Xuefei Ning

Enshu Liu, Junyi Zhu, Zinan Lin et al. · microsoft-research

The rapid advancement of large language models (LLMs) has led to architectures with billions to trillions of parameters, posing significant deployment challenges due to their substantial demands on memory, processing power, and energy consumption. Sparse Mixture-of-Experts (SMoE) architectures have emerged as a solution, activating only a subset of parameters per token, thereby achieving faster inference while maintaining performance. However, SMoE models still face limitations in broader deployment due to their large parameter counts and significant GPU memory requirements. In this work, we introduce a gradient-free evolutionary strategy named EEP (Efficient Expert P}runing) to enhance the pruning of experts in SMoE models. EEP relies solely on model inference (i.e., no gradient computation) and achieves greater sparsity while maintaining or even improving performance on downstream tasks. EEP can be used to reduce both the total number of experts (thus saving GPU memory) and the number of active experts (thus accelerating inference). For example, we demonstrate that pruning up to 75% of experts in Mixtral $8\times7$B-Instruct results in a substantial reduction in parameters with minimal performance loss. Remarkably, we observe improved performance on certain tasks, such as a significant increase in accuracy on the SQuAD dataset (from 53.4% to 75.4%), when pruning half of the experts. With these results, EEP not only lowers the barrier to deploying SMoE models,but also challenges the conventional understanding of model pruning by showing that fewer experts can lead to better task-specific performance without any fine-tuning. Code is available at https://github.com/imagination-research/EEP.

Enshu Liu, Xuefei Ning, Zinan Lin et al. · microsoft-research

Diffusion probabilistic models (DPMs) are a new class of generative models that have achieved state-of-the-art generation quality in various domains. Despite the promise, one major drawback of DPMs is the slow generation speed due to the large number of neural network evaluations required in the generation process. In this paper, we reveal an overlooked dimension -- model schedule -- for optimizing the trade-off between generation quality and speed. More specifically, we observe that small models, though having worse generation quality when used alone, could outperform large models in certain generation steps. Therefore, unlike the traditional way of using a single model, using different models in different generation steps in a carefully designed \emph{model schedule} could potentially improve generation quality and speed \emph{simultaneously}. We design OMS-DPM, a predictor-based search algorithm, to optimize the model schedule given an arbitrary generation time budget and a set of pre-trained models. We demonstrate that OMS-DPM can find model schedules that improve generation quality and speed than prior state-of-the-art methods across CIFAR-10, CelebA, ImageNet, and LSUN datasets. When applied to the public checkpoints of the Stable Diffusion model, we are able to accelerate the sampling by 2$\times$ while maintaining the generation quality.

Zixuan Zhou, Xuefei Ning, Yi Cai et al.

One-shot Neural Architecture Search (NAS) has been widely used to discover architectures due to its efficiency. However, previous studies reveal that one-shot performance estimations of architectures might not be well correlated with their performances in stand-alone training because of the excessive sharing of operation parameters (i.e., large sharing extent) between architectures. Thus, recent methods construct even more over-parameterized supernets to reduce the sharing extent. But these improved methods introduce a large number of extra parameters and thus cause an undesirable trade-off between the training costs and the ranking quality. To alleviate the above issues, we propose to apply Curriculum Learning On Sharing Extent (CLOSE) to train the supernet both efficiently and effectively. Specifically, we train the supernet with a large sharing extent (an easier curriculum) at the beginning and gradually decrease the sharing extent of the supernet (a harder curriculum). To support this training strategy, we design a novel supernet (CLOSENet) that decouples the parameters from operations to realize a flexible sharing scheme and adjustable sharing extent. Extensive experiments demonstrate that CLOSE can obtain a better ranking quality across different computational budget constraints than other one-shot supernets, and is able to discover superior architectures when combined with various search strategies. Code is available at https://github.com/walkerning/aw_nas.

Xuefei Ning, Zinan Lin, Zixuan Zhou et al. · microsoft-research

This work aims at decreasing the end-to-end generation latency of large language models (LLMs). One of the major causes of the high generation latency is the sequential decoding approach adopted by almost all state-of-the-art LLMs. In this work, motivated by the thinking and writing process of humans, we propose Skeleton-of-Thought (SoT), which first guides LLMs to generate the skeleton of the answer, and then conducts parallel API calls or batched decoding to complete the contents of each skeleton point in parallel. Not only does SoT provide considerable speed-ups across 12 LLMs, but it can also potentially improve the answer quality on several question categories. SoT is an initial attempt at data-centric optimization for inference efficiency, and showcases the potential of eliciting high-quality answers by explicitly planning the answer structure in language.

Zifu Wang, Xuefei Ning, Matthew B. Blaschko

Intersection over Union (IoU) losses are surrogates that directly optimize the Jaccard index. Leveraging IoU losses as part of the loss function have demonstrated superior performance in semantic segmentation tasks compared to optimizing pixel-wise losses such as the cross-entropy loss alone. However, we identify a lack of flexibility in these losses to support vital training techniques like label smoothing, knowledge distillation, and semi-supervised learning, mainly due to their inability to process soft labels. To address this, we introduce Jaccard Metric Losses (JMLs), which are identical to the soft Jaccard loss in standard settings with hard labels but are fully compatible with soft labels. We apply JMLs to three prominent use cases of soft labels: label smoothing, knowledge distillation and semi-supervised learning, and demonstrate their potential to enhance model accuracy and calibration. Our experiments show consistent improvements over the cross-entropy loss across 4 semantic segmentation datasets (Cityscapes, PASCAL VOC, ADE20K, DeepGlobe Land) and 13 architectures, including classic CNNs and recent vision transformers. Remarkably, our straightforward approach significantly outperforms state-of-the-art knowledge distillation and semi-supervised learning methods. The code is available at \href{https://github.com/zifuwanggg/JDTLosses}{https://github.com/zifuwanggg/JDTLosses}.

Tianchen Zhao, Xuefei Ning, Ke Hong et al. · tsinghua

Voxel-based methods have achieved state-of-the-art performance for 3D object detection in autonomous driving. However, their significant computational and memory costs pose a challenge for their application to resource-constrained vehicles. One reason for this high resource consumption is the presence of a large number of redundant background points in Lidar point clouds, resulting in spatial redundancy in both 3D voxel and dense BEV map representations. To address this issue, we propose an adaptive inference framework called Ada3D, which focuses on exploiting the input-level spatial redundancy. Ada3D adaptively filters the redundant input, guided by a lightweight importance predictor and the unique properties of the Lidar point cloud. Additionally, we utilize the BEV features' intrinsic sparsity by introducing the Sparsity Preserving Batch Normalization. With Ada3D, we achieve 40% reduction for 3D voxels and decrease the density of 2D BEV feature maps from 100% to 20% without sacrificing accuracy. Ada3D reduces the model computational and memory cost by 5x, and achieves 1.52x/1.45x end-to-end GPU latency and 1.5x/4.5x GPU peak memory optimization for the 3D and 2D backbone respectively.

Enshu Liu, Xuefei Ning, Yu Wang et al.

Discrete diffusion language models (dLLMs) have recently emerged as a promising alternative to traditional autoregressive approaches, offering the flexibility to generate tokens in arbitrary orders and the potential of parallel decoding. However, existing heuristic sampling strategies remain inefficient: they choose only a small part of tokens to sample at each step, leaving substantial room for improvement. In this work, we study the problem of token sampling order optimization and demonstrate its significant potential for acceleration. Specifically, we find that fully leveraging correct predictions at each step can reduce the number of sampling iterations by an order of magnitude without compromising accuracy. Based on this, we propose Neural Indicator Sampling (NI Sampling), a general sampling order optimization framework that utilize a neural indicator to decide which tokens should be sampled at each step. We further propose a novel trajectory-preserving objective to train the indicator. Experiments on LLaDA and Dream models across multiple benchmarks show that our method achieves up to 14.3$\times$ acceleration over full-step sampling with negligible performance drop, and consistently outperforms confidence threshold sampling in the accuracy-step trade-off. Code is available at https://github.com/imagination-research/NI-Sampling.

Junbo Zhao, Xuefei Ning, Enshu Liu et al. · tsinghua

Predictor-based Neural Architecture Search (NAS) employs an architecture performance predictor to improve the sample efficiency. However, predictor-based NAS suffers from the severe ``cold-start'' problem, since a large amount of architecture-performance data is required to get a working predictor. In this paper, we focus on exploiting information in cheaper-to-obtain performance estimations (i.e., low-fidelity information) to mitigate the large data requirements of predictor training. Despite the intuitiveness of this idea, we observe that using inappropriate low-fidelity information even damages the prediction ability and different search spaces have different preferences for low-fidelity information types. To solve the problem and better fuse beneficial information provided by different types of low-fidelity information, we propose a novel dynamic ensemble predictor framework that comprises two steps. In the first step, we train different sub-predictors on different types of available low-fidelity information to extract beneficial knowledge as low-fidelity experts. In the second step, we learn a gating network to dynamically output a set of weighting coefficients conditioned on each input neural architecture, which will be used to combine the predictions of different low-fidelity experts in a weighted sum. The overall predictor is optimized on a small set of actual architecture-performance data to fuse the knowledge from different low-fidelity experts to make the final prediction. We conduct extensive experiments across five search spaces with different architecture encoders under various experimental settings. Our method can easily be incorporated into existing predictor-based NAS frameworks to discover better architectures.

Tianchen Zhao, Niansong Zhang, Xuefei Ning et al. · tsinghua

Transformers have gained much attention by outperforming convolutional neural networks in many 2D vision tasks. However, they are known to have generalization problems and rely on massive-scale pre-training and sophisticated training techniques. When applying to 3D tasks, the irregular data structure and limited data scale add to the difficulty of transformer's application. We propose CodedVTR (Codebook-based Voxel TRansformer), which improves data efficiency and generalization ability for 3D sparse voxel transformers. On the one hand, we propose the codebook-based attention that projects an attention space into its subspace represented by the combination of "prototypes" in a learnable codebook. It regularizes attention learning and improves generalization. On the other hand, we propose geometry-aware self-attention that utilizes geometric information (geometric pattern, density) to guide attention learning. CodedVTR could be embedded into existing sparse convolution-based methods, and bring consistent performance improvements for indoor and outdoor 3D semantic segmentation tasks

Luning Wang, Shiyao Li, Xuefei Ning et al.

Large Language Models (LLMs) have been widely adopted to process long-context tasks. However, the large memory overhead of the key-value (KV) cache poses significant challenges in long-context scenarios. Existing training-free KV cache compression methods typically focus on quantization and token pruning, which have compression limits, and excessive sparsity can lead to severe performance degradation. Other methods design new architectures with less KV overhead but require significant training overhead. To address the above two drawbacks, we further explore the redundancy in the channel dimension and apply an architecture-level design with minor training costs. Therefore, we introduce CSKV, a training-efficient Channel Shrinking technique for KV cache compression: (1) We first analyze the singular value distribution of the KV cache, revealing significant redundancy and compression potential along the channel dimension. Based on this observation, we propose using low-rank decomposition for key and value layers and storing the low-dimension features. (2) To preserve model performance, we introduce a bi-branch KV cache, including a window-based full-precision KV cache and a low-precision compressed KV cache. (3) To reduce the training costs, we minimize the layer-wise reconstruction loss for the compressed KV cache instead of retraining the entire LLMs. Extensive experiments show that CSKV can reduce the memory overhead of the KV cache by 80% while maintaining the model's long-context capability. Moreover, we show that our method can be seamlessly combined with quantization to further reduce the memory overhead, achieving a compression ratio of up to 95%. Code is available at https://github.com/wln20/CSKV.

Tongcheng Fang, Hanling Zhang, Ruiqi Xie et al. · tsinghua

Diffusion Transformers have recently demonstrated remarkable performance in video generation. However, the long input sequences result in high computational latency due to the quadratic complexity of full attention. Various sparse attention mechanisms have been proposed. Training-free sparse attention is constrained by limited sparsity and thus offers modest acceleration, whereas training-based methods can reach much higher sparsity but demand substantial data and computation for training. In this work, we propose SALAD, introducing a lightweight linear attention branch in parallel with the sparse attention. By incorporating an input-dependent gating mechanism to finely balance the two branches, our method attains 90% sparsity and 1.72x inference speedup, while maintaining generation quality comparable to the full attention baseline. Moreover, our finetuning process is highly efficient, requiring only 2,000 video samples and 1,600 training steps with a batch size of 8.

Cheng Liu, Zhen Gao, Siting Liu et al.

With the rapid advancements of deep learning in the past decade, it can be foreseen that deep learning will be continuously deployed in more and more safety-critical applications such as autonomous driving and robotics. In this context, reliability turns out to be critical to the deployment of deep learning in these applications and gradually becomes a first-class citizen among the major design metrics like performance and energy efficiency. Nevertheless, the back-box deep learning models combined with the diverse underlying hardware faults make resilient deep learning extremely challenging. In this special session, we conduct a comprehensive survey of fault-tolerant deep learning design approaches with a hierarchical perspective and investigate these approaches from model layer, architecture layer, circuit layer, and cross layer respectively.

Lin Zhao, Hongxuan Li, Xuefei Ning et al.

Cross-modal Steganography is the practice of concealing secret signals in publicly available cover signals (distinct from the modality of the secret signals) unobtrusively. While previous approaches primarily concentrated on concealing a relatively small amount of information, we propose THInImg, which manages to hide lengthy audio data (and subsequently decode talking head video) inside an identity image by leveraging the properties of human face, which can be effectively utilized for covert communication, transmission and copyright protection. THInImg consists of two parts: the encoder and decoder. Inside the encoder-decoder pipeline, we introduce a novel architecture that substantially increase the capacity of hiding audio in images. Moreover, our framework can be extended to iteratively hide multiple audio clips into an identity image, offering multiple levels of control over permissions. We conduct extensive experiments to prove the effectiveness of our method, demonstrating that THInImg can present up to 80 seconds of high quality talking-head video (including audio) in an identity image with 160x160 resolution.

Kaiyi Huang, Yukun Huang, Yu Li et al.

Cinematic video production requires control over scene-subject composition and camera movement, but live-action shooting remains costly due to the need for constructing physical sets. To address this, we introduce the task of cinematic video generation with decoupled scene context: given multiple images of a static environment, the goal is to synthesize high-quality videos featuring dynamic subject while preserving the underlying scene consistency and following a user-specified camera trajectory. We present CineScene, a framework that leverages implicit 3D-aware scene representation for cinematic video generation. Our key innovation is a novel context conditioning mechanism that injects 3D-aware features in an implicit way: By encoding scene images into visual representations through VGGT, CineScene injects spatial priors into a pretrained text-to-video generation model by additional context concatenation, enabling camera-controlled video synthesis with consistent scenes and dynamic subjects. To further enhance the model's robustness, we introduce a simple yet effective random-shuffling strategy for the input scene images during training. To address the lack of training data, we construct a scene-decoupled dataset with Unreal Engine 5, containing paired videos of scenes with and without dynamic subjects, panoramic images representing the underlying static scene, along with their camera trajectories. Experiments show that CineScene achieves state-of-the-art performance in scene-consistent cinematic video generation, handling large camera movements and demonstrating generalization across diverse environments.

Shiyao Li, Xuefei Ning, Luning Wang et al.

Post-training quantization (PTQ) has emerged as a promising technique to reduce the cost of large language models (LLMs). Specifically, PTQ can effectively mitigate memory consumption and reduce computational overhead in LLMs. To meet the requirements of both high efficiency and performance across diverse scenarios, a comprehensive evaluation of quantized LLMs is essential to guide the selection of quantization methods. This paper presents a thorough evaluation of these factors by evaluating the effect of PTQ on Weight, Activation, and KV Cache on 11 model families, including OPT, LLaMA2, Falcon, Bloomz, Mistral, ChatGLM, Vicuna, LongChat, StableLM, Gemma, and Mamba, with parameters ranging from 125M to 180B. The evaluation encompasses five types of tasks: basic NLP, emergent ability, trustworthiness, dialogue, and long-context tasks. Moreover, we also evaluate the state-of-the-art (SOTA) quantization methods to demonstrate their applicability. Based on the extensive experiments, we systematically summarize the effect of quantization, provide recommendations to apply quantization techniques, and point out future directions. The code can be found in https://github.com/thu-nics/qllm-eval.

Yao Teng, Yue Wu, Han Shi et al.

Diffusion models have achieved great success in image generation, with the backbone evolving from U-Net to Vision Transformers. However, the computational cost of Transformers is quadratic to the number of tokens, leading to significant challenges when dealing with high-resolution images. In this work, we propose Diffusion Mamba (DiM), which combines the efficiency of Mamba, a sequence model based on State Space Models (SSM), with the expressive power of diffusion models for efficient high-resolution image synthesis. To address the challenge that Mamba cannot generalize to 2D signals, we make several architecture designs including multi-directional scans, learnable padding tokens at the end of each row and column, and lightweight local feature enhancement. Our DiM architecture achieves inference-time efficiency for high-resolution images. In addition, to further improve training efficiency for high-resolution image generation with DiM, we investigate "weak-to-strong" training strategy that pretrains DiM on low-resolution images ($256\times 256$) and then finetune it on high-resolution images ($512 \times 512$). We further explore training-free upsampling strategies to enable the model to generate higher-resolution images (e.g., $1024\times 1024$ and $1536\times 1536$) without further fine-tuning. Experiments demonstrate the effectiveness and efficiency of our DiM. The code of our work is available here: {\url{https://github.com/tyshiwo1/DiM-DiffusionMamba/}}.

Tao Yuan, Xuefei Ning, Dong Zhou et al.

State-of-the-art large language models (LLMs) are now claiming remarkable supported context lengths of 256k or even more. In contrast, the average context lengths of mainstream benchmarks are insufficient (5k-21k), and they suffer from potential knowledge leakage and inaccurate metrics, resulting in biased evaluation. This paper introduces LV-Eval, a challenging long-context benchmark with five length levels (16k, 32k, 64k, 128k, and 256k) reaching up to 256k words. LV-Eval features two main tasks, single-hop QA and multi-hop QA, comprising 11 bilingual datasets. The design of LV-Eval has incorporated three key techniques, namely confusing facts insertion, keyword and phrase replacement, and keyword-recall-based metric design. The advantages of LV-Eval include controllable evaluation across different context lengths, challenging test instances with confusing facts, mitigated knowledge leakage, and more objective evaluations. We evaluate 15 LLMs on LV-Eval and conduct ablation studies on the benchmarking techniques. The results reveal that: (i) Moonshot-v1 and recent large-scale open-source models, such as Qwen-2.5-72B and Llama-3.1-70B, achieve the highest performance on LV-Eval, particularly at lengths below 64k. (ii) Models exhibit distinct score trends. For example, GLM-4-9B-128k, Yi-6B-200k, and Llama3-8B-1M exhibit a relatively gentle degradation of performance, but their absolute performances may not necessarily be higher than those of LLMs with shorter context lengths. (iii) LLMs' performances can significantly degrade in the presence of confusing information, especially in the pressure test of "needle in a haystack". (iv) Issues related to knowledge leakage and inaccurate metrics introduce bias in evaluation, and these concerns are alleviated in LV-Eval. All datasets and evaluation codes are released at: https://github.com/infinigence/LVEval.

Yao Teng, Zhihuan Jiang, Han Shi et al.

Large autoregressive models can generate high-quality, high-resolution images but suffer from slow generation speed, because these models require hundreds to thousands of sequential forward passes for next-token prediction during inference. To accelerate autoregressive text-to-image generation, we propose Speculative Jacobi Decoding++ (SJD++), a training-free probabilistic parallel decoding algorithm. Unlike traditional next-token prediction, SJD++ performs multi-token prediction in each forward pass, drastically reducing generation steps. Specifically, it integrates the iterative multi-token prediction mechanism from Jacobi decoding, with the probabilistic drafting-and-verification mechanism from speculative sampling. More importantly, for further acceleration, SJD++ reuses high-confidence draft tokens after each verification phase instead of resampling them all. We conduct extensive experiments on several representative autoregressive text-to-image generation models and demonstrate that SJD++ achieves $2\times$ to $3\times$ inference latency reduction and $2\times$ to $7\times$ step compression, while preserving visual quality with no observable degradation.

Tianyu Fu, Tengxuan Liu, Qinghao Han et al. · tsinghua

The increasing demand to process long and high-resolution videos significantly burdens Large Vision-Language Models (LVLMs) due to the enormous number of visual tokens. Existing token reduction methods primarily prune tokens based on importance metrics, such as cumulative attention scores. However, even important tokens may exhibit high redundancy caused by similarity among adjacent video frames and repetitive visual elements. To address this limitation, we propose FrameFusion, a novel token reduction approach integrating similarity-based merging with importance-based pruning. We conduct a thorough study on token similarity characteristics, revealing three key insights: (1) spatially corresponding visual tokens between adjacent frames have higher cosine similarities compared to other token pairs; (2) high token similarities prominently decrease in deeper model layers; and (3) token similarity rankings are highly consistent across different layers. Guided by these observations, FrameFusion computes token similarities exclusively between corresponding visual tokens from adjacent frames, applies token merging at initial successive layers followed by pruning in deeper layers, and adopts a cascaded merging strategy to further enhance efficiency. We evaluate FrameFusion comprehensively across six diverse LVLMs, ranging from 2B to 72B parameters, using five video benchmarks encompassing video retrieval, question-answering, and spatial-temporal understanding tasks. Experiments show that FrameFusion reduces visual tokens by 70%, achieving 1.6-3.6x end-to-end speedups, with an average performance impact of less than 3%. Our code is available at: https://github.com/thu-nics/FrameFusion.

Lin Zhao, Tianchen Zhao, Zinan Lin et al. · microsoft-research, tsinghua

In recent years, there has been significant progress in the development of text-to-image generative models. Evaluating the quality of the generative models is one essential step in the development process. Unfortunately, the evaluation process could consume a significant amount of computational resources, making the required periodic evaluation of model performance (e.g., monitoring training progress) impractical. Therefore, we seek to improve the evaluation efficiency by selecting the representative subset of the text-image dataset. We systematically investigate the design choices, including the selection criteria (textural features or image-based metrics) and the selection granularity (prompt-level or set-level). We find that the insights from prior work on subset selection for training data do not generalize to this problem, and we propose FlashEval, an iterative search algorithm tailored to evaluation data selection. We demonstrate the effectiveness of FlashEval on ranking diffusion models with various configurations, including architectures, quantization levels, and sampler schedules on COCO and DiffusionDB datasets. Our searched 50-item subset could achieve comparable evaluation quality to the randomly sampled 500-item subset for COCO annotations on unseen models, achieving a 10x evaluation speedup. We release the condensed subset of these commonly used datasets to help facilitate diffusion algorithm design and evaluation, and open-source FlashEval as a tool for condensing future datasets, accessible at https://github.com/thu-nics/FlashEval.

Yuwei Yan, Yu Shang, Qingbin Zeng et al.

The AgentSociety Challenge is the first competition in the Web Conference that aims to explore the potential of Large Language Model (LLM) agents in modeling user behavior and enhancing recommender systems on web platforms. The Challenge consists of two tracks: the User Modeling Track and the Recommendation Track. Participants are tasked to utilize a combined dataset from Yelp, Amazon, and Goodreads, along with an interactive environment simulator, to develop innovative LLM agents. The Challenge has attracted 295 teams across the globe and received over 1,400 submissions in total over the course of 37 official competition days. The participants have achieved 21.9% and 20.3% performance improvement for Track 1 and Track 2 in the Development Phase, and 9.1% and 15.9% in the Final Phase, representing a significant accomplishment. This paper discusses the detailed designs of the Challenge, analyzes the outcomes, and highlights the most successful LLM agent designs. To support further research and development, we have open-sourced the benchmark environment at https://tsinghua-fib-lab.github.io/AgentSocietyChallenge.

Shiyao Li, Yingchun Hu, Xuefei Ning et al.

Vision-Language Models (VLMs) have enabled a variety of real-world applications. The large parameter size of VLMs brings large memory and computation overhead which poses significant challenges for deployment. Post-Training Quantization (PTQ) is an effective technique to reduce the memory and computation overhead. Existing PTQ methods mainly focus on large language models (LLMs), without considering the differences across other modalities. In this paper, we discover that there is a significant difference in sensitivity between language and vision tokens in large VLMs. Therefore, treating tokens from different modalities equally, as in existing PTQ methods, may over-emphasize the insensitive modalities, leading to significant accuracy loss. To deal with the above issue, we propose a simple yet effective method, Modality-Balanced Quantization (MBQ), for large VLMs. Specifically, MBQ incorporates the different sensitivities across modalities during the calibration process to minimize the reconstruction loss for better quantization parameters. Extensive experiments show that MBQ can significantly improve task accuracy by up to 4.4% and 11.6% under W3 and W4A8 quantization for 7B to 70B VLMs, compared to SOTA baselines. Additionally, we implement a W3 GPU kernel that fuses the dequantization and GEMV operators, achieving a 1.4x speedup on LLaVA-onevision-7B on the RTX 4090. The code is available at https://github.com/thu-nics/MBQ.

Enshu Liu, Junyi Zhu, Zinan Lin et al. · microsoft-research

Diffusion Models (DM) and Consistency Models (CM) are two types of popular generative models with good generation quality on various tasks. When training DM and CM, intermediate weight checkpoints are not fully utilized and only the last converged checkpoint is used. In this work, we find that high-quality model weights often lie in a basin which cannot be reached by SGD but can be obtained by proper checkpoint averaging. Based on these observations, we propose LCSC, a simple but effective and efficient method to enhance the performance of DM and CM, by combining checkpoints along the training trajectory with coefficients deduced from evolutionary search. We demonstrate the value of LCSC through two use cases: $\textbf{(a) Reducing training cost.}$ With LCSC, we only need to train DM/CM with fewer number of iterations and/or lower batch sizes to obtain comparable sample quality with the fully trained model. For example, LCSC achieves considerable training speedups for CM (23$\times$ on CIFAR-10 and 15$\times$ on ImageNet-64). $\textbf{(b) Enhancing pre-trained models.}$ Assuming full training is already done, LCSC can further improve the generation quality or speed of the final converged models. For example, LCSC achieves better performance using 1 number of function evaluation (NFE) than the base model with 2 NFE on consistency distillation, and decreases the NFE of DM from 15 to 9 while maintaining the generation quality on CIFAR-10. Our code is available at https://github.com/imagination-research/LCSC.

Enshu Liu, Qian Chen, Xuefei Ning et al.

Image Auto-regressive (AR) models have emerged as a powerful paradigm of visual generative models. Despite their promising performance, they suffer from slow generation speed due to the large number of sampling steps required. Although Distilled Decoding 1 (DD1) was recently proposed to enable few-step sampling for image AR models, it still incurs significant performance degradation in the one-step setting, and relies on a pre-defined mapping that limits its flexibility. In this work, we propose a new method, Distilled Decoding 2 (DD2), to further advances the feasibility of one-step sampling for image AR models. Unlike DD1, DD2 does not without rely on a pre-defined mapping. We view the original AR model as a teacher model which provides the ground truth conditional score in the latent embedding space at each token position. Based on this, we propose a novel \emph{conditional score distillation loss} to train a one-step generator. Specifically, we train a separate network to predict the conditional score of the generated distribution and apply score distillation at every token position conditioned on previous tokens. Experimental results show that DD2 enables one-step sampling for image AR models with an minimal FID increase from 3.40 to 5.43 on ImageNet-256. Compared to the strongest baseline DD1, DD2 reduces the gap between the one-step sampling and original AR model by 67%, with up to 12.3$\times$ training speed-up simultaneously. DD2 takes a significant step toward the goal of one-step AR generation, opening up new possibilities for fast and high-quality AR modeling. Code is available at https://github.com/imagination-research/Distilled-Decoding-2.

Zixuan Zhou, Xuefei Ning, Ke Hong et al. · tsinghua

Large Language Models (LLMs) have attracted extensive attention due to their remarkable performance across various tasks. However, the substantial computational and memory requirements of LLM inference pose challenges for deployment in resource-constrained scenarios. Efforts within the field have been directed towards developing techniques aimed at enhancing the efficiency of LLM inference. This paper presents a comprehensive survey of the existing literature on efficient LLM inference. We start by analyzing the primary causes of the inefficient LLM inference, i.e., the large model size, the quadratic-complexity attention operation, and the auto-regressive decoding approach. Then, we introduce a comprehensive taxonomy that organizes the current literature into data-level, model-level, and system-level optimization. Moreover, the paper includes comparative experiments on representative methods within critical sub-fields to provide quantitative insights. Last but not least, we provide some knowledge summary and discuss future research directions.

Shulin Zeng, Jun Liu, Guohao Dai et al. · tsinghua

Transformer-based Large Language Models (LLMs) have made a significant impact on various domains. However, LLMs' efficiency suffers from both heavy computation and memory overheads. Compression techniques like sparsification and quantization are commonly used to mitigate the gap between LLM's computation/memory overheads and hardware capacity. However, existing GPU and transformer-based accelerators cannot efficiently process compressed LLMs, due to the following unresolved challenges: low computational efficiency, underutilized memory bandwidth, and large compilation overheads. This paper proposes FlightLLM, enabling efficient LLMs inference with a complete mapping flow on FPGAs. In FlightLLM, we highlight an innovative solution that the computation and memory overhead of LLMs can be solved by utilizing FPGA-specific resources (e.g., DSP48 and heterogeneous memory hierarchy). We propose a configurable sparse DSP chain to support different sparsity patterns with high computation efficiency. Second, we propose an always-on-chip decode scheme to boost memory bandwidth with mixed-precision support. Finally, to make FlightLLM available for real-world LLMs, we propose a length adaptive compilation method to reduce the compilation overhead. Implemented on the Xilinx Alveo U280 FPGA, FlightLLM achieves 6.0$\times$ higher energy efficiency and 1.8$\times$ better cost efficiency against commercial GPUs (e.g., NVIDIA V100S) on modern LLMs (e.g., LLaMA2-7B) using vLLM and SmoothQuant under the batch size of one. FlightLLM beats NVIDIA A100 GPU with 1.2$\times$ higher throughput using the latest Versal VHK158 FPGA.

Zinan Lin, Enshu Liu, Xuefei Ning et al.

Generative modeling, representation learning, and classification are three core problems in machine learning (ML), yet their state-of-the-art (SoTA) solutions remain largely disjoint. In this paper, we ask: Can a unified principle address all three? Such unification could simplify ML pipelines and foster greater synergy across tasks. We introduce Latent Zoning Network (LZN) as a step toward this goal. At its core, LZN creates a shared Gaussian latent space that encodes information across all tasks. Each data type (e.g., images, text, labels) is equipped with an encoder that maps samples to disjoint latent zones, and a decoder that maps latents back to data. ML tasks are expressed as compositions of these encoders and decoders: for example, label-conditional image generation uses a label encoder and image decoder; image embedding uses an image encoder; classification uses an image encoder and label decoder. We demonstrate the promise of LZN in three increasingly complex scenarios: (1) LZN can enhance existing models (image generation): When combined with the SoTA Rectified Flow model, LZN improves FID on CIFAR10 from 2.76 to 2.59-without modifying the training objective. (2) LZN can solve tasks independently (representation learning): LZN can implement unsupervised representation learning without auxiliary loss functions, outperforming the seminal MoCo and SimCLR methods by 9.3% and 0.2%, respectively, on downstream linear classification on ImageNet. (3) LZN can solve multiple tasks simultaneously (joint generation and classification): With image and label encoders/decoders, LZN performs both tasks jointly by design, improving FID and achieving SoTA classification accuracy on CIFAR10. The code and trained models are available at https://github.com/microsoft/latent-zoning-networks. The project website is at https://zinanlin.me/blogs/latent_zoning_networks.html.

Tianyu Fu, Haofeng Huang, Xuefei Ning et al.

Sliding-window attention offers a hardware-efficient solution to the memory and throughput challenges of Large Language Models (LLMs) in long-context scenarios. Existing methods typically employ a single window length across all attention heads and input sizes. However, this uniform approach fails to capture the heterogeneous attention patterns inherent in LLMs, ignoring their distinct accuracy-latency trade-offs. To address this challenge, we propose *Mixture of Attention Spans* (MoA), which automatically tailors distinct sliding-window length configurations to different heads and layers. MoA constructs and navigates a search space of various window lengths and their scaling rules relative to input sizes. It profiles the model, evaluates potential configurations, and pinpoints the optimal length configurations for each head. MoA adapts to varying input sizes, revealing that some attention heads expand their focus to accommodate longer inputs, while other heads consistently concentrate on fixed-length local contexts. Experiments show that MoA increases the effective context length by 3.9x with the same average sliding-window length, boosting retrieval accuracy by 1.5-7.1x over the uniform-window baseline across Vicuna-{7B, 13B} and Llama3-{8B, 70B} models. Moreover, MoA narrows the performance gap with full attention, reducing the maximum relative performance drop from 9%-36% to within 5% across three long-context understanding benchmarks. MoA achieves a 1.2-1.4x GPU memory reduction, boosting decode throughput by 6.6-8.2x and 1.7-1.9x over FlashAttention2 and vLLM, with minimal performance impact. Our code is available at: https://github.com/thu-nics/MoA

Xuefei Ning, Zifu Wang, Shiyao Li et al.

Teaching to improve student models (e.g., knowledge distillation) is an extensively studied methodology in LLMs. However, for humans, teaching improves not only students but also teachers, by fostering more rigorous and clear reasoning as well as knowledge building. We ask: Can LLMs also learn by teaching (LbT) for better reasoning? If the answer is yes, we can potentially unlock the possibility of continuously advancing the models without solely relying on human-produced data or stronger models. In this paper, we provide a preliminary exploration on this question. We show that LbT ideas can be incorporated into existing LLM training/prompting pipelines and bring improvements. Specifically, we design three methods, each mimicking one of the three levels of LbT: observing students' feedback, learning from the feedback, and learning iteratively, with the goals of improving answer accuracy without training or improving models' inherent capability with fine-tuning. We reveal some findings: (1) Teaching materials that make it easier for students to learn have clearer and more accurate logic when using in-context learning as the student's "learning" method; (2) Weak-to-strong generalization: LbT might help improve strong models by teaching weak models; (3) Diversity in students might help: teaching multiple students could be better than teaching one student or the teacher itself. We hope that our exploration can inspire future research on LbT and more broadly adopting the advanced techniques in education to improve LLMs. The code and website are at https://github.com/imagination-research/lbt and https://sites.google.com/view/llm-learning-by-teaching.

Nianhui Guo, Joseph Bethge, Haojin Yang et al.

Recent works on Binary Neural Networks (BNNs) have made promising progress in narrowing the accuracy gap of BNNs to their 32-bit counterparts. However, the accuracy gains are often based on specialized model designs using additional 32-bit components. Furthermore, almost all previous BNNs use 32-bit for feature maps and the shortcuts enclosing the corresponding binary convolution blocks, which helps to effectively maintain the accuracy, but is not friendly to hardware accelerators with limited memory, energy, and computing resources. Thus, we raise the following question: How can accuracy and energy consumption be balanced in a BNN network design? We extensively study this fundamental problem in this work and propose a novel BNN architecture without most commonly used 32-bit components: \textit{BoolNet}. Experimental results on ImageNet demonstrate that BoolNet can achieve 4.6x energy reduction coupled with 1.2\% higher accuracy than the commonly used BNN architecture Bi-RealNet. Code and trained models are available at: https://github.com/hpi-xnor/BoolNet.

Xuefei Ning, Junbo Zhao, Wenshuo Li et al.

Convolutional neural networks (CNNs) are vulnerable to adversarial examples, and studies show that increasing the model capacity of an architecture topology (e.g., width expansion) can bring consistent robustness improvements. This reveals a clear robustness-efficiency trade-off that should be considered in architecture design. In this paper, considering scenarios with capacity budget, we aim to discover adversarially robust architecture at targeted capacities. Recent studies employed one-shot neural architecture search (NAS) to discover robust architectures. However, since the capacities of different topologies cannot be aligned in the search process, one-shot NAS methods favor topologies with larger capacities in the supernet. And the discovered topology might be suboptimal when augmented to the targeted capacity. We propose a novel multi-shot NAS method to address this issue and explicitly search for robust architectures at targeted capacities. At the targeted FLOPs of 2000M, the discovered MSRobNet-2000 outperforms the recent NAS-discovered architecture RobNet-large under various criteria by a large margin of 4%-7%. And at the targeted FLOPs of 1560M, MSRobNet-1560 surpasses another NAS-discovered architecture RobNet-free by 2.3% and 1.3% in the clean and PGD-7 accuracies, respectively. All codes are available at https://github.com/walkerning/aw\_nas.

Xuefei Ning, Changcheng Tang, Wenshuo Li et al.

Neural Architecture Search (NAS) has received extensive attention due to its capability to discover neural network architectures in an automated manner. aw_nas is an open-source Python framework implementing various NAS algorithms in a modularized manner. Currently, aw_nas can be used to reproduce the results of mainstream NAS algorithms of various types. Also, due to the modularized design, one can simply experiment with different NAS algorithms for various applications with awnas (e.g., classification, detection, text modeling, fault tolerance, adversarial robustness, hardware efficiency, and etc.). Codes and documentation are available at https://github.com/walkerning/aw_nas.

Xuefei Ning, Changcheng Tang, Wenshuo Li et al.

Conducting efficient performance estimations of neural architectures is a major challenge in neural architecture search (NAS). To reduce the architecture training costs in NAS, one-shot estimators (OSEs) amortize the architecture training costs by sharing the parameters of one "supernet" between all architectures. Recently, zero-shot estimators (ZSEs) that involve no training are proposed to further reduce the architecture evaluation cost. Despite the high efficiency of these estimators, the quality of such estimations has not been thoroughly studied. In this paper, we conduct an extensive and organized assessment of OSEs and ZSEs on five NAS benchmarks: NAS-Bench-101/201/301, and NDS ResNet/ResNeXt-A. Specifically, we employ a set of NAS-oriented criteria to study the behavior of OSEs and ZSEs and reveal that they have certain biases and variances. After analyzing how and why the OSE estimations are unsatisfying, we explore how to mitigate the correlation gap of OSEs from several perspectives. Through our analysis, we give out suggestions for future application and development of efficient architecture performance estimators. Furthermore, the analysis framework proposed in our work could be utilized in future research to give a more comprehensive understanding of newly designed architecture performance estimators. All codes are available at https://github.com/walkerning/aw_nas.

Xuefei Ning, Yin Zheng, Tianchen Zhao et al.

This work proposes a novel Graph-based neural ArchiTecture Encoding Scheme, a.k.a. GATES, to improve the predictor-based neural architecture search. Specifically, different from existing graph-based schemes, GATES models the operations as the transformation of the propagating information, which mimics the actual data processing of neural architecture. GATES is a more reasonable modeling of the neural architectures, and can encode architectures from both the "operation on node" and "operation on edge" cell search spaces consistently. Experimental results on various search spaces confirm GATES's effectiveness in improving the performance predictor. Furthermore, equipped with the improved performance predictor, the sample efficiency of the predictor-based neural architecture search (NAS) flow is boosted. Codes are available at https://github.com/walkerning/aw_nas.

Evelyn Zhang, Bang Xiao, Jiayi Tang et al.

Stable Diffusion has achieved remarkable success in the field of text-to-image generation, with its powerful generative capabilities and diverse generation results making a lasting impact. However, its iterative denoising introduces high computational costs and slows generation speed, limiting broader adoption. The community has made numerous efforts to reduce this computational burden, with methods like feature caching attracting attention due to their effectiveness and simplicity. Nonetheless, simply reusing features computed at previous timesteps causes the features across adjacent timesteps to become similar, reducing the dynamics of features over time and ultimately compromising the quality of generated images. In this paper, we introduce a dynamics-aware token pruning (DaTo) approach that addresses the limitations of feature caching. DaTo selectively prunes tokens with lower dynamics, allowing only high-dynamic tokens to participate in self-attention layers, thereby extending feature dynamics across timesteps. DaTo combines feature caching with token pruning in a training-free manner, achieving both temporal and token-wise information reuse. Applied to Stable Diffusion on the ImageNet, our approach delivered a 9$\times$ speedup while reducing FID by 0.33, indicating enhanced image quality. On the COCO-30k, we observed a 7$\times$ acceleration coupled with a notable FID reduction of 2.17.

Enshu Liu, Xuefei Ning, Huazhong Yang et al.

Recent years have witnessed the rapid progress and broad application of diffusion probabilistic models (DPMs). Sampling from DPMs can be viewed as solving an ordinary differential equation (ODE). Despite the promising performance, the generation of DPMs usually consumes much time due to the large number of function evaluations (NFE). Though recent works have accelerated the sampling to around 20 steps with high-order solvers, the sample quality with less than 10 NFE can still be improved. In this paper, we propose a unified sampling framework (USF) to study the optional strategies for solver. Under this framework, we further reveal that taking different solving strategies at different timesteps may help further decrease the truncation error, and a carefully designed \emph{solver schedule} has the potential to improve the sample quality by a large margin. Therefore, we propose a new sampling framework based on the exponential integral formulation that allows free choices of solver strategy at each step and design specific decisions for the framework. Moreover, we propose $S^3$, a predictor-based search method that automatically optimizes the solver schedule to get a better time-quality trade-off of sampling. We demonstrate that $S^3$ can find outstanding solver schedules which outperform the state-of-the-art sampling methods on CIFAR-10, CelebA, ImageNet, and LSUN-Bedroom datasets. Specifically, we achieve 2.69 FID with 10 NFE and 6.86 FID with 5 NFE on CIFAR-10 dataset, outperforming the SOTA method significantly. We further apply $S^3$ to Stable-Diffusion model and get an acceleration ratio of 2$\times$, showing the feasibility of sampling in very few steps without retraining the neural network.

Enshu Liu, Xuefei Ning, Yu Wang et al. · microsoft-research

Autoregressive (AR) models have achieved state-of-the-art performance in text and image generation but suffer from slow generation due to the token-by-token process. We ask an ambitious question: can a pre-trained AR model be adapted to generate outputs in just one or two steps? If successful, this would significantly advance the development and deployment of AR models. We notice that existing works that try to speed up AR generation by generating multiple tokens at once fundamentally cannot capture the output distribution due to the conditional dependencies between tokens, limiting their effectiveness for few-step generation. To address this, we propose Distilled Decoding (DD), which uses flow matching to create a deterministic mapping from Gaussian distribution to the output distribution of the pre-trained AR model. We then train a network to distill this mapping, enabling few-step generation. DD doesn't need the training data of the original AR model, making it more practical. We evaluate DD on state-of-the-art image AR models and present promising results on ImageNet-256. For VAR, which requires 10-step generation, DD enables one-step generation (6.3$\times$ speed-up), with an acceptable increase in FID from 4.19 to 9.96. For LlamaGen, DD reduces generation from 256 steps to 1, achieving an 217.8$\times$ speed-up with a comparable FID increase from 4.11 to 11.35. In both cases, baseline methods completely fail with FID>100. DD also excels on text-to-image generation, reducing the generation from 256 steps to 2 for LlamaGen with minimal FID increase from 25.70 to 28.95. As the first work to demonstrate the possibility of one-step generation for image AR models, DD challenges the prevailing notion that AR models are inherently slow, and opens up new opportunities for efficient AR generation. The project website is at https://imagination-research.github.io/distilled-decoding.

Rui Xie, Tianchen Zhao, Zhihang Yuan et al. · tsinghua

Visual Autoregressive (VAR) has emerged as a promising approach in image generation, offering competitive potential and performance comparable to diffusion-based models. However, current AR-based visual generation models require substantial computational resources, limiting their applicability on resource-constrained devices. To address this issue, we conducted analysis and identified significant redundancy in three dimensions of the VAR model: (1) the attention map, (2) the attention outputs when using classifier free guidance, and (3) the data precision. Correspondingly, we proposed efficient attention mechanism and low-bit quantization method to enhance the efficiency of VAR models while maintaining performance. With negligible performance lost (less than 0.056 FID increase), we could achieve 85.2% reduction in attention computation, 50% reduction in overall memory and 1.5x latency reduction. To ensure deployment feasibility, we developed efficient training-free compression techniques and analyze the deployment feasibility and efficiency gain of each technique.

Kaiyi Huang, Yukun Huang, Xintao Wang et al.

AI-driven content creation has shown potential in film production. However, existing film generation systems struggle to implement cinematic principles and thus fail to generate professional-quality films, particularly lacking diverse camera language and cinematic rhythm. This results in templated visuals and unengaging narratives. To address this, we introduce FilMaster, an end-to-end AI system that integrates real-world cinematic principles for professional-grade film generation, yielding editable, industry-standard outputs. FilMaster is built on two key principles: (1) learning cinematography from extensive real-world film data and (2) emulating professional, audience-centric post-production workflows. Inspired by these principles, FilMaster incorporates two stages: a Reference-Guided Generation Stage which transforms user input to video clips, and a Generative Post-Production Stage which transforms raw footage into audiovisual outputs by orchestrating visual and auditory elements for cinematic rhythm. Our generation stage highlights a Multi-shot Synergized RAG Camera Language Design module to guide the AI in generating professional camera language by retrieving reference clips from a vast corpus of 440,000 film clips. Our post-production stage emulates professional workflows by designing an Audience-Centric Cinematic Rhythm Control module, including Rough Cut and Fine Cut processes informed by simulated audience feedback, for effective integration of audiovisual elements to achieve engaging content. The system is empowered by generative AI models like (M)LLMs and video generation models. Furthermore, we introduce FilmEval, a comprehensive benchmark for evaluating AI-generated films. Extensive experiments show FilMaster's superior performance in camera language design and cinematic rhythm control, advancing generative AI in professional filmmaking.

Kaiyi Huang, Yukun Huang, Xuefei Ning et al. · microsoft-research

Text-to-video generation models have shown significant progress in the recent years. However, they still struggle with generating complex dynamic scenes based on compositional text prompts, such as attribute binding for multiple objects, temporal dynamics associated with different objects, and interactions between objects. Our key motivation is that complex tasks can be decomposed into simpler ones, each handled by a role-specialized MLLM agent. Multiple agents can collaborate together to achieve collective intelligence for complex goals. We propose GenMAC, an iterative, multi-agent framework that enables compositional text-to-video generation. The collaborative workflow includes three stages: Design, Generation, and Redesign, with an iterative loop between the Generation and Redesign stages to progressively verify and refine the generated videos. The Redesign stage is the most challenging stage that aims to verify the generated videos, suggest corrections, and redesign the text prompts, frame-wise layouts, and guidance scales for the next iteration of generation. To avoid hallucination of a single MLLM agent, we decompose this stage to four sequentially-executed MLLM-based agents: verification agent, suggestion agent, correction agent, and output structuring agent. Furthermore, to tackle diverse scenarios of compositional text-to-video generation, we design a self-routing mechanism to adaptively select the proper correction agent from a collection of correction agents each specialized for one scenario. Extensive experiments demonstrate the effectiveness of GenMAC, achieving state-of-the art performance in compositional text-to-video generation.

Lei Ke, Haohang Xu, Xuefei Ning et al.

Diffusion models have achieved significant progress in both image and video generation while still suffering from huge computation costs. As an effective solution, flow matching aims to reflow the diffusion process of diffusion models into a straight line for a few-step and even one-step generation. However, in this paper, we suggest that the original training pipeline of flow matching is not optimal and introduce two techniques to improve it. Firstly, we introduce progressive reflow, which progressively reflows the diffusion models in local timesteps until the whole diffusion progresses, reducing the difficulty of flow matching. Second, we introduce aligned v-prediction, which highlights the importance of direction matching in flow matching over magnitude matching. Experimental results on SDv1.5 and SDXL demonstrate the effectiveness of our method, for example, conducting on SDv1.5 achieves an FID of 10.70 on MSCOCO2014 validation set with only 4 sampling steps, close to our teacher model (32 DDIM steps, FID = 10.05).

Wenjin Guo, Donglai Liu, Weiying Xie et al.

Neural network training is a memory- and compute-intensive task. Quantization, which enables low-bitwidth formats in training, can significantly mitigate the workload. To reduce quantization error, recent methods have developed new data formats and additional pre-processing operations on quantizers. However, it remains quite challenging to achieve high accuracy and efficiency simultaneously. In this paper, we explore sub-8-bit integer training from its essence of gradient descent optimization. Our integer training framework includes two components: ShiftQuant to realize accurate gradient estimation, and L1 normalization to smoothen the loss landscape. ShiftQuant attains performance that approaches the theoretical upper bound of group quantization. Furthermore, it liberates group quantization from inefficient memory rearrangement. The L1 normalization facilitates the implementation of fully quantized normalization layers with impressive convergence accuracy. Our method frees sub-8-bit integer training from pre-processing and supports general devices. This framework achieves negligible accuracy loss across various neural networks and tasks ($0.92\%$ on 4-bit ResNets, $0.61\%$ on 6-bit Transformers). The prototypical implementation of ShiftQuant achieves more than $1.85\times/15.3\%$ performance improvement on CPU/GPU compared to its FP16 counterparts, and $33.9\%$ resource consumption reduction on FPGA than the FP16 counterparts. The proposed fully-quantized L1 normalization layers achieve more than $35.54\%$ improvement in throughout on CPU compared to traditional L2 normalization layers. Moreover, theoretical analysis verifies the advancement of our method.

Anyuan Zhuo, Xuefei Ning, Ningyuan Li et al.

This work investigates the resilience of contemporary LLMs against frequent and structured character-level perturbations, specifically through the insertion of noisy characters after each input character. We introduce UCC-Inj, a practical method that inserts invisible Unicode control characters into text to discourage LLM misuse in scenarios such as online exam systems. Surprisingly, despite strong obfuscation that fragments tokenization and reduces the signal-to-noise ratio significantly, many LLMs still maintain notable performance. Through comprehensive evaluation across model-, problem-, and noise-related configurations, we examine the extent and mechanisms of this robustness, exploring both the handling of character-level tokenization and implicit versus explicit denoising mechanism hypotheses of character-level noises. We hope our findings on the low-level robustness of LLMs will shed light on the risks of their misuse and on the reliability of deploying LLMs across diverse applications.

Federico Marcuzzi, Xuefei Ning, Roy Schwartz et al.

This work presents a comprehensive evaluation of how quantization affects model bias, with particular attention to its impact on individual demographic subgroups. We focus on weight and activation quantization strategies and examine their effects across a broad range of bias types, including stereotypes, toxicity, sentiment, and fairness. We employ both probabilistic and generated text-based metrics across nine benchmarks and evaluate models varying in architecture family and reasoning ability. Our findings show that quantization has a nuanced impact on bias: while it can reduce model toxicity and does not significantly impact sentiment, it tends to slightly increase stereotypes and unfairness in generative tasks, especially under aggressive compression. These trends are generally consistent across demographic categories and model types, although their magnitude depends on the specific setting. Overall, our results highlight the importance of carefully balancing efficiency and ethical considerations when applying quantization in practice.

Zhihang Yuan, Hanling Zhang, Pu Lu et al.

Diffusion Transformers (DiT) excel at image and video generation but face computational challenges due to the quadratic complexity of self-attention operators. We propose DiTFastAttn, a post-training compression method to alleviate the computational bottleneck of DiT. We identify three key redundancies in the attention computation during DiT inference: (1) spatial redundancy, where many attention heads focus on local information; (2) temporal redundancy, with high similarity between the attention outputs of neighboring steps; (3) conditional redundancy, where conditional and unconditional inferences exhibit significant similarity. We propose three techniques to reduce these redundancies: (1) Window Attention with Residual Sharing to reduce spatial redundancy; (2) Attention Sharing across Timesteps to exploit the similarity between steps; (3) Attention Sharing across CFG to skip redundant computations during conditional generation. We apply DiTFastAttn to DiT, PixArt-Sigma for image generation tasks, and OpenSora for video generation tasks. Our results show that for image generation, our method reduces up to 76% of the attention FLOPs and achieves up to 1.8x end-to-end speedup at high-resolution (2k x 2k) generation.

Tianchen Zhao, Tongcheng Fang, Haofeng Huang et al.

Diffusion transformers have demonstrated remarkable performance in visual generation tasks, such as generating realistic images or videos based on textual instructions. However, larger model sizes and multi-frame processing for video generation lead to increased computational and memory costs, posing challenges for practical deployment on edge devices. Post-Training Quantization (PTQ) is an effective method for reducing memory costs and computational complexity. When quantizing diffusion transformers, we find that existing quantization methods face challenges when applied to text-to-image and video tasks. To address these challenges, we begin by systematically analyzing the source of quantization error and conclude with the unique challenges posed by DiT quantization. Accordingly, we design an improved quantization scheme: ViDiT-Q (Video & Image Diffusion Transformer Quantization), tailored specifically for DiT models. We validate the effectiveness of ViDiT-Q across a variety of text-to-image and video models, achieving W8A8 and W4A8 with negligible degradation in visual quality and metrics. Additionally, we implement efficient GPU kernels to achieve practical 2-2.5x memory saving and a 1.4-1.7x end-to-end latency speedup.

Weilin Liu, Ye Mu, Chao Yu et al.

Discovering hazardous scenarios is crucial in testing and further improving driving policies. However, conducting efficient driving policy testing faces two key challenges. On the one hand, the probability of naturally encountering hazardous scenarios is low when testing a well-trained autonomous driving strategy. Thus, discovering these scenarios by purely real-world road testing is extremely costly. On the other hand, a proper determination of accident responsibility is necessary for this task. Collecting scenarios with wrong-attributed responsibilities will lead to an overly conservative autonomous driving strategy. To be more specific, we aim to discover hazardous scenarios that are autonomous-vehicle responsible (AV-responsible), i.e., the vulnerabilities of the under-test driving policy. To this end, this work proposes a Safety Test framework by finding Av-Responsible Scenarios (STARS) based on multi-agent reinforcement learning. STARS guides other traffic participants to produce Av-Responsible Scenarios and make the under-test driving policy misbehave via introducing Hazard Arbitration Reward (HAR). HAR enables our framework to discover diverse, complex, and AV-responsible hazardous scenarios. Experimental results against four different driving policies in three environments demonstrate that STARS can effectively discover AV-responsible hazardous scenarios. These scenarios indeed correspond to the vulnerabilities of the under-test driving policies, thus are meaningful for their further improvements.

Yi Cai, Xuefei Ning, Huazhong Yang et al.

Adversarial attacks have rendered high security risks on modern deep learning systems. Adversarial training can significantly enhance the robustness of neural network models by suppressing the non-robust features. However, the models often suffer from significant accuracy loss on clean data. Ensemble training methods have emerged as promising solutions for defending against adversarial attacks by diversifying the vulnerabilities among the sub-models, simultaneously maintaining comparable accuracy as standard training. However, existing ensemble methods are with poor scalability, owing to the rapid complexity increase when including more sub-models in the ensemble. Moreover, in real-world applications, it is difficult to deploy an ensemble with multiple sub-models, owing to the tight hardware resource budget and latency requirement. In this work, we propose ensemble-in-one (EIO), a simple but efficient way to train an ensemble within one random gated network (RGN). EIO augments the original model by replacing the parameterized layers with multi-path random gated blocks (RGBs) to construct a RGN. By diversifying the vulnerability of the numerous paths within the RGN, better robustness can be achieved. It provides high scalability because the paths within an EIO network exponentially increase with the network depth. Our experiments demonstrate that EIO consistently outperforms previous ensemble training methods with even less computational overhead.

Minxue Tang, Xuefei Ning, Yitu Wang et al.

Client-wise data heterogeneity is one of the major issues that hinder effective training in federated learning (FL). Since the data distribution on each client may vary dramatically, the client selection strategy can significantly influence the convergence rate of the FL process. Active client selection strategies are popularly proposed in recent studies. However, they neglect the loss correlations between the clients and achieve only marginal improvement compared to the uniform selection strategy. In this work, we propose FedCor -- an FL framework built on a correlation-based client selection strategy, to boost the convergence rate of FL. Specifically, we first model the loss correlations between the clients with a Gaussian Process (GP). Based on the GP model, we derive a client selection strategy with a significant reduction of expected global loss in each round. Besides, we develop an efficient GP training method with a low communication overhead in the FL scenario by utilizing the covariance stationarity. Our experimental results show that compared to the state-of-the-art method, FedCorr can improve the convergence rates by $34\%\sim 99\%$ and $26\%\sim 51\%$ on FMNIST and CIFAR-10, respectively.

Guyue Huang, Jingbo Hu, Yifan He et al.

With the down-scaling of CMOS technology, the design complexity of very large-scale integrated (VLSI) is increasing. Although the application of machine learning (ML) techniques in electronic design automation (EDA) can trace its history back to the 90s, the recent breakthrough of ML and the increasing complexity of EDA tasks have aroused more interests in incorporating ML to solve EDA tasks. In this paper, we present a comprehensive review of existing ML for EDA studies, organized following the EDA hierarchy.