # Characters
Gemma-3-4B-IT	1852.88
SFT-LLaMA-13B-Chat	121.26
MedGemma-4B-IT	1909.31
Qwen-4B-IT	1620.27
SFT-Qwen-4B	296.27

Appendix

Table 10: Average number of characters per answer for each generating model.

LLM-as-a-Judge

Large language model (LLM)-as-a-judge research uses language models to score, compare, and verify AI outputs as scalable alternatives to human evaluation.

See all 555 papers

Evaluating Large Language Model Raters for German Open-Response Clinical Questions: A Physician-Annotated Benchmark Study of Agreement, Evaluator Bias, and Abstention
Jul 1, 2026William Philipp, Finn Fassbender, Daniel Fister +12Multilingual Language Model Evaluation LLM-as-a-Judge
Section for Clinical Research-IT, Institute of Medical Biometry and Statistics, University of Luebeck, Luebeck, Germany · Medical Faculty, University of Tübingen, Tübingen, Germany · German Medical Students’ Association (bvmd), Berlin, Germany +9
Trade-offs in Medical LLM Adaptation: An Empirical Study in French QA
Jun 17, 2026Ikram Belmadani, Oumaima El Khettari, Carlos Ramisch +3LLM Fine-Tuning Domain-Adaptive Pretraining
Aix-Marseille Univ., CNRS, LIS UMR 7020, 13000 Marseille, France · Nantes Univ., École Centrale Nantes, CNRS, LS2N UMR 6004, 44000 Nantes, France · Grenoble Alpes Univ., CNRS, INRIA, Grenoble INP, LIG UMR 5217, 38000 Grenoble, France
Decision language models as affordable and rapid alternatives to LLM judges of clinical generative AI
Oct 20, 2026Rotem Gershon, Seffi Cohen, Dean Ariel +11LLM-as-a-Judge Clinical Language Model Evaluation
Innovation Division, Clalit Research Institute, Clalit Health Services, Ramat-Gan, Israel. · Stein Faculty of Computer and Information Science, Ben Gurion University of the Negev, Be’er Sheva, Israel. · Department of Family Medicine, Clalit Health Services, Tel Aviv, Israel. +4

Jul 1, 2026J/K move · Enter open · S save

Evaluating Large Language Model Raters for German Open-Response Clinical Questions: A Physician-Annotated Benchmark Study of Agreement, Evaluator Bias, and Abstention

William Philipp, Finn Fassbender, Daniel Fister, Thorsten Langer+11

Section for Clinical Research-IT, Institute of Medical Biometry and Statistics, University of Luebeck, Luebeck, Germany · Medical Faculty, University of Tübingen, Tübingen, Germany · German Medical Students’ Association (bvmd), Berlin, Germany+9

Multilingual Language Model Evaluation LLM-as-a-Judge Selective Prediction Clinical QA Language Model Bias Evaluation Clinical Language Model Evaluation

Open paper AskPDF arXiv

Background: Expert-annotated benchmarks for non-English open-response clinical questions are scarce. LLM-as-a-judge systems may scale evaluation but require validation. Objective: To introduce MedQADE, a standardized German open-response clinical benchmark with physician reference annotations, and evaluate LLM-as-a-judge alignment, self- and intra-family bias, and abstention. Methods: The benchmark contains 3,800 question-answer sets with answers from five student LLMs and annotations from 10 physicians. All 10 rated the 200-question core; two primary raters assessed each of 3,600 extension questions, with the tenth resolving disagreements. Nine LLM evaluators assessed all sets. We assessed physician reliability, student-model accuracy, evaluator alignment, bias, and abstention. Results: Physicians showed moderate-to-substantial agreement on answer correctness (unweighted mean pairwise Cohen's kappa = 0.612) but limited agreement on question difficulty (Krippendorff's alpha = 0.208 using squared numeric-score distances). Student-model accuracy was 17.8%-66.0% and generally decreased with physician-rated difficulty. Gemini 3 Flash approached the leave-one-out physician reference (kappa = 0.694 vs 0.709). Four of five models rated their own responses more favorably than out-of-family evaluators; five of six intra-family comparisons were positive. Physician abstention increased with perceived difficulty. Seven of nine LLM evaluators abstained in no more than 0.51% of evaluations; the two strongest evaluators assigned definitive labels to every response. Conclusions: Strong LLM evaluators approached physician agreement, but evaluator bias and low observed abstention warrant physician validation and further assessment of selective deferral before fully automated evaluation. These results do not establish clinical safety.

Abstract

LLM-as-a-Judge

Explore similar work

Evaluating Large Language Model Raters for German Open-Response Clinical Questions: A Physician-Annotated Benchmark Study of Agreement, Evaluator Bias, and Abstention

Trade-offs in Medical LLM Adaptation: An Empirical Study in French QA

Decision language models as affordable and rapid alternatives to LLM judges of clinical generative AI