Harsh Sharma

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.

Ammar Bhilwarawala, Likhamba Rongmei, Harsh Sharma et al.

IoT botnet detection has advanced, yet most published systems are validated on a single dataset and rarely generalise across environments. Heterogeneous feature spaces make multi-dataset training practically impossible without discarding semantic interpretability or introducing data integrity violations. No prior work has addressed both problems with a formally specified, reproducible methodology. This paper does. We introduce BRIDGE (Benchmark Reference for IoT Domain Generalisation Evaluation), the first formally specified heterogeneous multi-dataset benchmark for IoT intrusion detection, unifying CICIDS-2017, CIC-IoT-2023, Bot-IoT, Edge-IIoTset, and N-BaIoT through a 46-feature semantic canonical vocabulary grounded in CICFlowMeter nomenclature, with genuine-equivalence-only feature mapping, explicit zero-filling, and per-dataset coverage from 15% to 93%. A leave-one-dataset-out (LODO) protocol makes the generalisation gap precisely measurable: all five evaluated architectures achieve mean LODO F1 between 0.39 and 0.47, and we establish the first community generalisation baseline at mean LODO F1 = 0.5577, a result that shifts the agenda from single-benchmark optimisation toward cross-environment generalisation. We propose TCH-Net, a multi-branch network fusing a three-path Temporal branch (residual convolutional-BiGRU, stride-downsampled BiGRU, pre-LayerNorm Transformer), a provenance-conditioned Contextual branch, and a Statistical branch via Cross-Branch Gated Attention Fusion (CB-GAF) with learnable sigmoid gates for dynamic feature-wise mixing. Across five random seeds, TCH-Net achieves F1 = 0.8296 +/- 0.0028, AUC = 0.9380 +/- 0.0025, and MCC = 0.6972 +/- 0.0056, outperforming all twelve baselines (p < 0.05, Wilcoxon) and recording the highest LODO F1 overall. BRIDGE and the full pipeline are at https://github.com/Ammar-ss/TCH-Net.

Harsh Sharma, Mayuresh Patil, Craig Woolsey

The goal of this paper is to develop energy-preserving variational integrators for time-dependent mechanical systems with forcing. We first present the Lagrange-d'Alembert principle in the extended Lagrangian mechanics framework and derive the extended forced Euler-Lagrange equations in continuous-time. We then obtain the extended forced discrete Euler-Lagrange equations using the extended discrete mechanics framework and derive adaptive time step variational integrators for time-dependent Lagrangian systems with forcing. We consider two numerical examples to study the numerical performance of energy-preserving variational integrators. First, we consider the example of a nonlinear conservative system to illustrate the advantages of using adaptive time-stepping in variational integrators. We show a trade-off between energy-preserving performance and accurate discrete trajectories while choosing an initial time step. In addition, we demonstrate how the implicit equations become more ill-conditioned as the adaptive time step increases through a condition number analysis. As a second example, we numerically simulate a damped harmonic oscillator using the adaptive time step variational integrator framework. The adaptive time step increases monotonically for the dissipative system leading to unexpected energy behavior.

Harsh Sharma, Iman Adibnazari, Jacobo Cervera-Torralba et al.

Data-driven model reduction methods provide a nonintrusive way of constructing computationally efficient surrogates of high-fidelity models for real-time control of soft robots. This work leverages the Lagrangian nature of the model equations to derive structure-preserving linear reduced-order models via Lagrangian Operator Inference and compares their performance with prominent linear model reduction techniques through an anguilliform swimming soft robot model example with 231,336 degrees of freedom. The case studies demonstrate that preserving the underlying Lagrangian structure leads to learned models with higher predictive accuracy and robustness to unseen inputs.

Iyad Ait Hou, Shrenik Borad, Harsh Sharma et al.

Machine unlearning is increasingly important for clinical language models, where privacy regulations and institutional policies may require removing sensitive information from deployed systems without retraining from scratch. In practice, deletion requests must balance effective forgetting of targeted information with preservation of model utility and minimal parameter modification. We introduce Sparse Token Embedding Unlearning (STEU), a parameter-efficient method for behavioral class-level unlearning that updates only PMI-selected token embeddings together with a small classifier head while keeping all encoder layers frozen. Across experiments on MIMIC-IV, MIMIC-III, and eICU using BioClinicalBERT, BERT-base, and DistilBERT, STEU consistently suppresses the target class while largely preserving retained task performance. In the primary MIMIC-IV setting, STEU achieves near-complete forgetting (forget F1 = 0.0004) while maintaining competitive retained utility (retain avg F1 = 0.4766) after modifying only 0.19\% of model parameters. These results suggest that targeted behavioral unlearning can be achieved through sparse embedding edits without modifying deeper encoder representations.

Harsh Sharma, Nicholas Galioto, Alex A. Gorodetsky et al.

This paper proposes a probabilistic Bayesian formulation for system identification (ID) and estimation of nonseparable Hamiltonian systems using stochastic dynamic models. Nonseparable Hamiltonian systems arise in models from diverse science and engineering applications such as astrophysics, robotics, vortex dynamics, charged particle dynamics, and quantum mechanics. The numerical experiments demonstrate that the proposed method recovers dynamical systems with higher accuracy and reduced predictive uncertainty compared to state-of-the-art approaches. The results further show that accurate predictions far outside the training time interval in the presence of sparse and noisy measurements are possible, which lends robustness and generalizability to the proposed approach. A quantitative benefit is prediction accuracy with less than 10% relative error for more than 12 times longer than a comparable least-squares-based method on a benchmark problem.

Debaditya Pal, Kaustubh Chaudhari, Harsh Sharma

Hate speech classification has been a long-standing problem in natural language processing. However, even though there are numerous hate speech detection methods, they usually overlook a lot of hateful statements due to them being implicit in nature. Developing datasets to aid in the task of implicit hate speech classification comes with its own challenges; difficulties are nuances in language, varying definitions of what constitutes hate speech, and the labor-intensive process of annotating such data. This had led to a scarcity of data available to train and test such systems, which gives rise to high variance problems when parameter-heavy transformer-based models are used to address the problem. In this paper, we explore various optimization and regularization techniques and develop a novel RoBERTa-based model that achieves state-of-the-art performance.

Aboli Kathar, Aman Kumar, Anusha Kamath et al.

We present FiMI (Finance Model for India), a domain-specialized financial language model developed by National Payments Corporation of India (NPCI) for Indian digital payment systems. We develop two model variants: FiMI Base and FiMI Instruct. FiMI adapts the Mistral Small 24B architecture through a multi-stage training pipeline, beginning with continuous pre-training on 68 Billion tokens of curated financial, multilingual (English, Hindi, Hinglish), and synthetic data. This is followed by instruction fine-tuning and domain-specific supervised fine-tuning focused on multi-turn, tool-driven conversations that model real-world workflows, such as transaction disputes and mandate lifecycle management. Evaluations reveal that FiMI Base achieves a 20\% improvement over the Mistral Small 24B Base model on finance reasoning benchmark, while FiMI Instruct outperforms the Mistral Small 24B Instruct model by 87\% on domain-specific tool-calling. Moreover, FiMI achieves these significant domain gains while maintaining comparable performance to models of similar size on general benchmarks.

Eray Erturk, Fahad Kamran, Salar Abbaspourazad et al.

Wearable devices record physiological and behavioral signals that can improve health predictions. While foundation models are increasingly used for such predictions, they have been primarily applied to low-level sensor data, despite behavioral data often being more informative due to their alignment with physiologically relevant timescales and quantities. We develop foundation models of such behavioral signals using over 2.5B hours of wearable data from 162K individuals, systematically optimizing architectures and tokenization strategies for this unique dataset. Evaluated on 57 health-related tasks, our model shows strong performance across diverse real-world applications including individual-level classification and time-varying health state prediction. The model excels in behavior-driven tasks like sleep prediction, and improves further when combined with representations of raw sensor data. These results underscore the importance of tailoring foundation model design to wearables and demonstrate the potential to enable new health applications.

Shubham Mishra, Samyek Jain, Gorang Mehrishi et al.

Retrieval-Augmented Generation (RAG) grounds large language models (LLMs) in external evidence, but fails when retrieved sources conflict or contain outdated or subjective information. Prior work address these issues independently but lack unified reasoning supervision. We propose a reasoning-trace-augmented RAG framework that adds structured, interpretable reasoning across three stages : (1) document-level adjudication, (2) conflict analysis, and (3) grounded synthesis, producing citation-linked answers or justified refusals. A Conflict-Aware Trust-Score (CATS) pipeline is introduced which evaluates groundedness, factual correctness, refusal accuracy, and conflict-behavior alignment using an LLM-as-a-Judge. Our 539-query reasoning dataset and evaluation pipeline establish a foundation for conflict-aware, interpretable RAG systems. Experimental results demonstrate substantial gains over baselines, most notably with Qwen, where Supervised Fine-Tuning improved End-to-End answer correctness from 0.069 to 0.883 and behavioral adherence from 0.074 to 0.722.

Harsh Sharma, Juan Diego Draxl Giannoni, Boris Kramer

Existing model reduction techniques for high-dimensional models of conservative partial differential equations (PDEs) encounter computational bottlenecks when dealing with systems featuring non-polynomial nonlinearities. This work presents a nonlinear model reduction method that employs lifting variable transformations to derive structure-preserving quadratic reduced-order models for conservative PDEs with general nonlinearities. We present an energy-quadratization strategy that defines the auxiliary variable in terms of the nonlinear term in the energy expression to derive an equivalent quadratic lifted system with quadratic system energy. The proposed strategy combined with proper orthogonal decomposition model reduction yields quadratic reduced-order models that conserve the quadratized lifted energy exactly in high dimensions. We demonstrate the proposed model reduction approach on four nonlinear conservative PDEs: the one-dimensional wave equation with exponential nonlinearity, the two-dimensional sine-Gordon equation, the two-dimensional Klein-Gordon equation with parametric dependence, and the two-dimensional Klein-Gordon-Zakharov equations. The numerical results show that the proposed lifting approach is competitive with the state-of-the-art structure-preserving hyper-reduction method in terms of both accuracy and computational efficiency in the online stage while providing significant computational gains in the offline stage.

Daniel Lee, Nikhil Sharma, Donghoon Shin et al. · uw

Generative AI has made image creation more accessible, yet aligning outputs with nuanced creative intent remains challenging, particularly for non-experts. Existing tools often require users to externalize ideas through prompts or references, limiting fluid exploration. We introduce ThematicPlane, a system that enables users to navigate and manipulate high-level semantic concepts (e.g., mood, style, or narrative tone) within an interactive thematic design plane. This interface bridges the gap between tacit creative intent and system control. In our exploratory study (N=6), participants engaged in divergent and convergent creative modes, often embracing unexpected results as inspiration or iteration cues. While they grounded their exploration in familiar themes, differing expectations of how themes mapped to outputs revealed a need for more explainable controls. Overall, ThematicPlane fosters expressive, iterative workflows and highlights new directions for intuitive, semantics-driven interaction in generative design tools.

Nicholas Galioto, Harsh Sharma, Boris Kramer et al.

This paper presents a structure-preserving Bayesian approach for learning nonseparable Hamiltonian systems using stochastic dynamic models allowing for statistically-dependent, vector-valued additive and multiplicative measurement noise. The approach is comprised of three main facets. First, we derive a Gaussian filter for a statistically-dependent, vector-valued, additive and multiplicative noise model that is needed to evaluate the likelihood within the Bayesian posterior. Second, we develop a novel algorithm for cost-effective application of Bayesian system identification to high-dimensional systems. Third, we demonstrate how structure-preserving methods can be incorporated into the proposed framework, using nonseparable Hamiltonians as an illustrative system class. We assess the method's performance based on the forecasting accuracy of a model estimated from single-trajectory data. We compare the Bayesian method to a state-of-the-art machine learning method on a canonical nonseparable Hamiltonian model and a chaotic double pendulum model with small, noisy training datasets. The results show that using the Bayesian posterior as a training objective can yield upwards of 724 times improvement in Hamiltonian mean squared error using training data with up to 10% multiplicative noise compared to a standard training objective. Lastly, we demonstrate the utility of the novel algorithm for parameter estimation of a 64-dimensional model of the spatially-discretized nonlinear Schrödinger equation with data corrupted by up to 20% multiplicative noise.

Harsh Sharma, Gaurav Narang, Janardhan Rao Doppa et al.

Processing-in-memory (PIM) has emerged as an enabler for the energy-efficient and high-performance acceleration of deep learning (DL) workloads. Resistive random-access memory (ReRAM) is one of the most promising technologies to implement PIM. However, as the complexity of Deep convolutional neural networks (DNNs) grows, we need to design a manycore architecture with multiple ReRAM-based processing elements (PEs) on a single chip. Existing PIM-based architectures mostly focus on computation while ignoring the role of communication. ReRAM-based tiled manycore architectures often involve many Processing Elements (PEs), which need to be interconnected via an efficient on-chip communication infrastructure. Simply allocating more resources (ReRAMs) to speed up only computation is ineffective if the communication infrastructure cannot keep up with it. In this paper, we highlight the design principles of a dataflow-aware PIM-enabled manycore platform tailor-made for various types of DL workloads. We consider the design challenges with both 2.5D interposer- and 3D integration-enabled architectures.

Arnav Shukla, Harsh Sharma, Srikant Bharadwaj et al.

Heterogeneous chiplet-based systems improve scaling by disag-gregating CPUs/GPUs and emerging technologies (HBM/DRAM).However this on-package disaggregation introduces a latency inNetwork-on-Interposer(NoI). We observe that in modern large-modelinference, parameters and activations routinely move backand forth from HBM/DRAM, injecting large, bursty flows into theinterposer. These memory-driven transfers inflate tail latency andviolate Service Level Agreements (SLAs) across k-ary n-cube base-line NoI topologies. To address this gap we introduce an InterferenceScore (IS) that quantifies worst-case slowdown under contention.We then formulate NoI synthesis as a multi-objective optimization(MOO) problem. We develop PARL (Partition-Aware ReinforcementLearner), a topology generator that balances throughput, latency,and power. PARL-generated topologies reduce contention at the memory cut, meet SLAs, and cut worst-case slowdown to 1.2 times while maintaining competitive mean throughput relative to link-rich meshes. Overall, this reframes NoI design for heterogeneouschiplet accelerators with workload-aware objectives.

Harsh Sharma, Juan Diego Draxl Giannoni, Boris Kramer

This work presents structure-preserving Lift & Learn, a scientific machine learning method that employs lifting variable transformations to learn structure-preserving reduced-order models for nonlinear partial differential equations (PDEs) with conservation laws. We propose a hybrid learning approach based on a recently developed energy-quadratization strategy that uses knowledge of the nonlinearity at the PDE level to derive an equivalent quadratic lifted system with quadratic system energy. The lifted dynamics obtained via energy quadratization are linear in the old variables, making model learning very effective in the lifted setting. Based on the lifted quadratic PDE model form, the proposed method derives quadratic reduced terms analytically and then uses those derived terms to formulate a constrained optimization problem to learn the remaining linear reduced operators in a structure-preserving way. The proposed hybrid learning approach yields computationally efficient quadratic reduced-order models that respect the underlying physics of the high-dimensional problem. We demonstrate the generalizability of quadratic models learned via the proposed structure-preserving Lift & Learn method through three numerical examples: the one-dimensional wave equation with exponential nonlinearity, the two-dimensional sine-Gordon equation, and the two-dimensional Klein-Gordon-Zakharov equations. The numerical results show that the proposed learning approach is competitive with the state-of-the-art structure-preserving data-driven model reduction method in terms of both accuracy and computational efficiency.

Daniel Lee, Harsh Sharma, Jieun Han et al.

Translating knowledge-intensive and entity-rich text between English and Korean requires transcreation to preserve language-specific and cultural nuances beyond literal, phonetic or word-for-word conversion. We evaluate 13 models (LLMs and MT models) using automatic metrics and human assessment by bilingual annotators. Our findings show LLMs outperform traditional MT systems but struggle with entity translation requiring cultural adaptation. By constructing an error taxonomy, we identify incorrect responses and entity name errors as key issues, with performance varying by entity type and popularity level. This work exposes gaps in automatic evaluation metrics and hope to enable future work in completing culturally-nuanced machine translation.

Harsh Sharma, Hongliang Mu, Patrick Buchfink et al.

This work presents two novel approaches for the symplectic model reduction of high-dimensional Hamiltonian systems using data-driven quadratic manifolds. Classical symplectic model reduction approaches employ linear symplectic subspaces for representing the high-dimensional system states in a reduced-dimensional coordinate system. While these approximations respect the symplectic nature of Hamiltonian systems, linear basis approximations can suffer from slowly decaying Kolmogorov $N$-width, especially in wave-type problems, which then requires a large basis size. We propose two different model reduction methods based on recently developed quadratic manifolds, each presenting its own advantages and limitations. The addition of quadratic terms to the state approximation, which sits at the heart of the proposed methodologies, enables us to better represent intrinsic low-dimensionality in the problem at hand. Both approaches are effective for issuing predictions in settings well outside the range of their training data while providing more accurate solutions than the linear symplectic reduced-order models.

Stuti Sehgal, Harsh Sharma, Akshat Anand

People have the ability to make sensible assumptions about other people's emotional states by being sympathetic, and because of our common sense of knowledge and the ability to think visually. Over the years, much research has been done on providing machines with the ability to detect human emotions and to develop automated emotional intelligence systems. The computer's ability to detect human emotions is gaining popularity in creating sensitive systems such as learning environments, health care systems and real-world. Improving people's health has been the subject of much research. This paper describes the formation as conceptual evidence of emotional acquisition and control in intelligent health settings. The authors of this paper aim for an unconventional approach with a friendly look to get emotional scenarios from the system to establish a functional, non-intrusive and emotionally-sensitive environment where users can do their normal activities naturally and see the program only when pleasant mood activating services are received. The context-sensitive system interacts with users to detect and differentiate emotions through facial expressions or speech recognition, to make music recommendations and mood color treatments with the services installed on their IoT devices.

Debaditya Pal, Harsh Sharma, Kaustubh Chaudhari

Relational databases are among the most widely used architectures to store massive amounts of data in the modern world. However, there is a barrier between these databases and the average user. The user often lacks the knowledge of a query language such as SQL required to interact with the database. The NL2SQL task aims at finding deep learning approaches to solve this problem by converting natural language questions into valid SQL queries. Given the sensitive nature of some databases and the growing need for data privacy, we have presented an approach with data privacy at its core. We have passed RoBERTa embeddings and data-agnostic knowledge vectors into LSTM based submodels to predict the final query. Although we have not achieved state of the art results, we have eliminated the need for the table data, right from the training of the model, and have achieved a test set execution accuracy of 76.7%. By eliminating the table data dependency while training we have created a model capable of zero shot learning based on the natural language question and table schema alone.