Min Gu Kwak

Yuelyu Ji, Min Gu Kwak, Hang Zhang et al.

Medical RAG needs evidence-grounded claims, so plugging a claim-level NLI checker into retrieval-augmented RL is intuitive. \textbf{We find that the checker's \emph{output distribution} during training, not its held-out accuracy, decides whether it provides trainable gradient.} We compare four NLI checker back-ends as process rewards inside a GRPO-trained medical RAG agent (Qwen2.5-7B, replicated on Qwen3-4B and Llama-3.1-8B) across four held-out medical QA benchmarks. Three diagnostic findings emerge. \textbf{(i)} Signal collapse is log-prob-specific: LLM log-probability scoring labels over 97\% of claims neutral -- collapsing the RL gradient to zero -- while a calibrated MedNLI classifier scores the same pairs non-degenerately. \textbf{(ii)} Moderate signal beats strong signal on answer quality: a strong proprietary checker triggers a three-step reward-hacking cascade -- ultra-short answers, search avoidance, language collapse -- so a moderate-signal local classifier trains a higher-quality model (\textbf{+12\% BERTScore over zero-shot, no GPT dependency}). \textbf{(iii)} Signal strength is policy-dependent: the same checker registers as moderate on one policy but strong on another without triggering the cascade end-state. We frame these as boundary conditions for verifier-as-reward systems.

Min Gu Kwak, Hyungu Kahng, Seoung Bum Kim

Semi-supervised learning methods have shown promising results in solving many practical problems when only a few labels are available. The existing methods assume that the class distributions of labeled and unlabeled data are equal; however, their performances are significantly degraded in class distribution mismatch scenarios where out-of-distribution (OOD) data exist in the unlabeled data. Previous safe semi-supervised learning studies have addressed this problem by making OOD data less likely to affect training based on labeled data. However, even if the studies effectively filter out the unnecessary OOD data, they can lose the basic information that all data share regardless of class. To this end, we propose to apply a self-supervised contrastive learning approach to fully exploit a large amount of unlabeled data. We also propose a contrastive loss function with coefficient schedule to aggregate as an anchor the labeled negative examples of the same class into positive examples. To evaluate the performance of the proposed method, we conduct experiments on image classification datasets - CIFAR-10, CIFAR-100, Tiny ImageNet, and CIFAR-100+Tiny ImageNet - under various mismatch ratios. The results show that self-supervised contrastive learning significantly improves classification accuracy. Moreover, aggregating the in-distribution examples produces better representation and consequently further improves classification accuracy.

Min Gu Kwak, Yeonju Lee, Hairong Wang et al.

Brain tumors are among the most clinically significant neurological diseases and remain a major cause of morbidity and mortality due to their aggressive growth and structural heterogeneity. As tumors expand, they induce substantial anatomical deformation that disrupts both local tissue organization and global brain architecture, complicating diagnosis, treatment planning, and surgical navigation. Yet a subject-specific reference of how the brain would appear without tumor-induced changes is fundamentally unobtainable in clinical practice. We present BrainNormalizer, an anatomy-informed diffusion framework that reconstructs pseudo-healthy MRIs directly from tumorous scans by conditioning the generative process on boundary cues extracted from the subject's own anatomy. This boundary-guided conditioning enables anatomically plausible pseudo-healthy reconstruction without requiring paired non-tumorous and tumorous scans. BrainNormalizer employs a two-stage training strategy. The pretrained diffusion model is first adapted through inpainting-based fine-tuning on tumorous and non-tumorous scans. Next, an edge-map-guided ControlNet branch is trained to inject fine-grained anatomical contours into the frozen decoder while preserving learned priors. During inference, a deliberate misalignment strategy pairs tumorous inputs with non-tumorous prompts and mirrored contralateral edge maps, leveraging hemispheric correspondence to guide reconstruction. On the BraTS2020 dataset, BrainNormalizer achieves strong quantitative performance and qualitatively produces anatomically plausible reconstructions in tumor-affected regions while retaining overall structural coherence. BrainNormalizer provides clinically reliable anatomical references for treatment planning and supports new research directions in counterfactual modeling and tumor-induced deformation analysis.

Sonish Sivarajkumar, Hang Zhang, Yuelyu Ji et al.

Electronic health records (EHRs) are rich clinical data sources but complex repositories of patient data, spanning structured elements (demographics, vitals, lab results, codes), unstructured clinical notes and other modalities of data. Harnessing this heterogeneity is critical for improving patient outcomes. Recent advances in large language models (LLMs) have enabled foundation models that can learn from multiple data modalities and support clinical tasks. However, most current approaches simply serialize numeric EHR data into text, which risks losing temporal and quantitative detail. We introduce Generative Deep Patient (GDP), a multimodal foundation model that natively encodes structured EHR time-series via a CNN-Transformer encoder and fuses it with unstructured EHRs through cross-modal attention into a LLaMA-based decoder. GDP is trained in two stages: (1) generative pretraining, where it learns to produce clinical narratives from raw patient timelines while also performing masked feature prediction (MFP) and next time-step prediction (NTP) to capture temporal dynamics; and (2) multi-task fine-tuning for clinically meaningful predictions (e.g., heart failure, type 2 diabetes, 30-day readmission). In clinical prediction, GDP demonstrated superior performance on MIMIC-IV: heart failure AUROC = 0.923, type 2 diabetes AUROC = 0.817, and 30-day readmission AUROC = 0.627. For narrative generation, GDP achieved ROUGE-L = 0.135 and BERTScore-F1 = 0.545. In a blinded human evaluation, GDP-Instruct scored highest on faithfulness, fluency, and overall clinical utility, suggesting reduced hospital documentation workload without sacrificing accuracy. Our results demonstrate that a single multimodal foundation model can both predict clinically actionable events and generate high-quality clinical narratives. Furthermore, GDP's flexible architecture can be extended to additional modalities.