Eric Wu

Weixin Liang, Mert Yuksekgonul, Yining Mao et al. · stanford

The rapid adoption of generative language models has brought about substantial advancements in digital communication, while simultaneously raising concerns regarding the potential misuse of AI-generated content. Although numerous detection methods have been proposed to differentiate between AI and human-generated content, the fairness and robustness of these detectors remain underexplored. In this study, we evaluate the performance of several widely-used GPT detectors using writing samples from native and non-native English writers. Our findings reveal that these detectors consistently misclassify non-native English writing samples as AI-generated, whereas native writing samples are accurately identified. Furthermore, we demonstrate that simple prompting strategies can not only mitigate this bias but also effectively bypass GPT detectors, suggesting that GPT detectors may unintentionally penalize writers with constrained linguistic expressions. Our results call for a broader conversation about the ethical implications of deploying ChatGPT content detectors and caution against their use in evaluative or educational settings, particularly when they may inadvertently penalize or exclude non-native English speakers from the global discourse. The published version of this study can be accessed at: www.cell.com/patterns/fulltext/S2666-3899(23)00130-7

Yongchan Kwon, Eric Wu, Kevin Wu et al.

Quantifying the impact of training data points is crucial for understanding the outputs of machine learning models and for improving the transparency of the AI pipeline. The influence function is a principled and popular data attribution method, but its computational cost often makes it challenging to use. This issue becomes more pronounced in the setting of large language models and text-to-image models. In this work, we propose DataInf, an efficient influence approximation method that is practical for large-scale generative AI models. Leveraging an easy-to-compute closed-form expression, DataInf outperforms existing influence computation algorithms in terms of computational and memory efficiency. Our theoretical analysis shows that DataInf is particularly well-suited for parameter-efficient fine-tuning techniques such as LoRA. Through systematic empirical evaluations, we show that DataInf accurately approximates influence scores and is orders of magnitude faster than existing methods. In applications to RoBERTa-large, Llama-2-13B-chat, and stable-diffusion-v1.5 models, DataInf effectively identifies the most influential fine-tuning examples better than other approximate influence scores. Moreover, it can help to identify which data points are mislabeled.

Kevin Wu, Eric Wu, Ally Cassasola et al.

Large language models (LLMs) are currently being used to answer medical questions across a variety of clinical domains. Recent top-performing commercial LLMs, in particular, are also capable of citing sources to support their responses. In this paper, we ask: do the sources that LLMs generate actually support the claims that they make? To answer this, we propose three contributions. First, as expert medical annotations are an expensive and time-consuming bottleneck for scalable evaluation, we demonstrate that GPT-4 is highly accurate in validating source relevance, agreeing 88% of the time with a panel of medical doctors. Second, we develop an end-to-end, automated pipeline called \textit{SourceCheckup} and use it to evaluate five top-performing LLMs on a dataset of 1200 generated questions, totaling over 40K pairs of statements and sources. Interestingly, we find that between ~50% to 90% of LLM responses are not fully supported by the sources they provide. We also evaluate GPT-4 with retrieval augmented generation (RAG) and find that, even still, around 30\% of individual statements are unsupported, while nearly half of its responses are not fully supported. Third, we open-source our curated dataset of medical questions and expert annotations for future evaluations. Given the rapid pace of LLM development and the potential harms of incorrect or outdated medical information, it is crucial to also understand and quantify their capability to produce relevant, trustworthy medical references.

Eric Wu, Kevin Wu, Jason Hom et al.

Large language models (LLMs) are increasingly central to clinician workflows, spanning clinical decision support, medical education, and patient communication. However, current evaluation methods for medical LLMs rely heavily on static, templated benchmarks that fail to capture the complexity and dynamics of real-world clinical practice, creating a dissonance between benchmark performance and clinical utility. To address these limitations, we present MedArena, an interactive evaluation platform that enables clinicians to directly test and compare leading LLMs using their own medical queries. Given a clinician-provided query, MedArena presents responses from two randomly selected models and asks the user to select the preferred response. Out of 1571 preferences collected across 12 LLMs up to November 1, 2025, Gemini 2.0 Flash Thinking, Gemini 2.5 Pro, and GPT-4o were the top three models by Bradley-Terry rating. Only one-third of clinician-submitted questions resembled factual recall tasks (e.g., MedQA), whereas the majority addressed topics such as treatment selection, clinical documentation, or patient communication, with ~20% involving multi-turn conversations. Additionally, clinicians cited depth and detail and clarity of presentation more often than raw factual accuracy when explaining their preferences, highlighting the importance of readability and clinical nuance. We also confirm that the model rankings remain stable even after controlling for style-related factors like response length and formatting. By grounding evaluation in real-world clinical questions and preferences, MedArena offers a scalable platform for measuring and improving the utility and efficacy of medical LLMs.

Eric Wu, Kevin Wu, James Zou

There is great interest in fine-tuning frontier large language models (LLMs) to inject new information and update existing knowledge. While commercial LLM fine-tuning APIs from providers such as OpenAI and Google promise flexible adaptation for various applications, the efficacy of fine-tuning remains unclear. In this study, we introduce FineTuneBench, an evaluation framework and dataset for understanding how well commercial fine-tuning APIs can successfully learn new and updated knowledge. We analyze five frontier LLMs with commercially available fine-tuning APIs, including GPT-4o and Gemini 1.5 Pro, on their effectiveness in two settings: (1) ingesting novel information, such as recent news events and new people profiles, and (2) updating existing knowledge, such as updated medical guidelines and code frameworks. Our results reveal substantial shortcomings in all the models' abilities to effectively learn new information through fine-tuning, with an average generalization accuracy of 37% across all models. When updating existing knowledge, such as incorporating medical guideline updates, commercial fine-tuning APIs show even more limited capability (average generalization accuracy of 19%). Overall, fine-tuning GPT-4o mini is the most effective for infusing new knowledge and updating knowledge, followed by GPT-3.5 Turbo and GPT-4o. The fine-tuning APIs for Gemini 1.5 Flesh and Gemini 1.5 Pro are unable to learn new knowledge or update existing knowledge. These findings underscore a major shortcoming in using current commercial fine-tuning services to achieve reliable knowledge infusion in common scenarios. We open source the FineTuneBench dataset at https://github.com/kevinwu23/StanfordFineTuneBench.

Kevin Wu, Eric Wu, James Zou

Retrieval augmented generation (RAG) is frequently used to mitigate hallucinations and provide up-to-date knowledge for large language models (LLMs). However, given that document retrieval is an imprecise task and sometimes results in erroneous or even harmful content being presented in context, this raises the question of how LLMs handle retrieved information: If the provided content is incorrect, does the model know to ignore it, or does it recapitulate the error? Conversely, when the model's initial response is incorrect, does it always know to use the retrieved information to correct itself, or does it insist on its wrong prior response? To answer this, we curate a dataset of over 1200 questions across six domains (e.g., drug dosages, Olympic records, locations) along with content relevant to answering each question. We further apply precise perturbations to the answers in the content that range from subtle to blatant errors. We benchmark six top-performing LLMs, including GPT-4o, on this dataset and find that LLMs are susceptible to adopting incorrect retrieved content, overriding their own correct prior knowledge over 60% of the time. However, the more unrealistic the retrieved content is (i.e. more deviated from truth), the less likely the model is to adopt it. Also, the less confident a model is in its initial response (via measuring token probabilities), the more likely it is to adopt the information in the retrieved content. We exploit this finding and demonstrate simple methods for improving model accuracy where there is conflicting retrieved content. Our results highlight a difficult task and benchmark for LLMs -- namely, their ability to correctly discern when it is wrong in light of correct retrieved content and to reject cases when the provided content is incorrect.

Weixin Liang, Nazneen Rajani, Xinyu Yang et al. · salesforce, stanford

The rapid proliferation of AI models has underscored the importance of thorough documentation, as it enables users to understand, trust, and effectively utilize these models in various applications. Although developers are encouraged to produce model cards, it's not clear how much information or what information these cards contain. In this study, we conduct a comprehensive analysis of 32,111 AI model documentations on Hugging Face, a leading platform for distributing and deploying AI models. Our investigation sheds light on the prevailing model card documentation practices. Most of the AI models with substantial downloads provide model cards, though the cards have uneven informativeness. We find that sections addressing environmental impact, limitations, and evaluation exhibit the lowest filled-out rates, while the training section is the most consistently filled-out. We analyze the content of each section to characterize practitioners' priorities. Interestingly, there are substantial discussions of data, sometimes with equal or even greater emphasis than the model itself. To evaluate the impact of model cards, we conducted an intervention study by adding detailed model cards to 42 popular models which had no or sparse model cards previously. We find that adding model cards is moderately correlated with an increase weekly download rates. Our study opens up a new perspective for analyzing community norms and practices for model documentation through large-scale data science and linguistics analysis.

Kevin Wu, Eric Wu, Kit Rodolfa et al.

While the pace of development of AI has rapidly progressed in recent years, the implementation of safe and effective regulatory frameworks has lagged behind. In particular, the adaptive nature of AI models presents unique challenges to regulators as updating a model can improve its performance but also introduce safety risks. In the US, the Food and Drug Administration (FDA) has been a forerunner in regulating and approving hundreds of AI medical devices. To better understand how AI is updated and its regulatory considerations, we systematically analyze the frequency and nature of updates in FDA-approved AI medical devices. We find that less than 2% of all devices report having been updated by being re-trained on new data. Meanwhile, nearly a quarter of devices report updates in the form of new functionality and marketing claims. As an illustrative case study, we analyze pneumothorax detection models and find that while model performance can degrade by as much as 0.18 AUC when evaluated on new sites, re-training on site-specific data can mitigate this performance drop, recovering up to 0.23 AUC. However, we also observed significant degradation on the original site after re-training using data from new sites, providing insight from one example that challenges the current one-model-fits-all approach to regulatory approvals. Our analysis provides an in-depth look at the current state of FDA-approved AI device updates and insights for future regulatory policies toward model updating and adaptive AI.

Eric Wu, Kevin Wu, James Zou

Medical AI algorithms can often experience degraded performance when evaluated on previously unseen sites. Addressing cross-site performance disparities is key to ensuring that AI is equitable and effective when deployed on diverse patient populations. Multi-site evaluations are key to diagnosing such disparities as they can test algorithms across a broader range of potential biases such as patient demographics, equipment types, and technical parameters. However, such tests do not explain why the model performs worse. Our framework provides a method for quantifying the marginal and cumulative effect of each type of bias on the overall performance difference when a model is evaluated on external data. We demonstrate its usefulness in a case study of a deep learning model trained to detect the presence of pneumothorax, where our framework can help explain up to 60% of the discrepancy in performance across different sites with known biases like disease comorbidities and imaging parameters.

Eric Wu, Kevin Wu, William Lotter

Data scarcity and class imbalance are two fundamental challenges in many machine learning applications to healthcare. Breast cancer classification in mammography exemplifies these challenges, with a malignancy rate of around 0.5% in a screening population, which is compounded by the relatively small size of lesions (~1% of the image) in malignant cases. Simultaneously, the prevalence of screening mammography creates a potential abundance of non-cancer exams to use for training. Altogether, these characteristics lead to overfitting on cancer cases, while under-utilizing non-cancer data. Here, we present a novel generative adversarial network (GAN) model for data augmentation that can realistically synthesize and remove lesions on mammograms. With self-attention and semi-supervised learning components, the U-net-based architecture can generate high resolution (256x256px) outputs, as necessary for mammography. When augmenting the original training set with the GAN-generated samples, we find a significant improvement in malignancy classification performance on a test set of real mammogram patches. Overall, the empirical results of our algorithm and the relevance to other medical imaging paradigms point to potentially fruitful further applications.

William Lotter, Abdul Rahman Diab, Bryan Haslam et al.

Breast cancer remains a global challenge, causing over 1 million deaths globally in 2018. To achieve earlier breast cancer detection, screening x-ray mammography is recommended by health organizations worldwide and has been estimated to decrease breast cancer mortality by 20-40%. Nevertheless, significant false positive and false negative rates, as well as high interpretation costs, leave opportunities for improving quality and access. To address these limitations, there has been much recent interest in applying deep learning to mammography; however, obtaining large amounts of annotated data poses a challenge for training deep learning models for this purpose, as does ensuring generalization beyond the populations represented in the training dataset. Here, we present an annotation-efficient deep learning approach that 1) achieves state-of-the-art performance in mammogram classification, 2) successfully extends to digital breast tomosynthesis (DBT; "3D mammography"), 3) detects cancers in clinically-negative prior mammograms of cancer patients, 4) generalizes well to a population with low screening rates, and 5) outperforms five-out-of-five full-time breast imaging specialists by improving absolute sensitivity by an average of 14%. Our results demonstrate promise towards software that can improve the accuracy of and access to screening mammography worldwide.

Kevin Wu, Eric Wu, Yaping Wu et al.

A key promise of AI applications in healthcare is in increasing access to quality medical care in under-served populations and emerging markets. However, deep learning models are often only trained on data from advantaged populations that have the infrastructure and resources required for large-scale data collection. In this paper, we aim to empirically investigate the potential impact of such biases on breast cancer detection in mammograms. We specifically explore how a deep learning algorithm trained on screening mammograms from the US and UK generalizes to mammograms collected at a hospital in China, where screening is not widely implemented. For the evaluation, we use a top-scoring model developed for the Digital Mammography DREAM Challenge. Despite the change in institution and population composition, we find that the model generalizes well, exhibiting similar performance to that achieved in the DREAM Challenge, even when controlling for tumor size. We also illustrate a simple but effective method for filtering predictions based on model variance, which can be particularly useful for deployment in new settings. While there are many components in developing a clinically effective system, these results represent a promising step towards increasing access to life-saving screening mammography in populations where screening rates are currently low.

Eric Wu, Bin Kong, Xin Wang et al.

Computerized automatic methods have been employed to boost the productivity as well as objectiveness of hand bone age assessment. These approaches make predictions according to the whole X-ray images, which include other objects that may introduce distractions. Instead, our framework is inspired by the clinical workflow (Tanner-Whitehouse) of hand bone age assessment, which focuses on the key components of the hand. The proposed framework is composed of two components: a Mask R-CNN subnet of pixelwise hand segmentation and a residual attention network for hand bone age assessment. The Mask R-CNN subnet segments the hands from X-ray images to avoid the distractions of other objects (e.g., X-ray tags). The hierarchical attention components of the residual attention subnet force our network to focus on the key components of the X-ray images and generate the final predictions as well as the associated visual supports, which is similar to the assessment procedure of clinicians. We evaluate the performance of the proposed pipeline on the RSNA pediatric bone age dataset and the results demonstrate its superiority over the previous methods.

Eric Wu, Kevin Wu, David Cox et al.

Deep learning approaches to breast cancer detection in mammograms have recently shown promising results. However, such models are constrained by the limited size of publicly available mammography datasets, in large part due to privacy concerns and the high cost of generating expert annotations. Limited dataset size is further exacerbated by substantial class imbalance since "normal" images dramatically outnumber those with findings. Given the rapid progress of generative models in synthesizing realistic images, and the known effectiveness of simple data augmentation techniques (e.g. horizontal flipping), we ask if it is possible to synthetically augment mammogram datasets using generative adversarial networks (GANs). We train a class-conditional GAN to perform contextual in-filling, which we then use to synthesize lesions onto healthy screening mammograms. First, we show that GANs are capable of generating high-resolution synthetic mammogram patches. Next, we experimentally evaluate using the augmented dataset to improve breast cancer classification performance. We observe that a ResNet-50 classifier trained with GAN-augmented training data produces a higher AUROC compared to the same model trained only on traditionally augmented data, demonstrating the potential of our approach.

Kevin Wu, Eric Wu, Gabriel Kreiman

Advancements in convolutional neural networks (CNNs) have made significant strides toward achieving high performance levels on multiple object recognition tasks. While some approaches utilize information from the entire scene to propose regions of interest, the task of interpreting a particular region or object is still performed independently of other objects and features in the image. Here we demonstrate that a scene's 'gist' can significantly contribute to how well humans can recognize objects. These findings are consistent with the notion that humans foveate on an object and incorporate information from the periphery to aid in recognition. We use a biologically inspired two-part convolutional neural network ('GistNet') that models the fovea and periphery to provide a proof-of-principle demonstration that computational object recognition can significantly benefit from the gist of the scene as contextual information. Our model yields accuracy improvements of up to 50% in certain object categories when incorporating contextual gist, while only increasing the original model size by 5%. This proposed model mirrors our intuition about how the human visual system recognizes objects, suggesting specific biologically plausible constraints to improve machine vision and building initial steps towards the challenge of scene understanding.