Hoo-Chang Shin

Zhilin Wang, Jiaqi Zeng, Olivier Delalleau et al. · nvidia

Reinforcement Learning with Human Feedback (RLHF) and Reinforcement Learning with Verifiable Rewards (RLVR) are the main RL paradigms used in LLM post-training, each offering distinct advantages. However, RLHF struggles with interpretability and reward hacking because it relies on human judgments that usually lack explicit criteria, whereas RLVR is limited in scope by its focus on correctness-based verifiers. We propose Reinforcement Learning with Binary Flexible Feedback (RLBFF), which combines the versatility of human-driven preferences with the precision of rule-based verification, enabling reward models to capture nuanced aspects of response quality beyond mere correctness. RLBFF extracts principles that can be answered in a binary fashion (e.g. accuracy of information: yes, or code readability: no) from natural language feedback. Such principles can then be used to ground Reward Model training as an entailment task (response satisfies or does not satisfy an arbitrary principle). We show that Reward Models trained in this manner can outperform Bradley-Terry models when matched for data and achieve top performance on RM-Bench (86.2%) and JudgeBench (81.4%, #1 on leaderboard as of September 24, 2025). Additionally, users can specify principles of interest at inference time to customize the focus of our reward models, in contrast to Bradley-Terry models. Finally, we present a fully open source recipe (including data) to align Qwen3-32B using RLBFF and our Reward Model, to match or exceed the performance of o3-mini and DeepSeek R1 on general alignment benchmarks of MT-Bench, WildBench, and Arena Hard v2 (at <5% of the inference cost). Models: https://huggingface.co/collections/nvidia/reward-models-10-2025

Hoo-Chang Shin, Yang Zhang, Evelina Bakhturina et al.

There has been an influx of biomedical domain-specific language models, showing language models pre-trained on biomedical text perform better on biomedical domain benchmarks than those trained on general domain text corpora such as Wikipedia and Books. Yet, most works do not study the factors affecting each domain language application deeply. Additionally, the study of model size on domain-specific models has been mostly missing. We empirically study and evaluate several factors that can affect performance on domain language applications, such as the sub-word vocabulary set, model size, pre-training corpus, and domain transfer. We show consistent improvements on benchmarks with our larger BioMegatron model trained on a larger domain corpus, contributing to our understanding of domain language model applications. We demonstrate noticeable improvements over the previous state-of-the-art (SOTA) on standard biomedical NLP benchmarks of named entity recognition, relation extraction, and question answering. Model checkpoints and code are available at [https://ngc.nvidia.com] and [https://github.com/NVIDIA/NeMo].

Zhilin Wang, Jiaqi Zeng, Olivier Delalleau et al. · nvidia

Inference-Time Scaling has been critical to the success of recent models such as OpenAI o1 and DeepSeek R1. However, many techniques used to train models for inference-time scaling require tasks to have answers that can be verified, limiting their application to domains such as math, coding and logical reasoning. We take inspiration from how humans make first attempts, ask for detailed feedback from others and make improvements based on such feedback across a wide spectrum of open-ended endeavors. To this end, we collect HelpSteer3 data to train dedicated Feedback and Edit Models that are capable of performing inference-time scaling for open-ended general-domain tasks. In our setup, one model generates an initial response, which are given feedback by a second model, that are then used by a third model to edit the response. We show that performance on Arena Hard, a benchmark strongly predictive of Chatbot Arena Elo can be boosted by scaling the number of initial response drafts, effective feedback and edited responses. When scaled optimally, our setup based on 70B models from the Llama 3 family can reach SoTA performance on Arena Hard at 92.7 as of 5 Mar 2025, surpassing OpenAI o1-preview-2024-09-12 with 90.4 and DeepSeek R1 with 92.3.

Carol Anderson, Bo Liu, Anas Abidin et al.

Social media posts contain potentially valuable information about medical conditions and health-related behavior. Biocreative VII Task 3 focuses on mining this information by recognizing mentions of medications and dietary supplements in tweets. We approach this task by fine tuning multiple BERT-style language models to perform token-level classification, and combining them into ensembles to generate final predictions. Our best system consists of five Megatron-BERT-345M models and achieves a strict F1 score of 0.764 on unseen test data.

Virginia Adams, Hoo-Chang Shin, Carol Anderson et al.

The Biocreative VII Track-2 challenge consists of named entity recognition, entity-linking (or entity-normalization), and topic indexing tasks -- with entities and topics limited to chemicals for this challenge. Named entity recognition is a well-established problem and we achieve our best performance with BERT-based BioMegatron models. We extend our BERT-based approach to the entity linking task. After the second stage of pretraining BioBERT with a metric-learning loss strategy called self-alignment pretraining (SAP), we link entities based on the cosine similarity between their SAP-BioBERT word embeddings. Despite the success of our named entity recognition experiments, we find the chemical indexing task generally more challenging. In addition to conventional NER methods, we attempt both named entity recognition and entity linking with a novel text-to-text or "prompt" based method that uses generative language models such as T5 and GPT. We achieve encouraging results with this new approach.

Virginia Adams, Hoo-Chang Shin, Carol Anderson et al.

In Track-1 of the BioCreative VII Challenge participants are asked to identify interactions between drugs/chemicals and proteins. In-context named entity annotations for each drug/chemical and protein are provided and one of fourteen different interactions must be automatically predicted. For this relation extraction task, we attempt both a BERT-based sentence classification approach, and a more novel text-to-text approach using a T5 model. We find that larger BERT-based models perform better in general, with our BioMegatron-based model achieving the highest scores across all metrics, achieving 0.74 F1 score. Though our novel T5 text-to-text method did not perform as well as most of our BERT-based models, it outperformed those trained on similar data, showing promising results, achieving 0.65 F1 score. We believe a text-to-text approach to relation extraction has some competitive advantages and there is a lot of room for research advancement.

Hoo-Chang Shin, Alvin Ihsani, Ziyue Xu et al.

Positron Emission Tomography (PET) is now regarded as the gold standard for the diagnosis of Alzheimer's Disease (AD). However, PET imaging can be prohibitive in terms of cost and planning, and is also among the imaging techniques with the highest dosage of radiation. Magnetic Resonance Imaging (MRI), in contrast, is more widely available and provides more flexibility when setting the desired image resolution. Unfortunately, the diagnosis of AD using MRI is difficult due to the very subtle physiological differences between healthy and AD subjects visible on MRI. As a result, many attempts have been made to synthesize PET images from MR images using generative adversarial networks (GANs) in the interest of enabling the diagnosis of AD from MR. Existing work on PET synthesis from MRI has largely focused on Conditional GANs, where MR images are used to generate PET images and subsequently used for AD diagnosis. There is no end-to-end training goal. This paper proposes an alternative approach to the aforementioned, where AD diagnosis is incorporated in the GAN training objective to achieve the best AD classification performance. Different GAN lossesare fine-tuned based on the discriminator performance, and the overall training is stabilized. The proposed network architecture and training regime show state-of-the-art performance for three- and four- class AD classification tasks.

Hoo-Chang Shin, Alvin Ihsani, Swetha Mandava et al.

Synthesizing medical images, such as PET, is a challenging task due to the fact that the intensity range is much wider and denser than those in photographs and digital renderings and are often heavily biased toward zero. Above all, intensity values in PET have absolute significance, and are used to compute parameters that are reproducible across the population. Yet, usually much manual adjustment has to be made in pre-/post- processing when synthesizing PET images, because its intensity ranges can vary a lot, e.g., between -100 to 1000 in floating point values. To overcome these challenges, we adopt the Bidirectional Encoder Representations from Transformers (BERT) algorithm that has had great success in natural language processing (NLP), where wide-range floating point intensity values are represented as integers ranging between 0 to 10000 that resemble a dictionary of natural language vocabularies. BERT is then trained to predict a proportion of masked values images, where its "next sentence prediction (NSP)" acts as GAN discriminator. Our proposed approach, is able to generate PET images from MRI images in wide intensity range, with no manual adjustments in pre-/post- processing. It is a method that can scale and ready to deploy.

Ziyue Xu, Xiaosong Wang, Hoo-Chang Shin et al.

Radiogenomic map linking image features and gene expression profiles is useful for noninvasively identifying molecular properties of a particular type of disease. Conventionally, such map is produced in three separate steps: 1) gene-clustering to "metagenes", 2) image feature extraction, and 3) statistical correlation between metagenes and image features. Each step is independently performed and relies on arbitrary measurements. In this work, we investigate the potential of an end-to-end method fusing gene data with image features to generate synthetic image and learn radiogenomic map simultaneously. To achieve this goal, we develop a generative adversarial network (GAN) conditioned on both background images and gene expression profiles, synthesizing the corresponding image. Image and gene features are fused at different scales to ensure the realism and quality of the synthesized image. We tested our method on non-small cell lung cancer (NSCLC) dataset. Results demonstrate that the proposed method produces realistic synthetic images, and provides a promising way to find gene-image relationship in a holistic end-to-end manner.

Hoo-Chang Shin, Neil A Tenenholtz, Jameson K Rogers et al.

Data diversity is critical to success when training deep learning models. Medical imaging data sets are often imbalanced as pathologic findings are generally rare, which introduces significant challenges when training deep learning models. In this work, we propose a method to generate synthetic abnormal MRI images with brain tumors by training a generative adversarial network using two publicly available data sets of brain MRI. We demonstrate two unique benefits that the synthetic images provide. First, we illustrate improved performance on tumor segmentation by leveraging the synthetic images as a form of data augmentation. Second, we demonstrate the value of generative models as an anonymization tool, achieving comparable tumor segmentation results when trained on the synthetic data versus when trained on real subject data. Together, these results offer a potential solution to two of the largest challenges facing machine learning in medical imaging, namely the small incidence of pathological findings, and the restrictions around sharing of patient data.

Xiaosong Wang, Le Lu, Hoo-chang Shin et al.

The recent rapid and tremendous success of deep convolutional neural networks (CNN) on many challenging computer vision tasks largely derives from the accessibility of the well-annotated ImageNet and PASCAL VOC datasets. Nevertheless, unsupervised image categorization (i.e., without the ground-truth labeling) is much less investigated, yet critically important and difficult when annotations are extremely hard to obtain in the conventional way of "Google Search" and crowd sourcing. We address this problem by presenting a looped deep pseudo-task optimization (LDPO) framework for joint mining of deep CNN features and image labels. Our method is conceptually simple and rests upon the hypothesized "convergence" of better labels leading to better trained CNN models which in turn feed more discriminative image representations to facilitate more meaningful clusters/labels. Our proposed method is validated in tackling two important applications: 1) Large-scale medical image annotation has always been a prohibitively expensive and easily-biased task even for well-trained radiologists. Significantly better image categorization results are achieved via our proposed approach compared to the previous state-of-the-art method. 2) Unsupervised scene recognition on representative and publicly available datasets with our proposed technique is examined. The LDPO achieves excellent quantitative scene classification results. On the MIT indoor scene dataset, it attains a clustering accuracy of 75.3%, compared to the state-of-the-art supervised classification accuracy of 81.0% (when both are based on the VGG-VD model).

Hoo-Chang Shin, Kirk Roberts, Le Lu et al.

Despite the recent advances in automatically describing image contents, their applications have been mostly limited to image caption datasets containing natural images (e.g., Flickr 30k, MSCOCO). In this paper, we present a deep learning model to efficiently detect a disease from an image and annotate its contexts (e.g., location, severity and the affected organs). We employ a publicly available radiology dataset of chest x-rays and their reports, and use its image annotations to mine disease names to train convolutional neural networks (CNNs). In doing so, we adopt various regularization techniques to circumvent the large normal-vs-diseased cases bias. Recurrent neural networks (RNNs) are then trained to describe the contexts of a detected disease, based on the deep CNN features. Moreover, we introduce a novel approach to use the weights of the already trained pair of CNN/RNN on the domain-specific image/text dataset, to infer the joint image/text contexts for composite image labeling. Significantly improved image annotation results are demonstrated using the recurrent neural cascade model by taking the joint image/text contexts into account.

Xiaosong Wang, Le Lu, Hoo-chang Shin et al.

Obtaining semantic labels on a large scale radiology image database (215,786 key images from 61,845 unique patients) is a prerequisite yet bottleneck to train highly effective deep convolutional neural network (CNN) models for image recognition. Nevertheless, conventional methods for collecting image labels (e.g., Google search followed by crowd-sourcing) are not applicable due to the formidable difficulties of medical annotation tasks for those who are not clinically trained. This type of image labeling task remains non-trivial even for radiologists due to uncertainty and possible drastic inter-observer variation or inconsistency. In this paper, we present a looped deep pseudo-task optimization procedure for automatic category discovery of visually coherent and clinically semantic (concept) clusters. Our system can be initialized by domain-specific (CNN trained on radiology images and text report derived labels) or generic (ImageNet based) CNN models. Afterwards, a sequence of pseudo-tasks are exploited by the looped deep image feature clustering (to refine image labels) and deep CNN training/classification using new labels (to obtain more task representative deep features). Our method is conceptually simple and based on the hypothesized "convergence" of better labels leading to better trained CNN models which in turn feed more effective deep image features to facilitate more meaningful clustering/labels. We have empirically validated the convergence and demonstrated promising quantitative and qualitative results. Category labels of significantly higher quality than those in previous work are discovered. This allows for further investigation of the hierarchical semantic nature of the given large-scale radiology image database.

Hoo-Chang Shin, Holger R. Roth, Mingchen Gao et al.

Remarkable progress has been made in image recognition, primarily due to the availability of large-scale annotated datasets and the revival of deep CNN. CNNs enable learning data-driven, highly representative, layered hierarchical image features from sufficient training data. However, obtaining datasets as comprehensively annotated as ImageNet in the medical imaging domain remains a challenge. There are currently three major techniques that successfully employ CNNs to medical image classification: training the CNN from scratch, using off-the-shelf pre-trained CNN features, and conducting unsupervised CNN pre-training with supervised fine-tuning. Another effective method is transfer learning, i.e., fine-tuning CNN models pre-trained from natural image dataset to medical image tasks. In this paper, we exploit three important, but previously understudied factors of employing deep convolutional neural networks to computer-aided detection problems. We first explore and evaluate different CNN architectures. The studied models contain 5 thousand to 160 million parameters, and vary in numbers of layers. We then evaluate the influence of dataset scale and spatial image context on performance. Finally, we examine when and why transfer learning from pre-trained ImageNet (via fine-tuning) can be useful. We study two specific computer-aided detection (CADe) problems, namely thoraco-abdominal lymph node (LN) detection and interstitial lung disease (ILD) classification. We achieve the state-of-the-art performance on the mediastinal LN detection, with 85% sensitivity at 3 false positive per patient, and report the first five-fold cross-validation classification results on predicting axial CT slices with ILD categories. Our extensive empirical evaluation, CNN model analysis and valuable insights can be extended to the design of high performance CAD systems for other medical imaging tasks.

Holger R. Roth, Le Lu, Amal Farag et al.

Automatic organ segmentation is an important yet challenging problem for medical image analysis. The pancreas is an abdominal organ with very high anatomical variability. This inhibits previous segmentation methods from achieving high accuracies, especially compared to other organs such as the liver, heart or kidneys. In this paper, we present a probabilistic bottom-up approach for pancreas segmentation in abdominal computed tomography (CT) scans, using multi-level deep convolutional networks (ConvNets). We propose and evaluate several variations of deep ConvNets in the context of hierarchical, coarse-to-fine classification on image patches and regions, i.e. superpixels. We first present a dense labeling of local image patches via $P{-}\mathrm{ConvNet}$ and nearest neighbor fusion. Then we describe a regional ConvNet ($R_1{-}\mathrm{ConvNet}$) that samples a set of bounding boxes around each image superpixel at different scales of contexts in a "zoom-out" fashion. Our ConvNets learn to assign class probabilities for each superpixel region of being pancreas. Last, we study a stacked $R_2{-}\mathrm{ConvNet}$ leveraging the joint space of CT intensities and the $P{-}\mathrm{ConvNet}$ dense probability maps. Both 3D Gaussian smoothing and 2D conditional random fields are exploited as structured predictions for post-processing. We evaluate on CT images of 82 patients in 4-fold cross-validation. We achieve a Dice Similarity Coefficient of 83.6$\pm$6.3% in training and 71.8$\pm$10.7% in testing.

Hoo-Chang Shin, Le Lu, Lauren Kim et al.

Despite tremendous progress in computer vision, there has not been an attempt for machine learning on very large-scale medical image databases. We present an interleaved text/image deep learning system to extract and mine the semantic interactions of radiology images and reports from a national research hospital's Picture Archiving and Communication System. With natural language processing, we mine a collection of representative ~216K two-dimensional key images selected by clinicians for diagnostic reference, and match the images with their descriptions in an automated manner. Our system interleaves between unsupervised learning and supervised learning on document- and sentence-level text collections, to generate semantic labels and to predict them given an image. Given an image of a patient scan, semantic topics in radiology levels are predicted, and associated key-words are generated. Also, a number of frequent disease types are detected as present or absent, to provide more specific interpretation of a patient scan. This shows the potential of large-scale learning and prediction in electronic patient records available in most modern clinical institutions.

Holger R. Roth, Christopher T. Lee, Hoo-Chang Shin et al.

Automated classification of human anatomy is an important prerequisite for many computer-aided diagnosis systems. The spatial complexity and variability of anatomy throughout the human body makes classification difficult. "Deep learning" methods such as convolutional networks (ConvNets) outperform other state-of-the-art methods in image classification tasks. In this work, we present a method for organ- or body-part-specific anatomical classification of medical images acquired using computed tomography (CT) with ConvNets. We train a ConvNet, using 4,298 separate axial 2D key-images to learn 5 anatomical classes. Key-images were mined from a hospital PACS archive, using a set of 1,675 patients. We show that a data augmentation approach can help to enrich the data set and improve classification performance. Using ConvNets and data augmentation, we achieve anatomy-specific classification error of 5.9 % and area-under-the-curve (AUC) values of an average of 0.998 in testing. We demonstrate that deep learning can be used to train very reliable and accurate classifiers that could initialize further computer-aided diagnosis.