medical ai developers

Abstract: Foundation models have shown significant potential across diverse fields and have been applied to medical imaging, but their robustness across patient populations remains unclear. In this work, we investigate foundation models' ability to generalize from adult to pediatric chest X-ray classification. To the best of our knowledge, this is the first benchmark to compare multiple foundation models under a common adult-to-pediatric evaluation setting while also analyzing the effect of adaptation depth. We harmonize VinDr-CXR and PediCXR into a five-label classification task and evaluate six foundation models and two CNN baselines. All models are adapted using adult data only and tested in both the in-domain and pediatric domains using linear probing, partial fine-tuning, and full fine-tuning. The benchmark reveals a clear and consistent trend: all models suffer a substantial performance degradation when transferred from adult to pediatric data, with drops exceeding 10 percentage points. While Macro-AUROC values reach approximately 92% on the adult test set, performance decreases to around 78% on the pediatric domain. Another important finding is that increasing the depth of adaptation does not significantly improve cross-population generalization. Moreover, strong in-domain performance does not translate into better performance on pediatric data.

Abstract: Artificial intelligence systems deployed in high-risk settings increasingly require models that support meaningful human oversight. Concept-based models (CMs) offer a promising framework by enabling users to inspect and intervene on intermediate, human-understandable concepts. However, concept interventions can counterintuitively worsen downstream predictions, a failure mode we formalise as intervention harm. We distinguish harm after all concepts have been intervened on from harm under partial interventions, the more likely practical use case. Modern CMs often achieve high accuracy by encoding, or leaking, additional task-relevant information in their intermediate concept representations. Recent work argues that such leakage can be benign with respect to intervention harm when concept representations satisfy two conditions: sufficiency and localisation. We prove that these criteria are not sufficient to guarantee that concept interventions improve downstream performance. Across controlled and real-world settings, we find that concept supervision incompleteness is associated with increased intervention harm. Importantly, optimising the model when all concepts are intervened on, a previously proposed leakage mitigation objective, reduces endpoint harm but does not reliably prevent harmful partial corrections. Hence, we introduce a complementary monotonic intervention objective that directly penalises increases in task loss along partial intervention trajectories. Combining endpoint and monotonic intervention training substantially reduces the prevalence and magnitude of partial intervention harm while preserving concept and task predictive performance.

Abstract: Vision Transformers deployed in medical imaging exploit spurious correlations, yet the internal mechanisms remain unidentified and impossible to fix. We apply mechanistic interpretability to locate one such mechanism in ViT-Base 16: a single attention head (L8H10) that functions as a dedicated annotation-reader on chest X-rays, attending to corner text rather than lung pathology in 62% of 200 COVID-19 images. Causal validation via inpainting confirms annotation content rather than luminance or edge density drives the behaviour, increasing lung attention 1.52 times upon removal in affected images ($p{<}0.0001$). The finding replicates across NIH ChestX-ray14 (79%), DeiT-Base (88%), and ViT-Large 16 (80%), establishing spurious annotation-attending heads as a consistent property of ViT-family architectures across two datasets and three architectures. Full fine-tuning does not suppress the behaviour and may entrench it: 98.4% of spurious images remain spurious and the spurious rate rises marginally, as the shortcut is reinforced rather than removed. Critically, single-head ablation produces $p{=}0.340$ and multi-head ablation of all three identified carriers produces $p{=}0.532$, meaning a causally verified annotation-reading circuit is completely invisible to output-level evaluation. Mechanistic auditing finds what behavioural testing cannot.

Abstract: Rubrics as Rewards (RaR) extends reinforcement learning to non-verifiable domains by using an LLM judge to score responses against a structured rubric. Prior work has focused on improving the rubric or strengthening the single judge, with comparatively little attention to how different single or multiple judges might be used during RaR training. We first show that judge identity alone produces a substantial spread in scores when training Qwen3-4B-Base on RaR-Medicine and RaR-Science across five frontier and open-weight judges. Building on this, we propose RaRE (Rubrics as Rewards with Ensembles), a novel judge ensemble strategy for RaR with two mechanisms: a soft-Likert reward that uses judge logprobs to weight more confident judges more heavily, and a drop-on-disagreement mask that excludes high-disagreement rollouts from the GRPO update. We find that this design is critical: while a naive average of judge scores fails to outperform the best single judge, RaRE outperforms every single-judge RaR variant on both HealthBench and GPQA Diamond, with per-theme analysis showing that it inherits the complementary strengths of its constituent judges. Our findings establish judge ensembling as a compelling axis for improving RaR.

Abstract: Reliable medical reasoning requires LLMs to dynamically access external biomedical evidence to improve factual trustworthiness and reasoning accuracy. However, indiscriminate retrieval can introduce noise and latency, raising a critical retrieval-timing problem: at each reasoning step, the system must decide whether the model's internal knowledge is reliable or external evidence is needed. We propose PUlSE, a lightweight probing method that uses the model's internal representations as intrinsic signals for step-level retrieval control. Our posterior trajectory analysis reveals that erroneous steps exhibit substantially higher inter-layer volatility than correct ones, suggesting unstable internal knowledge. Based on this finding, PUlSE trains a compact probe on inter-layer geometric features to detect factual errors upon step completion, and uses it as an adaptive trigger for retrieval and step regeneration. Across four medical benchmarks, PUlSE-timed retrieval matches exhaustive per-step retrieval in accuracy while reducing retrieval calls by 72%, and can transfer to an unseen legal domain, indicating potential domain-agnostic generality. Overall, our study provides an initial step toward shifting retrieval-trigger design from surface-level confidence estimates to internal activation dynamics for adaptive access.

Abstract: Building foundation models for MRI has emerged as a promising direction for advancing medical image analysis and improving accessibility of task specific models. Existing approaches primarily focus on effective pretraining strategies and evaluating their performance across diverse downstream tasks. While these methods show performance gains, limited attention has been given to understanding task alignment between pretraining and downstream objectives. Unlike language models, where the two stages are naturally aligned, such alignment is less evident in vision based applications. Motivated by this gap, we explore continual pretraining. Specifically, continually pretraining a pretrained ViT encoder using Low-Rank Adaptation (LoRA) on a prostate segmentation task to enhance task alignment and downstream performance. We further perform layer-wise representational analysis using Centered Kernel Alignment (CKA) to assess representational shifts and their correlation with task outcomes. Our experiments show improved performance in 3 of 6 downstream tasks and comparable results in the others. Our Feature analysis reveals that, for segmentation tasks, continually pretrained encoder features exhibit greater similarity to those of the finetuned encoder than the original pretrained encoder, correlating with observed performance gains.

Abstract: Generating ultrasound reports from multiple images requires aggregating clinical evidence across views, yet archived key frames capture only part of the dynamic examination. Raw-report imitation is therefore misaligned with visual supervision: content that is clinically valid for the full examination may be unverifiable from the images available to a model. This gap creates a clinical behavior alignment problem. A model must preserve visible findings, avoid diagnostic reversals and unsupported completion, and not collapse into conservative templates. We propose CAMEO, a Clinically Aware Multi-image Evidence-grounded Orchestration framework for ultrasound report generation. Stage I learns ultrasound visual-language primitives; Stage II performs Cross-View Evidence Grounding by distilling trusted visible report points into multi-image QA and report-style supervision; and Stage III performs Clinically Aware Preference Alignment using clinical-error-oriented preference pairs. From USReport, we construct USReport-Distilled with 17,670 evidence-grounded paired-image training instances and USReport-Pref with 21,869 preference pairs; we additionally use 25,631 PubMedVision-US ultrasound instruction samples for domain adaptation and multi-image instruction tuning. On the primary USReport-Distilled benchmark, CAMEO improves over EchoVLM from 0.25 to 0.40 BLEU-1, 0.28 to 0.45 ROUGE-1, and 0.27 to 0.43 METEOR, while raising ClinicalScore from 55.02 to 74.20. These results underscore the value of evidence-grounded supervision, clinically aware alignment, and clinically structured evaluation for reliable ultrasound report generation.

Abstract: Recent advances in Vision-Language Models (VLMs) have led to rapid progress in video understanding across a wide range of benchmark tasks. However, existing evaluations largely focus on short-term reasoning, failing to assess a critical capability: maintaining cumulative temporal consistency over extended time horizons. To close this evaluation gap, we introduce HeiCo-FOCUS, a clinically grounded dataset for evaluating long-context video understanding through the task of Foreign Object Contextual Understanding in Surgery. Built on a dataset of Heidelberg Colorectal surgeries, this task requires models to continuously track multiple objects as they are inserted, manipulated, occluded, and removed over procedures lasting up to hours. HeiCo-FOCUS comprises 30,000 visual question answering (VQA) pairs covering five core capabilities: object recognition, temporal grounding, aggregation, event and procedural understanding, and complex reasoning. The dataset was constructed through a rigorous multi-stage annotation pipeline involving large-scale crowd annotation and 39 surgical domain experts to ensure high quality and clinical relevance. To systematically probe model behavior, we introduce a multi-track evaluation framework that progressively increases temporal and contextual demands from single frames to full procedures. Experiments with ten frontier VLMs show that HeiCo-FOCUS tasks are far from solved: only around half of the models clearly outperform a text-only baseline. Across the video tracks, models perform best on event and procedural understanding (mean Accuracy: 56.5% across all models), while temporal grounding remains particularly challenging for all evaluated models (mean Accuracy: 19.7%). We therefore expect HeiCo-FOCUS to serve as a catalyst for the development of models capable of reliable, temporally consistent reasoning over hours-long videos.

Abstract: While state-of-the-art automated models for medical image segmentation achieve high mean performance, they frequently suffer from localized, catastrophic failures that preclude safe clinical deployment, particularly in neuro-oncology. Interactive segmentation frameworks mitigate this by incorporating human oversight, but traditionally impose prohibitive cognitive and temporal workloads by requiring clinicians to manually search for errors. In this project, we present an efficient, Hybrid Structural-Aleatoric Human-in-the-Loop framework for glioma segmentation that bridges the gap between automated baseline performance and surgical-grade precision, achieving sub-2.0 mm HD95 on curated benchmarks while providing safety-net routing for structural failures across real-world clinical data. By extracting voxel-wise Test-Time Augmentation (TTA) uncertainty and applying hierarchical topological filtering, our method proactively isolates high-risk structural anomalies. We comprehensively evaluated our approach on a challenging out-of-distribution clinical stress-test cohort (N = 362). Operating under a simulated Human Oracle, the framework improved the Whole Tumor (WT) Dice score from 0.891 to 0.914 and reduced the 95th percentile Hausdorff Distance (HD95) from 5.82 mm to 4.76 mm. Critically for surgical safety, the system rescued severe boundary failures in the Tumor Core, reducing mean HD95 from 17.96 mm to 14.83 mm (improving absolute TC Dice to 0.356). These spatial rescues were achieved while demanding a median interactive workload of just 11.3% of the target volume. Acknowledging this as a simulated upper bound lacking real-world cognitive friction, the framework nevertheless demonstrates a highly Pareto-efficient pathway for safely deploying clinical AI.

Abstract: Multimodal large language models (MLLMs) are rapidly advancing clinical diagnosis, yet their adaptation pipelines remain anchored to accuracy-based objectives. Clinical data are heavily class-imbalanced: a constant-majority predictor can score above 90% accuracy while being clinically useless. We therefore evaluate and optimize for AUROC, a threshold-free score that ranks positives above negatives and is invariant to class balance. We focus on prompt optimization in MLLMs. Reflective methods such as GEPA use a binary scores matrix with one row per evaluation instance and one column per candidate prompt; cells record per-instance correctness, so the column average is accuracy and drives candidate selection. We introduce pair-level Pareto prompt evolution (Ranking-PE), which replaces each correctness row with a pairwise-ordering row over (positive, negative) instance pairs: the cell is 1 if the candidate scores the positive higher than the paired negative. The column average then equals empirical AUROC (by the Wilcoxon-Mann-Whitney identity). We apply this swap at all three layers the prompt evolution search reads from - the scores matrix that decides Pareto dominance, the per-example feedback to the reflection LM, and final candidate selection - at no extra model calls and with no surrogate loss. Across three diseases on MIMIC, accuracy-based prompt evolution can degrade ranking; Ranking-PE reverses this, beating the accuracy-based recipe by +5.8 AUROC pp on fine-tuned Qwen3-VL-8B and +16.2 pp on MedGemma-4B. Ablations examine each design component and show that a medical-grade visual backbone - via vision-encoder-tuned SFT or medical pretraining - is a prerequisite that prompt search cannot replace - our recipe extends reflective prompt evolution from text-only data to multimodal clinical decision-making.

Abstract: As vision language models are increasingly deployed in clinical diagnosis, understanding how they internally resolve competing visual and textual signals becomes a safety imperative. Existing mechanistic analyses remain confined to unimodal text and offer no explanation for why a single misleading sentence can override a correct image based diagnosis, or why a model commits to a confident answer despite insufficient visual evidence. We find that these two safety risks, arbitration failure where textual context overrides visual grounding and brake failure where the model commits without adequate evidence, are mediated by spatially disjoint attention head populations: arbitration heads form a mid-to-deep wideband reflecting cross-layer evidence competition, while brake heads concentrate in a narrow middle-to-late layer band that regulates evidence sufficiency and abstention behavior. To ground these observations in causal circuitry, we introduce CRAFT, which localizes each failure mode to a minimal causal head set via dual criteria and verifies necessity and sufficiency through temporal probes and Tuned Lens trajectory analysis. Excising arbitration heads sharply reduces conflict following with negligible degradation on clean inputs, while excising brake heads restores appropriate abstention under degraded visual evidence. The two interventions target spatially disjoint head sets and produce distinct corrective effects, underscoring the mechanistic separability of the failure modes. Experiments across multiple medical VQA benchmarks and VLM architectures validate both the localization and interventions, demonstrating that the identified heads causally drive each failure mode and that targeted modulation generalises without retraining. The code is available at GitHub repository.

Abstract: Constantly evolving real-world knowledge necessitates models to be updated continuously. Especially in medicine, as clinical evidence changes over time, outdated knowledge can pose safety risks. Existing evaluations of knowledge integration focus on factual recall, offering limited insight into whether newly integrated knowledge is actually usable. Our benchmark MedKIT (Medical Knowledge Integration and Transfer) provides a granular evaluation of how models integrate and apply knowledge under realistic sequences of clinical updates. Each instance corresponds to a factual update derived from clinical evidence, paired with targeted probes that assess transfer across lexical variation, relational transformations, compositional reasoning, and open-ended operationalization, as well as locality tests for knowledge preservation. Using MedKIT, we conduct a large-scale empirical study of 12 knowledge integration strategies across 5 diverse models, including both general-purpose and medical LLMs. Our results reveal a consistent gap between recall and usable knowledge: while most methods achieve strong gains on the original update task and under lexical variation, relational generalization is limited, and no method yields meaningful improvements on compositional or operational tasks. These findings highlight a fundamental challenge in knowledge integration and position MedKIT as a testbed for developing methods that make newly integrated knowledge more consistently usable across tasks and contexts.

Abstract: Uncertainty estimates tell us how unsure a model is, but not why. Without knowing which parts of an input influences a model's uncertainty, we cannot tell whether that uncertainty score depends on input features that are relevant for the task. We study this problem in randomset classifiers built using pretrained language models. These classifiers assign probability to individual answers and to groups of answers, producing lower and upper probabilities for each answer; The difference between these probabilities, called credal width, is used to represent epistemic uncertainty about an answer arising from limited training data. We propose XU-RS, a framework that attributes an answer's credal width to the input tokens (words or word pieces) supplied to a language model. XU-RS uses Expected Gradients (a standard feature attribution method) to estimate how input tokens contribute to credal width. The proposed framework is evaluated on a MedQA dataset using SmolLM3-3B and Llama-2-7B models, demonstrating that setting the embedding of a token ranked highly by XU-RS to zero (zero-masking) causes larger changes in credal width than zero-masking randomly selected tokens. In addition, we show that normalisation can cause other answer groups to influence an answer's width, reveal how token attribution can mask numerical errors, and provide diagnostic checks to verify whether a token ranked highly by XU-RS meaningfully explains model uncertainty.

Abstract: System One models such as Jev offer an efficient alternative to generative language models for tasks that require decisions rather than open-ended responses. However, existing Jev models exhibit limited Chinese-language decision accuracy, restricting their utility in both general and specialized settings. In this paper, we introduce Chinese-Jev, a System One model that addresses this gap through a unified data processing and training pipeline. Our data processing protocol converts heterogeneous Chinese-language annotations into probability targets over candidate options, enabling a shared training formulation across domains and question formats. To enable efficient inference, Chinese-Jev adopts a lightweight encoder-only backbone for text encoding and learns to score candidate answers through decision-oriented training. To address the misalignment between the pre-training distribution and downstream Chinese-language scenarios, we first train the model on a general-purpose corpus of 10 million examples, then fine-tune it separately for the medical, legal, and financial domains. To evaluate decision accuracy and calibration in both general and domain-specific Chinese-language settings, we introduce Chinese-Jev Bench (CJ-Bench). After first-stage pre-training, Chinese-Jev exceeds the accuracy of the closed-source Jev model by 1.24% on general-domain tasks while achieving a 20.3x speedup. Subsequent domain-specific fine-tuning yields a 4.0% accuracy improvement over Jev in medicine and achieves 92% of Jev's average accuracy across specialized domains, with a 17x speedup and an average latency of only 15 ms per example. We further demonstrate on-device deployment of an INT8-quantized model on mobile devices, achieving an inference latency of approximately 1.0 second per decision. The project is available at https://gulucaptain.github.io/Chinese-Jev/.

Abstract: Agentic reinforcement learning (ARL) with verifiable rewards improves the ability of large language models (LLMs) to tackle knowledge-intensive tasks by learning to interleave search and reasoning. However, most existing ARL methods optimize only LLM-generated tokens and treat retrieved evidence as environment observations. This creates an information-credit gap: failures caused by missing or misleading evidence are attributed to the LLM policy rather than to the retriever, which motivates training the LLM and the retriever jointly. In this paper, we show that retrieval and LLM policy learning are order-sensitive: adapting the retriever before optimizing the policy yields a larger reward gain than the reverse order. To preserve this hierarchy while allowing both components to co-adapt, we formulate retrieval-augmented agentic RL as a bilevel optimization problem. To solve it efficiently, we introduce BRIDGE, a memory-efficient first-order bilevel method motivated by a loss-landscape analysis of the RL and retrieval objectives. Across seven open-domain QA benchmarks, BRIDGE achieves the highest average accuracy with both 3B and 7B backbones, improving the multi-hop average over the strongest baseline by 9.6 and 3.4 EM points, respectively. It also achieves the best averaged answer accuracy and reasoning quality across medical QA benchmarks.

Abstract: Patients often have several co-occurring clinical conditions, and the findings needed to identify and disambiguate them emerge over the course of a consultation. Evaluating clinical reasoning in this setting requires both multi-turn interaction and multi-label diagnosis. We introduce CLIMB, a benchmark in which a doctor model interviews a simulated patient to recover a ground truth set of co-occurring clinical conditions. Cases are synthesized from clinical decision algorithms and diagnostic datasets, grounding multimorbid presentations in structured clinical knowledge. Across six frontier and open models, none recovers the exact set of conditions in more than 10% of interactive cases. Diagnostic performance declines when conditions co-occur, even when models receive the full clinical record and the true number of conditions. Interaction reduces performance further. In controlled experiments, models behave like single-hypothesis trackers: they anchor on the diagnosis suggested by the opening findings, keep questioning around it, and recover a second condition mainly when a finding in view points to it. Questioning them further does not complete the set but adds mostly wrong diagnoses. We formalise this pattern with a theoretical reference model of single-hypothesis tracking. The benchmark, generator, and evaluation code are available at https://anonymous.4open.science/r/CLIMB-8340.

Abstract: Radiology report generation (RRG) increasingly incorporates heterogeneous clinical evidence, such as multi-view radiographs and previous reports, whose availability varies across examinations. However, accommodating different input combinations does not ensure effective evidence use: generated reports may still omit or inaccurately describe clinically relevant findings. To address this problem, we propose PACER, a Progressive Availability-Conditioned Evidence Routing framework for structured incomplete-context RRG that follows a Refine-Calibrate-Commit pipeline. It first refines observed visual representations through endpoint-preserving patchwise routing across frozen encoder depths, incorporating complementary cues while retaining the pretrained terminal representation. It then calibrates the language-model prefix according to the observed evidence and availability state, adapting the shared generator's conditioning as the available source set changes. Finally, it generates polarity-structured clinical commitments before the report in the same autoregressive trajectory, providing structured clinical context for subsequent generation. Experiments demonstrate state-of-the-art clinical efficacy across all four MIMIC-RG4 settings and strong MIMIC-CXR performance, while maintaining competitive language-generation quality.

Abstract: Jev is a non-generative "System One" model that assigns probabilities to predefined answer options and cannot answer outside them. Its accuracy and calibration on medical question-answering and case-based diagnostic-reasoning tasks are unknown. We evaluated Jev 1.13 on four medical benchmarks: MetaMedQA, PubMedQA, DiagnosisArena-MCQ and the NEJM Case Challenges. GPT-6 Sol, with (medium) and without reasoning, was the reference. The primary outcome was top-1 accuracy; key secondary outcomes were calibration, selective prediction and recognition of unanswerable questions. All 8,469 requests returned a valid answer. Jev's accuracy was similar to that of GPT-6 Sol with medium reasoning on PubMedQA (78.4% vs 78.2%;), lower on MetaMedQA (74.8% vs 82.7%) and much lower on DiagnosisArena-MCQ (59.8% vs 82.4%;) and the NEJM cases (61.8% vs 82.4%). On MetaMedQA, Jev's probabilities were the best calibrated (expected calibration error 0.063 vs 0.146), and its answers with a probability of at least 0.9 (52.9% of questions) were 93.4% accurate, but GPT-6 Sol was as accurate when it accepted a similar proportion of questions. On DiagnosisArena-MCQ, Jev's probabilities discriminated poorly (AUROC 0.645 vs 0.768). Of the 162 questions whose correct answer was "I don't know or cannot answer", Jev chose that option for 10.5% (GPT-6 Sol, 8.6%). Median latency was 0.27-0.31 s; all 2,823 items cost USD 0.08. Jev was fast and inexpensive, and its accuracy was similar to that of a frontier LLM on research abstracts but lower on examination questions and much lower on complex diagnostic cases. Task-specific validation is required before clinical use.

Abstract: Medical question answering spans diverse specialties and modalities, and individual medical large language models (LLMs) exhibit distinct strengths across tasks and domains. This heterogeneity suggests that combining specialists may enable broader coverage of medical questions than relying on any single model. However, existing LLM routing methods primarily seek to balance answer quality and inference cost, leaving open how to exploit differences in specialist competence to improve medical reasoning. In this paper, we introduce MedRouter, an agentic system that uses an embedding-based multi-label router to select and query specialist LLMs, then passes their responses to a generator to produce the final answer. We further propose SCALE (Specialist Competence-Aware Learning), a two-stage training framework that first trains the Router with specialist correctness supervision and then optimizes its selections through reinforcement learning. The second stage uses a Performance Gain Reward (PGR) that measures how specialist information affects the generator's answer correctness relative to answering without that information. Experiments on eight text-based and multimodal medical QA benchmarks show that MedRouter outperforms the strongest routing baseline by 8% in average accuracy. Our analysis of specialist outputs further reveals distinct strengths and complementary question-level coverage, motivating learned routing to combine these capabilities for more comprehensive medical reasoning.

Abstract: Multidisciplinary tumor boards integrate multimodal clinical observations and longitudinal patient histories through specialist discussions, yet benchmarks rarely capture these real-world trajectories. We introduce OpenTumorBoard, a benchmark with 611 patient cases and 19,157 discussion turns across ten specialist roles, transcribed from 12,534 minutes of publicly available tumor board recordings on YouTube. The benchmark evaluates two settings: SPECIALIST TURN, in which an LLM responds to a clinically significant question posed during a real discussion, and BOARD SIMULATION, in which it generates an entire back-and-forth discussion and reaches a consensus on therapy recommendations, surgical plans, next actions and clinical trial matching. Evaluation of 14 general-purpose frontier and medical LLMs reveals substantial limitations: the best models score 3.43 out of 5 in clinical equivalence to specialist answers and 2.78 out of 5 in alignment with recorded board conclusions. Supervised finetuning and reinforcement learning improve performance on a held-out test set, suggesting that real-world discussion trajectories can support model adaptation. Three M.D. experts review a subset of the benchmark, finding high information coverage and factuality of patient cases and strong fidelity of extracted consensus conclusions. We will release OpenTumorBoard and its automated curation pipeline to support the development and evaluation of LLMs for multidisciplinary, personalized cancer decision-making.

Abstract: Agents that answer questions from compressed or retrieved memory must recognize when the evidence a query needs is no longer in memory. Benchmarks for this task usually create insufficient-evidence examples by deleting supporting passages. We show that this construction leaks the label through memory size: on MuSiQue, a classifier that only counts paragraphs reaches an area under the ROC curve (AUROC) of $0.979$ for detecting unsafe memory, higher than the lexical estimator we initially evaluated. We propose a size-matched construction that provably removes this shortcut, and use it to study MemSafe, an estimator that cross-encodes the query with each memory unit and aggregates the units with a set transformer. Across three multi-hop question answering datasets and five seeds, MemSafe reaches $0.968$ and $0.983$ AUROC on MuSiQue and HotpotQA, $0.26$ to $0.39$ above a lexical baseline, while the third dataset, 2WikiMultiHopQA, is saturated. A frozen pretrained cross-encoder with a logistic head already closes $41\%$ of the MuSiQue gap between the lexical baseline and MemSafe. At the same time, MemSafe degrades more than a weak baseline on the unanswerable questions released with MuSiQue, reaches only $0.639$ AUROC on SQuAD~2.0, and needs several thousand clinical training examples before it outperforms a feature-based estimator. Used as a gate for a 7B reader, it reduces the error rate on answered questions from $0.850$ to $0.631$ at $5\%$ coverage, outperforming both reader confidence and, on average, the ground-truth integrity label, although a 7B LLM judge is the better gate at $10\%$ coverage. These results indicate that the way insufficient evidence is constructed matters as much as the estimator that detects it.

Abstract: Audiology consultation requires structured history-taking, audiometric interpretation and patient-centred communication, yet real-world case material is scarce. We present a bilingual AI audiologist pairing a general-purpose large language model with rubric-guided playbook induction, multimodal audiogram interpretation and retrieval-augmented grounding, without fine-tuning the language-model backbone. Using a 21-item rubric and an AI patient simulator, we induced a 19-rule consultation policy from 73 training cases (43 English, 30 Chinese) and evaluated the system on 58 independent simulated cases (30 Chinese, 28 English) in a pre-specified, source-blinded comparison with 17 practising audiologists. The AI audiologist outperformed human audiologists on every case (58/58; mean paired $Δ$ = +1.35 on a 5-point composite, Cohen's d = 1.84, $P = 4.5 \times 10^{-20}$), on 20 of 21 rubric items and in both languages. Component ablation identified the playbook as the largest contributor, offering a practical route to specialist consultation agents in low-data medical domains.

Abstract: Medical vision-language model (VLM) post-training is commonly evaluated through answer accuracy. We examine how changes in accuracy and training objectives relate to image-conditioned decisions in a controlled Qwen2.5-VL-3B study on PMC-VQA. We compare supervised fine-tuning (SFT) with low-rank adaptation (LoRA) restricted to the language model, expanded multimodal adaptation scopes, standard answer-only Group Relative Policy Optimization (GRPO), and a counterfactual evidence objective. On 2,000 clean-test questions, language model LoRA SFT changes correct-image accuracy by +1.10 percentage points (95% paired bootstrap CI:-0.85 to +3.05), while visual-benefit events decrease by 2.40 points and image sensitivity decreases by 5.60 points. Paired records reveal 155 acquired and 203 lost visual-benefit events. Broader adaptation yields lower correct-image accuracy than language-model LoRA SFT. Standard GRPO produces mixed-reward groups and parameter updates, with an uncertain clean test accuracy change. A generation audit reveals that canonical option scores can follow a different token path from generated answers. With scores taken along the greedy generation path, the evidence target improves on the training set; its gains over standard GRPO remain inconsistent on validation data at matched training doses. Sample-level analyses trace how evidence scores, decision margins, and generated answers change during post-training. This empirical and measurement audit identifies gaps between optimization activity, target acquisition, and useful held-out visual behavior.

Abstract: Vision-Language Models (VLMs) exhibit remarkable zero-shot generalization, yet they often encode unwanted or hazardous stylistic domains such as idealized textbook diagrams in medical AI or cartoon vehicles in autonomous driving. Approximate Domain Unlearning (ADU) aims to selectively erase a model's recognition of a target visual domain while preserving accuracy on the remaining domains. However, existing ADU methods operate under a flawed closed-vocabulary assumption: they evaluate unlearning solely on the specific object classes seen during the unlearning fine-tuning phase. Consequently, these methods do not unlearn the domain itself; they merely overfit to seen class-domain pairs, leaving the domain easily recognizable for unseen classes and providing a false sense of removal. We argue that true domain erasure must be class-agnostic. To address this, we formalize Open-Vocabulary Domain Unlearning (OVDU), a rigorous protocol that mandates domain forgetting must transfer to held-out classes. To solve the OVDU challenge, we propose a surgical parameter-editing framework. First, a Fisher Information mask isolates domain-sensitive weights, mathematically protecting foundational zero-shot generalization. Second, our Targeted Manifold Scattering (TMS) objective uses preference-based mining to locally scatter the forget domain's stylistic geometry. Evaluated across PACS, OfficeHome, and DomainNet, our method vastly improves open-vocabulary generalization over existing baselines. Crucially, it delivers exceptional sample efficiency, outperforming peak 8-shot baseline results with only 4 shots.

Abstract: Glaucoma is the leading cause of irreversible blindness, and timely identification of fast progressors is essential to prevent disability. Current practice estimates progression by ordinary least-squares regression of mean deviation (MD) on time, requiring 6--10 visual field (VF) tests over several years to obtain a reliable slope. We present GLAM (Glaucoma Longitudinal Analysis Model), a deep learning framework that ingests longitudinal Humphrey 24-2 total deviation sequences with five clinical features and predicts MD and visual field index progression rates using attention-based fusion and aleatoric uncertainty. On the open-access University of Washington Humphrey Visual Field dataset (4,276 patient-eyes), GLAM achieved an MD-rate mean absolute error of 0.139 dB yr$^{-1}$ ($R^2 = 0.927$; 73.5% reduction over a ridge baseline) and an AUC of 0.990 for fast-progressor detection. VF-only deep learning can match multimodal pipelines for progression prognostication using routinely collected perimetry alone.

Abstract: Pathology vision-language foundation models vary in performance across diseases and tasks, with no single model consistently performing best. The high cost of expert pathology annotation can also limit the labeled data available for task-specific adaptation. Combining complementary pretrained representations is a potential approach to these limitations, yet learning an effective fusion from few labeled examples remains challenging. We introduce Few-shot Fusion Foundation Models of Computational Pathology (FFM-CP), which is a framework that combines multiple pathology vision-language models in the few-shot learning setting. The framework first aligns heterogeneous representations using a closed-form Orthogonal Procrustes transformation estimated from corresponding support images. This alignment preserves within-model feature geometry without training an additional alignment network. Within the aligned space, a unified graph enables information exchange across backbones by jointly refining support-image features and visual and textual class prototypes. These refined representations support complementary text-prototype and case-retrieval branches that capture semantic class knowledge and within-class visual variation, respectively. Each branch learns to combine predictions from all ordered backbone pairs, allowing queries encoded by one model to draw on evidence represented by another. We evaluate three backbone combinations on six histopathology datasets at 4, 8, and 16 shots per class. FFM-CP achieves higher mean macro-F1 than the strongest individually adapted member of each fused set in 50 of 54 comparisons. These findings suggest that combining complementary pretrained representations can improve histopathological classification when annotations are limited.

Abstract: An AI-generated radiology report can resemble a physician's report while omitting an abnormality, adding an unsupported finding, or reversing its presence. Measuring these factual differences is essential for evaluating report generators. We study Jev, a System One decision model, as a simple, low-cost judge of agreement with physician-written reference reports. Our evaluator checks whether each statement is supported by the other report and combines these judgments in both directions to capture unsupported claims and omissions. A single-question configuration reaches Kendall correlations of 0.573 on RadEvalX and 0.398 on RadEvalExpert with expert error counts, outperforming an open natural language inference judge under matched decomposition and aggregation. One support question per statement retains similar expert agreement to seven while using 43-45% fewer judgment input tokens. At the documented API price, judgments cost under three cents per hundred report pairs, excluding local decomposition. In a separate controlled-error test, Jev detects false negation with an AUROC of 0.977. Local RadMatch achieves stronger agreement on clinically significant errors in both expert datasets and on total errors in the shared RadEvalExpert subset. Finding-count and error-scope analyses show that benchmark agreement reflects report size and error definitions as well as medical error detection. These results support Jev as a practical judgment component for measuring factual differences in generated radiology reports and identify where more elaborate evaluation remains valuable.

Abstract: Surgical procedures follow a phase-to-step hierarchy, yet the video-language models used to recognize them are evaluated with flat per-level metrics that ignore cross-level coherence and error structure. In this paper we make two contributions to address this problem, (i) we introduce SurgHiBench, the first hierarchy-aware evaluation suite for surgical video understanding, with three tasks measuring recognition, consistency, and severity across granularity levels. We evaluate a general-purpose CLIP model, a Euclidean surgical model, and, as second contribution: (ii) HyperSurg, a new hyperbolic model that enforces phase-step containment via entailment cones, across four (existing) datasets spanning three procedure types. The suite reveals that two models with the same accuracy can produce predictions of very different error severity, ranging from sibling confusions within the correct phase to unrelated cross-phase predictions. Hyperbolic geometry shifts predictions toward the correct procedural neighborhood, and these gains scale with the tree-likeness of each dataset's annotation hierarchy, providing a principled indicator when hierarchy-aware geometry helps.

Abstract: Diffusion-based synthetic data generation offers a promising route for sharing medical imaging data without releasing sensitive patient records. However, generative models face a fundamental tension between privacy and fairness: they may memorize rare training samples, leading to privacy risks, or fail to reproduce underrepresented features, resulting in unfair synthetic distributions. While prior work has largely focused on either memorization or fairness in isolation, their interaction remains insufficiently understood. In this work, we introduce a data-interventional framework to systematically analyze privacy and fairness in diffusion models. We discuss synthetic anatomical fingerprints (SAFs), rare and manually injected image features, as controlled probes to study whether models generalize sensitive attributes across identities, memorize training samples, or suppress rare signals entirely. Across multiple conditioning modalities, we observe a consistent behavior: models either forget these fingerprints or memorize the entire image in which they appear, but do not generalize them to novel images. To support large-scale auditing where explicit sample extraction is infeasible, we further introduce the indicator metric t', which estimates a model's susceptibility to memorization by exploiting the internal structure of the diffusion process. By comparing conditioning signals of varying surprisal, we reveal a clear relationship between conditioning rarity and memorization behavior. Highly surprising conditioning signals act as retrieval keys that amplify memorization, whereas low-surprisal conditioning signals systematically suppress rare features, even when these appear repeatedly in the training data. Our findings provide actionable insights and concrete mitigation strategies for safe and fair synthetic medical data sharing. Code is available at https://github.com/MischaD/Privacy.

Abstract: Foundation model embeddings of screening mammograms may encode pre-diagnostic tissue change without task-specific adaptation. We tested whether embeddings move faster along a data-derived "cancer direction" in women later biopsied for cancer than in matched screen-negative controls, and whether this depends on pretraining domain. We studied 1,773 biopsied women (785 malignant, 988 biopsy-negative) and 1,773 matched controls, each with at least two annual screening exams before their index exam. An identical pipeline was applied to four 2D models: Mammo-CLIP (MC, out-of-distribution mammography), HOPPR (in-distribution mammography), MedImageInsight (MII, general medical imaging), and BiomedCLIP (biomedical vision-language pretraining on literature figures). Breast-level embeddings quantified longitudinal movement along the cancer direction. We compared cases and controls using a between-patient design with complementary mixed-effects analysis, and biopsied versus healthy contralateral breasts within patients. Under matched modality in MII embedding space, malignant cases drifted significantly faster than controls in the first two screening intervals preceding the index exam; biopsy-negative cases showed significance only in the first. MC differences were significant in the first interval for both biopsy groups. Within-patient comparisons showed a broadly similar pattern, with MC significance extending to the second interval in both groups and HOPPR showing significance at interval 1. BiomedCLIP showed no significant differences in either design or biopsy group. Overall, directional embedding velocity emerges as a property of clinically grounded rather than general biomedical pretraining, showing that foundation model embeddings can encode pre-diagnostic mammographic change without task-specific adaptation.

Abstract: Multi-agent language systems often treat agreement as evidence, yet heterogeneous agents can jointly repeat an unsupported claim or omit a correct specialist fact. We introduce C-MoA, an agreement-based conformal filter that turns inter-agent semantic support into a claim-level nonconformity score and calibrates a retention threshold at the example level, giving distribution-free within-domain factuality control for heterogeneous Mixture-of-Agents. C-MoA is effective: it nearly doubles retained-claim precision on long-form generation (from 0.41 to 0.75), certifies a human-labelled medical set, and transfers across domains without recalibration; its one failure mode is short-form answering, where consensus is cheap and the score is left near chance. We then ask whether counterfactual falsifiability can push past consensus, and introduce CONTRA-MoA, which adds a blinded near-miss tournament, leave-one-agent-out stability, and availability-aware fusion. This extension helps only where the verifier holds domain knowledge, dropping half of the false medical claims at 0.940 precision, whereas with a memory-only judge the added signals are near chance (AUC 0.531 and 0.511) and naive max fusion degrades the working agreement signal from 0.687 to 0.652. The message is twofold: agreement-based conformal calibration delivers reliable, transferable factuality control, while moving beyond consensus requires a knowledgeable verifier, availability-aware signals, and robust fusion.

Abstract: Applying large language models to Traditional Chinese Medicine (TCM) prescription generation reveals three clinically critical gaps: models produce end-to-end mappings without auditable reasoning following the li-fa-fang-yao paradigm (SR Gap), treat each encounter in isolation without follow-up adjustment via sui zheng jia jian (LA Gap), and fail to enforce absolute contraindication rules such as Shi Ba Fan (SC Gap). We propose a progressive four-stage framework (SFT $\to$ PG-CoT $\to$ Dynamic $\to$ K-RL) that addresses each gap: PG-CoT constrains CoT distillation under the li-fa-fang-yao paradigm to produce auditable diagnostic chains, Dynamic SFT models patient trajectories with explicit transition reasoning, and K-RL encodes deterministic pharmacological rules as rule-based DPO preference signals. Across 12 fine-tuned models and 6 zero-shot baselines, our framework substantially improves prescription quality over zero-shot baselines---with a 7B model (Mistral-7B) surpassing zero-shot GPT-5 on all three TCM evaluation metrics.

Abstract: Post-training quantization (PTQ) enables efficient deployment of large language models, and PTQ methods are usually optimized and evaluated with generic reconstruction, perplexity, or answer accuracy. But in explanation-critical domains, preserving only the final answer may be insufficient, since users may also inspect generated rationales to judge whether a prediction is trustworthy. We study this issue in medical multiple-choice question answering, where rationales should provide evidence that supports the selected answer. We propose an explanation-aware objective for transformation-based PTQ. Our method builds an offline faithfulness cache from full-precision teacher rationales and uses it during optimization to preserve answer-supporting evidence tokens and evidence-conditioned answer behavior. We instantiate it on OSTQuant under W4A4KV4 quantization and evaluate four 7B--8B medical and instruction-tuned LLMs on MedExQA, MedExpQA, and ChallengeClinicalQA. While a same-calibration OSTQuant baseline preserves task accuracy, it can substantially weaken answer-supporting rationales. Our objective is to preserve the full-precision model's answer-supporting behavior rather than improve gold-label accuracy, and our method better preserves the full-precision model's answer behavior and rationale-to-answer support. These results suggest that PTQ for explanation-critical settings should evaluate preservation of answer-supporting evidence, not only answer accuracy. Code and evaluation scripts are available at https://github.com/dut0817/EAQuant.

Abstract: The proposed pipeline integrates a representation-flow convolutional neural network with a dynamics- and uncertainty-aware tree-based Gaussian Process classifier. In this framework, latent motion dynamics are exploited both as discriminative representations and as a source of input uncertainty, enhancing robustness against temporal variations and abnormal motion transitions. Compared with conventional deep learning approaches, the proposed strategy requires less training data and offers improved computational efficiency. To further improve classification performance, we introduce novel semantic-aware compound kernels that effectively capture semantic, flow, and dynamic information embedded in surgical video features. In addition, uncertainty-aware kernels are developed to strengthen the robustness and practical applicability of the compound kernel framework. The proposed method is evaluated on two benchmark datasets, namely the JIGSAWS and the Cataract-LMM (Capsulorhexis) datasets. Experimental results demonstrate strong performance across both datasets, including the LOSO and LOUO evaluation protocols on JIGSAWS, including the subject-independent LOUO protocol on JIGSAWS, on which the framework attains a mean accuracy of \ph{96.9}\%; results under the within-subject LOSO protocol are reported for comparability with prior work, achieving competitive accuracy while substantially reducing computational cost. Overall, the proposed pipeline provides an efficient and accurate framework for video-based surgical skill assessment.

Abstract: How large language models (LLMs) integrate patient risk with clinical cost tradeoffs remains poorly understood. We investigated how four open-weight LLMs (Qwen-2.5-7B/32B and Llama-3.1-8B/70B) internally represent cost tradeoffs, how these representations relate to clinical predictions, and whether decisions shift as predicted by the specified cost direction and magnitude. Using a public diabetes dataset, we varied 11 false-negative (FN) to false-positive (FP) cost ratios across three phrasings and examined representations and behavioral outputs. Patient risk was linearly recoverable on par with conventional classifiers (AUC $\approx 0.83$), and cost direction was recoverable in every model. However, representational shifts in cost direction tracked output changes only in the two larger models, and responses to cost magnitude were predominantly direction-agnostic. Only 2 of 12 model-phrasings showed both opposing responses to increasing FN versus FP costs and cost-correct ordering. Representationally, a direction fitted on one cost side did not invert when transferred to the other, as expected under mirror-symmetric encoding. These findings suggest that LLMs encode risk and cost information but do not reliably integrate them into cost-correct decisions. Clinical evaluations should therefore include tradeoff tests, phrasing sensitivity, and default operating points alongside predictive performance.

Abstract: A medical model's benchmark score does not establish that the same conclusion holds under a different evaluation. This study tests whether claims about model ranking, score reliability and screening performance survive changes in cohort, prompt, negative spectrum, specified prevalence and operating threshold. We audit three medical vision-language models (BioMedCLIP, CheXficient, and MedSigLIP) and a general-domain OpenCLIP comparator on 12,200 chest radiograph records from four datasets (Montgomery, Shenzhen, TBX11K, and VinDr-CXR). Five fixed prompt families yield 244,000 model--image--prompt scores. No model leads every cohort and reliability criterion. Prompt-family changes alter AUROC in 21 of 48 multiplicity-controlled comparisons. Replacing healthy controls with sick non-tuberculosis controls reduces AUROC by 0.075--0.306 across all four models. On VinDr-CXR, the three medical models distinguish tuberculosis from no-finding controls substantially better than from pneumonia or lung tumor; their AUROC point estimates for both named diseases fall below 0.5. CheXficient has documented VinDr-CXR pretraining exposure, which limits the interpretation of its results. Thresholds chosen for 95\% sensitivity on TBX11K training retain that constraint by point estimate in only four of sixteen target evaluations. A five-seed supervised source model reaches 0.999 AUROC on TBX11K validation but 0.629 on each of two external cohorts. Conservative exclusion of perceptual-overlap candidates narrows this gap without closing it. These retrospective, single-task results show that discrimination, score reliability and threshold retention support different portability claims. Evidence for chest X-ray tuberculosis screening should identify the complete evaluation specification rather than attribute clinical portability to a checkpoint alone.

Abstract: Deep learning segmentation of renal tumors requires large annotated datasets, yet clinical deployments typically offer only a handful of tumor-positive cases from the target site. We propose a diffusion-based inpainting framework that synthesizes anatomically plausible renal tumors within healthy CT scans, requiring no additional annotation, and provide the first systematic comparison of 2D, 2.5D, and full 3D (MAISI) synthesis strategies for this task. Training the diffusion model on public data (KiTS23, KIRC) and evaluating nnU-Net segmentation on a internal cohort across three low-data regimes, we find that 2.5D and 3D augmentation substantially reduce false positives (from $\sim$18--20\% to $\sim$3--6\%) while maintaining Dice, whereas 2D provides no consistent benefit. Crucially, the proposed 2.5D method matches full 3D synthesis on every metric at substantially lower computational cost, indicating that local volumetric consistency alone is sufficient for effective augmentation in data- and resource-scarce clinical settings.

Abstract: Integrating diverse data modalities --- such as clinical notes, laboratory results, and medical imaging --- is essential for advancing clinical decision-making. While Large Language Models (LLMs) have shown remarkable performance in processing unstructured clinical text, their limited capacity to incorporate non-text modalities hinders their broader utility in healthcare applications. Here, we introduce M2G-LLM (Multimodal MedGraph-LLM), a novel framework that enhances LLMs with multimodal integration and alignment via Graph Neural Networks (GNNs). Our approach models temporal relationships between patient visits, propagates information across clinically similar patients, and aligns heterogeneous data sources to construct enriched multimodal context vectors. These vectors are injected into the intermediate layers of the LLM, enabling joint reasoning over textual and non-textual modalities. We evaluate M2G-LLM on the MIMIC-IV and MIMIC-CXR datasets, demonstrating improvements in clinical prediction tasks over strong baseline models. Our results highlight the promise of combining the language understanding of LLMs with the relational reasoning capabilities of GNNs for comprehensive, multimodal healthcare analysis.

Abstract: Cortical hypometabolism measured by Fluorodeoxyglucose Positron Emission Tomography (FDG-PET) is a highly sensitive biomarker for dementia diagnosis. However, high costs, radiation exposure, and limited accessibility constrain its clinical utility. While cross-modal synthesis from Magnetic Resonance Imaging (MRI) offers a promising alternative, existing volumetric generation methods do not explicitly account for the highly folded cortical geometry, where disease-related patterns predominantly reside. To address this, we introduce a novel surface-based diffusion bridge framework DB-SUiT for MRI-to-PET translation that operates natively on the cortical manifold. A conditional Spherical U-shaped vision Transformer (SUiT) is specifically designed to model the intricate cross-modal relationships while preserving surface topology. It combines spherical convolutional encoders for multi-scale surface feature extraction with bottleneck Transformers to capture long-range spatial dependencies, while incorporating demographic and subcortical conditions to refine the synthesis. Evaluated on two datasets, including subjects with different dementia types, DB-SUiT demonstrates high-fidelity synthesis that substantially outperforms other baselines. In automated dementia classification, synthesized PET surfaces improve performance over MRI by 14.2% and PET volumes by 11.3%, approaching the performance of real PET surfaces. In a blinded reader study, synthetic PET achieved 85.5% diagnostic accuracy, compared with 75.8% for MRI and 95.2% for real PET. This further demonstrates cross-cohort and cross-pathology generalization, as the model was evaluated without retraining on an external cohort that included a dementia subtype not represented during training. Our code is available at https://github.com/ai-med/DB-SUiT.

Abstract: Accurate estimation of left ventricular ejection fraction (EF) from echocardiography is central to cardiovascular care, and deep learning enables automated EF prediction from echocardiographic video. Because EF is clinically derived from left-ventricular (LV) volumes, a widely held intuition is that explicit LV segmentation should improve prediction. We introduce a quantitative criterion, the segmentation ceiling, that makes this testable: from EF as a normalized difference of end-diastolic and end-systolic volumes, we derive in closed form how per-frame segmentation area error propagates into EF error, and thus the accuracy a mask must reach before it can improve on direct regression. Using EchoNet-Dynamic, a UniFormer-S backbone, and the empirically measured within-patient error correlation, the criterion places the break-even near 10% per-frame area error, whereas a representative segmenter operates at roughly 14%, above the ceiling. Consistent with this, four strategies for injecting segmentation or area information (a predicted-mask channel, end-diastolic/end-systolic clip sampling, and per-bin and amplitude area-consistency objectives) fail to beat a raw-video baseline; ground-truth masks help only through label leakage. Input representation thus not being the limit, we identify generalization as the practical lever: weight averaging with strong augmentation attains a test R^2 of 0.806 (MAE 4.08) under a matched dense-clip protocol, comparable to an R(2+1)D baseline (0.811) while tightening the validation-to-test gap. Finally, a heteroscedastic beta-NLL formulation yields informative, well-calibrated per-prediction uncertainty, larger for clinically harder low-EF cases, where Monte-Carlo dropout does not. The segmentation ceiling gives a concrete design criterion for when mask-guided EF estimation is worthwhile, plus a simple, uncertainty-aware recipe for EF regression.

Abstract: Pathology vision-language models are commonly built by pretraining or fine-tuning large encoders on paired image-caption data. We asked whether a pathology vision-language model can instead be assembled by parameter-efficient alignment of frozen unimodal foundation models, leaving their pretrained representations untouched. Here we present Lumen, which aligns frozen Virchow2 and BioMedBERT backbones using rank-4 adapters and projection heads, training only 0.40% of the total parameters on the public QUILT-1M corpus. Across nine public zero-shot patch benchmarks, Lumen achieved the highest mean chance-corrected balanced accuracy, 0.546 versus 0.461 for the strongest baseline (paired difference 0.086, 95% CI 0.042-0.136). On lymph-node metastasis detection, Lumen reached an AUROC of 0.964 (95% CI 0.956-0.971) on 4,214 held-out internal slides and 0.955 (95% CI 0.942-0.966) on 2,368 slides across nine external cohorts and six organs. At the internally calibrated threshold, it outperformed all vision-language baselines, with a balanced accuracy of 0.909 (95% CI 0.896-0.923) internally and 0.915 (95% CI 0.902-0.929) externally. Lumen performed competitively across the evaluations, with the exception of cross-modal retrieval, where it ranked third behind CONCH and PathGen-L/14. Fully fine-tuning both encoders gave Lumen no consistent benefit over low-rank adaptation, although it improved retrieval. Aligning frozen unimodal foundation models therefore yields strong and transferable performance at patch and slide level while training only a small fraction of the parameters.

Abstract: Fine-tuning an aligned language model on narrow, flawed data can induce harmful behavior far outside the training domain, known as emergent misalignment (EM). Prior work has localized EM in model weights, activations, and training documents, but it remains unclear which training tokens carry the relevant fine-tuning signal. We introduce TAME (Token Attribution and Masking for Emergent Misalignment), a three-stage framework: token attribution scores how strongly the fine-tuning update raises each response token's likelihood, using forward passes through a released LoRA adapter; signal characterization finds patterns among high-attribution tokens; and causal validation tests them by attribution-guided loss masking. On released EM organisms and a 6,849-example medical-advice split, attribution is concentrated (the top 5% of tokens hold 32% of the mass) and, in Llama, depleted for medical vocabulary but enriched for a register of unwarranted certainty, even after controlling for token rarity. Masking high-attribution tokens during fresh fine-tuning cuts EM by 23x in Llama and 36x in Qwen, with the perplexity cost concentrated on the targeted register rather than on medical content; an equal random mask leaves EM unchanged. In Llama, the attribution pattern suggests that EM-relevant signal lies more in how confidently flawed content is expressed than in its domain vocabulary; the causal masking effect itself holds across both model families.

Abstract: Test-time reinforcement learning adapts a model on its own unlabeled test set using majority-vote pseudo-labels and has shown strong results in mathematics. We show that this recipe collapses on medical multiple-choice QA: accuracy stagnates while output diversity rapidly declines. Through a controlled experiment that keeps the questions, model, and optimizer fixed while changing only the answer space, we trace this failure to answer-space structure rather than domain difficulty. In small answer spaces, incorrect rollouts often collide on the same wrong pseudo-label and reinforce it; in large answer spaces, they disperse and receive little reward. This diagnosis motivates PROSE, Process Reward Guided Self-Training, which rewards reasoning quality instead of answer agreement. PROSE scores each reasoning step with a medical process reward model, assigns the trajectory reward as the minimum score across steps, and enforces answer-format constraints. Without labels, PROSE substantially improves a general Llama model, surpassing purpose-built medical models and matching much larger systems. Because the process signal is internalized into the policy, the adapted model requires no reward model at inference and transfers its gains to unseen datasets. We further show that the minimum aggregation is essential: mean aggregation can be exploited, saturating the proxy reward while degrading accuracy.

Abstract: A radiology report can already answer a clinical question, so it is hard to tell whether a vision-language model also uses the image. ModaLens, a paired image-swap audit, measures how report availability changes image sensitivity: MedGemma-27B on 3,199 paired MIMIC-CXR cases from 293 patients, all 14 questions per case (13 finding-specific and one composite), each image replaced by one from another study, usually of the same patient, with question and report fixed. Under an explicit answer instruction, the model's generated answer changes on 4.26 percent of trials with the report and 20.94 percent without it, a paired increase of 16.7 points (patient-clustered 95 percent CI 15.6 to 17.7), so report availability reduces image-swap sensitivity under this protocol; the original prompt with a lowercase first-token readout gives 4.70 percent against 17.07 percent, and substitutions also move continuous answer scores where the binary prediction does not change. The labels are derived from reports, which limits conclusions about visual correctness; the direction replicates in two further model lineages. Code, the exact prompts and a run record for every number are at https://github.com/criticaldata/MODALENS.

Abstract: We describe our submission to the MedReason 2026 challenge, covering multiple-choice (MCQ) and open-ended (OE) medical visual question answering (VQA) under fully offline, containerized inference. Our first finding is that MCQ retrieval must compare answer \emph{semantics} rather than answer labels: labels are independently assigned per question, so copying a retrieved neighbor's label transfers no useful information, whereas scoring each current option's text against correct-answer text from similar training cases raises retrieval-only accuracy from 20.0\% to 57.5\% on a 200-case retrieval-excluded development holdout. Our second finding attributes the submitted system's accuracy: holding the task-specific MCQ Low-Rank Adaptation (LoRA) adapter fixed and varying the number \(k\) of in-prompt retrieved examples changes accuracy by at most one case --- 187/200 (93.5\%) at both \(k=0\) and the adapter's training-time \(k=1\), 188/200 (94.0\%) at the packaged runtime's default \(k=3\) --- and the submitted confidence-gated override adds no net accuracy on top of \(k=3\), selecting the VLM in 198/200 cases. With the final MCQ adapter fixed, retrieval changes accuracy by at most one case, and gating provides no net gain. On 20 OE cases, token-F1 and RaTEScore~\cite{zhao2024ratescore} decrease as \(k\) grows, but paired sign tests on token-F1 differences are nonsignificant (\(p \ge 0.29\)); a single-annotator comparison found 6/20 wrong-anchor errors for the final configuration and 14/20 for an earlier configuration that jointly differed in routing, adapter, and prompting. The system reaches 94.0\% MCQ accuracy on the development holdout and 93.20\% on the organizer's official pre-evaluation, versus 29.43\% for the off-the-shelf reference baseline, while both of the organizer's open-ended scores are lower than that baseline's (ground-truth agreement 1.245 versus 1.588, visual accuracy 1.995 versus 2.696, each out of 4).

Abstract: Medical imaging modalities such as ultrasound and X-ray are widely used in clinical practice, where diagnosis follows a structured, evidence-driven workflow aligned with standardized criteria. While multimodal large language models (MLLMs) show promise for automated medical report generation, most existing systems rely on end-to-end multimodal fusion without modeling clinically defined intermediate attributes, leading to limited grounding and interpretability. To address this issue, we propose CORAL (COncept-grounded ReAsoning with Localization), a multimodal framework that integrates spatial grounding and concept-level supervision into a unified reasoning process. CORAL employs a prompt-driven medical segmentation model to localize lesions and predicts multi-class clinical attributes through a Concept Bottleneck module. The resulting textual concept tokens are combined with mask-modulated visual features within an MLLM to enable structured report generation and diagnostic prediction. Experiments on BUS-CoT and IU X-ray datasets demonstrate consistent improvements in diagnostic accuracy, concept consistency, and report quality over strong general-purpose and medical MLLMs, indicating that concept-grounded reasoning better aligns generation with clinical decision processes.

Abstract: Vision-language models are increasingly explored for clinical prediction from electronic health records and medical images, where identifying unreliable predictions is important for safe deployment. Uncertainty estimation (UE) enables detecting such predictions, but its evaluation depends on a correctness criterion that determines whether each model output is correct. If this criterion disagrees with human judgement or distorts downstream UE performance, conclusions about model reliability can be misleading. We introduce a two-axis framework that evaluates correctness criteria by their agreement with human judgements and fidelity to human-referenced UE performance. We assess eight criteria across three clinical prediction tasks and three models using 450 predictions annotated by two reviewers. Across the audited tasks, canonical exact matching (EM) achieved the highest observed human agreement and lowest UE distortion, while the BERT-based matching (BEM) and LLM-judge also showed strong human agreement. Across four UE methods and 23,254 clinical predictions, criterion choice changed error-detection AUROC by up to 0.146 and reversed the relative ranking of UE methods. The LLM-judge also selectively accepted invalid or uncertain outputs, accepting 16 of 30 such human-identified errors. These results demonstrate that correctness assessment is an integral component of clinical UE evaluation and should be validated before UE methods are compared.

Abstract: Dental caries and endodontic disease are among the most common health conditions worldwide, and intraoral periapical radiographs are central to their detection, treatment planning, and follow-up. Automated tooth-level classification of these images, however, lacks reproducible benchmarks, is often evaluated with image-level splits that leak patients between training and test, and is rarely validated across clinics. This paper presents the first patient-level, leakage-aware classification benchmark for single-tooth intraoral periapical radiographs on the DentIRO dataset, which comprises 5,300 images from 3,243 patients across two clinics and four classes: Healthy, Caries, Crowned, and Root Canal. Five transfer-learning models are compared with patient-grouped stratified cross-validation, so that every patient remains within a single fold. DenseNet121 gave the strongest and most stable result at a mean macro-F1 of 0.9787, while the four ImageNet-initialized backbones performed comparably. A controlled comparison on a fixed architecture showed that chest-radiograph pretraining transferred less effectively than ImageNet initialization. Bidirectional cross-center validation produced a small average generalization gap of 0.0077, and Grad-CAM confirmed that predictions rely on clinically meaningful tooth regions rather than acquisition artifacts. The benchmark offers a rigorous and reproducible baseline for intraoral radiograph classification.

Abstract: A correct diagnosis can still be reached for the wrong reasons. In clinical reasoning, every disease hypothesis creates obligations: key evidence must be checked, alternatives must be ruled out, contradictions must be resolved, useful tests must be considered, and closure must be justified. Current evaluations of medical LLMs mostly focus on final answers, local steps, or isolated facts, and therefore miss these hypothesis-induced commitments. We introduce \textbf{VeriDx}, a disease-centric verification framework that links free-form diagnostic reasoning to structured disease profiles. VeriDx tracks whether each hypothesis is satisfied, unresolved, or violated its clinical obligations, exposing failures such as missing critical tests, unresolved differentials, ignored contradictions, unsupported claims, and premature closure. We instantiate VeriDx for complex respiratory diagnosis using guideline-derived disease profiles and expert-annotated longitudinal cases. Our results show that many diagnostic errors are not isolated mistakes, but broken commitments made earlier in the reasoning process.

Abstract: Multimodal clinical AI typically assumes that the waveform, report, metadata, and downstream predictions attached to a record belong to the same patient. In practice, linkage failures can silently assemble individually plausible but cross-patient components, creating a safety problem that standard predictive models are not designed to detect. We study this problem as multimodal record integrity triage: given an assembled record, should its modalities be trusted to belong together? We introduce SHIFT-M3, a lightweight text-based pre-fusion screen that measures alignment-based consistency between two separately produced ECG text views: an LLM-generated interpretation and a clinical report summary. On 784,680 MEETI ECG records, SHIFT-M3 achieves 97.6% TPR@5% FPR for full text-view swaps (AUROC 0.996), 90.3% for partial swaps (AUROC 0.974), and 97.7% for label-matched hard negatives (AUROC 0.996) with only 573,569 parameters. Compared with same-dataset lexical baselines, the gains are largest on partial swaps and hard negatives, suggesting that the model is learning more than surface overlap. We also introduce the CMST (Conflict-type Multimodal Stress Test) evaluation taxonomy, a three-seed stability study, a loss ablation, a temporal-tolerance sweep, and a shared-token masking control. The main remaining failure mode is longitudinal ambiguity: at the default operating point, same-patient cross-visit pairs still produce 87.0% Type-II false positives.

Abstract: Hallucinations can undermine clinician trust in LLMs, making it important that evaluation methods capture clinically relevant errors. Rubric-based evaluation has become the leading approach for assessing LLMs in medicine, but it is unclear whether rubric scores reflect such errors. We first study this in a controlled setting using MedHallu, finding that more specific rubrics better distinguish correct from hallucinated responses. To test this systematically, we develop a taxonomy of medical hallucination types and a clinician-validated error-injection pipeline that creates matched correct and error-injected responses. Across HealthBench, HealthBench Professional, and LiveMedBench, our clinically relevant hallucinations are missed by rubrics, often leaving scores unchanged. We find that rubrics are most effective when explicitly checking facts, and are less effective for additional or unexpected errors they do not anticipate. A preliminary retrieval-based factuality check recovers some of the rubric-blind errors, suggesting a complementary approach. These findings reveal systematic blind spots in current medical evaluation of LLMs and suggest that rubric scores alone are insufficient to establish clinical reliability, potentially undermining clinician trust and confidence in clinical deployment.

Abstract: Clinical electroencephalography (EEG) data are valuable for healthcare research and for developing artificial intelligence (AI)-based clinical decision-support systems, but EEG recordings and derived features may contain sensitive patient-specific information. This creates privacy risks when data are reused, analyzed, or shared across clinical and research environments. Conventional anonymization methods are often insufficient for high-dimensional biomedical signals, since removing direct identifiers does not necessarily prevent re-identification, linkage, or inference risks. At the same time, strong privacy protection may distort clinically relevant signal characteristics and reduce data utility. This paper studies subject-level differential privacy for protecting clinical EEG-derived feature representations using Gaussian and Laplace perturbations. The proposed framework considers three deployment scenarios: client-side anonymization, centralized server-side anonymization, and decentralized local training. Following EEG preprocessing and feature extraction, Gaussian and Laplace perturbations are applied to the resulting patient-level EEG feature representations. The Laplace experiments evaluate the implemented noise scales, while the scales required for formal full-vector calibration are derived separately. The effects of both perturbations are assessed using statistical utility measures and a downstream machine-learning-based utility check. The results show that differentially private perturbation can be integrated into EEG processing workflows, but the selected mechanism, privacy parameters, and sensitivity calibration strongly influence data utility. The study highlights the practical privacy-utility trade-off in DP-based EEG feature anonymization and the challenges of preserving downstream utility in small and imbalanced clinical EEG datasets.

Abstract: Early diagnosis of melanoma is critical for improving patient survival rates. However, accurately distinguishing melanoma from other skin lesions remains a significant clinical challenge due to the high visual similarity among lesion types and variability in image acquisition conditions. Artificial intelligence, particularly machine learning, has emerged as a promising tool to support dermatological diagnosis by automating feature extraction from medical images. Among the available approaches, convolutional neural networks (CNNs) have demonstrated strong performance in image classification tasks, making them well-suited for analyzing both dermatoscopic and histopathological images, given their ability to capture hierarchical visual patterns relevant to lesion characterization. Nevertheless, despite numerous pre-trained CNN architectures having been proposed, selecting the most appropriate one for a given imaging modality remains an open challenge. In this study, we evaluate pre-trained convolutional neural networks (CNNs) for skin lesion classification using dermatoscopic and histopathological image datasets. Experiments were conducted on the HAM10000, ISIC 2018, and CR-AI4SkIN datasets, evaluating the ResNet50, VGG16, VGG19, MobileNet, and InceptionV3 architectures under the same training protocol. The experimental evaluation showed that the models achieved accuracies ranging from 71% (InceptionV3 on ISIC 2018) to 84% (ResNet50 on HAM10000) on dermatoscopic images. For histopathological images, accuracies ranged from 72% (VGG19) to 83% (ResNet50) on the CR-AI4SkIN dataset. The results demonstrate that model performance differs between dermatoscopic and histopathological image modalities, showing that architectures exhibiting similar performance on dermatoscopic images exhibit different performance on histopathological data.

Abstract: Evidence-based medicine demands strict logical consistency, yet current evaluations of large language models (LLMs) prioritize superficial label matching over genuine reasoning. We introduce LogiMed-RoB, a benchmark grounded in Cochrane Risk of Bias (RoB) 2.0 expert logic, comprising 860 randomized controlled trials (RCTs) and 14,820 queries. It evaluates models under the Hierarchical Logical Consistency (HLC) framework across four dimensions: Atomic Consistency, Domain Consistency, Aggregation Consistency, and Evidential Faithfulness. Experiments on 10 state-of-the-art LLMs reveal a catastrophic Error Compounding Effect: despite the top model reaching 98.88% Atomic Consistency, its end-to-end consistency collapses to 45.13%, with several open-weight architectures plummeting to nearly 0%. We further uncover a systematic evidence-reasoning gap: even when models retrieve high-quality evidence, they fail to deduce correct outcomes in 18.63-40.05% of cases, while Blind Guess Rates reach 48.28%. LogiMed-RoB demonstrates that high outcome accuracy can conceal critical reasoning flaws, underscoring the necessity of white-box logical verification for clinical deployment.

Abstract: Machine learning models can achieve strong test performance while relying on demographic or acquisition-related shortcuts. We propose counterfactual (CF) marginalisation as a test-time evaluation procedure for assessing robustness of classification models to such variables. Given a CF image generator, we intervene on nuisance parent variables such as age or sex, generate CF versions of each test image, and average predictions over a target intervention distribution. This produces intervention-aware predictions that marginalise demographic effects while preserving patient-specific latent information. We use these predictions to define metrics for CF risk, calibration, stability and worst-case sensitivity. We demonstrate this framework's utility for quantitative robustness evaluation.

Abstract: Clinical diagnosis agents must decide not only what test to request next, but also when to diagnose or defer. Existing agent benchmarks largely evaluate accuracy after fixed or unconstrained interaction, leaving autonomous stopping reliability implicit. We present Cros, a risk-constrained stopping layer combining state-wise error ranking, policy design on disjoint development splits, and LTT-style exact tests of selective diagnostic error and minimum autonomous coverage for complete sequential policies. Its finite-sample guarantee requires the candidate family, testing rule, and any randomization to be frozen before calibration labels are accessed. On a 1,834-episode MIMIC-derived abdominal-pain benchmark, the full ranker achieves exploratory state-error AUROC 0.853, compared with 0.715 for maximum class probability and 0.552 for the backbone's native stop score. On the previously viewed 367-episode evaluation split, analytically averaging over the frozen Cros weights yields 16.9% selective error at 78.8% coverage, cost 5.57, and 0.68 tests, versus 30.8% error at 100% coverage, cost 8.14, and 1.53 tests under native stopping. Forced continuation is non-monotone: error is 28.3% with HPI alone and 34.3% after full workup. However, the uniform-weight mixture ablation is cheaper on this viewed split despite missing the locked development margins, and Cros nominally satisfies the joint criterion in only 6 of 20 development resplits. Because evaluation labels were inspected during earlier development, these findings provide exploratory feasibility and audit evidence, not a confirmatory safety certificate.

Abstract: Brain-age prediction models are commonly evaluated by predictive accuracy, yet accurate predictions alone do not establish that a model relies on reproducible or neurobiologically supported mechanisms. Post-hoc explanation methods can expose these mechanisms, but existing workflows typically stop at diagnosis or require correction targets to be specified before model analysis. We propose XAI-Refine, an automated explanation-knowledge loop for brain-age prediction from resting-state functional connectivity. At each iteration, XAI-Refine consolidates complementary post-hoc analyses across repeated training runs into reliable, structured model explanations. It converts each reliable explanation into a neutral neurobiological question, retrieves and verifies relevant literature, and compiles the verified evidence into an admissible set in the same typed explanation space. The target for refinement is defined as the minimal projection of the current model explanation onto the admissible set induced by applicable verified knowledge. This revised explanation is then translated into a differentiable constraint while preserving the originating model variable, measurement operator, and applicable scope. Candidate updates are promoted only when multi-seed validation confirms target-directed explanatory movement, predictive performance remains within a prespecified guardrail, and non-target explanatory drift remains bounded. Experiments on functional-connectivity-based brain-age prediction evaluate predictive performance, explanation reliability, literature alignment, and target-specific model revision, illustrating a structured route from post-hoc analysis to evidence-guided model refinement.

Abstract: Despite recent advances in surgical vision-language models (VLMs), temporal reasoning remains limited because existing supervision is largely frame-centric. Frame-level scene graphs (SGs) have proven effective in providing structured representations of surgical environments but do not explicitly model the dynamics of surgical workflows. To explicitly model how surgical states evolve across time, we introduce a multi-level structured temporal supervision methodology that augments frame-level surgical SGs with object-level continuity, event-level interaction continuity, and procedure-level connectivity. We then execute temporal queries over the resulting spatio-temporal scene graphs (STSGs) to generate evidence-grounded question-answer pairs, which together form the STSG-VQA benchmark. Each question is linked to the temporal interval and STSG evidence used to derive its reference answer, enabling traceable verification. The benchmark contains 18,458 question-answer pairs across seven temporal categories. Fine-tuning Qwen3-VL-4B and Hulu-Med-4B with STSG-derived supervision improves question-level micro accuracy by 24.39 and 19.56 percentage points over their zero-shot baselines and by 16.50 and 14.25 points over static scene-graph supervision, respectively. These gains span all temporal categories, indicating that STSG-derived supervision helps surgical VLMs reason over temporally grounded interactions rather than isolated frames. The code and dataset will be made publicly available upon acceptance.

Abstract: Medical imaging artificial intelligence (AI) is commonly developed as separate mappings from radiographs to diagnostic outputs or from clinical descriptions to generated images, although both arise from the same underlying radiographic state. A world-model formulation instead seeks to learn an internal representation of this state that can support both clinical readout and conditional simulation of radiographic observations. Here we introduce MedDream, a radiographic world model that learns a shared continuous latent state from paired chest radiograph-text observations for diagnostic reasoning and report-conditioned evidence generation. MedDream was pretrained on 2.65 million leakage-controlled chest radiograph-text pairs curated from 4.40 million candidates. Across eight clinical datasets and two independent reader cohorts, MedDream outperformed leading diagnostic and generative comparators. For diagnostic reasoning, MedDream showed strong generalization across disease recognition, label-scarce adaptation, severity assessment, and localization, while MedDream-supported review increased mean resident concordance with independent radiologist consensus from 56.3% to 63.0%. For evidence generation, MedDream produced radiographs that preserved clinically relevant pathology and improved downstream performance on held-out real data, with synthetic augmentation increasing external VinDr-CXR macro-AUROC from 76.4% to 81.4%. More importantly, conditioning generation on prespecified subgroup performance gaps enabled targeted evidence construction, increasing weighted F1 by 3.1 percentage points in Asian patients, whereas matched-volume unguided augmentation decreased it by 2.3 points. These findings establish radiographic world models as a path toward medical AI that learns clinically meaningful internal states for interpreting, simulating, and constructing evidence for clinical use.

Abstract: Should multilingual LLMs answer medical questions consistently across input languages, or adapt responses to cultural cues? Existing multilingual medical benchmarks usually assume that medically correct answers should remain consistent across languages and treat cross-lingual variation as model error. In contrast, cultural adaptation research argues that appropriate medical answers may legitimately differ across contexts. We review the multilingual medical NLP literature through these two perspectives, we identify three gaps: limited stakeholder perspectives (e.g., of medical professionals), a lack of empirical evidence on which approach better serves users, and no benchmarks capable of distinguishing universally correct from culture-specific cases. To address the first gap, we survey 356 participants across three stakeholder groups (medical, NLP, and anthropology professionals) in three countries (Germany, Spain, and the United States). Anthropologists consistently favor adaptation, while medical and NLP respondents remain divided, with notable divergence between U.S. and European medical professionals. LLMs prompted with profession and country personas fail to reproduce this variation, overestimating cross-lingual consistency preference among NLP and medical personas. We conclude that neither consistency nor adaptation can currently be considered clearly preferable, highlighting the need for empirical evidence on which approach better serves users across cultural contexts.