The Illusion of Readiness: Stress Testing Large Frontier Models on Multimodal Medical Benchmarks (2509.18234v1)

Abstract: Large frontier models like GPT-5 now achieve top scores on medical benchmarks. But our stress tests tell a different story. Leading systems often guess correctly even when key inputs like images are removed, flip answers under trivial prompt changes, and fabricate convincing yet flawed reasoning. These aren't glitches; they expose how today's benchmarks reward test-taking tricks over medical understanding. We evaluate six flagship models across six widely used benchmarks and find that high leaderboard scores hide brittleness and shortcut learning. Through clinician-guided rubric evaluation, we show that benchmarks vary widely in what they truly measure yet are treated interchangeably, masking failure modes. We caution that medical benchmark scores do not directly reflect real-world readiness. If we want AI to earn trust in healthcare, we must demand more than leaderboard wins and must hold systems accountable for robustness, sound reasoning, and alignment with real medical demands.

• The paper shows that high benchmark scores may mask models’ reliance on superficial cues instead of genuine multimodal reasoning.
• It employs stress tests such as modality sensitivity, distractor perturbation, and visual substitution to expose vulnerabilities in model performance.
• The study recommends clinician-guided rubrics and dynamic evaluation methods to ensure more reliable and clinically meaningful health AI systems.

Stress Testing Large Frontier Models on Multimodal Medical Benchmarks: Revealing the Illusion of Readiness

Introduction

This paper presents a comprehensive evaluation of large multimodal models (LMMs), including GPT-5 and other flagship systems, on widely used medical benchmarks. The authors argue that high leaderboard scores on these benchmarks do not equate to real-world readiness, as models often exploit superficial patterns and shortcut learning rather than demonstrating robust medical understanding. Through a series of targeted stress tests and clinician-guided benchmark profiling, the paper exposes critical fragilities in current evaluation paradigms and proposes a framework for more trustworthy assessment of health AI systems.

Stress Test Methodology and Findings

Modality Sensitivity and Necessity

The authors designed stress tests to probe the robustness of LMMs under conditions that simulate real-world uncertainty and incomplete information. Two key tests—modality sensitivity and modality necessity—demonstrate that models retain much of their accuracy even when essential inputs (e.g., images) are removed. For example, on the NEJM benchmark, GPT-5's accuracy dropped only 13.3 percentage points when images were withheld, and models performed well above chance on visually required questions even without images. This indicates reliance on dataset artifacts, frequency priors, and memorized associations rather than genuine multimodal reasoning.

Shortcut Pattern Disruption

Further stress tests involved perturbing answer formats and distractor options. Reordering answer choices in text-only settings led to consistent accuracy drops, revealing that models exploit positional biases and learned response patterns. Replacing distractors with irrelevant options or the token "Unknown" also affected performance, with models treating "Unknown" as a weakened distractor rather than a legitimate abstention signal. These behaviors highlight brittle dependencies on superficial cues and elimination heuristics.

Visual Substitution

A critical test involved substituting the original image with one corresponding to a distractor answer. Top-performing models exhibited sharp declines in accuracy (e.g., GPT-5: -31.6 pp), demonstrating that their predictions are often driven by learned visual-answer pairings rather than robust visual-textual integration. This exposes a fundamental limitation in current multimodal medical benchmarks, which may reward shortcut strategies over true visual reasoning.

Reasoning Signal Integrity

Chain-of-thought (CoT) prompting and manual audits of model-generated explanations revealed that explicit reasoning scaffolding does not consistently improve accuracy. Models frequently produce plausible but factually incorrect rationales, hallucinate visual findings, and amplify initial misinterpretations. Even when reasoning appears structured, it often lacks functional linkage to the final answer, undermining the reliability of current reasoning signals in medical tasks.

Benchmark Profiling and Implications

Clinician-Guided Rubric

The paper introduces a structured, clinician-guided rubric to profile benchmarks along axes of reasoning complexity and visual dependency. Analysis of six representative benchmarks reveals substantial heterogeneity: NEJM tasks require high reasoning and visual complexity, while JAMA is largely text-solvable. VQA-RAD and PMC-VQA are visually dependent but low in inference complexity. This heterogeneity explains benchmark-specific failure modes and underscores the risk of treating all benchmarks as interchangeable measures of readiness.

Diagnostic Use of Benchmarks

The authors advocate for treating benchmarks as diagnostic instruments rather than performance goals. Leaderboard scores should be disaggregated by medically meaningful axes, and benchmark selection should reflect intended deployment contexts. Without such practices, benchmark-driven optimization risks reinforcing narrow capabilities and masking critical fragilities.

Recommendations for Evaluation Reform

Stress Testing as Core Evaluation

Static benchmark scores are insufficient for assessing real-world readiness. The paper proposes modular stress testing as a foundation for trustworthy evaluation, targeting vulnerabilities such as neglect of visual input, dependence on spurious answer patterns, and overconfident reasoning. This approach enables systematic comparison across models and tasks, revealing brittleness that static metrics obscure.

Metadata and Reporting Standards

Benchmarks should be accompanied by metadata characterizing their reasoning and visual complexity. Evaluation protocols must incorporate adversarial and stress-based assessments, especially for high-stakes medical applications. The field should shift its mindset to view benchmarks as tools for assessment, not endpoints for optimization.

Implications for Model Development and Deployment

The findings have significant implications for both theoretical and practical aspects of AI in medicine. The demonstrated fragilities challenge the assumption that benchmark success signals medical viability. Real-world medical decision-making requires models that are robust to missing or noisy data, capable of justifying decisions in clinically interpretable terms, and able to reason across modalities and contexts. Advances in model architecture must be matched by advances in evaluation methodology to ensure alignment with clinical realities.

Future Directions

The paper suggests several avenues for future research:

• Development of benchmarks with explicit documentation of reasoning and visual demands.
• Integration of stress testing into model development pipelines.
• Exploration of training strategies that mitigate shortcut learning and enhance robustness.
• Investigation of abstention and uncertainty quantification mechanisms for safer deployment.

Conclusion

This paper provides a rigorous analysis of the limitations of current multimodal medical benchmarks and the models evaluated on them. Through targeted stress tests and clinician-guided profiling, the authors reveal that high leaderboard scores often mask brittle, shortcut-driven behavior. The proposed framework for stress testing and benchmark profiling offers a path toward more reliable and clinically relevant evaluation of health AI systems. Progress in medical AI must be measured not by artificial test scores, but by robustness, sound reasoning, and alignment with real-world medical demands.