Zhaoyi Yan

Xin Zhong, Zhaoyi Yan, Jing Qin et al.

In recent years, crowd counting has become an important issue in computer vision. In most methods, the density maps are generated by convolving with a Gaussian kernel from the ground-truth dot maps which are marked around the center of human heads. Due to the fixed geometric structures in CNNs and indistinct head-scale information, the head features are obtained incompletely. Deformable convolution is proposed to exploit the scale-adaptive capabilities for CNN features in the heads. By learning the coordinate offsets of the sampling points, it is tractable to improve the ability to adjust the receptive field. However, the heads are not uniformly covered by the sampling points in the deformable convolution, resulting in loss of head information. To handle the non-uniformed sampling, an improved Normed-Deformable Convolution (\textit{i.e.,}NDConv) implemented by Normed-Deformable loss (\textit{i.e.,}NDloss) is proposed in this paper. The offsets of the sampling points which are constrained by NDloss tend to be more even. Then, the features in the heads are obtained more completely, leading to better performance. Especially, the proposed NDConv is a light-weight module which shares similar computation burden with Deformable Convolution. In the extensive experiments, our method outperforms state-of-the-art methods on ShanghaiTech A, ShanghaiTech B, UCF\_QNRF, and UCF\_CC\_50 dataset, achieving 61.4, 7.8, 91.2, and 167.2 MAE, respectively. The code is available at https://github.com/bingshuangzhuzi/NDConv

Yuanyi Wang, Yanggan Gu, Su Lu et al.

Weight-space model merging is usually formulated as an algebraic operation on checkpoints, yet at LLM scale the limiting resource is often the set of expert weights that must be read. We introduce MergePipe, a budget-aware execution layer that casts LLM merging as an \emph{expert access-set} problem: given a merge operator and a checkpoint family in a shared weight coordinate system, choose which expert delta blocks to access under an explicit I/O budget. MergePipe indexes parameter blocks, builds deterministic access plans, and executes the induced budgeted merge with replayable manifests. The plan is budget-sound by construction and recovers the full-read merge at full budget; for fixed-coefficient additive operators, the omitted-update error is bounded by the norm of omitted deltas. Across Qwen and Llama merging workloads, MergePipe reduces expert-read I/O by up to an order of magnitude and achieves up to $11\times$ speedups. Representative budget sweeps show $O(10^{-3})$ parameter deviation from full-read merges and no monotonic degradation on downstream benchmarks.

Yuanyi Wang, Su Lu, Yanggan Gu et al.

On-policy distillation (OPD) trains a student on its own rollouts with token-level teacher supervision. Recent selective OPD methods exploit the non-uniformity of OPD signals by prioritizing high-entropy or high-disagreement tokens. We revisit this principle and ask: which token-level teacher signals are actually learnable? Using a fixed-context diagnostic that measures same-context teacher-student KL reduction, we show that raw KL disagreement is a coarse proxy for learning value. It conflates learnable disagreement, where the teacher assigns corrective mass to the student's top-K candidates, with incompatible disagreement, where the teacher places mass mostly off the student's current support. We formalize this local compatibility as token teachability and show that it better predicts fixed-context improvement than raw KL alone. Motivated by this finding, we propose Teachability-Aware OPD (TA-OPD), a lightweight token-position selection method that applies OPD loss to high-teachability positions without reward models or verifiers. Across Qwen2.5 and Qwen 3 teacher-student settings, TA-OPD often surpasses full-token OPD with only 5% retained tokens and improves over entropy- and divergence-based baselines. Our results reframe selective OPD as selecting learnable teacher signals rather than merely salient tokens.

Mingyue Guo, Binghui Chen, Zhaoyi Yan et al.

Multidomain crowd counting aims to learn a general model for multiple diverse datasets. However, deep networks prefer modeling distributions of the dominant domains instead of all domains, which is known as domain bias. In this study, we propose a simple-yet-effective Modulating Domain-specific Knowledge Network (MDKNet) to handle the domain bias issue in multidomain crowd counting. MDKNet is achieved by employing the idea of `modulating', enabling deep network balancing and modeling different distributions of diverse datasets with little bias. Specifically, we propose an Instance-specific Batch Normalization (IsBN) module, which serves as a base modulator to refine the information flow to be adaptive to domain distributions. To precisely modulating the domain-specific information, the Domain-guided Virtual Classifier (DVC) is then introduced to learn a domain-separable latent space. This space is employed as an input guidance for the IsBN modulator, such that the mixture distributions of multiple datasets can be well treated. Extensive experiments performed on popular benchmarks, including Shanghai-tech A/B, QNRF and NWPU, validate the superiority of MDKNet in tackling multidomain crowd counting and the effectiveness for multidomain learning. Code is available at \url{https://github.com/csguomy/MDKNet}.

Yuanyi Wang, Zhaoyi Yan, Yiming Zhang et al.

Recent advances in large language models (LLMs) have intensified efforts to fuse heterogeneous open-source models into a unified system that inherits their complementary strengths. Existing logit-based fusion methods maintain inference efficiency but treat vocabulary dimensions independently, overlooking semantic dependencies encoded by cross-dimension interactions. These dependencies reflect how token types interact under a model's internal reasoning and are essential for aligning models with diverse generation behaviors. To explicitly model these dependencies, we propose \textbf{InfiGFusion}, the first structure-aware fusion framework with a novel \textit{Graph-on-Logits Distillation} (GLD) loss. Specifically, we retain the top-$k$ logits per output and aggregate their outer products across sequence positions to form a global co-activation graph, where nodes represent vocabulary channels and edges quantify their joint activations. To ensure scalability and efficiency, we design a sorting-based closed-form approximation that reduces the original $O(n^4)$ cost of Gromov-Wasserstein distance to $O(n \log n)$, with provable approximation guarantees. Experiments across multiple fusion settings show that GLD consistently improves fusion quality and stability. InfiGFusion outperforms SOTA models and fusion baselines across 11 benchmarks spanning reasoning, coding, and mathematics. It shows particular strength in complex reasoning tasks, with +35.6 improvement on Multistep Arithmetic and +37.06 on Causal Judgement over SFT, demonstrating superior multi-step and relational inference.

Yuanyi Wang, Yifan Yang, Su Lu et al.

Continual post-training aims to extend large language models (LLMs) with new knowledge, skills, and behaviors, yet it remains unclear when sequential updates enable capability transfer and when they cause catastrophic forgetting. Existing methods mitigate forgetting through sequential fine-tuning, replay, regularization, or model merging, but offer limited criteria for determining when incorporating new updates is beneficial or harmful. In this work, we study LLM continual post-training through three questions: What drives forgetting? When do sequentially acquired capabilities transfer or interfere? How can compatibility be used to control update integration? We address these questions through task geometry: we represent each post-training task by its parameter update and study the covariance geometry induced by the update. Our central finding is that: forgetting can be considered as a state-relative update-integration failure, it arises when the covariance geometries induced by tasks misalign with the geometry of the evolving model state. Sequential updates transfer when they remain compatible with the model state shaped by previous updates, and interfere when state-relative geometry conflict becomes high. Motivated by this finding, we propose Geometry-Conflict Wasserstein Merging (GCWM), a data-free update-integration method that constructs a shared Wasserstein metric via Gaussian Wasserstein barycenters and uses geometry conflict to gate geometry-aware correction. Across Qwen3 0.6B--14B on domain-continual and capability-continual settings, GCWM consistently outperforms data-free baselines, improving retention and final performance without replay data. These results identify geometry conflict as both an explanatory signal for forgetting and a practical control signal for LLM continual post-training.

Yuanyi Wang, Yanggan Gu, Zihao Wang et al.

Large language model (LLM) merging has become a key technique in modern LLM development pipelines, enabling the integration of multiple task- or domain-specific expert models without retraining. However, as the number of experts grows, existing merging implementations treat model parameters as unstructured files and execute merges in a stateless, one-shot manner, leading to excessive disk I/O, redundant parameter scans, and poor scalability. In this paper, we present \textbf{MergePipe}, a parameter management system for scalable LLM merging. MergePipe is the first system that treats LLM merging as a data management and execution problem, and introduces a catalog-driven abstraction over model parameters, merge plans, and execution lineage. At its core, MergePipe employs a cost-aware planner that explicitly models expert parameter I/O and enforces user-specified I/O budgets, followed by a streaming execution engine that materializes merged models under transactional guarantees. Our key insight is that while base model reads and output writes are unavoidable, expert parameter reads dominate merge cost and constitute the primary optimization target. By making expert access budget-aware throughout planning and execution, MergePipe mitigates the $O(K)$ I/O growth of naive pipelines and achieves predictable scaling behavior. Experiments show that MergePipe reduces total I/O by up to an order of magnitude and delivers up to $11\times$ end-to-end speedups (up to 90\% wall-time reduction) over state-of-the-art LLM merging pipelines.

Zhaoyi Yan, Binghui Chen, Yunfan Liu et al.

Knowledge distillation (KD) aims to transfer knowledge from a large-scale teacher model to a lightweight one, significantly reducing computational and storage requirements. However, the inherent learning capacity gap between the teacher and student often hinders the sufficient transfer of knowledge, motivating numerous studies to address this challenge. Inspired by the progressive approximation principle in the Stone-Weierstrass theorem, we propose Expandable Residual Approximation (ERA), a novel KD method that decomposes the approximation of residual knowledge into multiple steps, reducing the difficulty of mimicking the teacher's representation through a divide-and-conquer approach. Specifically, ERA employs a Multi-Branched Residual Network (MBRNet) to implement this residual knowledge decomposition. Additionally, a Teacher Weight Integration (TWI) strategy is introduced to mitigate the capacity disparity by reusing the teacher's head weights. Extensive experiments show that ERA improves the Top-1 accuracy on the ImageNet classification benchmark by 1.41% and the AP on the MS COCO object detection benchmark by 1.40, as well as achieving leading performance across computer vision tasks. Codes and models are available at https://github.com/Zhaoyi-Yan/ERA.

Zhaoyi Yan, Xiaoming Li, Mu Li et al.

Deep convolutional networks (CNNs) have exhibited their potential in image inpainting for producing plausible results. However, in most existing methods, e.g., context encoder, the missing parts are predicted by propagating the surrounding convolutional features through a fully connected layer, which intends to produce semantically plausible but blurry result. In this paper, we introduce a special shift-connection layer to the U-Net architecture, namely Shift-Net, for filling in missing regions of any shape with sharp structures and fine-detailed textures. To this end, the encoder feature of the known region is shifted to serve as an estimation of the missing parts. A guidance loss is introduced on decoder feature to minimize the distance between the decoder feature after fully connected layer and the ground-truth encoder feature of the missing parts. With such constraint, the decoder feature in missing region can be used to guide the shift of encoder feature in known region. An end-to-end learning algorithm is further developed to train the Shift-Net. Experiments on the Paris StreetView and Places datasets demonstrate the efficiency and effectiveness of our Shift-Net in producing sharper, fine-detailed, and visually plausible results. The codes and pre-trained models are available at https://github.com/Zhaoyi-Yan/Shift-Net.

Mingyue Guo, Li Yuan, Zhaoyi Yan et al.

Crowd counting has achieved significant progress by training regressors to predict instance positions. In heavily crowded scenarios, however, regressors are challenged by uncontrollable annotation variance, which causes density map bias and context information inaccuracy. In this study, we propose mutual prompt learning (mPrompt), which leverages a regressor and a segmenter as guidance for each other, solving bias and inaccuracy caused by annotation variance while distinguishing foreground from background. In specific, mPrompt leverages point annotations to tune the segmenter and predict pseudo head masks in a way of point prompt learning. It then uses the predicted segmentation masks, which serve as spatial constraint, to rectify biased point annotations as context prompt learning. mPrompt defines a way of mutual information maximization from prompt learning, mitigating the impact of annotation variance while improving model accuracy. Experiments show that mPrompt significantly reduces the Mean Average Error (MAE), demonstrating the potential to be general framework for down-stream vision tasks.

Zhaoyi Yan, Yiming Zhang, Baoyi He et al.

We introduce InfiFusion, an efficient training pipeline designed to integrate multiple domain-specialized Large Language Models (LLMs) into a single pivot model, effectively harnessing the strengths of each source model. Traditional fusion methods either merge model parameters directly or rely on knowledge distillation with rigid assumptions, limiting their flexibility and efficiency. InfiFusion overcomes these limitations by enhancing Universal Logit Distillation (ULD) with Top-K selection and Logits Standardization. We propose two fusion strategies: Pairwise Fusion (InfiFusion$_p$), where each source model knowledge is distilled individually into the pivot model followed by merging and Unified Fusion (InfiFusion$_u$), where knowledge from all source models is distilled simultaneously into the pivot model. InfiFusion outperforms the state-of-the-art models, such as Qwen-2.5-14B-Instruct and Phi-4, across 11 widely applied benchmarks covering reasoning, coding, mathematics, and instruction-following tasks. Notably, InfiFusion achieves this superior performance while significantly reduces computational costs, completing full training with only 160 H800 GPU hours compared to the millions typically required for traditional LLM training.

Yanggan Gu, Yuanyi Wang, Zhaoyi Yan et al.

Model fusion combines multiple Large Language Models (LLMs) with different strengths into a more powerful, integrated model through lightweight training methods. Existing works on model fusion focus primarily on supervised fine-tuning (SFT), leaving preference alignment (PA) --a critical phase for enhancing LLM performance--largely unexplored. The current few fusion methods on PA phase, like WRPO, simplify the process by utilizing only response outputs from source models while discarding their probability information. To address this limitation, we propose InfiFPO, a preference optimization method for implicit model fusion. InfiFPO replaces the reference model in Direct Preference Optimization (DPO) with a fused source model that synthesizes multi-source probabilities at the sequence level, circumventing complex vocabulary alignment challenges in previous works and meanwhile maintaining the probability information. By introducing probability clipping and max-margin fusion strategies, InfiFPO enables the pivot model to align with human preferences while effectively distilling knowledge from source models. Comprehensive experiments on 11 widely-used benchmarks demonstrate that InfiFPO consistently outperforms existing model fusion and preference optimization methods. When using Phi-4 as the pivot model, InfiFPO improve its average performance from 79.95 to 83.33 on 11 benchmarks, significantly improving its capabilities in mathematics, coding, and reasoning tasks.

Yuanyi Wang, Yanggan Gu, Yiming Zhang et al.

We study empirical scaling laws for language model merging measured by cross-entropy. Despite its wide practical use, merging lacks a quantitative rule that predicts returns as we add experts or scale the model size. We identify a compact power law that links model size and expert number: the size-dependent floor decreases with model capacity, while the merging tail exhibits clear diminishing returns in the number of experts. The law holds in-domain and cross-domain, tightly fits measured curves across diverse architectures and methods (Average, TA, TIES, DARE), and explains two robust regularities: most gains arrive early, and variability shrinks as more experts are included. Building on this, we present a simple theory that explains why gains fall roughly as 1/k and links the floor and tail to properties of the base model and the diversity across domains. This law enables predictive planning: estimate how many experts are needed to reach a target loss, decide when to stop adding experts, and trade off scaling the base model versus adding experts under a fixed budget--turning merging from heuristic practice into a computationally efficient, planable alternative to multitask training. This suggests a scaling principle for distributed generative AI: predictable gains can be achieved by composing specialists, offering a complementary path toward AGI-level systems.

Zhaoyi Yan, Kangjun Liu, Qixiang Ye

Knowledge distillation has become a cornerstone technique in deep learning, facilitating the transfer of knowledge from complex models to lightweight counterparts. Traditional distillation approaches focus on transferring knowledge at the instance level, but fail to capture nuanced semantic relationships within the data. In response, this paper introduces a novel methodology, Semantics-based Relation Knowledge Distillation (SeRKD), which reimagines knowledge distillation through a semantics-relation lens among each sample. By leveraging semantic components, \ie, superpixels, SeRKD enables a more comprehensive and context-aware transfer of knowledge, which skillfully integrates superpixel-based semantic extraction with relation-based knowledge distillation for a sophisticated model compression and distillation. Particularly, the proposed method is naturally relevant in the domain of Vision Transformers (ViTs), where visual tokens serve as fundamental units of representation. Experimental evaluations on benchmark datasets demonstrate the superiority of SeRKD over existing methods, underscoring its efficacy in enhancing model performance and generalization capabilities.

Binghui Chen, Zhaoyi Yan, Ke Li et al.

In crowd counting, due to the problem of laborious labelling, it is perceived intractability of collecting a new large-scale dataset which has plentiful images with large diversity in density, scene, etc. Thus, for learning a general model, training with data from multiple different datasets might be a remedy and be of great value. In this paper, we resort to the multi-domain joint learning and propose a simple but effective Domain-specific Knowledge Propagating Network (DKPNet)1 for unbiasedly learning the knowledge from multiple diverse data domains at the same time. It is mainly achieved by proposing the novel Variational Attention(VA) technique for explicitly modeling the attention distributions for different domains. And as an extension to VA, Intrinsic Variational Attention(InVA) is proposed to handle the problems of over-lapped domains and sub-domains. Extensive experiments have been conducted to validate the superiority of our DKPNet over several popular datasets, including ShanghaiTech A/B, UCF-QNRF and NWPU.

Zhaoyi Yan, Ruimao Zhang, Hongzhi Zhang et al.

Crowd counting is critical for numerous video surveillance scenarios. One of the main issues in this task is how to handle the dramatic scale variations of pedestrians caused by the perspective effect. To address this issue, this paper proposes a novel convolution neural network-based crowd counting method, termed Perspective-guided Fractional-Dilation Network (PFDNet). By modeling the continuous scale variations, the proposed PFDNet is able to select the proper fractional dilation kernels for adapting to different spatial locations. It significantly improves the flexibility of the state-of-the-arts that only consider the discrete representative scales. In addition, by avoiding the multi-scale or multi-column architecture that used in other methods, it is computationally more efficient. In practice, the proposed PFDNet is constructed by stacking multiple Perspective-guided Fractional-Dilation Convolutions (PFC) on a VGG16-BN backbone. By introducing a novel generalized dilation convolution operation, the PFC can handle fractional dilation ratios in the spatial domain under the guidance of perspective annotations, achieving continuous scales modeling of pedestrians. To deal with the problem of unavailable perspective information in some cases, we further introduce an effective perspective estimation branch to the proposed PFDNet, which can be trained in either supervised or weakly-supervised setting once the branch has been pre-trained. Extensive experiments show that the proposed PFDNet outperforms state-of-the-art methods on ShanghaiTech A, ShanghaiTech B, WorldExpo'10, UCF-QNRF, UCF_CC_50 and TRANCOS dataset, achieving MAE 53.8, 6.5, 6.8, 84.3, 205.8, and 3.06 respectively.

Zhaoyi Yan, Yuchen Yuan, Wangmeng Zuo et al.

In this paper, we propose a novel perspective-guided convolution (PGC) for convolutional neural network (CNN) based crowd counting (i.e. PGCNet), which aims to overcome the dramatic intra-scene scale variations of people due to the perspective effect. While most state-of-the-arts adopt multi-scale or multi-column architectures to address such issue, they generally fail in modeling continuous scale variations since only discrete representative scales are considered. PGCNet, on the other hand, utilizes perspective information to guide the spatially variant smoothing of feature maps before feeding them to the successive convolutions. An effective perspective estimation branch is also introduced to PGCNet, which can be trained in either supervised setting or weakly-supervised setting when the branch has been pre-trained. Our PGCNet is single-column with moderate increase in computation, and extensive experimental results on four benchmark datasets show the improvements of our method against the state-of-the-arts. Additionally, we also introduce Crowd Surveillance, a large scale dataset for crowd counting that contains 13,000+ high-resolution images with challenging scenarios.