Yubo Gao

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.

Yubo Gao, Haotian Wu, Hong Chen et al.

Chain-of-Thought (CoT) has significantly enhanced LLM reasoning, yet often incurs substantial computational overhead due to "overthinking": generating excessively long rationales without commensurate accuracy gains. Existing efficiency methods typically apply uniform compression, which overlooks a critical observation that reasoning complexity is heterogeneous at two distinct granularity: across different problems and within individual reasoning steps. This motivates our principle of Thinking Economically: intelligently allocating computational resources based on intrinsic task and step demands rather than pursuing uniform brevity. We propose Hierarchical Adaptive Budgeter (HAB), a training framework that operationalizes this principle through coarse-to-fine budgeting. At the inter-step level, HAB predicts the optimal reasoning depth for each problem. At the intra-step level, HAB learns step-specific token budgeting signals from PPL-derived step comparisons and an adaptive Pareto optimization objective that captures the local quality-efficiency trade-off, while a Fisher Information-based pruner further provides fine-grained training-time guidance, thereby encouraging the generator to internalize more economical reasoning patterns. Experiments on GSM8K and MATH500 show that HAB not only surpasses standard CoT in accuracy but also reduces token usage, achieving a stronger performance-efficiency trade-off than the compared baselines.

Junquan Huang, Haotian Wu, Yubo Gao et al.

Large language models (LLMs) with Chain-of-Thought (CoT) prompting achieve strong reasoning but often produce unnecessarily long explanations, increasing cost and sometimes reducing accuracy. Fair comparison of efficiency-oriented approaches is hindered by fragmented evaluation practices. We introduce EffiReason-Bench, a unified benchmark for rigorous cross-paradigm evaluation of efficient reasoning methods across three categories: Reasoning Blueprints, Dynamic Execution, and Post-hoc Refinement. To enable step-by-step evaluation, we construct verified CoT annotations for CommonsenseQA and LogiQA via a pipeline that enforces standardized reasoning structures, comprehensive option-wise analysis, and human verification. We evaluate 7 methods across 6 open-source LLMs (1B-70B) on 4 datasets spanning mathematics, commonsense, and logic, and propose the E3-Score, a principled metric inspired by economic trade-off modeling that provides smooth, stable evaluation without discontinuities or heavy reliance on heuristics. Experiments show that no single method universally dominates; optimal strategies depend on backbone scale, task complexity, and architecture.

Hong Chen, Xiang Liu, Yubo Gao et al.

Long-context language models are limited by the memory footprint of the key-value (KV) cache. Existing training-free KV compression methods usually rank tokens by one importance signal -- attention, recency, layer-wise allocation, or key distinctiveness -- which becomes brittle when useful context is globally distinctive, locally episodic, or immediately relevant. We introduce NestedKV, a key-only KV cache compression method inspired by the Continuum Memory System in Nested Learning. NestedKV maintains global, block-level, and sliding-window key anchors, scores tokens by multi-time-scale cosine anomaly, and combines the resulting rankings with a training-free outer learner using head-adaptive mixing and surprise-gated token routing. The score is paired with adaptive per-head budgets and requires no training or LLM modification. Across RULER (4k--32k), LooGLE, LongBench, LongBench-E, InfiniteBench, and MMLU-Pro on Qwen3 and Llama-3.2 models, NestedKV is strongest when the retained cache is small. On Qwen3-4B, it improves over KeyDiff by up to 19.10 points on RULER and 19.29 on LongBench at $r=0.75$; at $r=0.95$, it retains 37.32 on LongBench versus 17.55 for KeyDiff.

Yubo Gao, Haotian Wu

knowledge graph-based recommendation methods have achieved great success in the field of recommender systems. However, over-reliance on high-quality knowledge graphs is a bottleneck for such methods. Specifically, the long-tailed distribution of entities of KG and noise issues in the real world will make item-entity dependent relations deviate from reflecting true characteristics and significantly harm the performance of modeling user preference. Contrastive learning, as a novel method that is employed for data augmentation and denoising, provides inspiration to fill this research gap. However, the mainstream work only focuses on the long-tail properties of the number of items clicked, while ignoring that the long-tail properties of total number of clicks per user may also affect the performance of the recommendation model. Therefore, to tackle these problems, motivated by the Debiased Contrastive Learning of Unsupervised Sentence Representations (DCLR), we propose Two-Level Debiased Contrastive Graph Learning (TDCGL) model. Specifically, we design the Two-Level Debiased Contrastive Learning (TDCL) and deploy it in the KG, which is conducted not only on User-Item pairs but also on User-User pairs for modeling higher-order relations. Also, to reduce the bias caused by random sampling in contrastive learning, with the exception of the negative samples obtained by random sampling, we add a noise-based generation of negation to ensure spatial uniformity. Considerable experiments on open-source datasets demonstrate that our method has excellent anti-noise capability and significantly outperforms state-of-the-art baselines. In addition, ablation studies about the necessity for each level of TDCL are conducted.

Yubo Gao, Maryam Haghifam, Christina Giannoula et al.

Deep learning (DL) models have revolutionized numerous domains, yet optimizing them for computational efficiency remains a challenging endeavor. Development of new DL models typically involves two parties: the model developers and performance optimizers. The collaboration between the parties often necessitates the model developers exposing the model architecture and computational graph to the optimizers. However, this exposure is undesirable since the model architecture is an important intellectual property, and its innovations require significant investments and expertise. During the exchange, the model is also vulnerable to adversarial attacks via model stealing. This paper presents Proteus, a novel mechanism that enables model optimization by an independent party while preserving the confidentiality of the model architecture. Proteus obfuscates the protected model by partitioning its computational graph into subgraphs and concealing each subgraph within a large pool of generated realistic subgraphs that cannot be easily distinguished from the original. We evaluate Proteus on a range of DNNs, demonstrating its efficacy in preserving confidentiality without compromising performance optimization opportunities. Proteus effectively hides the model as one alternative among up to $10^{32}$ possible model architectures, and is resilient against attacks with a learning-based adversary. We also demonstrate that heuristic based and manual approaches are ineffective in identifying the protected model. To our knowledge, Proteus is the first work that tackles the challenge of model confidentiality during performance optimization. Proteus will be open-sourced for direct use and experimentation, with easy integration with compilers such as ONNXRuntime.

Geoffrey X. Yu, Yubo Gao, Pavel Golikov et al.

Deep learning researchers and practitioners usually leverage GPUs to help train their deep neural networks (DNNs) faster. However, choosing which GPU to use is challenging both because (i) there are many options, and (ii) users grapple with competing concerns: maximizing compute performance while minimizing costs. In this work, we present a new practical technique to help users make informed and cost-efficient GPU selections: make performance predictions with the help of a GPU that the user already has. Our technique exploits the observation that, because DNN training consists of repetitive compute steps, predicting the execution time of a single iteration is usually enough to characterize the performance of an entire training process. We make predictions by scaling the execution time of each operation in a training iteration from one GPU to another using either (i) wave scaling, a technique based on a GPU's execution model, or (ii) pre-trained multilayer perceptrons. We implement our technique into a Python library called Habitat and find that it makes accurate iteration execution time predictions (with an average error of 11.8%) on ResNet-50, Inception v3, the Transformer, GNMT, and DCGAN across six different GPU architectures. Habitat supports PyTorch, is easy to use, and is open source.

Junyan Zhang, Shuliang Liu, Aiwei Liu et al. · tsinghua

Watermarking technology is a method used to trace the usage of content generated by large language models. Sentence-level watermarking aids in preserving the semantic integrity within individual sentences while maintaining greater robustness. However, many existing sentence-level watermarking techniques depend on arbitrary segmentation or generation processes to embed watermarks, which can limit the availability of appropriate sentences. This limitation, in turn, compromises the quality of the generated response. To address the challenge of balancing high text quality with robust watermark detection, we propose CoheMark, an advanced sentence-level watermarking technique that exploits the cohesive relationships between sentences for better logical fluency. The core methodology of CoheMark involves selecting sentences through trained fuzzy c-means clustering and applying specific next sentence selection criteria. Experimental evaluations demonstrate that CoheMark achieves strong watermark strength while exerting minimal impact on text quality.

Song Dai, Yibo Yan, Jiamin Su et al.

Multimodal Large Language Models (MLLMs) have demonstrated remarkable capabilities in diverse reasoning tasks, yet their application to complex physics reasoning remains underexplored. Physics reasoning presents unique challenges, requiring grounding in physical conditions and the interpretation of multimodal information. Current physics benchmarks are limited, often focusing on text-only inputs or solely on problem-solving, thereby overlooking the critical intermediate steps of variable identification and process formulation. To address these limitations, we introduce PhysicsArena, the first multimodal physics reasoning benchmark designed to holistically evaluate MLLMs across three critical dimensions: variable identification, physical process formulation, and solution derivation. PhysicsArena aims to provide a comprehensive platform for assessing and advancing the multimodal physics reasoning abilities of MLLMs.

Junyan Zhang, Yiming Huang, Shuliang Liu et al.

The rapid adoption of LLMs has overshadowed the potential advantages of traditional BERT-like models in text classification. This study challenges the prevailing "LLM-centric" trend by systematically comparing three category methods, i.e., BERT-like models fine-tuning, LLM internal state utilization, and zero-shot inference across six high-difficulty datasets. Our findings reveal that BERT-like models often outperform LLMs. We further categorize datasets into three types, perform PCA and probing experiments, and identify task-specific model strengths: BERT-like models excel in pattern-driven tasks, while LLMs dominate those requiring deep semantics or world knowledge. Based on this, we propose TaMAS, a fine-grained task selection strategy, advocating for a nuanced, task-driven approach over a one-size-fits-all reliance on LLMs.

Sicheng Tao, Jungang Li, Yibo Yan et al.

Video reasoning has emerged as a critical capability for multimodal large language models (MLLMs), requiring models to move beyond static perception toward coherent understanding of temporal dynamics in complex scenes. Yet existing MLLMs often exhibit process inconsistency, where intermediate reasoning drifts from video dynamics even when the final answer is correct, undermining interpretability and robustness. To address this issue, we introduce MOSS-ChatV, a reinforcement learning framework with a Dynamic Time Warping (DTW)-based process reward. This rule-based reward aligns reasoning traces with temporally grounded references, enabling efficient process supervision without auxiliary reward models. We further identify dynamic state prediction as a key measure of video reasoning and construct MOSS-Video, a benchmark with annotated reasoning traces, where the training split is used to fine-tune MOSS-ChatV and the held-out split is reserved for evaluation. MOSS-ChatV achieves 87.2\% on MOSS-Video (test) and improves performance on general video benchmarks such as MVBench and MMVU. The framework consistently yields gains across different architectures, including Qwen2.5-VL and Phi-2, confirming its broad applicability. Evaluations with GPT-4o-as-judge further show that MOSS-ChatV produces more consistent and stable reasoning traces.

Junyan Zhang, Yubo Gao, Yibo Yan et al.

The finetuning of Large Language Models (LLMs) has significantly advanced their instruction-following capabilities, yet the underlying computational mechanisms driving these improvements remain poorly understood. This study systematically examines how fine-tuning reconfigures LLM computations by isolating and analyzing instruction-specific sparse components, i.e., neurons in dense models and both neurons and experts in Mixture-of-Experts (MoE) architectures. In particular, we introduce HexaInst, a carefully curated and balanced instructional dataset spanning six distinct categories, and propose SPARCOM, a novel analytical framework comprising three key contributions: (1) a method for identifying these sparse components, (2) an evaluation of their functional generality and uniqueness, and (3) a systematic comparison of their alterations. Through experiments, we demonstrate functional generality, uniqueness, and the critical role of these components in instruction execution. By elucidating the relationship between fine-tuning-induced adaptations and sparse computational substrates, this work provides deeper insights into how LLMs internalize instruction-following behavior for the trustworthy LLM community.

Yubo Gao, Haotian Wu, Lei Zhang

Multimodal Sentiment Analysis (MSA) aims to recognize human emotions by exploiting textual, acoustic, and visual modalities, and thus how to make full use of the interactions between different modalities is a central challenge of MSA. Interaction contains alignment and conflict aspects. Current works mainly emphasize alignment and the inherent differences between unimodal modalities, neglecting the fact that there are also potential conflicts between bimodal combinations. Additionally, multi-task learning-based conflict modeling methods often rely on the unstable generated labels. To address these challenges, we propose a novel multi-level conflict-aware network (MCAN) for multimodal sentiment analysis, which progressively segregates alignment and conflict constituents from unimodal and bimodal representations, and further exploits the conflict constituents with the conflict modeling branch. In the conflict modeling branch, we conduct discrepancy constraints at both the representation and predicted output levels, avoiding dependence on the generated labels. Experimental results on the CMU-MOSI and CMU-MOSEI datasets demonstrate the effectiveness of the proposed MCAN.

Yubo Gao, Renbo Tu, Gennady Pekhimenko et al.

Differentially-Private SGD (DP-SGD) is a powerful technique to protect user privacy when using sensitive data to train neural networks. During training, converting model weights and activations into low-precision formats, i.e., quantization, can drastically reduce training times, energy consumption, and cost, and is thus a widely used technique. In this work, we demonstrate that quantization causes significantly higher accuracy degradation in DP-SGD compared to regular SGD. We observe that this is caused by noise injection in DP-SGD, which amplifies quantization variance, leading to disproportionately large accuracy degradation. To address this challenge, we present QPQuant, a dynamic quantization framework that adaptively selects a changing subset of layers to quantize at each epoch. Our method combines two key ideas that effectively reduce quantization variance: (i) probabilistic sampling of the layers that rotates which layers are quantized every epoch, and (ii) loss-aware layer prioritization, which uses a differentially private loss sensitivity estimator to identify layers that can be quantized with minimal impact on model quality. This estimator consumes a negligible fraction of the overall privacy budget, preserving DP guarantees. Empirical evaluations on ResNet18, ResNet50, and DenseNet121 across a range of datasets demonstrate that DPQuant consistently outperforms static quantization baselines, achieving near Pareto-optimal accuracy-compute trade-offs and up to 2.21x theoretical throughput improvements on low-precision hardware, with less than 2% drop in validation accuracy.

Honghua Dong, Qidong Su, Yubo Gao et al.

Large Language Models (LLMs) have become increasingly capable of handling diverse tasks with the aid of well-crafted prompts and integration of external tools, but as task complexity rises, the workflow involving LLMs can be complicated and thus challenging to implement and maintain. To address this challenge, we propose APPL, A Prompt Programming Language that acts as a bridge between computer programs and LLMs, allowing seamless embedding of prompts into Python functions, and vice versa. APPL provides an intuitive and Python-native syntax, an efficient parallelized runtime with asynchronous semantics, and a tracing module supporting effective failure diagnosis and replaying without extra costs. We demonstrate that APPL programs are intuitive, concise, and efficient through three representative scenarios: Chain-of-Thought with self-consistency (CoT-SC), ReAct tool use agent, and multi-agent chat. Experiments on three parallelizable workflows further show that APPL can effectively parallelize independent LLM calls, with a significant speedup ratio that almost matches the estimation.