Patient education is a major NLP application, but current evaluation largely focuses on text-only assistance. In radiology, this overlooks an important part of clinical communication: patients often need both a plain-language explanation and a visual aid showing where to look and what it refers to. We introduce MediSketch, a multimodal patient-education task in which an agent must produce an image–text bundle: an educational image and a grounded plain-language explanation. We instantiate this setting with a clinician-curated radiology testbed, a structured multimodal rubric, and a rubric-following judge that evaluates bundle quality rather than text-only correctness. Under a matched image-generation budget, we compare three test-time scaling strategies: parallel sampling, sequential revision, and hybrid search. Across five backbone models, the no-scaling baseline remains weak, while naive parallel search often provides little gain despite using the full budget. Hybrid performs best on 4 of 5 backbones, suggesting that test-time compute works best when exploration and revision are combined. Further analysis identifies weak multimodal self-critique as a major bottleneck: many models stop revising too early and overestimate the quality of their own image–text outputs. MediSketch exposes both an evaluation gap in multimodal patient education and a concrete failure mode in test-time scaling.

A benchmark score credits final answers, but not the route by which an item can be answered. In medical multimodal multiple-choice questions (MCQs), this distinction matters because a correct answer can be supported by the intended image finding or by benchmark-preserved cues in the wording of answers, non-visual clinical text, visible image text, artificial annotations, or device/context artifacts. We call the resulting score-level overinterpretation reasoning inflation. Here, a route is an observable input path that can support answer selection, not a claim about the model’s hidden cognition. Across six medical multimodal MCQ datasets, we separate candidate cues from behavioral evidence through prompt- and image-side audits, modality ablations, and matched repairs that preserve the medical target and answer key. In a 13-configuration open-model panel, full-input accuracy is 62.63%, while text-only and options-only settings achieve 53.96% and 29.71%, respectively. Removing length-gap, absolute/conspicuous, and spatial/prepositional cues lowers accuracy by 6.58, 3.50, and 4.77 percentage points. We also construct MedQA-MM, a 1,000-item shortcut-mitigated subset, where text-only and options-only accuracy fall to 5.21% and 12.33%. This does not imply that models never use images; it shows that medical image-reasoning claims require route-level evidence.

We present CPTCoder, a human-in-the-loop system that predicts standardized medical procedure codes from clinical text. Clinical procedure coding is an extreme multi-label classification problem over a long-tailed space of short numeric identifiers, where a single-digit difference denotes an entirely different procedure. CPTCoder adapts an instruction-tuned LLM with a code-aware vocabulary and constrained decoding that guarantees all outputs are valid codes. To support human review, we derive per-code posterior inclusion probabilities from n-best reweighting, producing interpretable confidence scores that rank predictions and flag uncertain cases. A post-decoding constraint repair step enforces mutual-exclusion rules between conflicting codes. To enable reproducible research in this underexplored setting, we release MIMIC-CPT, a PhysioNet-accessible benchmark of 37,885 expert-cleaned report–code pairs with a deliberately hardened test split: 88% of test examples contain label combinations unseen during training, and over a third include codes with five or fewer training occurrences. We additionally provide 413,085 weakly aligned pairs and evaluate on a separate live dataset from a hospital, which includes out-of-domain radiology reports with billing-expert-verified labels. CPTCoder achieves 0.61 and 0.51 micro-F1 on the hardened MIMIC split and Hospital-298 respectively, outperforming the strongest baseline by 12 and 5 absolute points while reducing digit-level near-miss errors.

Large language models (LLMs) have demonstrated promising performance on medical benchmarks; however, their ability to perform medical calculations, a crucial aspect of clinical decision-making, remains underexplored and poorly evaluated. Existing benchmarks often assess only the final answer with a wide numerical tolerance, overlooking systematic reasoning failures and potentially causing serious clinical misjudgments. In this work, we revisit medical calculation evaluation with a stronger focus on clinical trustworthiness. First, we clean and restructure the MedCalc-Bench dataset and propose a new step-by-step evaluation pipeline that independently assesses formula selection, entity extraction, and arithmetic computation. Under this granular framework, the accuracy of GPT-4o drops from 62.7% to 43.6%, revealing errors masked by prior evaluations. Second, we introduce an automatic error analysis framework that generates structured attribution for each failure mode. Human evaluation confirms its alignment with expert judgment, enabling scalable and explainable diagnostics. Finally, we propose a modular agentic pipeline, MedRaC, that combines retrieval-augmented generation and Python-based code execution. Without any fine-tuning, MedRaC improves the accuracy of different LLMs from 16.35% up to 53.19%. Our work highlights the limitations of current benchmark practices and proposes a more clinically faithful methodology. By enabling transparent and transferable reasoning evaluation, we move closer to making LLM-based systems trustworthy for real-world medical applications.