Searcharxiv⌕ Search

arXiv subjects

Shiwei Liu

Publications and source records attributed to Shiwei Liu.

At least 37 records · Page 2Linked to original sources
arXiv2026

AlphaQ: Calibration-Free Bit Allocation for Mixture-of-Experts Quantization

Mixture-of-Experts (MoE) architectures scale model capacity through sparse expert activation, but their deployment remains memory-bound because all expert weights must reside in memory. Mixed-precision quantization can substantially reduce this footprint by assigning different bit-widths to different experts. Existing approaches, however, typically rely on calibration data to estimate expert importance and determine bit allocation. For frontier MoE LLMs, the original training data, and hence the true training distribution, is proprietary and inaccessible. As a result, calibration sets are inevitably imperfect surrogates, and this can misestimate expert utilization and lead to suboptimal bit allocation. Motivated by the substantial cross-expert quality variability observed in modern MoE models, and by the success of Heavy-Tailed Self-Regularization (HT-SR) theory at predicting neural network model quality without access to training or testing data, we propose AlphaQ, a calibration-free bit-allocation method for MoE quantization. AlphaQ draws on HT-SR theory and follows a simple principle: experts with more heavy-tailed weight spectra are typically better trained and hence should receive higher bit-widths, while experts with weaker heavy-tailed structure can be quantized more aggressively. AlphaQ operationalizes this principle by measuring expert-wise spectral heavy-tailedness and solving a budget-constrained optimization problem that minimizes total quantization error under a global bit-budget constraint. Across several MoE models, AlphaQ consistently outperforms calibration-based baselines under matched bit budgets. Notably, on Qwen1.5-MoE, AlphaQ achieves near full-precision accuracy with an average expert precision of only 3.5 bits, while delivering more than 4$\times$ memory compression. Our code is available at https://github.com/Superone77/AlphaQ.

cs.LG↗
arXiv2026

Leave it to the Specialist: Repair Sparse LLMs with Sparse Fine-Tuning via Sparsity Evolution

Sparse large language models (LLMs) offer an attractive direction toward efficient deployment, but adapting them to downstream tasks remains challenging. The central difficulty is to enable effective task adaptation without sacrificing the efficiency advantages of sparsity. Existing fine-tuning methods are not well-suited to this setting, as they either introduce additional dense parameters or assume a fixed sparse topology, limiting their compatibility with sparse LLMs. In this paper, we propose Sparsity Evolution Fine-Tuning (SEFT), a fine-tuning framework designed specifically for sparse LLMs. SEFT allows sparse structure to evolve during fine-tuning by periodically reallocating sparse task-specific updates and reactivating previously pruned weights when beneficial. At the same time, SEFT preserves the efficiency advantages of sparsity through topology adaptation based on parameter importance. Experiments on LLaMA, DeepSeek, and Mistral models across multiple benchmarks show that SEFT delivers stronger performance while offering superior memory and time efficiency compared to existing baselines. Our code is publicly available at: https://github.com/QiaoXiao7282/SEFT.

cs.AI↗
arXiv2026

One LR Doesn't Fit All: Heavy-Tail Guided Layerwise Learning Rates for LLMs

Learning rate configuration is a fundamental aspect of modern deep learning. The prevailing practice of applying a uniform learning rate across all layers overlooks the structural heterogeneity of Transformers, potentially limiting their effectiveness as the backbone of Large Language Models (LLMs). In this paper, we introduce Layerwise Learning Rate (LLR), an adaptive scheme that assigns distinct learning rates to individual Transformer layers. Our method is grounded in Heavy-Tailed Self-Regularization (HT-SR) theory, which characterizes the empirical spectral density (ESD) of weight correlation matrices to quantify heavy-tailedness. Layers with weaker heavy-tailedness are assigned larger learning rates to accelerate training, while layers with stronger heavy-tailedness receive smaller learning rates. By tailoring learning rates in this manner, LLR promotes more balanced training across layers, leading to faster convergence and improved generalization. Extensive experiments across architectures ranging from LLaMA to GPT-nano, optimizers including AdamW and Muon, and model scales from 60M to 3B parameters with up to 100B training tokens demonstrate the effectiveness of LLR. LLR achieves up to 1.5x training speedup and consistently outperforms uniform-learning-rate baselines. In particular, it improves the average zero-shot accuracy of 1B models from 47.09% to 49.02%, and that of 3B models from 48.58% to 50.61%. A key advantage of LLR is its low tuning overhead: it can transfer nearly optimal learning-rate settings directly from the uniform baseline. Code is available at https://github.com/hed-ucas/Layer-wise-Learning-Rate.

cs.LG↗
arXiv2026

Motion-Aware Caching for Efficient Autoregressive Video Generation

Autoregressive video generation paradigms offer theoretical promise for long video synthesis, yet their practical deployment is hindered by the computational burden of sequential iterative denoising. While cache reuse strategies can accelerate generation by skipping redundant denoising steps, existing methods rely on coarse-grained chunk-level skipping that fails to capture fine-grained pixel dynamics. This oversight is critical: pixels with high motion require more denoising steps to prevent error accumulation, while static pixels tolerate aggressive skipping. We formalize this insight theoretically by linking cache errors to residual instability, and propose MotionCache, a motion-aware cache framework that exploits inter-frame differences as a lightweight proxy for pixel-level motion characteristics. MotionCache employs a coarse-to-fine strategy: an initial warm-up phase establishes semantic coherence, followed by motion-weighted cache reuse that dynamically adjusts update frequencies per token. Extensive experiments on state-of-the-art models like SkyReels-V2 and MAGI-1 demonstrate that MotionCache achieves significant speedups of $\textbf{6.28}\times$ and $\textbf{1.64}\times$ respectively, while effectively preserving generation quality (VBench: $1\%\downarrow$ and $0.01\%\downarrow$ respectively). The code is available at https://github.com/ywlq/MotionCache.

cs.CV↗
arXiv2026

Layer Collapse in Diffusion Language Models

Diffusion language models (DLMs) have recently emerged as competitive alternatives to autoregressive (AR) language models, yet differences in their activation dynamics remain poorly understood. We characterize these dynamics in LLaDA-8B and identify a striking layer-collapse property: a few early layers exhibit highly similar, collapsed activation patterns dominated by a single large super-outlier persisting over a long token range. Despite its apparent redundancy, this outlier is critical: pruning it causes outputs to degrade into repetitive random token loops. Paradoxically, layers in LLaDA contain more redundant representations overall, with redundancy most pronounced in earlier layers -- the reverse of AR models, where deeper layers grow redundant due to undertraining. Our analysis indicates that layer collapse in DLMs is not driven by undertraining but by overtraining: a dominant outlier becomes an indispensable information carrier while remaining representations collapse into redundant structure. These findings have strong practical implications, verified through controlled pre-training experiments. DLMs are surprisingly robust to compression: LLaDA under 3-bit GPTQ quantization drops only -1.8% on GSM8K, whereas Llama-3.1-8B drops -64.7%. Optimal sparsity allocation also reverses between families: at 50% average sparsity, allocating more to early layers in LLaDA yields +8.4% over the reverse strategy, while the same allocation costs Llama -8.4%. Our findings reveal that the DLM training objective fundamentally reshapes layer dynamics relative to AR models, with direct consequences for compression and deployment. Code: github.com/Conzel/super-outlier-dlm.

cs.LG↗
arXiv2026

ELAS: Efficient Pre-Training of Low-Rank Large Language Models via 2:4 Activation Sparsity

Large Language Models (LLMs) have achieved remarkable capabilities, but their immense computational demands during training remain a critical bottleneck for widespread adoption. Low-rank training has received attention in recent years due to its ability to significantly reduce training memory usage. Meanwhile, applying 2:4 structured sparsity to weights and activations to leverage NVIDIA GPU support for 2:4 structured sparse format has become a promising direction. However, existing low-rank methods often leave activation matrices in full-rank, which dominates memory consumption and limits throughput during large-batch training. Furthermore, directly applying sparsity to weights often leads to non-negligible performance degradation. To achieve efficient pre-training of LLMs, this paper proposes ELAS: Efficient pre-training of Low-rank LLMs via 2:4 Activation Sparsity, a novel framework for low-rank models via 2:4 activation sparsity. ELAS applies squared ReLU activation functions to the feed-forward networks in low-rank models and implements 2:4 structured sparsity on the activations after the squared ReLU operation. We evaluated ELAS through pre-training experiments on LLaMA models ranging from 60M to 1B parameters. The results demonstrate that ELAS maintains performance with minimal degradation after applying 2:4 activation sparsity, while achieving training and inference acceleration. Moreover, ELAS reduces activation memory overhead, particularly with large batch sizes. Code is available at ELAS Repo.

cs.LG↗
arXiv2026

Junk DNA Hypothesis: Pruning Small Pre-Trained Weights Irreversibly and Monotonically Impairs "Difficult" Downstream Tasks in LLMs

We present Junk DNA Hypothesis by adopting a novel task-centric angle for the pre-trained weights of large language models (LLMs). It has been believed that weights in LLMs contain significant redundancy, leading to the conception that a considerable chunk of the parameters can be removed by pruning without compromising performance. Contrary to this belief, this paper presents a counter-argument: small-magnitude weights of pre-trained model weights encode vital knowledge essential for tackling difficult downstream tasks - manifested as the monotonic relationship between the performance drop of downstream tasks across the difficulty spectrum, as we prune more pre-trained weights by magnitude. Moreover, we reveal that these seemingly inconsequential weights can result in irreparable loss of knowledge and performance degradation in difficult tasks, even when downstream continual training is allowed. Interestingly, our evaluations show that the other popular compression, namely quantization, fails to exhibit similar monotonic effect and does not as convincingly disentangle this task-difficulty information. To study formally, we introduce several quantifiable metrics to gauge the downstream task difficulty: (1) within the same task category, and (2) across different task categories. Our extensive experiments substantiate the Junk DNA Hypothesis across a diverse range of model sizes, tasks, datasets, and even pruning methods. Codes are available at: https://github.com/VITA-Group/Junk_DNA_Hypothesis.git.

cs.LG↗
arXiv2026

Condense, Don't Just Prune: Enhancing Efficiency and Performance in MoE Layer Pruning

Mixture-of-Experts (MoE) has garnered significant attention for its ability to scale up neural networks while utilizing the same or even fewer active parameters. However, MoE does not alleviate the massive memory requirements of networks, which limits their practicality in real-world applications, especially in the era of large language models (LLMs). While recent work explores the possibility of removing entire layers of MoE to reduce memory, the performance degradation is still notable. In this paper, we propose ConDense-MoE (CD-MoE), which, instead of dropping the entire MoE layer, condenses the large, sparse MoE layer into a smaller, denser layer with only a few experts activated for all tokens, while maintaining hardware friendliness. Our approach is specifically designed for fine-grained MoE with shared experts, where Feed-Forward Networks are split into many small experts, with certain experts isolated to serve as shared experts that are always activated, such as DeepSeekMoE and QwenMoE. We demonstrate the effectiveness of our method. Specifically, for the DeepSeekMoE-16B model, our approach maintains 90% of the average accuracy while reducing memory usage by 27.5% and increasing inference speed by 1.26 times. Moreover, we show that by applying lightweight expert fine-tuning -- only to the condensed layers -- and using 5 hours on a single 80G A100 GPU, we can successfully recover 98% of the original performance. Our code is available at: https://github.com/duterscmy/CD-MoE/tree/main.

cs.LG↗
arXiv2026

LIFT the Veil for the Truth: Principal Weights Emerge after Rank Reduction for Reasoning-Focused Supervised Fine-Tuning

Recent studies have shown that supervised fine-tuning of LLMs on a small number of high-quality datasets can yield strong reasoning capabilities. However, full fine-tuning (Full FT), while powerful, is computationally expensive and susceptible to overfitting and catastrophic forgetting, particularly when data is limited. Sparse fine-tuning, which previously achieved notable success by updating only a small subset of model parameters, offers a promising trade-off between efficiency and effectiveness. Yet, it has lagged behind in the LLM era due to the difficulty of identifying parameters truly critical for reasoning. In this work, we state that weights with the largest magnitude after low-rank approximation are critical weights for fine-tuning, which we call Principal Weights. Surprisingly, while magnitude-based sparse fine-tuning performs poorly as a baseline on LLM fine-tuning, it becomes highly effective after rank reduction. These insights motivate our method: Low-rank Informed Sparse Fine-Tuning (LIFT). LIFT only updates the top 5% Principal Weights throughout training and consistently achieves better performance on reasoning tasks than Full FT, while maintaining memory efficiency on par with popular parameter-efficient fine-tuning methods. In addition to strong performance on target domains such as arithmetic reasoning, LIFT also retains up to 20% more source-domain knowledge, compared to Full FT and LoRA. Our code is available at: https://github.com/zihanghliu/LIFT.

cs.LG↗
arXiv2026

Diffusion Language Models Know the Answer Before Decoding

Diffusion language models (DLMs) have recently emerged as an alternative to autoregressive approaches, offering parallel sequence generation and flexible token orders. However, their inference remains slower than that of autoregressive models, primarily due to the cost of bidirectional attention and the large number of refinement steps required for high quality outputs. In this work, we highlight and leverage an overlooked property of DLMs early answer convergence: in many cases, the correct answer can be internally identified by half steps before the final decoding step, both under semi-autoregressive and random remasking schedules. For example, on GSM8K and MMLU, up to 97% and 99% of instances, respectively, can be decoded correctly using only half of the refinement steps. Building on this observation, we introduce Prophet, a training-free fast decoding paradigm that enables early commit decoding. Specifically, Prophet dynamically decides whether to continue refinement or to go "all-in" (i.e., decode all remaining tokens in one step), using the confidence gap between the top-2 prediction candidates as the criterion. It integrates seamlessly into existing DLM implementations, incurs negligible overhead, and requires no additional training. Empirical evaluations of LLaDA-8B and Dream-7B across multiple tasks show that Prophet reduces the number of decoding steps by up to 3.4x while preserving high generation quality. These results recast DLM decoding as a problem of when to stop sampling, and demonstrate that early decode convergence provides a simple yet powerful mechanism for accelerating DLM inference, complementary to existing speedup techniques. Our code is publicly available at https://github.com/pixeli99/Prophet.

cs.CL↗
arXiv2026

Mask-Enhanced Autoregressive Prediction: Pay Less Attention to Learn More

Large Language Models (LLMs) are discovered to suffer from accurately retrieving key information. To address this, we propose Mask-Enhanced Autoregressive Prediction (MEAP), a simple yet effective training paradigm that seamlessly integrates Masked Language Modeling (MLM) into Next-Token Prediction (NTP) to enhance the latter's in-context retrieval capabilities. Specifically, MEAP first randomly masks a small fraction of input tokens and then directly performs the standard next-token prediction autoregressive using a decoder-only Transformer. MEAP eliminates the need for bidirectional attention or encoder-decoder architectures for MLM, incurring no additional computational overhead during pre-training or inference. Intensive experiments demonstrate that MEAP substantially outperforms NTP on key information retrieval and long-context reasoning tasks, while performing on par or better on commonsense reasoning tasks. The benefits of MEAP also extend to supervised fine-tuning, where it shows remarkable advantages in lost-in-the-middle scenarios, outperforming NTP by 11.77 percentage points. Our analysis indicates that MEAP's effectiveness arises from its ability to promote more distinguishable attention scores by concentrating on a reduced set of non-masked tokens. This mechanism improves the model's focus on task-relevant signals while mitigating the influence of peripheral context. These findings position MEAP as a promising training paradigm for large language models.

cs.CL↗
arXiv2026

Why Diffusion Language Models Struggle with Truly Parallel (Non-Autoregressive) Decoding?

Diffusion Language Models (DLMs) are often advertised as enabling parallel token generation, yet practical fast DLMs frequently converge to left-to-right, autoregressive (AR)-like decoding dynamics. In contrast, genuinely non-AR generation is promising because it removes AR's sequential bottleneck, better exploiting parallel hardware to reduce synchronization/communication overhead and improve latency scaling with output length. We argue that a primary driver of AR-like decoding is a mismatch between DLM objectives and the highly sequential structure of widely used training data, including standard pretraining corpora and long chain-of-thought (CoT) supervision. Motivated by this diagnosis, we propose NAP (Non-Autoregressive Parallel DLMs), a proof-of-concept, data-centric approach that better aligns supervision with non-AR parallel decoding. NAP curates examples as multiple independent reasoning trajectories and couples them with a parallel-forced decoding strategy that encourages multi-token parallel updates. Across math reasoning benchmarks, NAP yields stronger performance under parallel decoding than DLMs trained on standard long CoT data, with gains growing as parallelism increases. Our results suggest that revisiting data and supervision is a principled direction for mitigating AR-like behavior and moving toward genuinely non-autoregressive parallel generation in DLMs. Our code is available at https://github.com/pixeli99/NAP.

cs.CL↗
arXiv2026

Search or Accelerate: Confidence-Switched Position Beam Search for Diffusion Language Models

Diffusion Language Models (DLMs) generate text by iteratively denoising a masked sequence, repeatedly deciding which positions to commit at each step. Standard decoding follows a greedy rule: unmask the most confident positions, yet this local choice can lock the model into a suboptimal unmasking order, especially on reasoning-heavy prompts. We present SOAR, a training-free decoding algorithm that adapts its behavior to the model's uncertainty. When confidence is low, SOAR briefly widens the search over alternative unmasking decisions to avoid premature commitments; when confidence is high, it collapses the search and decodes many positions in parallel to reduce the number of denoising iterations. Across mathematical reasoning and code generation benchmarks (GSM8K, MBPP, HumanEval) on Dream-7B and LLaDA-8B, SOAR improves generation quality while maintaining competitive inference speed, offering a practical way to balance quality and efficiency in DLM decoding. Our Code is available at https://github.com/duterscmy/SOAR

cs.CL↗
arXiv2026

(PASS) Visual Prompt Locates Good Structure Sparsity through a Recurrent HyperNetwork

Large-scale neural networks have demonstrated remarkable performance in different domains like vision and language processing, although at the cost of massive computation resources. As illustrated by compression literature, structural model pruning is a prominent algorithm to encourage model efficiency, thanks to its acceleration-friendly sparsity patterns. One of the key questions of structural pruning is how to estimate the channel significance. In parallel, work on data-centric AI has shown that prompting-based techniques enable impressive generalization of large language models across diverse downstream tasks. In this paper, we investigate a charming possibility - \textit{leveraging visual prompts to capture the channel importance and derive high-quality structural sparsity}. To this end, we propose a novel algorithmic framework, namely \texttt{PASS}. It is a tailored hyper-network to take both visual prompts and network weight statistics as input, and output layer-wise channel sparsity in a recurrent manner. Such designs consider the intrinsic channel dependency between layers. Comprehensive experiments across multiple network architectures and six datasets demonstrate the superiority of \texttt{PASS} in locating good structural sparsity. For example, at the same FLOPs level, \texttt{PASS} subnetworks achieve $1\%\sim 3\%$ better accuracy on Food101 dataset; or with a similar performance of $80\%$ accuracy, \texttt{PASS} subnetworks obtain $0.35\times$ more speedup than the baselines.

cs.CV↗
arXiv2026

ActTail: Global Activation Sparsity in Large Language Models

Activation sparsity is a promising approach for accelerating large language model (LLM) inference by reducing computation and memory movement. However, existing activation sparsity methods typically apply uniform sparsity across projections, ignoring the heterogeneous statistical properties of Transformer weights and thereby amplifying performance degradation. In this paper, we propose ActTail, a TopK magnitude-based activation sparsity method with global activation sparsity allocation grounded in Heavy-Tailed Self-Regularization (HT-SR) theory. Specifically, we capture this heterogeneity via the heavy-tail exponent computed from each projection's empirical spectral density (ESD), which is used as a quantitative indicator to assign projection-specific sparsity budgets. Importantly, we provide a theoretical analysis that establishes an explicit relationship between the activation sparsity ratio and the heavy-tail exponent under the HT-SR regime, offering principled guidance for sparsity allocation beyond heuristic design. Experiments on LLaMA and Mistral models show that our method improves both perplexity and downstream task performance at high sparsity compared to uniform allocation. At 80% sparsity, perplexity is reduced by 21.8% on LLaMA-2-7B, 40.1% on LLaMA-2-13B, and 9.4% on Mistral-7B.

cs.CL↗
arXiv2026

Deep Ensembling with No Overhead for either Training or Testing: The All-Round Blessings of Dynamic Sparsity

The success of deep ensembles on improving predictive performance, uncertainty estimation, and out-of-distribution robustness has been extensively studied in the machine learning literature. Albeit the promising results, naively training multiple deep neural networks and combining their predictions at inference leads to prohibitive computational costs and memory requirements. Recently proposed efficient ensemble approaches reach the performance of the traditional deep ensembles with significantly lower costs. However, the training resources required by these approaches are still at least the same as training a single dense model. In this work, we draw a unique connection between sparse neural network training and deep ensembles, yielding a novel efficient ensemble learning framework called FreeTickets. Instead of training multiple dense networks and averaging them, we directly train sparse subnetworks from scratch and extract diverse yet accurate subnetworks during this efficient, sparse-to-sparse training. Our framework, FreeTickets, is defined as the ensemble of these relatively cheap sparse subnetworks. Despite being an ensemble method, FreeTickets has even fewer parameters and training FLOPs than a single dense model. This seemingly counter-intuitive outcome is due to the ultra training/inference efficiency of dynamic sparse training. FreeTickets surpasses the dense baseline in all the following criteria: prediction accuracy, uncertainty estimation, out-of-distribution (OoD) robustness, as well as efficiency for both training and inference. Impressively, FreeTickets outperforms the naive deep ensemble with ResNet50 on ImageNet using around only 1/5 of the training FLOPs required by the latter. We have released our source code at https://github.com/VITA-Group/FreeTickets.

cs.LG↗
arXiv2026

Demystifying the Roles of LLM Layers in Retrieval, Knowledge, and Reasoning

Recent studies suggest that the deeper layers of Large Language Models (LLMs) contribute little to representation learning and can often be removed without significant performance loss. However, such claims are typically drawn from narrow evaluations and may overlook important aspects of model behavior. In this work, we present a systematic study of depth utilization across diverse dimensions, including evaluation protocols, task categories, and model architectures. Our analysis confirms that very deep layers are generally less effective than earlier ones, but their contributions vary substantially with the evaluation setting. Under likelihood-based metrics without generation, pruning most layers preserves performance, with only the initial few being critical. By contrast, generation-based evaluation uncovers indispensable roles for middle and deeper layers in enabling reasoning and maintaining long-range coherence. We further find that knowledge and retrieval are concentrated in shallow components, whereas reasoning accuracy relies heavily on deeper layers -- yet can be reshaped through distillation. These results highlight that depth usage in LLMs is highly heterogeneous and context-dependent, underscoring the need for task-, metric-, and model-aware perspectives in both interpreting and compressing large models.

cs.AI↗
arXiv2025

Effective Learning for Small Reasoning Models: An Empirical Study on 0.5B Reasoning LLMs

The ongoing evolution of language models has led to the development of large-scale architectures that demonstrate exceptional performance across a wide range of tasks. However, these models come with significant computational and energy demands, as well as potential privacy implications. In this context, Small Reasoning Language Models (SRLMs) with approximately 0.5 billion parameters present a compelling alternative due to their remarkable computational efficiency and cost-effectiveness, particularly in resource-constrained environments. Despite these advantages, the limited capacity of 0.5 billion parameter models poses challenges in handling complex tasks such as mathematical reasoning. This research investigates various training strategies, including supervised fine-tuning (SFT), knowledge distillation (KD), and reinforcement learning (RL), as well as their hybrid implementations, to enhance the performance of 0.5B SRLMs. We analyze effective methodologies to bridge the performance gap between SRLMS and larger models and present insights into optimal training pipelines tailored for these smaller architectures. Through extensive experimental validation and analysis, our work aims to provide actionable recommendations for maximizing the reasoning capabilities of 0.5B models.

cs.AI↗