Ying Lin

Aaron Blakeman, Aarti Basant, Abhinav Khattar et al. · nvidia

As inference-time scaling becomes critical for enhanced reasoning capabilities, it is increasingly becoming important to build models that are efficient to infer. We introduce Nemotron-H, a family of 8B and 56B/47B hybrid Mamba-Transformer models designed to reduce inference cost for a given accuracy level. To achieve this goal, we replace the majority of self-attention layers in the common Transformer model architecture with Mamba layers that perform constant computation and require constant memory per generated token. We show that Nemotron-H models offer either better or on-par accuracy compared to other similarly-sized state-of-the-art open-sourced Transformer models (e.g., Qwen-2.5-7B/72B and Llama-3.1-8B/70B), while being up to 3$\times$ faster at inference. To further increase inference speed and reduce the memory required at inference time, we created Nemotron-H-47B-Base from the 56B model using a new compression via pruning and distillation technique called MiniPuzzle. Nemotron-H-47B-Base achieves similar accuracy to the 56B model, but is 20% faster to infer. In addition, we introduce an FP8-based training recipe and show that it can achieve on par results with BF16-based training. This recipe is used to train the 56B model. We are releasing Nemotron-H base model checkpoints with support in Hugging Face and NeMo.

Aakshita Chandiramani, Aaron Blakeman, Abdullahi Olaoye et al. · amazon-science, cmu

We describe the pre-training, post-training, and quantization of Nemotron 3 Super, a 120 billion (active 12 billion) parameter hybrid Mamba-Attention Mixture-of-Experts model. Nemotron 3 Super is the first model in the Nemotron 3 family to 1) be pre-trained in NVFP4, 2) leverage LatentMoE, a new Mixture-of-Experts architecture that optimizes for both accuracy per FLOP and accuracy per parameter, and 3) include MTP layers for inference acceleration through native speculative decoding. We pre-trained Nemotron 3 Super on 25 trillion tokens followed by post-training using supervised fine tuning (SFT) and reinforcement learning (RL). The final model supports up to 1M context length and achieves comparable accuracy on common benchmarks, while also achieving up to 2.2x and 7.5x higher inference throughput compared to GPT-OSS-120B and Qwen3.5-122B, respectively. Nemotron 3 Super datasets, along with the base, post-trained, and quantized checkpoints, are open-sourced on HuggingFace.

Aarti Basant, Abhijit Khairnar, Abhijit Paithankar et al. · nvidia

We introduce Nemotron-Nano-9B-v2, a hybrid Mamba-Transformer language model designed to increase throughput for reasoning workloads while achieving state-of-the-art accuracy compared to similarly-sized models. Nemotron-Nano-9B-v2 builds on the Nemotron-H architecture, in which the majority of the self-attention layers in the common Transformer architecture are replaced with Mamba-2 layers, to achieve improved inference speed when generating the long thinking traces needed for reasoning. We create Nemotron-Nano-9B-v2 by first pre-training a 12-billion-parameter model (Nemotron-Nano-12B-v2-Base) on 20 trillion tokens using an FP8 training recipe. After aligning Nemotron-Nano-12B-v2-Base, we employ the Minitron strategy to compress and distill the model with the goal of enabling inference on up to 128k tokens on a single NVIDIA A10G GPU (22GiB of memory, bfloat16 precision). Compared to existing similarly-sized models (e.g., Qwen3-8B), we show that Nemotron-Nano-9B-v2 achieves on-par or better accuracy on reasoning benchmarks while achieving up to 6x higher inference throughput in reasoning settings like 8k input and 16k output tokens. We are releasing Nemotron-Nano-9B-v2, Nemotron-Nano12B-v2-Base, and Nemotron-Nano-9B-v2-Base checkpoints along with the majority of our pre- and post-training datasets on Hugging Face.

Aaron Blakeman, Aaron Grattafiori, Aarti Basant et al. · nvidia

We present Nemotron 3 Nano 30B-A3B, a Mixture-of-Experts hybrid Mamba-Transformer language model. Nemotron 3 Nano was pretrained on 25 trillion text tokens, including more than 3 trillion new unique tokens over Nemotron 2, followed by supervised fine tuning and large-scale RL on diverse environments. Nemotron 3 Nano achieves better accuracy than our previous generation Nemotron 2 Nano while activating less than half of the parameters per forward pass. It achieves up to 3.3x higher inference throughput than similarly-sized open models like GPT-OSS-20B and Qwen3-30B-A3B-Thinking-2507, while also being more accurate on popular benchmarks. Nemotron 3 Nano demonstrates enhanced agentic, reasoning, and chat abilities and supports context lengths up to 1M tokens. We release both our pretrained Nemotron 3 Nano 30B-A3B Base and post-trained Nemotron 3 Nano 30B-A3B checkpoints on Hugging Face.

Aaron Blakeman, Aaron Grattafiori, Aarti Basant et al. · nvidia

We introduce the Nemotron 3 family of models - Nano, Super, and Ultra. These models deliver strong agentic, reasoning, and conversational capabilities. The Nemotron 3 family uses a Mixture-of-Experts hybrid Mamba-Transformer architecture to provide best-in-class throughput and context lengths of up to 1M tokens. Super and Ultra models are trained with NVFP4 and incorporate LatentMoE, a novel approach that improves model quality. The two larger models also include MTP layers for faster text generation. All Nemotron 3 models are post-trained using multi-environment reinforcement learning enabling reasoning, multi-step tool use, and support granular reasoning budget control. Nano, the smallest model, outperforms comparable models in accuracy while remaining extremely cost-efficient for inference. Super is optimized for collaborative agents and high-volume workloads such as IT ticket automation. Ultra, the largest model, provides state-of-the-art accuracy and reasoning performance. Nano is released together with its technical report and this white paper, while Super and Ultra will follow in the coming months. We will openly release the model weights, pre- and post-training software, recipes, and all data for which we hold redistribution rights.

Akhiad Bercovich, Itay Levy, Izik Golan et al. · nvidia

We introduce the Llama-Nemotron series of models, an open family of heterogeneous reasoning models that deliver exceptional reasoning capabilities, inference efficiency, and an open license for enterprise use. The family comes in three sizes -- Nano (8B), Super (49B), and Ultra (253B) -- and performs competitively with state-of-the-art reasoning models such as DeepSeek-R1 while offering superior inference throughput and memory efficiency. In this report, we discuss the training procedure for these models, which entails using neural architecture search from Llama 3 models for accelerated inference, knowledge distillation, and continued pretraining, followed by a reasoning-focused post-training stage consisting of two main parts: supervised fine-tuning and large scale reinforcement learning. Llama-Nemotron models are the first open-source models to support a dynamic reasoning toggle, allowing users to switch between standard chat and reasoning modes during inference. To further support open research and facilitate model development, we provide the following resources: 1. We release the Llama-Nemotron reasoning models -- LN-Nano, LN-Super, and LN-Ultra -- under the commercially permissive NVIDIA Open Model License Agreement. 2. We release the complete post-training dataset: Llama-Nemotron-Post-Training-Dataset. 3. We also release our training codebases: NeMo, NeMo-Aligner, and Megatron-LM.

Yuxuan Li, Ying Lin, Chenang Liu

Investigating the causal relationships between characteristics and expressions plays a critical role in healthcare analytics. Effective synthesis for expressions using given characteristics can make great contributions to health risk management and medical decision-making. For example, predicting the resulting physiological symptoms on patients from given treatment characteristics is helpful for the disease prevention and personalized treatment strategy design. Therefore, the objective of this study is to effectively synthesize the expressions based on given characteristics. However, the mapping from characteristics to expressions is usually from a relatively low dimension space to a high dimension space, but most of the existing methods such as regression models could not effectively handle such mapping. Besides, the relationship between characteristics and expressions may contain not only deterministic patterns, but also stochastic patterns. To address these challenges, this paper proposed a novel selective ensemble characteristic-to-expression synthesis (SE-CTES) approach inspired by generative adversarial network (GAN). The novelty of the proposed method can be summarized into three aspects: (1) GAN-based architecture for deep neural networks are incorporated to learn the relatively low dimensional mapping to high dimensional mapping containing both deterministic and stochastic patterns; (2) the weights of the two mismatching errors in the GAN-based architecture are proposed to be different to reduce the learning bias in the training process; and (3) a selective ensemble learning framework is proposed to reduce the prediction bias and improve the synthesis stability. To validate the effectiveness of the proposed approach, extensive numerical simulation studies and a real-world healthcare case study were applied and the results demonstrated that the proposed method is very promising.

Shiva Afshar, Yinghan Chen, Shizhong Han et al.

Combining multiple predictors obtained from distributed data sources to an accurate meta-learner is promising to achieve enhanced performance in lots of prediction problems. As the accuracy of each predictor is usually unknown, integrating the predictors to achieve better performance is challenging. Conventional ensemble learning methods assess the accuracy of predictors based on extensive labeled data. In practical applications, however, the acquisition of such labeled data can prove to be an arduous task. Furthermore, the predictors under consideration may exhibit high degrees of correlation, particularly when similar data sources or machine learning algorithms were employed during their model training. In response to these challenges, this paper introduces a novel structured unsupervised ensemble learning model (SUEL) to exploit the dependency between a set of predictors with continuous predictive scores, rank the predictors without labeled data and combine them to an ensembled score with weights. Two novel correlation-based decomposition algorithms are further proposed to estimate the SUEL model, constrained quadratic optimization (SUEL.CQO) and matrix-factorization-based (SUEL.MF) approaches. The efficacy of the proposed methods is rigorously assessed through both simulation studies and real-world application of risk genes discovery. The results compellingly demonstrate that the proposed methods can efficiently integrate the dependent predictors to an ensemble model without the need of ground truth data.

Dan Su, Kezhi Kong, Ying Lin et al.

Recent English Common Crawl datasets like FineWeb-Edu and DCLM achieved significant benchmark gains via aggressive model-based filtering, but at the cost of removing 90% of data. This limits their suitability for long token horizon training, such as 15T tokens for Llama 3.1. In this paper, we show how to achieve better trade-offs between accuracy and data quantity by a combination of classifier ensembling, synthetic data rephrasing, and reduced reliance on heuristic filters. When training 8B parameter models for 1T tokens, using a high-quality subset of our data improves MMLU by 5.6 over DCLM, demonstrating the efficacy of our methods for boosting accuracies over a relatively short token horizon. Furthermore, our full 6.3T token dataset matches DCLM on MMLU, but contains four times more unique real tokens than DCLM. This unlocks state-of-the-art training over a long token horizon: an 8B parameter model trained for 15T tokens, of which 7.2T came from our dataset, is better than the Llama 3.1 8B model: +5 on MMLU, +3.1 on ARC-Challenge, and +0.5 on average across ten diverse tasks. The dataset is available at https://data.commoncrawl.org/contrib/Nemotron/Nemotron-CC/index.html

Muyun Lu, Haoyang Hong, Huazheng Wang et al.

As an extension of the classical multi-armed bandit problem, multi-fidelity multi-armed bandits (MF-MAB) enable individual arms to be evaluated using diverse feedback sources that vary in both cost and accuracy. Prior stochastic models typically assume fixed low-to-high fidelity discrepancies, whereas modern proxy sources, such as learning-based simulators and Large Language Models (LLMs), can be improved using additional calibration. We investigate adaptive MF-MAB with improving proxy sources, and focus on the canonical two-fidelity case in which the low-fidelity source becomes more informative with repeated use. To capture this dynamic, we introduce a selected-average mismatch bound that converts dynamic low-fidelity observations into improvement-aware confidence bounds for the high-fidelity target. We propose the Threshold-Based Adaptive Continuation Companion (TACC), an optimistic algorithm that uses a bounded continuation rule to decide when low-fidelity sampling remains cost-effective and when to escalate. We prove an instance-dependent regret bound showing that, for detected intermediate arms, adaptive continuation replaces logarithmic high-fidelity confirmation with bounded low-fidelity continuation. Experiments on synthetic bandits and an LLM-as-a-judge policy-evaluation task examine when continuation improves cost-weighted regret.

Jian Cao, Nian Liu, Ying Lin

We introduce a novel framework for constructing scalable and flexible covariance kernels for Gaussian processes (GPs) by directly learning the covariance structure under a regression-type parameterization induced by Vecchia approximations, using deep neural architectures. Specifically, we model kriging coefficients and conditional standard deviations, deterministic quantities that uniquely characterize the covariance, providing stable and informative learning targets. Exploiting the permutation-equivariant structure of conditioning sets in the Vecchia factorization, we derive a universal representation for permutation-preserving functions and design neural architectures that respect this symmetry, leading to improved training stability and data efficiency. The proposed approach enables expressive, non-stationary kernel learning while maintaining computational scalability, thereby bridging classical GP methodology with modern deep learning.

Kaituo Zhang, Zhen Xiong, Mingyu Zhong et al.

Tool-augmented reasoning has become a popular direction for LLM-based agents, and it is widely assumed to improve reasoning and reliability. However, we demonstrate that this consensus does not always hold: in the presence of semantic distractors, tool-augmented reasoning does not necessarily outperform native CoT. To explain this performance gap, we propose a Factorized Intervention Framework that isolates the cost of prompt formatting, the overhead of the tool-calling protocol, and the actual gain from executing tools. Our analysis reveals a critical tradeoff: under semantic noise, the gains from tools often fail to offset the "tool-use tax", which is the performance degradation introduced by the tool-calling protocol itself. To address this, we introduce G-STEP, a lightweight inference-time gate to mitigate protocol-induced errors. While this yields partial recovery, our findings suggest that more substantial improvements still require strengthening the model's intrinsic reasoning and tool-interaction capabilities.

Syeda Nahida Akter, Shrimai Prabhumoye, Matvei Novikov et al. · stanford

Large Language Models (LLMs) have shown strong reasoning capabilities, particularly when enhanced through Reinforcement Learning (RL). While prior work has successfully applied RL to mathematical reasoning -- where rules and correctness are well-defined -- generalizing these methods to broader reasoning domains remains challenging due to limited data, the lack of verifiable reward structures, and diverse task requirements. In this work, we propose NEMOTRON-CROSSTHINK, a framework that systematically incorporates multi-domain corpora, including both synthetic and real-world question-answer pairs, into RL training to improve generalization across diverse reasoning tasks. NEMOTRON-CROSSTHINK addresses key challenges by (1) incorporating data from varied sources spanning STEM, humanities, social sciences, etc.; (2) applying structured templates (e.g., multiple-choice and open-ended) to control answer-space complexity; (3) filtering for verifiable answers; and (4) optimizing data blending strategies that utilizes data from multiple sources effectively. Our approach enables scalable and verifiable reward modeling beyond mathematics and demonstrates improved accuracies on both math (MATH-500: +30.1%, AMC23:+27.5%) and non-math reasoning benchmarks (MMLU-PRO: +12.8%, GPQA-DIAMOND: +11.3%, AGIEVAL: +15.1%, SUPERGPQA: +3.8%). Moreover, NEMOTRON-CROSSTHINK exhibits significantly improved response efficiency -- using 28% fewer tokens for correct answers -- highlighting more focused and effective reasoning. Through NEMOTRON-CROSSTHINK, we demonstrate that integrating multi-domain, multi-format data in RL leads to more accurate, efficient, and generalizable LLMs.

Tanapol Kosolwattana, Huazheng Wang, Ying Lin

Adaptive monitoring of a large population of dynamic processes is critical for the timely detection of abnormal events under limited resources in many healthcare and engineering systems. Examples include the risk-based disease screening and condition-based process monitoring. However, existing adaptive monitoring models either ignore the dependency among processes or overlook the uncertainty in process modeling. To design an optimal monitoring strategy that accurately monitors the processes with poor health conditions and actively collects information for uncertainty reduction, a novel online collaborative learning method is proposed in this study. The proposed method designs a collaborative learning-based upper confidence bound (CL-UCB) algorithm to optimally balance the exploitation and exploration of dependent processes under limited resources. Efficiency of the proposed method is demonstrated through theoretical analysis, simulation studies and an empirical study of adaptive cognitive monitoring in Alzheimer's disease.

Kaituo Zhang, Mingzhi Hu, Hoang Anh Duy Le et al.

Large Language Models (LLMs) have emerged as powerful tools for generating data across various modalities. By transforming data from a scarce resource into a controllable asset, LLMs mitigate the bottlenecks imposed by the acquisition costs of real-world data for model training, evaluation, and system iteration. However, ensuring the high quality of LLM-generated synthetic data remains a critical challenge. Existing research primarily focuses on generation methodologies, with limited direct attention to the quality of the resulting data. Furthermore, most studies are restricted to single modalities, lacking a unified perspective across different data types. To bridge this gap, we propose the \textbf{LLM Data Auditor framework}. In this framework, we first describe how LLMs are utilized to generate data across six distinct modalities. More importantly, we systematically categorize intrinsic metrics for evaluating synthetic data from two dimensions: quality and trustworthiness. This approach shifts the focus from extrinsic evaluation, which relies on downstream task performance, to the inherent properties of the data itself. Using this evaluation system, we analyze the experimental evaluations of representative generation methods for each modality and identify substantial deficiencies in current evaluation practices. Based on these findings, we offer concrete recommendations for the community to improve the evaluation of data generation. Finally, the framework outlines methodologies for the practical application of synthetic data across different modalities.

Ying Lin, Yao Kuang, Ahmet Alacaoglu et al.

Distributed optimization requires nodes to coordinate, yet full synchronization scales poorly. When $n$ nodes collaborate through $m$ pairwise regularizers, standard methods demand $\mathcal{O}(m)$ communications per iteration. This paper proposes randomized local coordination: each node independently samples one regularizer uniformly and coordinates only with nodes sharing that term. This exploits partial separability, where each regularizer $G_j$ depends on a subset $S_j \subseteq \{1,\ldots,n\}$ of nodes. For graph-guided regularizers where $|S_j|=2$, expected communication drops to exactly 2 messages per iteration. This method achieves $\tilde{\mathcal{O}}(\varepsilon^{-2})$ iterations for convex objectives and under strong convexity, $\mathcal{O}(\varepsilon^{-1})$ to an $\varepsilon$-solution and $\mathcal{O}(\log(1/\varepsilon))$ to a neighborhood. Replacing the proximal map of the sum $\sum_j G_j$ with the proximal map of a single randomly selected regularizer $G_j$ preserves convergence while eliminating global coordination. Experiments validate both convergence rates and communication efficiency across synthetic and real-world datasets.

Zhishen Nie, Ying Lin, Sp Ren et al.

Adversarial training has become the primary method to defend against adversarial samples. However, it is hard to practically apply due to many shortcomings. One of the shortcomings of adversarial training is that it will reduce the recognition accuracy of normal samples. Adaptive perturbation adversarial training is proposed to alleviate this problem. It uses marginal adversarial samples that are close to the decision boundary but does not cross the decision boundary for adversarial training, which improves the accuracy of model recognition while maintaining the robustness of the model. However, searching for marginal adversarial samples brings additional computational costs. This paper proposes a method for finding marginal adversarial samples based on reinforcement learning, and combines it with the latest fast adversarial training technology, which effectively speeds up training process and reduces training costs.

Ying Lin, Han Wang, Jiangning Chen et al.

The growing popularity of Virtual Assistants poses new challenges for Entity Resolution, the task of linking mentions in text to their referent entities in a knowledge base. Specifically, in the shopping domain, customers tend to use implicit utterances (e.g., "organic milk") rather than explicit names, leading to a large number of candidate products. Meanwhile, for the same query, different customers may expect different results. For example, with "add milk to my cart", a customer may refer to a certain organic product, while some customers may want to re-order products they regularly purchase. To address these issues, we propose a new framework that leverages personalized features to improve the accuracy of product ranking. We first build a cross-source heterogeneous knowledge graph from customer purchase history and product knowledge graph to jointly learn customer and product embeddings. After that, we incorporate product, customer, and history representations into a neural reranking model to predict which candidate is most likely to be purchased for a specific customer. Experiments show that our model substantially improves the accuracy of the top ranked candidates by 24.6% compared to the state-of-the-art product search model.

Qingyun Wang, Manling Li, Xuan Wang et al.

To combat COVID-19, both clinicians and scientists need to digest vast amounts of relevant biomedical knowledge in scientific literature to understand the disease mechanism and related biological functions. We have developed a novel and comprehensive knowledge discovery framework, COVID-KG to extract fine-grained multimedia knowledge elements (entities and their visual chemical structures, relations, and events) from scientific literature. We then exploit the constructed multimedia knowledge graphs (KGs) for question answering and report generation, using drug repurposing as a case study. Our framework also provides detailed contextual sentences, subfigures, and knowledge subgraphs as evidence.

Ronald Cardenas, Ying Lin, Heng Ji et al.

Unsupervised part of speech (POS) tagging is often framed as a clustering problem, but practical taggers need to \textit{ground} their clusters as well. Grounding generally requires reference labeled data, a luxury a low-resource language might not have. In this work, we describe an approach for low-resource unsupervised POS tagging that yields fully grounded output and requires no labeled training data. We find the classic method of Brown et al. (1992) clusters well in our use case and employ a decipherment-based approach to grounding. This approach presumes a sequence of cluster IDs is a `ciphertext' and seeks a POS tag-to-cluster ID mapping that will reveal the POS sequence. We show intrinsically that, despite the difficulty of the task, we obtain reasonable performance across a variety of languages. We also show extrinsically that incorporating our POS tagger into a name tagger leads to state-of-the-art tagging performance in Sinhalese and Kinyarwanda, two languages with nearly no labeled POS data available. We further demonstrate our tagger's utility by incorporating it into a true `zero-resource' variant of the Malopa (Ammar et al., 2016) dependency parser model that removes the current reliance on multilingual resources and gold POS tags for new languages. Experiments show that including our tagger makes up much of the accuracy lost when gold POS tags are unavailable.

Ying Lin, Joe Hoover, Morteza Dehghani et al.

In this paper, we address the problem of detecting expressions of moral values in tweets using content analysis. This is a particularly challenging problem because moral values are often only implicitly signaled in language, and tweets contain little contextual information due to length constraints. To address these obstacles, we present a novel approach to automatically acquire background knowledge from an external knowledge base to enrich input texts and thus improve moral value prediction. By combining basic text features with background knowledge, our overall context-aware framework achieves performance comparable to a single human annotator. To the best of our knowledge, this is the first attempt to incorporate background knowledge for the prediction of implicit psychological variables in the area of computational social science.

Ying Lin, Zhengzhang Chen, Cheng Cao et al.

Given a large number of low-level heterogeneous categorical alerts from an anomaly detection system, how to characterize complex relationships between different alerts, filter out false positives, and deliver trustworthy rankings and suggestions to end users? This problem is motivated by and generalized from applications in enterprise security and attack scenario reconstruction. While existing techniques focus on either reconstructing abnormal scenarios or filtering out false positive alerts, it can be more advantageous to consider the two perspectives simultaneously in order to improve detection accuracy and better understand anomaly behaviors. In this paper, we propose CAR, a collaborative alerts ranking framework that exploits both temporal and content correlations from heterogeneous categorical alerts. CAR first builds a tree-based model to capture both short-term correlations and long-term dependencies in each alert sequence, which identifies abnormal action sequences. Then, an embedding-based model is employed to learn the content correlations between alerts via their heterogeneous categorical attributes. Finally, by incorporating both temporal and content dependencies into one optimization framework, CAR ranks both alerts and their corresponding alert patterns. Our experiments, using real-world enterprise monitoring data and real attacks launched by professional hackers, show that CAR can accurately identify true positive alerts and successfully reconstruct attack scenarios at the same time.

Boxiang Dong, Zhengzhang Chen, Hui Wang et al.

Intrusion detection system (IDS) is an important part of enterprise security system architecture. In particular, anomaly-based IDS has been widely applied to detect abnormal process behaviors that deviate from the majority. However, such abnormal behavior usually consists of a series of low-level heterogeneous events. The gap between the low-level events and the high-level abnormal behaviors makes it hard to infer which single events are related to the real abnormal activities, especially considering that there are massive "noisy" low-level events happening in between. Hence, the existing work that focus on detecting single entities/events can hardly achieve high detection accuracy. Different from previous work, we design and implement GID, an efficient graph-based intrusion detection technique that can identify abnormal event sequences from a massive heterogeneous process traces with high accuracy. GID first builds a compact graph structure to capture the interactions between different system entities. The suspiciousness or anomaly score of process paths is then measured by leveraging random walk technique to the constructed acyclic directed graph. To eliminate the score bias from the path length, the Box-Cox power transformation based approach is introduced to normalize the anomaly scores so that the scores of paths of different lengths have the same distribution. The efficiency of suspicious path discovery is further improved by the proposed optimization scheme. We fully implement our GID algorithm and deploy it into a real enterprise security system, and it greatly helps detect the advanced threats, and optimize the incident response. Executing GID on system monitoring datasets showing that GID is efficient (about 2 million records per minute) and accurate (higher than 80% in terms of detection rate).