We propose Hymba, a family of small language models featuring a hybrid-head parallel architecture that integrates attention mechanisms and state space models (SSMs) within the same layer, offering parallel and complementary processing of the same inputs. In this hybrid-head module, attention heads provide high-resolution recall, while SSM heads facilitate efficient context summarization. Additionally, we introduce learnable meta tokens, which are prepended to prompts to store critical meta information, guiding subsequent tokens and alleviating the “forced-to-attend” burden associated with attention mechanisms. Thanks to the global context summarized by SSMs, the attention heads in our model can be further optimized through cross-layer key-value (KV) sharing and a mix of global and local attention, resulting in a compact cache size without compromising accuracy. Notably, Hymba achieves state-of-the-art performance among small LMs: Our Hymba-1.5B-Base model surpasses all sub-2B public models and even outperforms Llama-3.2-3B, achieving 1.32\% higher average accuracy, an 11.67$\times$ reduction in cache size, and 3.49$\times$ higher throughput.

State space models (SSMs) have emerged as an efficient alternative to Transformer models for language modeling, offering linear computational complexity and constant memory usage as context length increases. However, despite their efficiency in handling long contexts, recent studies have shown that SSMs, such as Mamba models, generally underperform compared to Transformers in long-context understanding tasks. To address this significant shortfall and achieve both efficient and accurate long-context understanding, we propose LongMamba, a training-free technique that significantly enhances the long-context capabilities of Mamba models. LongMamba builds on our discovery that the hidden channels in Mamba can be categorized into local and global channels based on their receptive field lengths, with global channels primarily responsible for long-context capability. These global channels can become the key bottleneck as the input context lengthens. Specifically, when input lengths largely exceed the training sequence length, global channels exhibit limitations in adaptively extend their receptive fields, leading to Mamba’s poor long-context performance. The key idea of LongMamba is to mitigate the hidden state memory decay in these global channels by preventing the accumulation of unimportant tokens in their memory. This is achieved by first identifying critical tokens in the global channels and then applying token filtering to accumulate only those critical tokens. Through extensive benchmarking across synthetic and real-world long-context scenarios, LongMamba sets a new standard for Mamba’s long-context performance, significantly extending its operational range without requiring additional training. Our code is available at https://github.com/GATECH-EIC/LongMamba.

Recent advancements in Large Language Models (LLMs) have spurred interest in numerous applications requiring robust long-range capabilities, essential for processing extensive input contexts and continuously generating extended outputs. As sequence lengths increase, the number of Key-Value (KV) pairs in LLMs escalates, creating a significant efficiency bottleneck. In this paper, we propose a new KV cache optimization paradigm called LaCache, a training-free method for efficient and accurate generative inference of LLMs. LaCache enables LLMs to simultaneously address both of the critical challenges in long-range modeling: robust long-range capabilities and continuous generation without running out-of-memory (OOM). Specifically, LaCache integrates two key innovations: (1) a ladder-shaped KV cache pattern that stores KV pairs not only sequentially (left-to-right within each layer) but also across layers (from shallow to deep), providing an extended span for capturing long-range dependencies under a fixed storage budget, thereby boosting long-range capabilities; and (2) an iterative compaction mechanism that progressively compresses older caches, freeing up space for new tokens within a fixed cache size. This token distance-based dynamic compression enables more effective continuous generation under constrained cache budgets. Experiments across various tasks, benchmarks, and LLM models consistently validate LaCache’s effectiveness in enhancing LLMs’ long-range capabilities. Our code is available at https://github.com/GATECH-EIC/LaCache.

Pre-training datasets are typically collected from web content and lack inherent domain divisions. For instance, widely used datasets like Common Crawl do not include explicit domain labels, while manually curating labeled datasets such as The Pile is labor-intensive. Consequently, identifying an optimal pre-training data mixture remains a challenging problem, despite its significant benefits for pre-training performance. To address these challenges, we propose CLustering-based Iterative Data Mixture Bootstrapping (Nemotron-CLIMB), an automated framework that discovers, evaluates, and refines data mixtures in a pre-training setting. Specifically, Nemotron-CLIMB embeds and clusters large-scale datasets in a semantic space and then iteratively searches for optimal mixtures using a smaller proxy model and a predictor. This strategy enables effective domain adaptation without relying solely on curated data. When continuously trained on 400B tokens with this mixture, our 1B model exceeds the state-of-the-art Llama-3.2-1B by 2.0%. Moreover, we observe that optimizing for a specific domain (e.g., Social Sciences) yields a 5% improvement over random sampling. Finally, we introduce Nemotron-ClimbLab, a filtered 1.2-trillion-token corpus with 20 clusters as a research playground, and Nemotron-ClimbMix, a compact yet powerful 400-billion-token dataset designed for efficient pre-training that delivers superior performance under an equal token budget. We analyze the final data mixture, elucidating the characteristics of an optimal data mixture.

Efficient deployment of small language models (SLMs) is essential for numerous real-world applications with stringent latency constraints.While previous work on SLM design has primarily focused on reducing the number of parameters to achieve parameter-optimal SLMs, parameter efficiency does not necessarily translate into proportional real-device speed-ups. This work aims to identify the key determinants of SLMs' real-device latency and offer generalizable principles and methodologies for SLM design and training when real-device latency is the primary consideration. Specifically, we identify two central architectural factors: depth–width ratios and operator choices. The former is crucial for small-batch-size latency, while the latter affects both latency and large-batch-size throughput. In light of this, we first study latency-optimal depth–width ratios, with the key finding that although deep–thin models generally achieve better accuracy under the same parameter budget, they may not lie on the accuracy–latency trade-off frontier. Next, we explore emerging efficient attention alternatives to evaluate their potential as candidate building operators. Using the identified promising operators, we construct an evolutionary search framework to automatically discover latency-optimal combinations of these operators within hybrid SLMs, thereby advancing the accuracy–latency frontier. In addition to architectural improvements, we further enhance SLM training using a weight normalization technique that enables more effective weight updates and improves final convergence. This technique can serve as a generalizable component for future SLMs. Combining these methods, we introduce a new family of hybrid SLMs, called Nemotron-Flash, which significantly advances the accuracy–efficiency frontier of state-of-the-art SLMs, e.g., achieving over +5.5\% average accuracy, 1.3$\times$/1.9$\times$ lower latency, and 18.7$\times$/45.6$\times$ higher throughput compared to Qwen3-1.7B/0.6B, respectively.

Recent advances in reasoning-centric language models have highlighted reinforcement learning (RL) as a promising method for aligning models with verifiable rewards. However, it remains contentious whether RL truly expands a model’s reasoning capabilities or merely amplifies high-reward outputs already latent in the base model’s distribution, and whether continually scaling up RL compute reliably leads to improved reasoning performance. In this work, we challenge prevailing assumptions by demonstrating that prolonged RL (ProRL) training can uncover novel reasoning strategies that are inaccessible to base models, even under extensive sampling. We introduce ProRL, a novel training methodology that incorporates KL divergence control, reference policy resetting, and a diverse suite of tasks. Our empirical analysis reveals that RL-trained models consistently outperform base models across a wide range of pass@$k$ evaluations, including scenarios where base models fail entirely regardless of the number of attempts. We further show that reasoning boundary improvements correlates strongly with task competence of base model and training duration, suggesting that RL can explore and populate new regions of solution space over time. These findings offer new insights into the conditions under which RL meaningfully expands reasoning boundaries in language models and establish a foundation for future work on long-horizon RL for reasoning. We will release model weights and data to support further research.

Federated learning (FL) is an effective machine learning paradigm where multiple clients can train models based on heterogeneous data in a decentralized manner without accessing their private data. However, existing FL systems undergo performance deterioration due to feature-level test-time shifts, which are well investigated in centralized settings but rarely studied in FL. The common non-IID issue in FL usually refers to inter-client heterogeneity during training phase, while the test-time shift refers to the intra-client heterogeneity during test phase. Although the former is always deemed to be notorious for FL, there is still a wealth of useful information delivered by heterogeneous data sources, which may potentially help alleviate the latter issue. To explore the possibility of using inter-client heterogeneity in handling intra-client heterogeneity, we firstly propose a contrastive learning-based FL framework, namely FedICON, to capture invariant knowledge among heterogeneous clients and consistently tune the model to adapt to test data. In FedICON, each client performs sample-wise supervised contrastive learning during the local training phase, which enhances sample-wise invariance encoding ability. Through global aggregation, the invariance extraction ability can be mutually boosted among inter-client heterogeneity. During the test phase, our test-time adaptation procedure leverages unsupervised contrastive learning to guide the model to smoothly generalize to test data under intra-client heterogeneity. Extensive experiments validate the effectiveness of the proposed FedICON in taming heterogeneity to handle test-time shift problems.

Various approaches have been proposed for out-of-distribution (OOD) detection by augmenting models, input examples, training sets, and optimization objectives. Deviating from existing work, we have a simple hypothesis that standard off-the-shelf models may already contain sufficient information about the training set distribution which can be leveraged for reliable OOD detection. Our empirical study on validating this hypothesis, which measures the model activation's mean for OOD and in-distribution (ID) minibatches, surprisingly finds that activation means of OOD mini-batches consistently deviate more from those of the training data. In addition, training data's activation means can be computed offline efficiently or retrieved from batch normalization layers as a ‘free lunch’. Based upon this observation, we propose a novel metric called Neural Mean Discrepancy (NMD), which compares neural means of the input examples and training data. Leveraging the simplicity of NMD, we propose an efficient OOD detector that computes neural means by a standard forward pass followed by a lightweight classifier. Extensive experiments show that NMD outperforms state-of-the-art OOD approaches across multiple datasets and model architectures in terms of both detection accuracy and computational cost.

We design deep neural networks (DNNs) and corresponding networks' splittings to distribute DNNs' workload to camera sensors and a centralized aggregator on head mounted devices to meet system performance targets in inference accuracy and latency under the given hardware resource constraints. To achieve an optimal balance among computation, communication, and performance, a split-aware neural architecture search framework, SplitNets, is introduced to conduct model designing, splitting, and communication reduction simultaneously. We further extend the framework to multi-view systems for learning to fuse inputs from multiple camera sensors with optimal performance and systemic efficiency. We validate SplitNets for single-view system on ImageNet as well as multi-view system on 3D classification, and show that the SplitNets framework achieves state-of-the-art (SOTA) performance and system latency compared with existing approaches.

Low-resolution uniform quantization (e.g., 4-bit bitwidth) for both Deep Neural Network (DNN) weights and data has emerged as an important technique for efficient inference. Departing from conventional quantization, we describe a novel training approach to support inference at multiple resolutions by reusing a single set of quantization terms (the same set of nonzero bits in values). The proposed approach streamlines the training and supports dynamic selection of resolution levels during inference. We evaluate the method on a diverse range of applications including multiple CNNs on ImageNet, an LSTM on Wikitext-2, and YOLO-v5 on COCO. We show that models resulting from our multi-resolution training can support up to 10 resolutions with only a moderate performance reduction (e.g., ≤ 1%) compared to training them individually. Lastly, using an FPGA, we compare our multi-resolution multiplier-accumulator (mMAC) against other conventional MAC designs and evaluate the inference performance. We show that the mMAC design broadens the choices in trading off cost, efficiency, and latency across a range of computational budgets.

User data confidentiality protection is becoming a rising challenge in the present deep learning research. Without access to data, conventional data-driven model compression faces a higher risk of performance degradation. Recently, some works propose to generate images from a specific pretrained model to serve as training data. However, the inversion process only utilizes biased feature statistics stored in one model and is from low-dimension to high-dimension. As a consequence, it inevitably encounters the difficulties of generalizability and inexact inversion, which leads to unsatisfactory performance. To address these problems, we propose MixMix based on two simple yet effective techniques: (1) Feature Mixing: utilizes various models to construct a universal feature space for generalized inversion; (2) Data Mixing: mixes the synthesized images and labels to generate exact label information. We prove the effectiveness of MixMix from both theoretical and empirical perspectives. Extensive experiments show that MixMix outperforms existing methods on the mainstream compression tasks, including quantization, knowledge distillation and pruning. Specifically, MixMix achieves up to 4% and 20% accuracy uplift on quantization and pruning, respectively, compared to existing data-free compression work.

Spiking Neural Network (SNN) has been recognized as one of the next generation of neural networks. Conventionally, SNN can be converted from a pre-trained ANN by only replacing the ReLU activation to spike activation while keeping the parameters intact. Perhaps surprisingly, in this work we show that a proper way to calibrate the parameters during the conversion of ANN to SNN can bring significant improvements. We introduce SNN Calibration, a cheap but extraordinarily effective method by leveraging the knowledge within a pre-trained Artificial Neural Network (ANN). Starting by analyzing the conversion error and its propagation through layers theoretically, we propose the calibration algorithm that can correct the error layer-by-layer. The calibration only takes a handful number of training data and several minutes to finish. Moreover, our calibration algorithm can produce SNN with state-of-the-art architecture on the large-scale ImageNet dataset, including MobileNet and RegNet. Extensive experiments demonstrate the effectiveness and efficiency of our algorithm. For example, our advanced pipeline can increase up to 69% top-1 accuracy when converting MobileNet on ImageNet compared to baselines. Codes are released at https://github.com/yhhhli/SNN_Calibration.

To deploy deep neural networks on resource-limited devices, quantization has been widely explored. In this work, we study the extremely low-bit networks which have tremendous speed-up, memory saving with quantized activation and weights. We first bring up three omitted issues in extremely low-bit networks: the squashing range of quantized values; the gradient vanishing during backpropagation and the unexploited hardware acceleration of ternary networks. By reparameterizing quantized activation and weights vector with full precision scale and offset for fixed ternary vector, we decouple the range and magnitude from direction to extenuate above problems. Learnable scale and offset can automatically adjust the range of quantized values and sparsity without gradient vanishing. A novel encoding and computation pattern are designed to support efficient computing for our reparameterized ternary network (RTN). Experiments on ResNet-18 for ImageNet demonstrate that the proposed RTN finds a much better efficiency between bitwidth and accuracy and achieves up to 26.76% relative accuracy improvement compared with state-of-the-art methods. Moreover, we validate the proposed computation pattern on Field Programmable Gate Arrays (FPGA), and it brings 46.46 × and 89.17 × savings on power and area compared with the full precision convolution.

We present the Maestro memory-on-logic 3D-IC architecture for coordinated parallel use of a plurality of systolic arrays (SAs) in performing deep neural network (DNN) inference. Maestro reduces under-utilization common for a single large SA by allowing parallel use of many smaller SAs on DNN weight matrices of varying shapes and sizes. In order to buffer immediate results in memory blocks (MBs) and provide coordinated high-bandwidth communication between SAs and MBs in transferring weights and results Maestro employs three innovations. (1) An SA on the logic die can access its corresponding MB on the memory die in short distance using 3D-IC interconnects, (2) through an efficient switch based on H-trees, an SA can access any MB with low latency, and (3) the switch can combine partial results from SAs in an elementwise fashion before writing back to a destination MB. We describe the Maestro architecture, including a circuit and layout design, detail scheduling of the switch, analyze system performance for real-time inference applications using input with batch size equal to one, and showcase applications for deep learning inference, with ShiftNet for computer vision and recent Transformer models for natural language processing. For the same total number of systolic cells, Maestro, with multiple smaller SAs, leads to 16x and 12x latency improvements over a single large SA on ShiftNet and Transformer, respectively. Compared to a floating-point GPU implementation of ShiftNet and Transform, a baseline Maestro system with 4,096 SAs (each with 8x8 systolic cells) provides significant latency improvements of 30x and 47x, respectively.

To reduce memory footprint and run-time latency, techniques such as neural net-work pruning and binarization have been explored separately. However, it is un-clear how to combine the best of the two worlds to get extremely small and efficient models. In this paper, we, for the first time, define the filter-level pruning problem for binary neural networks, which cannot be solved by simply migrating existing structural pruning methods for full-precision models. A novel learning-based approach is proposed to prune filters in our main/subsidiary network frame-work, where the main network is responsible for learning representative features to optimize the prediction performance, and the subsidiary component works as a filter selector on the main network. To avoid gradient mismatch when training the subsidiary component, we propose a layer-wise and bottom-up scheme. We also provide the theoretical and experimental comparison between our learning-based and greedy rule-based methods. Finally, we empirically demonstrate the effectiveness of our approach applied on several binary models, including binarizedNIN, VGG-11, and ResNet-18, on various image classification datasets. For bi-nary ResNet-18 on ImageNet, we use 78.6% filters but can achieve slightly better test error 49.87% (50.02%-0.15%) than the original model

Binary neural networks (BNN) have been studied extensively since they run dramatically faster at lower memory and power consumption than floating-point networks, thanks to the efficiency of bit operations. However, contemporary BNNs whose weights and activations are both single bits suffer from severe accuracy degradation. To understand why, we investigate the representation ability, speed and bias/variance of BNNs through extensive experiments. We conclude that the error of BNNs is predominantly caused by intrinsic instability (training time) and non-robustness (train & test time). Inspired by this investigation, we propose the Binary Ensemble Neural Network (BENN) which leverages ensemble methods to improve the performance of BNNs with limited efficiency cost. While ensemble techniques have been broadly believed to be only marginally helpful for strong classifiers such as deep neural networks, our analysis and experiments show that they are naturally a perfect fit to boost BNNs. We find that our BENN, which is faster and more robust than state-of-the-art binary networks, can even surpass the accuracy of the full-precision floating number network with the same architecture.

We present a full-stack optimization framework for accelerating inference of CNNs (Convolutional Neural Networks) and validate the approach with a field-programmable gate array (FPGA) implementation. By jointly optimizing CNN models, computing architectures, and hardware implementations, our full-stack approach achieves unprecedented performance in the trade-off space characterized by inference latency, energy efficiency, hardware utilization, and inference accuracy. An FPGA implementation is used as the validation vehicle for our design, achieving a 2.28ms inference latency for the ImageNet benchmark. Our implementation shines in that it has 9x higher energy efficiency compared to other implementations while achieving comparable latency. A highlight of our approach which contributes to the achieved high energy efficiency is an efficient Selector-Accumulator (SAC) architecture for implementing CNNs with powers-of-two weights. Compared to an FPGA implementation for a traditional 8-bit MAC, SAC substantially reduces required hardware resources (4.85x fewer lookup tables) and power consumption (2.48x).

How to develop slim and accurate deep neural networks has become crucial for real- world applications, especially for those employed in embedded systems. Though previous work along this research line has shown some promising results, most existing methods either fail to significantly compress a well-trained deep network or require a heavy retraining process for the pruned deep network to re-boost its prediction performance. In this paper, we propose a new layer-wise pruning method for deep neural networks. In our proposed method, parameters of each individual layer are pruned independently based on second order derivatives of a layer-wise error function with respect to the corresponding parameters. We prove that the final prediction performance drop after pruning is bounded by a linear combination of the reconstructed errors caused at each layer. By controlling layer-wise errors properly, one only needs to perform a light retraining process on the pruned network to resume its original prediction performance. We conduct extensive experiments on benchmark datasets to demonstrate the effectiveness of our pruning method compared with several state-of-the-art baseline methods. Codes of our work are released at: https://github.com/csyhhu/L-OBS.