Judge-dependent safety gains and model-specific helpfulness costs of evidence-sufficiency prompting in clinical LLMs

• URL: http://arxiv.org/abs/2607.18086v1
• Date: Mon, 20 Jul 2026 15:53:02 GMT
• Title: Judge-dependent safety gains and model-specific helpfulness costs of evidence-sufficiency prompting in clinical LLMs
• Abstract summary: We tested whether a structured evidence-sufficiency prompt reduces unsafe overconfident answers.<n>Blinded clinicians characterized the primary judge as a high-sensitivity (1.00), low-specificity (0.55) screen, not a calibrated scaffold gain.<n>Matched controls showed genuine behavior change, not judge circularity.
• Score: 0.0
• License: http://creativecommons.org/licenses/by/4.0/
• Abstract: Background: LLM judges increasingly score whether clinical language models give overconfident answers under incomplete evidence, yet whether a measured "safety gain" reflects real behavior change or the judge's calibration is unresolved. Using a structured evidence-sufficiency prompt as a test case, we asked whether it reduces unsafe overconfident answers, how far that effect depends on the scoring judge, and what it costs in helpfulness. Methods: In a retrospective public-data benchmark (Real-POCQi, HealthBench, MedRBench), four models (GPT-5.5, Claude Opus 4.8, Gemini 3.5 Flash, Grok 4.3) answered a fully paired common panel (1,200 cells) with a standard prompt and the wrapper. The pre-specified endpoint was the paired reduction in unsafe overconfidence scored by the primary judge (GPT-5.4-nano); secondary analyses added a different-family judge (Claude Sonnet 5), a correctness judge, matched scaffold controls, and a blinded three-clinician review. Results: Unsafe overconfidence fell from 49.3% to 24.7%, a paired reduction of 24.7 points (95% CI 21.8-27.7; p<0.001), robust in direction across models and paraphrases. Magnitude was judge-dependent: Sonnet agreed on direction but nearly halved the effect (+13.1 points), with one-directional disagreement. Blinded clinicians characterized the primary judge as a high-sensitivity (1.00), low-specificity (0.55) screen, not a calibrated rate. The gain carried a model-specific helpfulness cost (correct diagnosis 80.3% to 50.3%): near-free for GPT-5.5, near-total for Gemini (-58 points). Matched scaffold controls showed genuine behavior change, not judge circularity. Conclusions: LLM-judged clinical safety effects should be reported as directional and relative, anchored to human review and evaluated jointly with helpfulness, not as calibrated absolute rates. This does not establish clinical deployment readiness.

Related papers

• Does Reasoning Improve Psychological Depth in Large Language Models? It Depends on Who's Judging [41.62710134204231]
  We study the failure mode through psychological depth in short stories.<n>A judge's correlation with human ratings on its development set may not guarantee valid measurement when outputs are closely matched and human preferences are subjective.<n>These results suggest that development-set performance is insufficient evidence for deployment validity on a shifted distribution.
  arXiv  Detail & Related papers  (2026-09-12T07:34:54Z)
• NovGauge: A Fine-Grained Benchmark for Diagnosing LLMs' Capability in Paper Novelty Assessment [54.4265627247104]
  Large language models (LLMs) are increasingly used in peer review at major AI conferences.<n>Existing benchmarks assess novelty as a single holistic score.<n>We present NovGauge, a human-anchored benchmark for fine-grained novelty assessment diagnosis.
  arXiv  Detail & Related papers  (2026-09-10T08:34:56Z)
• Evaluating medical AI under missing information: same-provider judges and human raters change apparent safety [0.0]
  We stress-test four models by deleting half the final user turn in HealthBench conversations, grading responses with a four-provider LLM-judge panel and a blinded clinician-anchored reference.<n>Two evaluator-facing results are robust. First, judge choice materially changes apparent safety.<n>Second, LLM judges are more permissive than clinicians on a blinded 50-item subsample.
  arXiv  Detail & Related papers  (2026-07-21T08:05:35Z)
• Trivial Prompt Reframing Bypasses Safety Guardrails in Googleś MedGemma-4B [0.0]
  Open-weight medical language models are increasingly used as the base of patient-facing and clinician-support applications.<n>We quantify that gap for MedGemma-4B-it under attacks that require no technical sophistication.
  arXiv  Detail & Related papers  (2026-07-09T18:31:43Z)
• CLExEval: A Human-in-the-Loop Framework for Qualitative Evaluation of LLM Clinical Reasoning [37.91882261875461]
  Large Language Models (LLMs) achieve strong results on many medical benchmarks, but their clinical reasoning remains difficult to evaluate reliably.<n>We introduce CLExEval, a human-in-the-loop framework for evaluating LLM clinical reasoning under progressive information masking.
  arXiv  Detail & Related papers  (2026-06-30T12:56:42Z)
• ClinicalBench: Stress-Testing Assertion-Aware Retrieval for Cross-Admission Clinical QA on MIMIC-IV [0.0]
  Reasoning benchmarks measure clinical performance on clean inputs.<n>We evaluate the step before reasoning: retrieval over real EHR notes, where negation, temporality, and family-versus-patient attribution can flip a correct answer to a wrong one.<n>EpiKG carries assertion label and a temporality tag with every fact in a patient knowledge graph, then routes retrieval by question intent.
  arXiv  Detail & Related papers  (2026-05-11T18:47:52Z)
• Benchmarking Multi-turn Medical Diagnosis: Hold, Lure, and Self-Correction [72.89352076103889]
  Large language models (LLMs) achieve high accuracy in medical diagnosis when all clinical information is provided in a single turn.<n>We introduce MINT, a high-fidelity, multi-turn medical diagnosis benchmark comprising 1,035 cases.<n>We uncover three persistent behavioral patterns that significantly impact diagnostic decisions.
  arXiv  Detail & Related papers  (2026-04-06T00:23:10Z)
• From Black Box to Glass Box: Cross-Model ASR Disagreement to Prioto Review in Ambient AI Scribe Documentation [43.148402136307716]
  Cross-model disagreement among heterogeneous ASR systems can act as a reference-free uncertainty signal.<n>We transcribed 50 publicly available medical audio clips with eight ASR systems spanning commercial APIs and open-source engines.<n>Low-agreement regions were enriched for content disagreements, with the content fraction increasing from 53.9% to 73.9% across quintiles of high-risk mass.
  arXiv  Detail & Related papers  (2026-03-02T13:02:13Z)
• PanCanBench: A Comprehensive Benchmark for Evaluating Large Language Models in Pancreatic Oncology [48.732366302949515]
  Large language models (LLMs) have achieved expert-level performance on standardized examinations, yet multiple-choice accuracy poorly reflects real-world clinical utility and safety.<n>We developed a human-in-the-loop pipeline to create expert rubrics for de-identified patient questions.<n>We evaluated 22 proprietary and open-source LLMs using an LLM-as-a-judge framework, measuring clinical completeness, factual accuracy, and web-search integration.
  arXiv  Detail & Related papers  (2026-03-02T00:50:39Z)
• The Value Sensitivity Gap: How Clinical Large Language Models Respond to Patient Preference Statements in Shared Decision-Making [0.6768558752130311]
  Large language models (LLMs) are entering clinical as decision support tools, yet how they respond to explicit patient value statements remains unmeasured.<n>We conducted a factorial experiment using clinical vignettes derived from 98,759 de-identified Medicaid encounter notes.<n>All models acknowledged patient values in 100% of non-control trials, yet actual recommendation shifting remained modest.
  arXiv  Detail & Related papers  (2026-02-12T21:42:21Z)
• EchoBench: Benchmarking Sycophancy in Medical Large Vision-Language Models [82.43729208063468]
  Recent benchmarks for medical Large Vision-Language Models (LVLMs) emphasize leaderboard accuracy, overlooking reliability and safety.<n>We study sycophancy -- models' tendency to uncritically echo user-provided information.<n>We introduce EchoBench, a benchmark to systematically evaluate sycophancy in medical LVLMs.
  arXiv  Detail & Related papers  (2025-09-24T14:09:55Z)
• Evaluating Generative AI as an Educational Tool for Radiology Resident Report Drafting [1.5457333450799497]
  This study evaluates a HIPAA-compliant GPT-4o system that delivers automated feedback on breast imaging reports drafted by residents in real clinical settings.<n>Three common error types were identified: (1) omission or addition of key findings, (2) incorrect use or omission of technical descriptors, and (3) final assessment inconsistent with findings.<n>GPT-4o showed strong agreement with attending consensus: 90.5%, 78.3%, and 90.4% across error types.
  arXiv  Detail & Related papers  (2025-09-22T20:51:09Z)
• Benchmarking GPT-5 in Radiation Oncology: Measurable Gains, but Persistent Need for Expert Oversight [1.0471566053937098]
  GPT-5 is a large language model that has been specifically marketed towards oncology use.<n>On the TXIT benchmark, GPT-5 achieved a mean accuracy of 92.8%, outperforming GPT-4 (78.8%) and GPT-3.5 (62.1%)<n>In the vignette evaluation, GPT-5's treatment recommendations were rated highly for correctness (mean 3.24/4, 95% CI: 3.11-3.38) and comprehensiveness (3.59/4, 95% CI: 3.49-3.69)<n>While hallucinations were infrequent, the presence of substantive errors underscores that GPT-5-generated recommendations require rigorous expert oversight before clinical implementation
  arXiv  Detail & Related papers  (2025-08-29T16:55:25Z)
• Evaluating Large Language Models for Multimodal Simulated Ophthalmic Decision-Making in Diabetic Retinopathy and Glaucoma Screening [37.69303106863453]
  Large language models (LLMs) can simulate clinical reasoning based on natural language prompts, but their utility in ophthalmology is largely unexplored.<n>This study evaluated GPT-4's ability to interpret structured textual descriptions of retinal fundus photographs.<n>We conducted a retrospective diagnostic validation study using 300 annotated fundus images.
  arXiv  Detail & Related papers  (2025-07-02T01:35:59Z)

This list is automatically generated from the titles and abstracts of the papers in this site.

This site does not guarantee the quality of this site (including all information) and is not responsible for any consequences.