Shih-Cheng Huang’s research while affiliated with Stanford University and other places

Ensuring that large language models (LLMs) are both helpful and harmless is a critical challenge, as overly strict constraints can lead to excessive refusals, while permissive models risk generating harmful content. Existing approaches, such as reinforcement learning from human feedback (RLHF) and direct preference optimization (DPO), attempt to balance these trade-offs but suffer from performance conflicts, limited controllability, and poor extendability. To address these issues, we propose Preference Vector, a novel framework inspired by task arithmetic. Instead of optimizing multiple preferences within a single objective, we train separate models on individual preferences, extract behavior shifts as preference vectors, and dynamically merge them at test time. This modular approach enables fine-grained, user-controllable preference adjustments and facilitates seamless integration of new preferences without retraining. Experiments show that our proposed Preference Vector framework improves helpfulness without excessive conservatism, allows smooth control over preference trade-offs, and supports scalable multi-preference alignment.

Large foundation models show promise in biomedicine but face challenges in clinical use due to performance gaps, accessibility, cost, and lack of scalable evaluation. Here we show that open-source small multimodal models can bridge these gaps in radiology by generating free-text findings from chest X-ray images. Our data-centric approach leverages 697K curated radiology image-text pairs to train a specialized, domain-adapted chest X-ray encoder. We integrate this encoder with pre-trained language models via a lightweight adapter that aligns image and text modalities. To enable robust, clinically relevant evaluation, we develop and validate CheXprompt, a GPT-4-based metric for assessing factual accuracy aligned with radiologists’ evaluations. Benchmarked with CheXprompt and other standard factuality metrics, LLaVA-Rad (7B) achieves state-of-the-art performance, outperforming much larger models like GPT-4V and Med-PaLM M (84B). While not immediately ready for real-time clinical deployment, LLaVA-Rad is a scalable, privacy-preserving and cost-effective step towards clinically adaptable multimodal AI for radiology.

Artificial Intelligence (AI) holds immense potential to transform healthcare, yet progress is often hindered by the reliance on large labeled datasets and unimodal data. Multimodal Foundation Models (FMs), particularly those leveraging Self-Supervised Learning (SSL) on multimodal data, offer a paradigm shift towards label-efficient, holistic patient modeling. However, the rapid emergence of these complex models has created a fragmented landscape. Here, we provide a systematic review of multimodal FMs for medical imaging applications. Through rigorous screening of 1,144 publications (2012–2024) and in-depth analysis of 48 studies, we establish a unified terminology and comprehensively assess the current state-of-the-art. Our review aggregates current knowledge, critically identifies key limitations and underexplored opportunities, and culminates in actionable guidelines for researchers, clinicians, developers, and policymakers. This work provides a crucial roadmap to navigate and accelerate the responsible development and clinical translation of next-generation multimodal AI in healthcare.

Advancements in artificial intelligence (AI) offer promising solutions for enhancing clinical workflows and patient care, potentially revolutionizing healthcare delivery. However, the traditional paradigm of AI integration in healthcare is limited by models that rely on single input modalities during training and require extensive labeled data, failing to capture the multimodal nature of medical practice. Multimodal foundation models, particularly Large Vision Language Models (VLMs), have the potential to overcome these limitations by processing diverse data types and learning from large-scale unlabeled datasets or natural pairs of different modalities, thereby significantly contributing to the development of more robust and versatile AI systems in healthcare. In this review, we establish a unified terminology for multimodal foundation models for medical imaging applications and provide a systematic analysis of papers published between 2012 and 2024. In total, we screened 1,144 papers from medical and AI domains and extracted data from 97 included studies. Our comprehensive effort aggregates the collective knowledge of prior work, evaluates the current state of multimodal AI in healthcare, and delineates both prevailing limitations and potential growth areas. We provide implementation guidelines and actionable recommendations for various stakeholders, including model developers, clinicians, policymakers, and dataset curators.

Since the release of the original CheXpert paper five years ago, CheXpert has become one of the most widely used and cited clinical AI datasets. The emergence of vision language models has sparked an increase in demands for sharing reports linked to CheXpert images, along with a growing interest among AI fairness researchers in obtaining demographic data. To address this, CheXpert Plus serves as a new collection of radiology data sources, made publicly available to enhance the scaling, performance, robustness, and fairness of models for all subsequent machine learning tasks in the field of radiology. CheXpert Plus is the largest text dataset publicly released in radiology, with a total of 36 million text tokens, including 13 million impression tokens. To the best of our knowledge, it represents the largest text de-identification effort in radiology, with almost 1 million PHI spans anonymized. It is only the second time that a large-scale English paired dataset has been released in radiology, thereby enabling, for the first time, cross-institution training at scale. All reports are paired with high-quality images in DICOM format, along with numerous image and patient metadata covering various clinical and socio-economic groups, as well as many pathology labels and RadGraph annotations. We hope this dataset will boost research for AI models that can further assist radiologists and help improve medical care. Data is available at the following URL: https://stanfordaimi.azurewebsites.net/datasets/5158c524-d3ab-4e02-96e9-6ee9efc110a1 Models are available at the following URL: https://github.com/Stanford-AIMI/chexpert-plus

Synthesizing information from various data sources plays a crucial role in the practice of modern medicine. Current applications of artificial intelligence in medicine often focus on single-modality data due to a lack of publicly available, multimodal medical datasets. To address this limitation, we introduce INSPECT, which contains de-identified longitudinal records from a large cohort of pulmonary embolism (PE) patients, along with ground truth labels for multiple outcomes. INSPECT contains data from 19,402 patients, including CT images, radiology report impression sections, and structured electronic health record (EHR) data (including demographics, diagnoses, procedures, and vitals). Using our provided dataset, we develop and release a benchmark for evaluating several baseline model-ing approaches on a variety of important PE related tasks. We evaluate image-only, EHR-only, and fused models. Trained models and the de-identified dataset are made available for non-commercial use under a data use agreement. To the best our knowledge, INSPECT is the largest multimodal dataset for enabling reproducible research on strategies for integrating 3D medical imaging and EHR data.

Pre-trained multi-modal vision-language models (VLMs) are becoming increasingly popular due to their exceptional performance on downstream vision applications, particularly in the few- and zero-shot settings. However, selecting the best-performing VLM for some downstream applications is non-trivial, as it is dataset and task-dependent. Meanwhile, the exhaustive evaluation of all available VLMs on a novel application is not only time and computationally demanding but also necessitates the collection of a labeled dataset for evaluation. As the number of open-source VLM variants increases, there is a need for an efficient model selection strategy that does not require access to a curated evaluation dataset. This paper proposes a novel task and benchmark for efficiently evaluating VLMs' zero-shot performance on downstream applications without access to the downstream task dataset. Specifically, we introduce a new task LOVM: Language-Only Vision Model Selection, where methods are expected to perform both model selection and performance prediction based solely on a text description of the desired downstream application. We then introduced an extensive LOVM benchmark consisting of ground-truth evaluations of 35 pre-trained VLMs and 23 datasets, where methods are expected to rank the pre-trained VLMs and predict their zero-shot performance.