Mixture-of-Experts (MoE) has emerged as an effective and efficient scaling mechanism for large language models (LLMs) and vision-language models (VLMs). By expanding a single feed-forward network into multiple expert branches, MoE increases model capacity while maintaining efficiency through sparse activation. However, despite this sparsity, the need to preload all experts into memory and activate multiple experts per input introduces significant computational and memory overhead. The expert module becomes the dominant contributor to model size and inference cost, posing a major challenge for deployment. To address this, we propose MC# (Mixture-Compressor-sharp), a unified framework that combines static quantization and dynamic expert pruning by leveraging the significance of both experts and tokens to achieve aggressive compression of MoE-LLMs/VLMs. To reduce storage and loading overhead, we introduce Pre-Loading Mixed-Precision Quantization (PMQ), which formulates adaptive bit allocation as a linear programming problem. The objective function jointly considers expert importance and quantization error, producing a Pareto-optimal trade-off between model size and performance. To reduce runtime computation, we further introduce Online Top-any Pruning (OTP), which models expert activation per token as a learnable distribution via Gumbel-Softmax sampling. During inference, OTP dynamically selects a subset of experts for each token, allowing fine-grained control over activation. By combining PMQ's static bit-width optimization with OTP's dynamic routing, MC# achieves extreme compression with minimal accuracy degradation. On DeepSeek-VL2, MC# achieves a 6.2 × weight reduction at an average of 2.57 bits, with only a 1.7% drop across five multimodal benchmarks compared to the 16-bit baseline. Moreover, OTP further reduces expert activation by 20% with less than 1% performance loss, demonstrating strong potential for efficient deployment of MoE-based models.

The advancement of Large Vision Language Models (LVLMs) has significantly improved multimodal understanding, yet challenges remain in video reasoning tasks due to the scarcity of high-quality, large-scale datasets. Existing video question-answering (VideoQA) datasets often rely on costly manual annotations with insufficient granularity or automatic construction methods with redundant frame-by-frame analysis, limiting their scalability and effectiveness for complex reasoning. To address these challenges, we introduce VideoEspresso, a novel dataset that features VideoQA pairs preserving essential spatial details and temporal coherence, along with multimodal annotations of intermediate reasoning steps. Our construction pipeline employs a semantic-aware method to reduce redundancy, followed by generating QA pairs using GPT-4o. We further develop video Chain-of-Thought (CoT) annotations to enrich reasoning processes, guiding GPT-4o in extracting logical relationships from QA pairs and video content. To exploit the potential of high-quality VideoQA pairs, we propose a Hybrid LVLMs Collaboration framework, featuring a Frame Selector and a two-stage instruction fine-tuned reasoning LVLM. This framework adaptively selects core frames and performs CoT reasoning using multimodal evidence. Evaluated on our proposed benchmark with 14 tasks against 9 popular LVLMs, our method outperforms existing baselines on most tasks, demonstrating superior video reasoning capabilities. Our code and dataset have been released at: https://github.com/hshjerry/VideoEspresso

Mixture-of-Experts large language models (MoE-LLMs) marks a significant step forward of language models, however, they encounter two critical challenges in practice: 1) expert parameters lead to considerable memory consumption and loading latency; and 2) the current activated experts are redundant, as many tokens may only require a single expert. Motivated by these issues, we investigate the MoE-LLMs and make two key observations: a) different experts exhibit varying behaviors on activation reconstruction error, routing scores, and activated frequencies, highlighting their differing importance, and b) not all tokens are equally important-- only a small subset is critical. Building on these insights, we propose MC, a training-free Mixture-Compressor for MoE-LLMs, which leverages the significance of both experts and tokens to achieve an extreme compression. First, to mitigate storage and loading overheads, we introduce Pre-Loading Mixed-Precision Quantization (PMQ), which formulates the adaptive bit-width allocation as a Linear Programming (LP) problem, where the objective function balances multi-factors reflecting the importance of each expert. Additionally, we develop Online Dynamic Pruning (ODP), which identifies important tokens to retain and dynamically select activated experts for other tokens during inference to optimize efficiency while maintaining performance. Our MC integrates static quantization and dynamic pruning to collaboratively achieve extreme compression for MoE-LLMs with less accuracy loss, ensuring an optimal trade-off between performance and efficiency Extensive experiments confirm the effectiveness of our approach. For instance, at 2.54 bits, MC compresses 76.6% of the model, with only a 3.8% average accuracy loss. During dynamic inference, we further reduce activated parameters by 15%, with a performance drop of less than 0.6%. Remarkably, MC even surpasses floating-point 13b dense LLMs with significantly smaller parameter sizes, suggesting that mixture compression in MoE-LLMs has the potential to outperform both comparable and larger dense LLMs. Our code is available at https://github.com/Aaronhuang-778/MC-MoE

In this paper, we present InfoMax, a novel data pruning method, also known as coreset selection, designed to maximize the information content of selected samples while minimizing redundancy. By doing so, InfoMax enhances the overall informativeness of the coreset. The information of individual samples is measured by importance scores, which capture their influence or difficulty in model learning. To quantify redundancy, we use pairwise sample similarities, based on the premise that similar samples contribute similarly to the learning process. We formalize the coreset selection problem as a discrete quadratic programming (DQP) task, with the objective of maximizing the total information content, represented as the sum of individual sample contributions minus the redundancies introduced by similar samples within the coreset. To ensure practical scalability, we introduce an efficient gradient-based solver, complemented by sparsification techniques applied to the similarity matrix and dataset partitioning strategies. This enables InfoMax to seamlessly scale to datasets with millions of samples. Extensive experiments demonstrate the superior performance of InfoMax in various data pruning tasks, including image classification, vision-language pre-training, and instruction tuning for large language models.

Post-training quantization (PTQ) is an effective technique for compressing large language models (LLMs). However, while uniform-precision quantization is computationally efficient, it often compromises model performance. To address this, we propose SliM-LLM, a salience-driven mixed-precision quantization framework that allocates bit-widths at the group-wise with high accuracy. Our approach leverages the observation that important weights follow a structured distribution and introduces two key components: 1) Salience-Determined Bit Allocation adaptively assigns bit-widths to groups within each layer based on their salience; and 2) Salience-Weighted Quantizer Calibration optimizes quantizer parameters by incorporating element-level salience, retain essential information. With its structured group-wise partitioning, SliM-LLM provides a hardware-friendly solution that matches the efficiency of uniform quantization methods while significantly improving accuracy. Experiments show that SliM-LLM achieves superior performance across various LLMs at low bit-widths. For example, a 2-bit quantized LLaMA-7B model reduces memory usage by nearly 6x compared to the floating-point baseline, decreases perplexity by 48% compared to state-of-the-art gradient-free PTQ methods, and maintains GPU inference speed. Additionally, the extended version, SliM-LLM+, which incorporates gradient-based quantization, further reduces perplexity by 35.1%. Our code is available at https://github.com/Aaronhuang-778/SliM-LLM.

We introduce a full-stack framework that scales up reasoning in vision-language models (VLMs) to long videos, leveraging reinforcement learning. We address the unique challenges of long video reasoning by integrating three critical components: (1) a large-scale dataset, LongVideo-Reason, comprising 104K long video QA pairs with high-quality reasoning annotations across diverse domains such as sports, games, and vlogs; (2) a two-stage training pipeline that extends VLMs with chain-of-thought supervised fine-tuning (CoT-SFT) and reinforcement learning (RL); and (3) a training infrastructure for long video RL, named Multi-modal Reinforcement Sequence Parallelism (MR-SP), which incorporates sequence parallelism and a vLLM-based engine tailored for long video, using cached video embeddings for efficient rollout and prefilling. In our experiments, LongVILA-R1-7B achieves strong performance on video benchmarks, reaching 65.1% and 71.1% accuracy on VideoMME without and with subtitles, respectively, and consistently outperforming LongVILA-7B across multiple benchmarks. Moreover, LongVILA-R1-7B supports processing up to 8,192 video frames per video, and configurable FPS settings. Notably, our MR-SP system achieves up to 2.1x speedup on long video RL training. In addition, we release our training system for public availability that supports RL training on various modalities (video, text, and audio), various models (VILA and Qwen series), and even image and video generation models. On a single A100 node (8 GPUs), it supports RL training on hour-long videos (e.g., 3,600 frames). Code and models are available at https://github.com/NVlabs/Long-RL

Recent text-to-image systems face limitations in handling multimodal inputs and complex reasoning tasks. We introduce MindOmni, a unified multimodal large language model that addresses these challenges by incorporating reasoning generation through reinforcement learning. MindOmni leverages a three-phase training strategy: i) design of a unified vision language model with a decoder-only diffusion module, ii) supervised fine-tuning with Chain-of-Thought (CoT) instruction data, and iii) our proposed Reasoning Generation Policy Optimization (RGPO) algorithm, utilizing multimodal feedback to effectively guide policy updates. Experimental results demonstrate that MindOmni outperforms existing models, achieving impressive performance on both understanding and generation benchmarks, meanwhile showcasing advanced fine-grained reasoning generation capabilities, especially with mathematical reasoning instruction. All codes will be made public.

Biologically plausible Spiking Neural Networks (SNNs), characterized by spike sparsity, are growing tremendous attention over intellectual edge devices and critical bio-medical applications as compared to artificial neural networks (ANNs). However, there is a considerable risk from malicious attempts to extract white-box information (i.e., weights) from SNNs, as attackers could exploit well-trained SNNs for profit and white-box adversarial concerns. There is a dire need for intellectual property (IP) protective measures. In this paper, we present a novel secure software-hardware co-designed RRAM-based neuromorphic accelerator for protecting the IP of SNNs. Software-wise, we design a tailored genetic algorithm with classic XOR encryption to target the least number of weights that need encryption. From a hardware perspective, we develop a low-energy decryption module, meticulously designed to provide zero decryption latency. Extensive results from various datasets, including NMNIST, DVSGesture, EEGMMIDB, Braille Letter, and SHD, demonstrate that our proposed method effectively secures SNNs by encrypting a minimal fraction of stealthy weights, only 0.00005% to 0.016% weight bits. Additionally, it achieves a substantial reduction in energy consumption, ranging from ×59 to ×6780, and significantly lowers decryption latency, ranging from ×175 to ×4250. Moreover, our method requires as little as one sample per class in dataset for encryption and addresses hessian/gradient-based search insensitive problems. This strategy offers a highly efficient and flexible solution for securing SNNs in diverse applications1.

Pretrained large language models (LLMs) exhibit exceptional general language processing capabilities but come with significant demands on memory and computational resources. As a powerful compression technology, binarization can extremely reduce model weights to a mere 1 bit, lowering the expensive computation and memory requirements. However, existing quantization techniques fall short of maintaining LLM performance under ultra-low bit-widths. In response to this challenge, we present BiLLM, a groundbreaking 1-bit post-training quantization scheme tailored for pretrained LLMs. Based on the weight distribution of LLMs, BiLLM first identifies and structurally selects salient weights, and minimizes the compression loss through an effective binary residual approximation strategy. Moreover, considering the bell-shaped distribution of the non-salient weights, we propose an optimal splitting search to group and binarize them accurately. BiLLM, for the first time, achieves high-accuracy inference (e.g. 8.41 perplexity on LLaMA2-70B) with only 1.08-bit weights across various LLM families and evaluation metrics, outperforms SOTA quantization methods of LLM by significant margins. Moreover, BiLLM enables the binarization process of a 7-billion LLM within 0.5 hours on a single GPU, demonstrating satisfactory time efficiency. Our code is available at https://github.com/Aaronhuang-778/BiLLM .

In order to cultivate students' software and hardware collaborative design thinking and system construction ability, this paper proposes an innovative experimental teaching method of software and hardware collaborative design to better develop students'sys-tem view. We create a student-centered learning environment and adopt a mixed teaching method whereby experiment courses are conducted in a “traditional + discussion + scientific research-driven” manner. The course will focus on the design of the “Convolutional Neural Network for Handwritten Digit Recognition”, and run through curriculums that involve from software to hardware, algorithms to systems, and design to verification. We explore teaching methods that revolve around principles of using experiments to derive theories, and using theories to guide experiments. Besides, we establish an online laboratory that can provide an experiment platform (MOOE) for “PC+ARM+FPGA”, which realizes the seamless integration of physical operations, semi-physical emulators, and virtual models. Through investigation and research, most students who have completed this course have formed a good knowledge of computer systems and the necessary comprehensive development ability.