arXiv subjects

Ziang Wu

Publications and source records attributed to Ziang Wu.

8 recordsLinked to original sources

arXiv2026

Relax Within, Balance Across: Geometry-Guided Load Balancing for Vision-Language Mixture-of-Experts

Vision-language MoE batches contain different numbers of image and text tokens. Image resolution, image count, tiling, and prompt length all change this token mix. We call the standard token-level Switch auxiliary loss Std-Aux. Std-Aux balances only the mixed load, so large image and text load errors can cancel at one mix. On our main model, the same trained router shows more than a fivefold change in load imbalance across image resolutions. We hold the image and text load profiles fixed and derive the exact load curve as the token mix varies. The image-text load gap controls sensitivity to the token mix. Physical preprocessing can also change the conditional profiles. The fixed-profile law excludes such changes. To design a remedy, we examine the router input structure. Image and text occupy distinct regions, while visual tokens group strongly by source image. The modality boundary motivates separate image and text terms. The image boundary motivates one equal-weight routing instance per image. ReBA, or Relax Within, Balance Across, implements both choices. Across four split backbones, ReBA lowers load on every reported benchmark input while keeping mean task accuracy comparable to Std-Aux. ReBA also lowers average load over the tested range and worst physical load under resolution and tiling shifts. Code is available at https://github.com/ZiangWu-77/ReBA.

cs.CV↗

arXiv2026

Divergence Decoding: Training-Free Capability Fusion

While large language models excel in reasoning, these generalists often lack knowledge for specialized scientific domains. Conversely, domain models~(specialists), while knowledgeable, suffer from specialization side-effects including diminished logic and reduced robustness.To address this dilemma, we introduce Divergence Decoding, a training-free framework for capability fusion. It reconstructs the "draft-and-verify" skeleton of speculative decoding into an adaptive routing mechanism. The core is using Jensen-Shannon divergence to monitor the distributional disagreement between the two models at each token. When the specialist exhibits significant divergence, our method identifies it as a potential reasoning risk and instantaneously routes control to the generalist. This allows the dynamic injection of general reasoning while preserving domain expertise, achieving inference-time policy composition of the generalist and the specialist.We evaluate Divergence Decoding across diverse model families (Qwen and Llama series) on challenging scientific benchmarks (GPQA, ChemBench, and ChemCoTBench). Experimental results demonstrate that Divergence Decoding outperforms both the domain-specialized and general-purpose models, effectively surpassing the performance of most single-model baseline. This suggests that Divergence Decoding provides a general, training-free paradigm for fusing diverse LLM capabilities through adaptive inference-time collaboration.

cs.AI↗

arXiv2025

Measuring the splashback feature: Dependence on halo properties and history

In this study, we define the novel splashback depth $\mathcal{D}$ and width $\mathcal{W}$ to examine how the splashback features of dark matter haloes are affected by the physical properties of haloes themselves. We use the largest simulation run in the hydrodynamic MillenniumTNG project. By stacking haloes in bins of halo mass, redshift, mass-dependent properties such as peak height and concentration, and halo formation history, we measure the shape of the logarithmic slope of the density profile of dark matter haloes. Our results show that the splashback depth has a strong dependence on the halo mass which follows a power law $\mathcal{D}\propto\left(\log_{10}M\right)^{2.8}$. Properties with strong correlation with halo mass demonstrate similar dependence. The splashback width has the strongest dependence on halo peak height and follows a power law $\mathcal{W}\propto\nu^{-0.87}$. We provide the fitting functions of the splashback depth and width in terms of halo mass, redshift, peak height, concentrations and halo formation time. The depth and width are therefore considered to be a long term memory tracker of haloes since they depend more on accumulative physical properties, e.g., halo mass, peak height and halo formation time. They are shaped primarily by the halo's assembly history, which exerts a stronger influence on the inner density profile than short-term dynamical processes. In contrast, the splashback features have little dependence on the short term factors such as halo mass accretion rate and most recent major merger time. The splashback depth and width can therefore be used to complement information gained from quantities like the point of steepest slope or truncation radius to characterise the halo's history and inner structure.

astro-ph.CO↗

arXiv2025

Core collapse in resonant self-interacting dark matter across two decades in halo mass

Core collapse, a process associated with self-interacting dark matter (SIDM) models, can increase the central density of halos by orders of magnitude with observable consequences for dwarf galaxy properties and gravitational lensing. Resonances in the self-interaction cross section, features of hidden-sector models with light mediators and attractive potentials, can boost the strength of self-interactions near specific relative velocities, accelerating collapse in halos with central velocity dispersions near the resonance. To explore this phenomenon, we present a suite of idealized N-body simulations of isolated halos with masses $10^7$-$10^9 \ \rm{M_\odot}$ evolved under two resonant cross section (RCS) models with localized enhancement to the cross section on scales $v \sim 5$-$50 \ \rm{km} \ \rm{s^{-1}}$. We show that the change in halo internal structure depends on how the velocity distribution of bound particles moves across resonances in the cross section during core formation and collapse. The interplay between the velocity distribution of bound particles and localized features of the cross section causes deviations from self-similar evolution, a characteristic of velocity-independent cross sections, at the level of up to $20\%$. Depending on the alignment with resonant features, halos of different masses reach different evolutionary stages after a fixed physical time and develop diverse density profiles and rotation curves.

astro-ph.GA↗

arXiv2024

LLaVA-CoT: Let Vision Language Models Reason Step-by-Step

Large language models have demonstrated substantial advancements in reasoning capabilities. However, current Vision-Language Models (VLMs) often struggle to perform systematic and structured reasoning, especially when handling complex visual question-answering tasks. In this work, we introduce LLaVA-CoT, a large VLM designed to conduct autonomous multistage reasoning. Unlike chain-of-thought prompting, LLaVA-CoT independently engages in sequential stages of summarization, visual interpretation, logical reasoning, and conclusion generation. This structured approach enables LLaVA-CoT to achieve marked improvements on reasoning-intensive tasks. To accomplish this, we construct the LLaVA-CoT-100k dataset, integrating samples from various visual question answering sources and providing structured reasoning annotations. Besides, we propose a test-time stage-wise retracing search method (SWIRES), which enables effective and efficient test-time scaling. Remarkably, with only 100k training samples and test-time scaling, LLaVA-CoT not only outperforms its base model by 9.4% on a wide range of multimodal reasoning benchmarks, but also surpasses the performance of larger and even closed-source models, such as Gemini-1.5-pro, GPT-4o-mini, and Llama-3.2-90B-Vision-Instruct. The code, dataset, and pre-trained weights are publicly available at https://github.com/PKU-YuanGroup/LLaVA-CoT.

cs.CV↗

arXiv2024

LOOK-M: Look-Once Optimization in KV Cache for Efficient Multimodal Long-Context Inference

Long-context Multimodal Large Language Models (MLLMs) demand substantial computational resources for inference as the growth of their multimodal Key-Value (KV) cache, in response to increasing input lengths, challenges memory and time efficiency. Unlike single-modality LLMs that manage only textual contexts, the KV cache of long-context MLLMs includes representations from multiple images with temporal and spatial relationships and related textual contexts. The predominance of image tokens means traditional optimizations for LLMs' KV caches are unsuitable for multimodal long-context settings, and no prior works have addressed this challenge. In this work, we introduce LOOK-M, a pioneering, fine-tuning-free approach that efficiently reduces the multimodal KV cache size while maintaining performance comparable to a full cache. We observe that during prompt prefill, the model prioritizes more textual attention over image features, and based on the multimodal interaction observation, a new proposed text-prior method is explored to compress the KV cache. Furthermore, to mitigate the degradation of image contextual information, we propose several compensatory strategies using KV pairs merging. LOOK-M demonstrates that with a significant reduction in KV Cache memory usage, such as reducing it by 80% in some cases, it not only achieves up to 1.5x faster decoding but also maintains or even enhances performance across a variety of long context multimodal tasks.

cs.CL↗

arXiv2024

Adaptive Convolutional Neural Network for Image Super-resolution

Convolutional neural networks can automatically learn features via deep network architectures and given input samples. However, the robustness of obtained models may face challenges in varying scenes. Bigger differences in network architecture are beneficial to extract more diversified structural information to strengthen the robustness of an obtained super-resolution model. In this paper, we proposed a adaptive convolutional neural network for image super-resolution (ADSRNet). To capture more information, ADSRNet is implemented by a heterogeneous parallel network. The upper network can enhance relation of context information, salient information relation of a kernel mapping and relations of shallow and deep layers to improve performance of image super-resolution. That can strengthen adaptability of an obtained super-resolution model for different scenes. The lower network utilizes a symmetric architecture to enhance relations of different layers to mine more structural information, which is complementary with a upper network for image super-resolution. The relevant experimental results show that the proposed ADSRNet is effective to deal with image resolving. Codes are obtained at https://github.com/hellloxiaotian/ADSRNet.

eess.IV↗

arXiv2022

Generative Adversarial Networks for Image Super-Resolution: A Survey

Single image super-resolution (SISR) has played an important role in the field of image processing. Recent generative adversarial networks (GANs) can achieve excellent results on low-resolution images. However, there are little literatures summarizing different GANs in SISR. In this paper, we conduct a comparative study of GANs from different perspectives. We begin by surveying the development of GANs and popular GAN variants for image-related applications, and then analyze motivations, implementations and differences of GANs based optimization methods and discriminative learning for image super-resolution in terms of supervised, semi-supervised and unsupervised manners, where these GANs are analyzed via integrating different network architectures, prior knowledge, loss functions and multiple tasks. Secondly, we compare the performances of these popular GANs on public datasets via quantitative and qualitative analysis in SISR. Finally, we highlight challenges of GANs and potential research points for SISR.

eess.IV↗