Beyond Correctness: Validity-Oriented Evaluation of Biomedical LLM Judges

Rodrigo de Oliveira; Federico Pittino; James Gwinnutt; Jay Nanavati

cs.AIAug 29, 2026

Beyond Correctness: Validity-Oriented Evaluation of Biomedical LLM Judges

Authors: Rodrigo de Oliveira, Federico Pittino, James Gwinnutt, Jay Nanavati

Organizations: Independent Researcher∗, London, UK · IQVIA, Barcelona, Spain · IQVIA, Reading, UK · IQVIA, Cambridge, UK

Abstract

We propose a scalable, validity-oriented pipeline for evaluating biomedical LLM judges when high-quality human judgments are scarce. First, we augment existing human-labelled biomedical benchmarks with deterministic, metric-grounded mutations that produce auditable preference pairs. Second, we evaluate judges beyond aggregate correctness using three deployment-relevant dimensions: correctness against metric-derived gold labels, robustness under repeated stochastic sampling, and compliance with the requested output format. We use this pipeline to assess Llama-3.1-8B-Instruct under four regimes: (1) base, using the instruct model as is; (2) SFT, distillation-based supervised fine-tuning only; (3) RL, GRPO-based reinforcement learning only; and (4) SFT $\rightarrow$ RL, SFT followed by RL. The base and single-stage regimes struggle on structured medical discrimination such as PICO extraction and clinical calculations, whereas SFT $\rightarrow$ RL performs best across correctness, compliance, and robustness; gains concentrate on decomposable tasks (PICO, MedCalc), at times matching or outperforming frontier models.

Figures & tables

Figure 1: Response rates (correct, invalid, and incorrect) by dataset . Each dataset compares base, tuned, and frontier judge models ( $n$ in legend). Variance whiskers show robustness of correct response rate (Eq. ( 6 )). Exact values in Appendix C .

Figure 2: Response rates (correct, invalid, and incorrect) by rank-gap difficulty . Decomposable items only (PICO + MedCalc), stratified by Hard , Medium , and Easy tertiles of $\delta$ per source example (hard = smallest gaps; $n$ in legend). Variance whiskers show robustness of correct response rate (Eq. ( 6 )). Exact values in Appendix C .

Appendix figures & tables5 assets

Supplementary material from the paper’s appendix.

Appendix

Suite	Type	$y$	$o$	$m$	$s_{k}$
PICO	Decomposable	Span sets (P/I/C/O)	Add/delete spans	macro-F1 exact	$[0,1]\uparrow$
MedCalc	Decomposable	Scalar (clearance)	Perturb formula inputs	MSE	$\geq 0\downarrow$
QA tasks	Atomic	Multiple-choice label	Swap wrong option	boolean	$\{0,1\}$

Appendix

Table 1: Task-specific operators, metrics, and ground-truth types (Algorithms 1 and 2). MedQA and PubMedQA use the same atomic row as MedBullets.

	base	SFT	RL	SFT $\rightarrow$ RL	O4-mini	GPT-5	Opus 4.6
Decomposable ( $n{=}1026$ )	0.75	0.79* (+0.04)	0.67* ( $-$ 0.08)	0.84* (+0.09)	0.82	0.87	0.86
PICO ( $n{=}187$ )	0.84	0.88 (+0.04)	0.78 ( $-$ 0.06)	0.94* (+0.10)	0.86	0.94	0.94
MedCalc ( $n{=}839$ )	0.65	0.71* (+0.06)	0.57* ( $-$ 0.08)	0.75* (+0.10)	0.79	0.80	0.79
Atomic ( $n{=}2081$ )	0.76	0.79* (+0.03)	0.69* ( $-$ 0.07)	0.78 (+0.02)	0.91	0.91	0.91
MedBullets ( $n{=}308$ )	0.69	0.72 (+0.03)	0.63* ( $-$ 0.06)	0.73 (+0.04)	0.94	0.93	0.94
MedQA ( $n{=}1273$ )	0.79	0.82* (+0.03)	0.71* ( $-$ 0.08)	0.78 ( $-$ 0.01)	0.96	0.97	0.96

Appendix

Table 2: Pairwise judging correctness with rows labeled Decomposable and Atomic as unweighted family means, indented rows as individual benchmarks, Avg as the unweighted mean across all five suites, Bold as best-in-class among Llama-8B or frontier columns, underlined bold as overall row best, and asterisk as McNemar significance [ 20 ] vs. baseline (

p<0.05

	base	SFT	RL	SFT $\rightarrow$ RL	O4-mini	GPT-5	Opus 4.6
PICO ( $n{=}187$ )	0.14	0.06 ( $-$ 0.08)	0.03 ( $-$ 0.11)	0.06 ( $-$ 0.08)	0.09	0.03	0.02
MedCalc ( $n{=}839$ )	0.07	0.05 ( $-$ 0.02)	0.03 ( $-$ 0.04)	0.02 ( $-$ 0.05)	0.10	0.01	0.02
MedBullets ( $n{=}308$ )	0.05	0.04 ( $-$ 0.01)	0.06 (+0.01)	0.07 (+0.02)	0.11	0.01	0.01
MedQA ( $n{=}1273$ )	0.04	0.09 (+0.05)	0.09 (+0.05)	0.04 (+0.00)	0.07	0.01	0.00
PubMedQA ( $n{=}500$ )	0.08	0.06 ( $-$ 0.03)	0.11 (+0.03)	0.06 ( $-$ 0.02)	0.03	0.01	0.00
Avg	0.08	0.06 ( $-$ 0.02)	0.07 ( $-$ 0.01)	0.05 ( $-$ 0.03)	0.08	0.01	0.01

Appendix

Table 3: Verdict variance is the mean

|\mathrm{high}-\mathrm{low}|

across datasets over majority-vote and any-/all-sample correctness rates (lower is better), with the same columns and formatting as Table 2 .

	base	SFT	RL	SFT $\rightarrow$ RL	O4-mini	GPT-5	Opus 4.6
PICO ( $n{=}187$ )	0.0%	4.5% (+4.5)	0.0% (+0.0)	0.0% (+0.0)	3.7%	0.0%	0.0%
MedCalc ( $n{=}839$ )	5.5%	6.5% (+1.0)	4.7% ( $-$ 0.8)	4.2% ( $-$ 1.3)	0.0%	0.0%	0.0%
MedBullets ( $n{=}308$ )	1.1%	0.0% ( $-$ 1.1)	4.3% (+3.2)	0.0% ( $-$ 1.1)	0.0%	0.0%	0.0%
MedQA ( $n{=}1273$ )	0.4%	0.9% (+0.5)	3.8% (+3.4)	0.7% (+0.3)	12.8%	0.0%	0.0%
PubMedQA ( $n{=}500$ )	0.0%	0.0% (+0.0)	0.0% (+0.0)	0.0% (+0.0)	1.2%	0.0%	0.0%
Avg	2.3%	2.8% (+0.5)	3.6% (+1.3)	1.6% ( $-$ 0.7)	2.2%	0.0%	0.0%

Appendix

Table 4: Invalid-response share is invalid / (invalid + legitimate errors), lower is better, with the same columns and formatting as Table 2 .

	base	SFT	RL	SFT $\rightarrow$ RL	O4-mini	GPT-5	Opus 4.6
Hard ( $n{=}439$ )	0.62	0.65 (+0.03)	0.54* ( $-$ 0.08)	0.69* (+0.07)	0.69	0.72	0.72
Medium ( $n{=}440$ )	0.72	0.79* (+0.07)	0.63* ( $-$ 0.09)	0.83* (+0.11)	0.86	0.88	0.87
Easy ( $n{=}147$ )	0.79	0.86 (+0.07)	0.72 ( $-$ 0.07)	0.90* (+0.11)	0.94	0.94	0.95

Appendix

Table 5: Rank-gap difficulty on decomposable items (PICO + MedCalc), where Hard / Medium / Easy are three tertiles of

\delta

per source example with boundaries over

[1,K{-}1]

that depend on that example’s candidate count

K

(hard = smallest gaps), not fixed absolute integers across tasks, using the same columns and formatting as Table 2 .

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

Same Verdict, Different Reasons: LLM-as-a-Judge and Clinician Disagreement on Medical Chatbot Completeness
Mar 26, 2026Alexandra DeLucia, Heyuan Huang, Sonal Joshi +3LLM-as-a-Judge LLM Reliability
Data Science and AI Institute, Johns Hopkins University
Reassessing High-Performing LLMs on Polish Medical Exams: True Competence or Bias-Driven Performance?
Jun 10, 2026Antoni Lasik, Jakub Pokrywka, Łukasz Grzybowski +7LLM Evaluation LLM Reliability
NASK National Research Institute · Adam Mickiewicz University · ARAAI Poland +4
Accounting for Bias Enables Sustainable LLM Evaluation
Sep 25, 2026Harshita Katoch, David Antony Selby, Gerrit Großmann +1LLM Evaluation Bias in LLM-as-a-Judge
Department of Computer Science, University of Kaiserslautern–Landau (RPTU), Germany · Data Science and its Applications, German Research Centre for Artificial Intelligence (DFKI), Germany

Mar 26, 2026J/K move · Enter open · S save

Same Verdict, Different Reasons: LLM-as-a-Judge and Clinician Disagreement on Medical Chatbot Completeness

Alexandra DeLucia, Heyuan Huang, Sonal Joshi, Mahsa Yarmohammadi+2

Data Science and AI Institute, Johns Hopkins University

LLM-as-a-Judge LLM Reliability Clinical Language Model Evaluation

Open paper AskPDF arXiv

LLM-as-a-Judge frameworks are increasingly trusted to automate evaluation in place of human experts, yet their reliability in high-stakes medical contexts remains unproven. We stress-test this assumption for detecting incomplete patient-facing medical responses, evaluating three rubric granularities (General-Likert, Analytical-Rubric, Dynamic-Checklist) and three backbone models across two clinician-annotated datasets, including HealthBench, the largest publicly available benchmark for medical response evaluation. LLM Judges discriminate complete from incomplete responses at and slightly above near chance (AUC

0.49

0.66

); at the threshold required to recall

90\%

of incomplete responses, clinicians must still review the vast majority of the dataset, offering no triage utility. Even when model and clinician verdicts agree, they rarely cite the same explanation; and when they diverge, false positives stem from over-flagging non-essential gaps while false negatives reflect outright detection failures. These results reveal that LLM Judges and clinicians apply fundamentally different completeness standards; a finding that undermines their use as autonomous evaluators or triage filters in clinical settings.

Abstract

LLM-as-a-Judge

Explore similar work

Same Verdict, Different Reasons: LLM-as-a-Judge and Clinician Disagreement on Medical Chatbot Completeness

Reassessing High-Performing LLMs on Polish Medical Exams: True Competence or Bias-Driven Performance?

Accounting for Bias Enables Sustainable LLM Evaluation