Shengkun Tang

Shengkun Tang, Yaqing Wang, Caiwen Ding et al. · baidu

Diffusion models achieve great success in generating diverse and high-fidelity images, yet their widespread application, especially in real-time scenarios, is hampered by their inherently slow generation speed. The slow generation stems from the necessity of multi-step network inference. While some certain predictions benefit from the full computation of the model in each sampling iteration, not every iteration requires the same amount of computation, potentially leading to inefficient computation. Unlike typical adaptive computation challenges that deal with single-step generation problems, diffusion processes with a multi-step generation need to dynamically adjust their computational resource allocation based on the ongoing assessment of each step's importance to the final image output, presenting a unique set of challenges. In this work, we propose AdaDiff, an adaptive framework that dynamically allocates computation resources in each sampling step to improve the generation efficiency of diffusion models. To assess the effects of changes in computational effort on image quality, we present a timestep-aware uncertainty estimation module (UEM). Integrated at each intermediate layer, the UEM evaluates the predictive uncertainty. This uncertainty measurement serves as an indicator for determining whether to terminate the inference process. Additionally, we introduce an uncertainty-aware layer-wise loss aimed at bridging the performance gap between full models and their adaptive counterparts.

Shengkun Tang, Yaqing Wang, Zhenglun Kong et al. · baidu, harvard

Large-scale Transformer models bring significant improvements for various downstream vision language tasks with a unified architecture. The performance improvements come with increasing model size, resulting in slow inference speed and increased cost for severing. While some certain predictions benefit from the full complexity of the large-scale model, not all of inputs need the same amount of computation to conduct, potentially leading to computation resource waste. To handle this challenge, early exiting is proposed to adaptively allocate computational power in term of input complexity to improve inference efficiency. The existing early exiting strategies usually adopt output confidence based on intermediate layers as a proxy of input complexity to incur the decision of skipping following layers. However, such strategies cannot apply to encoder in the widely-used unified architecture with both encoder and decoder due to difficulty of output confidence estimation in the encoder. It is suboptimal in term of saving computation power to ignore the early exiting in encoder component. To handle this challenge, we propose a novel early exiting strategy for unified visual language models, which allows dynamically skip the layers in encoder and decoder simultaneously in term of input layer-wise similarities with multiple times of early exiting, namely \textbf{MuE}. By decomposing the image and text modalities in the encoder, MuE is flexible and can skip different layers in term of modalities, advancing the inference efficiency while minimizing performance drop. Experiments on the SNLI-VE and MS COCO datasets show that the proposed approach MuE can reduce expected inference time by up to 50\% and 40\% while maintaining 99\% and 96\% performance respectively.

Aidar Myrzakhan, Tianyi Li, Bowei Guo et al.

Diffusion Language Models (DLMs) incur high inference cost due to iterative denoising, motivating efficient pruning. Existing pruning heuristics largely inherited from autoregressive (AR) LLMs, typically preserve attention sink tokens because AR sinks serve as stable global anchors. We show that this assumption does not hold for DLMs: the attention-sink position exhibits substantially higher variance over the full generation trajectory (measured by how the dominant sink locations shift across timesteps), indicating that sinks are often transient and less structurally essential than in AR models. Based on this observation, we propose ${\bf \texttt{Sink-Aware Pruning}}$, which automatically identifies and prunes unstable sinks in DLMs (prior studies usually keep sinks for AR LLMs). Without retraining, our method achieves a better quality-efficiency trade-off and outperforms strong prior pruning baselines under matched compute. Our code is available at https://github.com/VILA-Lab/Sink-Aware-Pruning.

Shengkun Tang, Oliver Sieberling, Eldar Kurtic et al.

Large Language Models (LLMs) have achieved significant success across various NLP tasks. However, their massive computational costs limit their widespread use, particularly in real-time applications. Structured pruning offers an effective solution by compressing models and directly providing end-to-end speed improvements, regardless of the hardware environment. Meanwhile, different components of the model exhibit varying sensitivities towards pruning, calling for non-uniform model compression. However, a pruning method should not only identify a capable substructure, but also account for post-compression training. To this end, we propose DarwinLM, a method for training-aware structured pruning. DarwinLM builds upon an evolutionary search process, generating multiple offspring models in each generation through mutation, and selecting the fittest for survival. To assess the effect of post-training, we incorporate a lightweight, multistep training process within the offspring population, progressively increasing the number of tokens and eliminating poorly performing models in each selection stage. We validate our method through extensive experiments on Llama-2-7B, Llama-3.1-8B and Qwen-2.5-14B-Instruct, achieving state-of-the-art performance for structured pruning. For instance, DarwinLM surpasses ShearedLlama while requiring 5x less training data during post-compression training. Code is at: https://github.com/IST-DASLab/DarwinLM

Shengkun Tang, Liqun Ma, Haonan Li et al.

The typical Selective State-Space Model (SSM) used in Mamba addresses several limitations of Transformers, such as the quadratic computational complexity with respect to sequence length and the significant memory requirements during inference due to the key-value (KV) cache. However, the increasing size of Mamba models continues to pose challenges for training and deployment, particularly due to their substantial computational demands during both training and inference. In this work, we introduce $\texttt{Bi-Mamba}$, a scalable and powerful 1-bit Mamba architecture designed to enable more efficient large language models (LLMs), with model sizes of 780M, 1.3B, and 2.7B parameters. $\texttt{Bi-Mamba}$ models are trained from scratch on a standard LLM-scale dataset using an autoregressive distillation loss. Extensive experiments on language modeling benchmarks demonstrate that $\texttt{Bi-Mamba}$ achieves performance comparable to its full-precision (FP16 or BF16) counterparts, while outperforming post-training binarization (PTB) Mamba and binarization-aware training (BAT) Transformer baselines. Moreover, $\texttt{Bi-Mamba}$ drastically reduces memory usage and computational cost compared to the original Mamba. Our work pioneers a new line of linear-complexity LLMs under low-bit representation and provides the way for the design of specialized hardware optimized for efficient 1-bit Mamba-based models. Code and the pre-trained weights are available at https://github.com/Tangshengku/Bi-Mamba.

Zongfang Liu, Shengkun Tang, Yifan Shen et al.

Mixture-of-Experts (MoE) language models increase parameter capacity without proportional per-token compute, but the deployment still requires storing all experts, making expert pruning important for reducing memory and serving overhead. Existing task-agnostic expert pruning methods are typically calibration-dependent: they estimate expert importance from routing or activation statistics on a calibration set, which makes pruning outcomes sensitive to the choice of calibration set and adds substantial preprocessing cost. We introduce AIMER (\textbf{A}bsolute mean over root mean square \textbf{IM}portance for \textbf{E}xpert \textbf{R}anking), a simple calibration-free criterion that yields clear within-layer score separation and distinct expert stratification. Across 7B to 30B MoE language models at 25\% and 50\% pruning ratios over 16 benchmarks, AIMER consistently delivers competitive or stronger overall performance against state-of-the-art calibration-based expert pruning baselines with only 0.22--1.27 seconds for scoring the experts.

Sondos Mahmoud Bsharat, Mukul Ranjan, Aidar Myrzakhan et al.

Rapid advancements in large language models (LLMs) have increased interest in deploying them on mobile devices for on-device AI applications. Mobile users interact differently with LLMs compared to desktop users, creating unique expectations and data biases. Current benchmark datasets primarily target at server and desktop environments, and there is a notable lack of extensive datasets specifically designed for mobile contexts. Additionally, mobile devices face strict limitations in storage and computing resources, constraining model size and capabilities, thus requiring optimized efficiency and prioritized knowledge. To address these challenges, we introduce Mobile-MMLU, a large-scale benchmark dataset tailored for mobile intelligence. It consists of 16,186 questions across 80 mobile-related fields, designed to evaluate LLM performance in realistic mobile scenarios. A challenging subset, Mobile-MMLU-Pro, provides advanced evaluation similar in size to MMLU-Pro but significantly more difficult than our standard full set. Both benchmarks use multiple-choice, order-invariant questions focused on practical mobile interactions, such as recipe suggestions, travel planning, and essential daily tasks. The dataset emphasizes critical mobile-specific metrics like inference latency, energy consumption, memory usage, and response quality, offering comprehensive insights into model performance under mobile constraints. Moreover, it prioritizes privacy and adaptability, assessing models' ability to perform on-device processing, maintain user privacy, and adapt to personalized usage patterns. Mobile-MMLU family offers a standardized framework for developing and comparing mobile-optimized LLMs, enabling advancements in productivity and decision-making within mobile computing environments. Our code and data are available at: https://github.com/VILA-Lab/Mobile-MMLU.

Shengkun Tang, Zekun Wang, Bo Zheng et al.

Structured pruning and knowledge distillation (KD) are typical techniques for compressing large language models, but it remains unclear how they should be applied at pretraining scale, especially to recent mixture-of-experts (MoE) models. In this work, we systematically study MoE compression in large-scale pretraining, focusing on three key questions: whether pruning provides a better initialization than training from scratch, how expert compression choices affect the final model after continued training, and which training strategy is most effective. We have the following findings: First, across depth, width, and expert compression, pruning a pretrained MoE consistently outperforms training the target architecture from scratch under the same training budget. Second, different one-shot expert compression methods converge to similar final performance after large-scale continual pretraining. Motivated by this, we introduce a simple partial-preservation expert merging strategy that improves downstream performance across most benchmarks. Third, combining KD with the language modeling loss outperforms KD alone, particularly on knowledge-intensive tasks. We further propose multi-token prediction (MTP) distillation, which yields consistent gains. Finally, given the same training tokens, progressive pruning schedules outperform one-shot compression, suggesting that gradual architecture transitions lead to better optimization trajectories. Putting it all together, we compress Qwen3-Next-80A3B to a 23A2B model that retains competitive performance. These results offer practical guidance for efficient MoE compression at scale.

Puyuan Yi, Shengkun Tang, Jian Yao

To obtain high-resolution depth maps, some previous learning-based multi-view stereo methods build a cost volume pyramid in a coarse-to-fine manner. These approaches leverage fixed depth range hypotheses to construct cascaded plane sweep volumes. However, it is inappropriate to set identical range hypotheses for each pixel since the uncertainties of previous per-pixel depth predictions are spatially varying. Distinct from these approaches, we propose a Dynamic Depth Range Network (DDR-Net) to determine the depth range hypotheses dynamically by applying a range estimation module (REM) to learn the uncertainties of range hypotheses in the former stages. Specifically, in our DDR-Net, we first build an initial depth map at the coarsest resolution of an image across the entire depth range. Then the range estimation module (REM) leverages the probability distribution information of the initial depth to estimate the depth range hypotheses dynamically for the following stages. Moreover, we develop a novel loss strategy, which utilizes learned dynamic depth ranges to generate refined depth maps, to keep the ground truth value of each pixel covered in the range hypotheses of the next stage. Extensive experimental results show that our method achieves superior performance over other state-of-the-art methods on the DTU benchmark and obtains comparable results on the Tanks and Temples benchmark. The code is available at https://github.com/Tangshengku/DDR-Net.

Jiacheng Liu, Shengkun Tang, Jiacheng Cui et al.

Acceleration methods for diffusion models (e.g., token merging or downsampling) typically optimize synthesis quality under reduced compute, yet often ignore discriminative capacity. We revisit token compression with a joint objective and present BiGain, a training-free, plug-and-play framework that preserves generation quality while improving classification in accelerated diffusion models. Our key insight is frequency separation: mapping feature-space signals into a frequency-aware representation disentangles fine detail from global semantics, enabling compression that respects both generative fidelity and discriminative utility. BiGain reflects this principle with two frequency-aware operators: (1) Laplacian-gated token merging, which encourages merges among spectrally smooth tokens while discouraging merges of high-contrast tokens, thereby retaining edges and textures; and (2) Interpolate-Extrapolate KV Downsampling, which downsamples keys/values via a controllable interextrapolation between nearest and average pooling while keeping queries intact, thereby conserving attention precision. Across DiT- and U-Net-based backbones and ImageNet-1K, ImageNet-100, Oxford-IIIT Pets, and COCO-2017, our operators consistently improve the speed-accuracy trade-off for diffusion-based classification, while maintaining or enhancing generation quality under comparable acceleration. For instance, on ImageNet-1K, with 70% token merging on Stable Diffusion 2.0, BiGain increases classification accuracy by 7.15% while improving FID by 0.34 (1.85%). Our analyses indicate that balanced spectral retention, preserving high-frequency detail and low/mid-frequency semantics, is a reliable design rule for token compression in diffusion models. To our knowledge, BiGain is the first framework to jointly study and advance both generation and classification under accelerated diffusion, supporting lower-cost deployment.

Zongfang Liu, Shengkun Tang, Boyang Sun et al.

Sparse Mixture-of-Experts (SMoE) language models achieve strong capability at low per-token compute, yet deployment remains memory- and throughput-bound because the full expert pool must be stored and served. Post-training expert pruning reduces this cost, but most methods focus on which experts to prune within each layer and default to a uniform layer-wise sparsity allocation, even though the allocation can strongly affect performance. We decouple pruning into within-layer expert ranking and across-layer budget allocation, and introduce \textbf{E}xpected \textbf{S}peculative \textbf{A}cceptance \textbf{P}roxy (\textbf{ESAP}), a speculative-decoding-inspired, teacher-forced metric that measures how well a pruned model matches the full model. ESAP is bounded and stable, enabling cheap comparison of many candidates without costly autoregressive decoding. Building on ESAP, we propose EvoESAP, an evolutionary searching framework that optimizes a non-uniform layer-wise sparsity allocation under a fixed global budget while holding the within-layer pruning order fixed, making it a plug-and-play method with criteria such as Frequency, EAN, SEER, and REAP. Across 7B--30B SMoE LLMs at 25\% and 50\% sparsity, EvoESAP consistently discovers non-uniform allocations that improve open-ended generation (up to \textbf{+19.6\%} on MATH-500 at 50\% sparsity) while preserving competitive multiple-choice accuracy compared with uniform pruning at the same sparsity.

Bowei Guo, Shengkun Tang, Cong Zeng et al.

Diffusion models are renowned for their generative capabilities, yet their pretraining processes exhibit distinct phases of learning speed that have been entirely overlooked in prior post-training acceleration efforts in the community. In this study, we introduce a novel framework called MosaicDiff that aligns diffusion pretraining dynamics with post-training sampling acceleration via trajectory-aware structural pruning. Our approach leverages the observation that the middle, fast-learning stage of diffusion pretraining requires more conservative pruning to preserve critical model features, while the early and later, slow-learning stages benefit from a more aggressive pruning strategy. This adaptive pruning mechanism is the first to explicitly mirror the inherent learning speed variations of diffusion pretraining, thereby harmonizing the model's inner training dynamics with its accelerated sampling process. Extensive experiments on DiT and SDXL demonstrate that our method achieves significant speed-ups in sampling without compromising output quality, outperforming previous state-of-the-art methods by large margins, also providing a new viewpoint for more efficient and robust training-free diffusion acceleration.

Cong Zeng, Shengkun Tang, Yuanzhou Chen et al.

The rapid advancement of large language models (LLMs) such as ChatGPT, DeepSeek, and Claude has significantly increased the presence of AI-generated text in digital communication. This trend has heightened the need for reliable detection methods to distinguish between human-authored and machine-generated content. Existing approaches both zero-shot methods and supervised classifiers largely conceptualize this task as a binary classification problem, often leading to poor generalization across domains and models. In this paper, we argue that such a binary formulation fundamentally mischaracterizes the detection task by assuming a coherent representation of human-written texts. In reality, human texts do not constitute a unified distribution, and their diversity cannot be effectively captured through limited sampling. This causes previous classifiers to memorize observed OOD characteristics rather than learn the essence of `non-ID' behavior, limiting generalization to unseen human-authored inputs. Based on this observation, we propose reframing the detection task as an out-of-distribution (OOD) detection problem, treating human-written texts as distributional outliers while machine-generated texts are in-distribution (ID) samples. To this end, we develop a detection framework using one-class learning method including DeepSVDD and HRN, and score-based learning techniques such as energy-based method, enabling robust and generalizable performance. Extensive experiments across multiple datasets validate the effectiveness of our OOD-based approach. Specifically, the OOD-based method achieves 98.3% AUROC and AUPR with only 8.9% FPR95 on DeepFake dataset. Moreover, we test our detection framework on multilingual, attacked, and unseen-model and -domain text settings, demonstrating the robustness and generalizability of our framework. Code, pretrained weights, and demo will be released.

Zongfang Liu, Shengkun Tang, Zongliang Wu et al.

Diffusion models have achieved remarkable success in high-fidelity image generation but remain computationally demanding due to their multi-step denoising process and large model sizes. Although prior work improves efficiency either by reducing sampling steps or by compressing model parameters, existing structured pruning approaches still struggle to balance real acceleration and image quality preservation. In particular, prior methods such as MosaicDiff rely on heuristic, manually tuned stage-wise sparsity schedules and stitch multiple independently pruned models during inference, which increases memory overhead. However, the importance of diffusion steps is highly non-uniform and model-dependent. As a result, schedules derived from simple heuristics or empirical observations often fail to generalize and may lead to suboptimal performance. To this end, we introduce \textbf{Diff-ES}, a stage-wise structural \textbf{Diff}usion pruning framework via \textbf{E}volutionary \textbf{S}earch, which optimizes the stage-wise sparsity schedule and executes it through memory-efficient weight routing without model duplication. Diff-ES divides the diffusion trajectory into multiple stages, automatically discovers an optimal stage-wise sparsity schedule via evolutionary search, and activates stage-conditioned weights dynamically without duplicating model parameters. Our framework naturally integrates with existing structured pruning methods for diffusion models including depth and width pruning. Extensive experiments on DiT and SDXL demonstrate that Diff-ES consistently achieves wall-clock speedups while incurring minimal degradation in generation quality, establishing state-of-the-art performance for structured diffusion model pruning.

Cong Zeng, Shengkun Tang, Xianjun Yang et al.

The advent of Large Language Models (LLMs) has revolutionized text generation, producing outputs that closely mimic human writing. This blurring of lines between machine- and human-written text presents new challenges in distinguishing one from the other a task further complicated by the frequent updates and closed nature of leading proprietary LLMs. Traditional logits-based detection methods leverage surrogate models for identifying LLM-generated content when the exact logits are unavailable from black-box LLMs. However, these methods grapple with the misalignment between the distributions of the surrogate and the often undisclosed target models, leading to performance degradation, particularly with the introduction of new, closed-source models. Furthermore, while current methodologies are generally effective when the source model is identified, they falter in scenarios where the model version remains unknown, or the test set comprises outputs from various source models. To address these limitations, we present Distribution-Aligned LLMs Detection (DALD), an innovative framework that redefines the state-of-the-art performance in black-box text detection even without logits from source LLMs. DALD is designed to align the surrogate model's distribution with that of unknown target LLMs, ensuring enhanced detection capability and resilience against rapid model iterations with minimal training investment. By leveraging corpus samples from publicly accessible outputs of advanced models such as ChatGPT, GPT-4 and Claude-3, DALD fine-tunes surrogate models to synchronize with unknown source model distributions effectively.