PaperlayerView exact revision on arXiv

Source-linked AI summary

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

Rotem Gershon, Seffi Cohen, Dean Ariel, Havatzelet Bilavsky-Yarden, Amiel Sberro, Shai Yitzhaki, Osnat Tausky, Naama Golan, Maayan Rabinovitch, Ariel Avraham Hasidim, Itay Arianu, Maya Dagan, Ran Balicer, Noa Dagan

arXiv:2610.28826v1cs.AI

TL;DR

Physician grading is the reference standard for clinical AI output, but its cost and time limit how often it can be repeated. The study compared Jev with frontier, low-cost and open-weight LLM judges across four clinical AI output evaluation tasks. Across four clinical evaluation tasks, Jev judged clinical AI output as well as or better than several LLM judges at a small fraction of their cost and time.

  • Problem

    Physician grading is the reference standard for clinical AI output, but its cost and time limit how often it can be repeated.

    Source: Discussion · p. 15

  • Method

    The study compared Jev with frontier, low-cost and open-weight LLM judges across four clinical AI output evaluation tasks.

    Source: Discussion · p. 15

  • Results

    Across four clinical evaluation tasks, Jev judged clinical AI output as well as or better than several LLM judges at a small fraction of their cost and time.

    Source: Discussion · p. 15

  • Takeaways & Limitations

    Clinical AI output evaluation may be particularly well-suited to decision models because judgments can be structured around explicit criteria and supplied context.

    Source: Discussion · p. 17

    Jev returns scores and probabilities without a rationale explaining why an answer failed or guiding targeted improvements.

    Source: Discussion · p. 15

Abstract

from arXiv · show

Automated judges of clinical artificial intelligence (AI) outputs are necessary, but large language model (LLM) judges are costly and slow at scale. We compared Jev, a decision language model that returns structured judgments, with frontier and open-weight LLM judges in four use cases. On 118 pediatric answers, Jev's agreement with physicians' ordinal quality ratings (Gwet's AC2 0.50; expected score 0.53) matched or exceeded that of seven of 12 LLM judges, including frontier models, Claude Opus 5.5 (0.49) and GPT-6 (0.38). On physician-graded binary HealthBench criteria, Jev matched or exceeded four of six LLM judges. On MedHELM's open-ended tasks, Jev's agreement with a three-LLM jury approached or exceeded the jurors' internal agreement. On HealthBench Professional, Jev matched the benchmark's reference judge, GPT-5.4, on 91% of criteria (other judges, 91-93%). Frontier judges cost 73-294 times more than Jev and took 7-57 times longer. Jev may enable efficient clinical AI evaluation at scale.

Results

Across four clinical evaluation tasks, Jev’s agreement with reference raters was comparable to or better than several LLM judges, while costing and taking less time. Results also show stable repeated ratings, modest gains from deferral or hybrids, and differences across rubric types.

Stability of Jev’s ratings across repeated runs in the four use cases

Use case 1: agreement of Jev and the LLM judges with physicians on 118 pediatric answers

  • For total score, Jev’s agreement with physicians was AC2 0.50 for its most probable level and 0.53 for its expected score; under either reading, it matched or exceeded several LLM judges.It matched Claude Opus 5.5 (0.49), MiMo-V2.6-Pro (0.50), the published run of gpt-oss-120b (0.50) and o3 (0.55), and exceeded GPT-6 (0.38), Claude Sonnet 4.5 (0.41) and GPT-5 (0.44).Source: Use case 1: agreement of Jev and the… · pp. 4–5

Use case 2: agreement of Jev and the LLM judges with physicians on HealthBench

  • On HealthBench, Jev’s macro F1 with physicians was 0.665, level with gpt-oss-120b (0.667) and above GPT-5-nano (0.627), but below MiMo-V2.6-Pro and GLM-5.3.On the stratified half sample, Jev scored 0.664, versus 0.578 for Claude Opus 5.5 and 0.535 for GPT-6.Source: Use case 2: agreement of Jev and the… · p. 8

Use case 3: agreement of Jev with clinicians and the MedHELM LLM jury

  • On 56 clinician-rated MedHELM responses, Jev’s expected score agreed with clinicians at ICC 0.49, near the jury’s 0.47, while its most probable level scored 0.31.The expected score ranked third among the seven graders, after GLM-5.3 and Claude Opus 5.5 (0.52 each).Source: Agreement of Jev, the MedHELM LLM jury… · p. 10

Agreement of Jev with each juror of the MedHELM LLM jury on 45,384 responses

  • Across 45,384 MedHELM responses, Jev agreed with the jury nearly as closely as jurors agreed with one another, with expected-score AC2 of 0.87 versus 0.79 on completeness.Jev’s expected score also agreed more closely on accuracy (0.92 versus 0.90), and model-ranking correlations with the jury ranged from 0.90 to 0.99.Source: Agreement of Jev with each juror of… · p. 11

Use case 4: agreement of Jev with GPT-5.4 on HealthBench Professional

  • On HealthBench Professional, Jev and GPT-5.4 gave the same label on 90.8% of 1,135 criteria, level with gpt-oss-120b (90.9%) and within 2.5 percentage points of every other judge.The release has no physician labels, so comparisons use GPT-5.4, its default grader, as the reference.Source: Use case 4: agreement of Jev with… · p. 12

Cost and time of Jev and LLM judges across the four use cases

  • Jev cost US$22 to US$89 per million judgments; frontier judges cost 73 to 294 times as much and took 7 to 57 times longer under comparable timing conditions.Jev’s median time per request was 0.12 to 0.55 s; open-weight judges cost 1.1 to 163 times as much and took 31 to 190 times longer.Source: Cost and time of Jev and LLM judges… · p. 12

Performance-cost trade-off of hybrid evaluation by Jev and an LLM judge

  • At the 0.4 cutoff, hybrids added at most 0.04 to Jev’s agreement in use cases 1 and 2 and 1.0 to 1.8 percentage points in use case 4, at 1.2 to 76 times Jev’s cost.In use case 3, the jury hybrid added less than 0.01 to accuracy agreement, while no hybrid agreed detectably better than Jev’s expected score on the 56 clinician-rated responses.Source: Performance-cost trade-off of hybrid… · p. 14

Discussion

Across four clinical evaluation tasks, Jev matched or exceeded several LLM judges while costing substantially less than frontier judges. Its results also show task-specific weaknesses and practical trade-offs, including limited failure explanations and possible benchmark-data overlap.

Discussion

  • Across four tasks, Jev judged clinical AI output as well as or better than several LLM judges, while frontier judges cost 73–294 times more.Jev’s median time was 0.12–0.55 seconds per request.Source: Discussion · pp. 15, 17
  • Jev fell short on faithfulness, medical reasoning and actionability in use case 1; long contexts and multi-step reasoning may contribute to these shortfalls.In use case 4, low-confidence judgments can be set aside or routed elsewhere; the expected score preserves uncertainty between adjacent levels that the most probable level discards.Source: Discussion · p. 16
  • Routing only the safety rubric to Claude Opus 5.5 or GPT-6 raised unsafe-answer detection from 31% to 49% or 51%, respectively, at a third of their cost.The authors caution that the performance–cost trade-off should remain realistic.Source: Discussion · p. 16
  • On 56 clinician-rated MedHELM responses, Jev’s expected score raised ICC with clinicians from 0.31 to 0.49 and completeness agreement with the jury to 0.87, versus 0.79 between jurors.Because both readings come from the same request, the expected score adds no cost or time.Source: Discussion · p. 16
  • Frontier judges sometimes agreed less closely with physicians than weaker or older judges, supporting selection by measured clinician agreement rather than general capability.Why frontier judges judged fewer criteria met was not tested.Source: Discussion · pp. 16–17
  • The study’s evidence is bounded by one institution for use case 1, possible training overlap with public data in use cases 2–4, and a MedHELM jury that may favor its own models.Jev returns scores and probabilities without rationales that could explain failed responses or guide targeted improvements.Source: Discussion · p. 17

Methods

The study compared Jev with LLM judges across four clinical evaluation settings, using physician ratings as the reference in the first two. Judge selection and evaluation procedures varied by use case.

Study design

  • The four use cases covered pediatric decision-support answers, physician-graded HealthBench conversations, open-ended MedHELM tasks and HealthBench Professional clinician tasks.Physician ratings were the reference in use cases 1 and 2; use case 3 also compared Jev with the jury, and use case 4 used GPT-5.4 as reference.Source: Study design · p. 17

LLM judges

  • Five judges graded every use case: GPT-6 and Claude Opus 5.5, plus MiMo-V2.6-Pro, GLM-5.3 and gpt-oss-120b.Each use case added judges specific to its setting, including the published judges, the MedHELM jury, GPT-5-nano or GPT-5.4.Source: LLM judges · p. 18
  • LLM judges received the use case’s original prompt template without labels, with HealthBench outputs parsed using its published parser.Use case 2 evaluated a stratified half-sample for Claude Opus 5.5 and GPT-6, while GPT-5-nano and open-weight models graded all rows.Source: LLM judges · p. 18

Use case 1: pediatric decision-support answers (published in npj Digital Medicine16)

The first two use cases evaluated Jev on pediatric answers and HealthBench conversations using physician judgments as reference. Jev applied structured questions adapted from each evaluation’s grading instructions.

Use case 1: pediatric decision-support answers (published in npj Digital Medicine16)

  • Five physicians rated 118 pediatric decision-support answers on six rubrics and a total score from 1 to 5.Answers were in Hebrew and had senior-physician reference answers; Jev and the new judges rated English translations.Source: Use case 1: pediatric decision-support… · p. 18

Jev in use case 1

  • Jev evaluated each typed question independently and returned probabilities for permitted answers, using seven ordinal questions and six binary questions for each pediatric answer.Ordinal instructions retained the original rubric definitions and calibration examples; binary answers passed at probability of yes ≥ 0.5, and analyses used the first of three runs.Source: Jev in use case 1 · pp. 18–19

Use case 2: physician-graded health conversations (HealthBench, a public benchmark)

  • HealthBench paired multi-turn conversations and model responses with 34 consensus criteria, judged by two to five physicians across 29,511 rows.Jev received unlabeled conversations and final responses, answered one yes/no question per criterion using the published grader instructions, and majority labels from three runs were used.Source: Use case 2: physician-graded health… · p. 19

Use case 3: open-ended medical tasks graded by an LLM jury (MedHELM, a public benchmark)

The MedHELM evaluation compared Jev’s ratings with an LLM jury and clinician ratings on a subset of responses. The same passage set also describes HealthBench Professional, which lacks physician labels.

Use case 3: open-ended medical tasks graded by an LLM jury (MedHELM, a public benchmark)

  • MedHELM’s three-LLM jury scored open-ended medical responses from 13 models, and Jev graded 45,385 responses with all nine jury scores.Jev used the jury’s prompt blocks and five-level score questions for each rating axis.Source: Use case 3: open-ended medical tasks… · pp. 19–20
  • On 56 clinician-rated responses, the analysis compared Jev and jury scores with clinician means using the consistency intraclass correlation ICC(3,2).The 56 responses came from ACI-Bench and MEDIQA and were rated by 13 clinicians; confidence intervals used bootstrap resampling.Source: Use case 3: open-ended medical tasks… · p. 20

Use case 4: real clinician tasks graded by GPT-5.4 (HealthBench Professional, a public benchmark)

  • HealthBench Professional contains 525 clinician tasks with 1,135 physician-written criteria and no released grader outputs or physician labels.Jev and LLM judges graded physician-written responses against their criteria, with GPT-5.4 serving as the benchmark’s default grader and reference.Source: Use case 3: open-ended medical tasks… · p. 20

Rubric groupings

The appendix groups evaluation criteria by the kind of judgment required and defines two readings of Jev’s ordinal outputs. It also specifies how those readings enter agreement analyses.

Rubric groupings

  • Faithfulness, completeness and conciseness require text comparison, whereas actionability, medical reasoning and safety require clinical inference.The former compare answers largely with context and reference answers.Source: Rubric groupings · p. 21
  • HealthBench’s 34 criteria divide into 16 knowledge criteria and 18 behaviour criteria.Knowledge criteria concern medical accuracy, safety or appropriateness; behaviour criteria concern context seeking, hedging, detail, tailoring and instruction following.Source: Rubric groupings · p. 21

Two readings of Jev’s ordinal scores

  • For each 1-to-5 score, Jev returns a probability for every level, allowing both a most-probable-level reading and an expected-score reading.Ties in the most probable level are resolved to the lower level.Source: Two readings of Jev’s ordinal scores · p. 21
  • Jev (expected) is the sum of levels weighted by their probabilities, plus 1 because Jev numbers levels from 0 to 4.The expected score can be used unrounded or rounded, depending on the analysis.Source: Two readings of Jev’s ordinal scores · p. 21
  • AC2, exact and within-one agreement analyses use expected scores rounded half-up, while rank and correlation analyses use unrounded scores.Mean ratings in Supplementary Table 3 use rounded expected scores.Source: Two readings of Jev’s ordinal scores · p. 21
  • Binary questions have one reading: expected value is probability of yes, and rounding it gives the most probable label.Both ordinal readings came from the same Jev outputs, analysis code, random seeds and bootstrap resamples.Source: Two readings of Jev’s ordinal scores · p. 21

Agreement metrics

The authors used Gwet’s agreement coefficients for ordinal and binary judgments, comparing judges with physician references and with one another. Metrics varied by use case and included class-sensitive performance measures and task-score correlations.

Agreement metrics

  • Gwet’s AC2 with quadratic weights measured ordinal ratings, while Gwet’s AC1 measured binary ratings.The authors selected these measures because ratings in use case 1 clustered at the top of the scale and at pass, where other agreement coefficients are unstable.Source: Agreement metrics · p. 21
  • Each judge was compared with the physician panel’s median rating, while physician agreement used each physician’s comparison with the median of the other four.The panel median was rounded up when a missing rating left an even number of physician ratings.Source: Agreement metrics · p. 21
  • Binary use case 1 comparisons used the majority of five dichotomized physician ratings as reference and reported AC1, sensitivity, specificity, and predictive values with failure as positive.Bias was the judge’s mean rating minus the panel’s; additional analyses included exact and within-one agreement and Spearman correlation.Source: Agreement metrics · pp. 21–22
  • Use case 2’s primary metric was macro-averaged F1, averaging met and not-met class F1 scores across grader–physician label pairs.The physician reference applied the same metric to every ordered pair of physicians labeling the same row, pooled across pairs.Source: Agreement metrics · p. 22
  • In use case 3, Jev’s AC2 with each juror and Spearman correlations were averaged and compared with the corresponding mean across juror pairs.Model ranking correlated Jev’s mean rating with mean jury score across 13 models within each benchmark set, excluding the three juror models.Source: Agreement metrics · p. 22
  • Use case 4 reported label agreement, macro F1, AC1, class-sensitive metrics, task-score correlation, and stability across runs.Task scores were met-criteria points divided by positive points and clipped to 0–1; stability used AC2 in use cases 1 and 3 and identical-label share across three runs in use cases 2 and 4.Source: Agreement metrics · p. 22

Comparisons between judges

Judge comparisons used paired bootstrap intervals, with shared inputs or tasks potentially making some intervals too narrow. The authors also tested confidence-based deferral and hybrids that route low-confidence judgments to an LLM judge.

Comparisons between judges

  • Judge differences were assessed using paired bootstrap resamples, with detectability defined by a two-sided 95% confidence interval excluding zero.Resampling used answers, conversations, clinician-rated responses, or criteria depending on the use case; intervals for MedHELM and HealthBench Professional may be too narrow because inputs or tasks were shared.Source: Comparisons between judges · p. 22
  • Per-rubric, per-model, per-theme, and rubric-group comparisons were not adjusted for multiplicity.Differences were recomputed within each resample, pairing raters on the same items.Source: Comparisons between judges · p. 23

Deferral and hybrid

  • For deferral tests, judgments were kept when confidence met a cutoff from 0 to 0.9, and comparisons examined agreement on kept versus deferred judgments.Confidence measures differed by use case, so score-confidence cutoffs and |2p − 1| cutoffs are not comparable; hybrids deferred judgments below 0.4.Source: Deferral and hybrid · p. 23
  • Hybrids kept Jev’s first-run judgment above the cutoff and otherwise substituted a first-run LLM rating, with fallback choices varying by use case.Hybrid cost combined Jev’s cost for all judgments with LLM cost for deferred judgments; estimated time combined Jev’s median request time and the LLM judge’s time weighted by the deferred share.Source: Deferral and hybrid · p. 23
  • Use case 4 timing comparisons were limited to GPT-5.4 because Jev and GPT-5.4 ran at 16 concurrent requests while other judges ran at six.The use case 2 published graders could not be fallbacks because their row-level labels were unavailable.Source: Deferral and hybrid · p. 23
  • In the safety-only use case 1 test, routing binary judgments at a 0.2 cutoff deferred one-third of safety judgments.A 0.4 cutoff would have deferred 58%; the evaluated outcome was sensitivity on answers the physician majority failed on safety, alongside relative cost.Source: Deferral and hybrid · p. 24

Cost and runtime

Costs were estimated from billed tokens or published-prompt tokenization, and runtime was measured as client-side end-to-end request time under specified concurrency conditions.

Cost and runtime

  • Billed input and output tokens were converted using October 2026 prices, with batch pricing for specified models and OpenRouter-reported costs for open-weight models.Standard prices were used otherwise; supplementary material listed standard-price comparisons.Source: Cost and runtime · p. 24
  • For seven published judges, input-token costs were estimated by tokenizing original prompts with each answer at list prices and exclude output and reasoning tokens, making them lower bounds.The tokenizer was approximate for non-OpenAI models.Source: Cost and runtime · p. 24
  • Runtime was client-side end-to-end wall-clock time per synchronous request, measured from a central-US cloud container under use-case-specific concurrency and sampling conditions.Use case 2 timing used 150 random rows for most judges, while GPT-5-nano answered 36 of 100 sequential requests and Jev was also timed on 500 sequential requests.Source: Cost and runtime · p. 24
  • The MedHELM jury’s unreported cost was estimated by tokenizing its published prompts and outputs at list prices, including a stated model-rate proxy for GPT-4o.The estimate used Amazon Bedrock’s on-demand rate for Llama 3.3 70B.Source: Cost and runtime · p. 24

Software

The authors used custom implementations of agreement and F1 metrics and checked them against published examples or reference functions. The use cases relied on waived or publicly released data, and Claude assisted with drafting and analysis under author review.

Software

  • Custom implementations of AC1, AC2, and HealthBench macro F1 were checked against published examples, the irrCAC Python package, or HealthBench reference functions.The AC2 implementation reproduced Gwet’s published two-rater example of 0.9402.Source: Software · p. 24

Ethics

  • Use case 1 had waivers of formal institutional approval and informed consent; its physician-written queries contained no patient data.Use case 2 was publicly released and anonymized, use cases 3 and 4 were publicly released, and no new human participants took part.Source: Ethics · p. 25

Use of AI tools

  • Claude assisted with manuscript drafting and analysis, and the corresponding author reviewed and validated all AI-generated content.The corresponding author takes full responsibility for the manuscript content.Source: Use of AI tools · p. 25

Data availability

The HealthBench meta-evaluation data and HealthBench Professional data are openly available; MedHELM responses and jury scores, plus clinician ratings, are also publicly available. Use case 1 ratings and study-generated judge outputs are available on request or via Zenodo.

Data availability

  • HealthBench meta-evaluation and HealthBench Professional data are openly available; use case 1 physician and LLM ratings are available from the corresponding author on reasonable request.Source: Data availability · p. 25
  • MedHELM v4.0.0 responses and jury scores are openly available from a public bucket, and clinician ratings are available through the MedHELM repository.Source: Data availability · p. 25
  • Jev and LLM judge outputs generated in this study are available at Zenodo.Source: Data availability · p. 25

Extended Data

The extended data describe confidence-based deferral and hybrid judging across use cases, alongside tables listing the study’s use cases and raters.

Extended Data

  • Extended Data Fig. 1 plots agreement and retained judgment shares as Jev confidence cutoffs rise in use cases 1, 2 and 4.The figure uses a 0.1 cutoff grid and defines confidence differently across use cases.Source: Figure 1 · p. 29
  • Extended Data Fig. 2 compares agreement and cost when low-confidence Jev judgments are delegated to an LLM judge across four use cases.A cutoff of zero represents Jev alone; hybrid costs include Jev on every judgment.Source: Figure 2 · p. 30
  • Extended Data Table 1 lists the study’s use cases, raters and analyses.Source: Table 1 · p. 31, Extended Data · p. 31

Use case 4: real clinician tasks graded by GPT-5.4 (HealthBench Professional, a public benchmark)

Use case 4 evaluates Jev and LLM judges on clinician-task criteria using GPT-5.4 as the reference. The supplied results describe run stability and judging cost and runtime.

Use case 4

Related papers

Loading 2610.28826v1…