arXiv subjects

Mengyang Zhao

Publications and source records attributed to Mengyang Zhao.

At least 19 recordsLinked to original sources

arXiv2026

CRISP: Pre-LLM Yet Text-Driven Visual Token Pruning for Efficient LVLM Inference

Large Vision-Language Models (LVLMs) typically require processing hundreds to thousands of visual tokens, leading to substantial inference overhead. Existing visual token pruning methods either operate before the LLM using text-agnostic heuristics or prune inside the LLM at the cost of efficiency and noisy cross-modal attention. To address these limitations, we propose CRISP, a pre-LLM yet text-driven visual token pruning framework that preserves both instruction-relevant evidence and essential scene context. CRISP works in a two-stage pipeline: Stage 1 first identifies text-aligned visual tokens, and Stage 2 enhances contextual completeness through semantic diversity. Extensive experiments on LLaVA-1.5 and LLaVA-NeXT demonstrate that CRISP achieves superior performance retention under aggressive pruning ratios, maintaining up to 99.5% accuracy while reducing inference cost and latency by more than 2 times. CRISP serves as a practical solution for efficient LVLM inference, especially in resource-constrained scenarios.

cs.CV↗

arXiv2026

Cheeger--Hodge Contrastive Learning for Structurally Robust Graph Representation Learning

Graph Contrastive Learning (GCL) has emerged as a prominent framework for unsupervised graph representation learning. However, relying on augmentation design alone to define the invariances learned by GCL can be brittle under structural perturbations. To address this issue, we propose Cheeger--Hodge Contrastive Learning (CHCL), a framework that aligns a perturbation-stable Cheeger--Hodge joint signature across augmented views for robust graph representation learning. The proposed signature combines a Cheeger-inspired connectivity signature derived from the algebraic connectivity \(λ_2\) with the low-frequency spectrum of the 1-Hodge Laplacian, thereby capturing both global connectivity and higher-order structural information. By aligning encoder representations with the proposed Cheeger--Hodge joint signature across augmented views, CHCL learns graph embeddings that are robust to local structural perturbations. Extensive experiments on standard benchmarks, transfer settings demonstrate that CHCL consistently improves performance, robustness, and generalization.

cs.LG↗

arXiv2026

DocCogito: Aligning Layout Cognition and Step-Level Grounded Reasoning for Document Understanding

Document understanding with multimodal large language models (MLLMs) requires not only accurate answers but also explicit, evidence-grounded reasoning, especially in high-stakes scenarios. However, current document MLLMs still fall short of forming a complete, human-like reasoning process, because even when they improve both layout encoding and CoT-style prompting, the interaction between the two is typically learned implicitly and remains loosely coupled rather than being enforced as a systematic mechanism. So we propose DocCogito, a unified framework that integrates global layout perception with structured, region-grounded reasoning. DocCogito introduces a lightweight layout tower that distills page structure into learnable global layout prior tokens, and a deterministic Visual-Semantic Chain (VSC)-a concise structured representation less ambiguous than free-form natural-language CoT-to supervise fine-grained intermediate reasoning aligned with evidence regions. Training follows a progressive recipe, including layout perception pretraining, VSC-guided cold start, rejection sampling, and GRPO. To further strengthen the internal coupling between layout priors and VSC execution, we augment standard rewards with a fine-grained region-confidence signal that encourages reasoning traces to stay aligned with corresponding evidence regions. Extensive experiments on six benchmarks (DocVQA, WTQ, ChartQA, TextVQA, OCRBench, and InfoVQA) demonstrate strong generalization, achieving state-of-the-art results on four benchmarks.

cs.CV↗

arXiv2026

Logic-Guided Multistage Inference for Explainable Multidefendant Judgment Prediction

Crime disrupts societal stability, making law essential for balance. In multidefendant cases, assigning responsibility is complex and challenges fairness, requiring precise role differentiation. However, judicial phrasing often obscures the roles of the defendants, hindering effective AI-driven analyses. To address this issue, we incorporate sentencing logic into a pretrained Transformer encoder framework to enhance the intelligent assistance in multidefendant cases while ensuring legal interpretability. Within this framework an oriented masking mechanism clarifies roles and a comparative data construction strategy improves the model's sensitivity to culpability distinctions between principals and accomplices. Predicted guilt labels are further incorporated into a regression model through broadcasting, consolidating crime descriptions and court views. Our proposed masked multistage inference (MMSI) framework, evaluated on the custom IMLJP dataset for intentional injury cases, achieves significant accuracy improvements, outperforming baselines in role-based culpability differentiation. This work offers a robust solution for enhancing intelligent judicial systems, with publicly code available.

cs.AI↗

arXiv2025

Multi-Scale Harmonic Encoding for Feature-Wise Graph Message Passing

Most Graph Neural Networks (GNNs) propagate messages by treating node embeddings as holistic feature vectors, implicitly assuming uniform relevance across feature dimensions. This limits their ability to selectively transmit informative components, especially when graph structures exhibit distinct frequency characteristics. We propose MSH-GNN (Multi-Scale Harmonic Graph Neural Network), a frequency-aware message passing framework that performs feature-wise adaptive propagation. Each node projects incoming messages onto node-conditioned feature subspaces derived from its own representation, enabling selective extraction of frequency-relevant components. Learnable multi-scale harmonic modulations further allow the model to capture both smooth and oscillatory structural patterns. A frequency-aware attention pooling mechanism is introduced for graph-level readout. We show that MSH-GNN admits an interpretation as a learnable Fourier-feature approximation of kernelized message functions and matches the expressive power of the 1-Weisfeiler-Lehman (1-WL) test. Extensive experiments on node- and graph-level benchmarks demonstrate consistent improvements over state-of-the-art methods, particularly in joint structure-frequency analysis tasks.

cs.LG↗

arXiv2025

EVE: Towards End-to-End Video Subtitle Extraction with Vision-Language Models

Video subtitles play a crucial role in short videos and movies, as they not only help models better understand video content but also support applications such as video translation and content retrieval. Existing video subtitle extraction methods typically rely on multi-stage frameworks, where errors accumulate across stages and temporal dependencies are underutilized due to frame-wise processing. Moreover, although some Large Vision-Language Models (LVLMs) possess strong OCR capabilities, predicting accurate timestamps for subtitle texts remains challenging. To this end, we propose an End-to-end Video subtitle Extraction framework based on LVLMs, named EVE, which can output subtitles and their timestamps simultaneously. Specifically, we introduce a dual-branch Spatiotemporal Subtitle-Salient (S\textsuperscript{3}) Module that serves as an adapter for LVLMs, capable of representing subtitle-related content and considering inter-frame correlations using only a small number of tokens. Within this module, the Spatial Semantic Context Aggregate branch aggregates high-level global semantics to provide spatial visual contextual information, while the Temporal Subtitle Token Query branch explicitly queries subtitle-relevant tokens while considering temporal correlation across frames. The small number of tokens retained by the S\textsuperscript{3} module are fed to the language model, which then directly outputs the subtitle text along with its timestamps. Furthermore, we construct the first large-scale dataset dedicated to video subtitle extraction, ViSa, containing over 2.5M videos with timestamped and bilingual annotation, thereby providing the community with a well-organized training and evaluation benchmark.

cs.CV↗

arXiv2025

OmniPT: Unleashing the Potential of Large Vision Language Models for Pedestrian Tracking and Understanding

LVLMs have been shown to perform excellently in image-level tasks such as VQA and caption. However, in many instance-level tasks, such as visual grounding and object detection, LVLMs still show performance gaps compared to previous expert models. Meanwhile, although pedestrian tracking is a classical task, there have been a number of new topics in combining object tracking and natural language, such as Referring MOT, Cross-view Referring MOT, and Semantic MOT. These tasks emphasize that models should understand the tracked object at an advanced semantic level, which is exactly where LVLMs excel. In this paper, we propose a new unified Pedestrian Tracking framework, namely OmniPT, which can track, track based on reference and generate semantic understanding of tracked objects interactively. We address two issues: how to model the tracking task into a task that foundation models can perform, and how to make the model output formatted answers. To this end, we implement a training phase consisting of RL-Mid Training-SFT-RL. Based on the pre-trained weights of the LVLM, we first perform a simple RL phase to enable the model to output fixed and supervisable bounding box format. Subsequently, we conduct a mid-training phase using a large number of pedestrian-related datasets. Finally, we perform supervised fine-tuning on several pedestrian tracking datasets, and then carry out another RL phase to improve the model's tracking performance and enhance its ability to follow instructions. We conduct experiments on tracking benchmarks and the experimental results demonstrate that the proposed method can perform better than the previous methods.

cs.CV↗

arXiv2025

CReFT-CAD: Boosting Orthographic Projection Reasoning for CAD via Reinforcement Fine-Tuning

Computer-Aided Design (CAD) plays a pivotal role in industrial manufacturing. Orthographic projection reasoning underpins the entire CAD workflow, encompassing design, manufacturing, and simulation. However, prevailing deep-learning approaches employ standard 3D reconstruction pipelines as an alternative, which often introduce imprecise dimensions and limit the parametric editability required for CAD workflows. Recently, some researchers adopt vision-language models (VLMs), particularly supervised fine-tuning (SFT), to tackle CAD-related challenges. SFT shows promise but often devolves into pattern memorization, yielding poor out-of-distribution performance on complex reasoning tasks. To address these gaps, we introduce CReFT-CAD, a two-stage fine-tuning paradigm that first employs a curriculum-driven reinforcement learning stage with difficulty-aware rewards to build reasoning ability steadily, and then applies supervised post-tuning to hone instruction following and semantic extraction. Complementing this, we release TriView2CAD, the first large-scale, open-source benchmark for orthographic projection reasoning, comprising 200,000 synthetic and 3,000 real-world orthographic projections with precise dimension annotations and six interoperable data modalities. We benchmark leading VLMs on orthographic projection reasoning and demonstrate that CReFT-CAD substantially improves reasoning accuracy and out-of-distribution generalizability in real-world scenarios, offering valuable insights for advancing CAD reasoning research.

cs.CV↗

arXiv2025

GAZE:Governance-Aware pre-annotation for Zero-shot World Model Environments

Training robust world models requires large-scale, precisely labeled multimodal datasets, a process historically bottlenecked by slow and expensive manual annotation. We present a production-tested GAZE pipeline that automates the conversion of raw, long-form video into rich, task-ready supervision for world-model training. Our system (i) normalizes proprietary 360-degree formats into standard views and shards them for parallel processing; (ii) applies a suite of AI models (scene understanding, object tracking, audio transcription, PII/NSFW/minor detection) for dense, multimodal pre-annotation; and (iii) consolidates signals into a structured output specification for rapid human validation. The GAZE workflow demonstrably yields efficiency gains (~19 minutes saved per review hour) and reduces human review volume by >80% through conservative auto-skipping of low-salience segments. By increasing label density and consistency while integrating privacy safeguards and chain-of-custody metadata, our method generates high-fidelity, privacy-aware datasets directly consumable for learning cross-modal dynamics and action-conditioned prediction. We detail our orchestration, model choices, and data dictionary to provide a scalable blueprint for generating high-quality world model training data without sacrificing throughput or governance.

cs.CV↗

arXiv2025

Human + AI for Accelerating Ad Localization Evaluation

Adapting advertisements for multilingual audiences requires more than simple text translation; it demands preservation of visual consistency, spatial alignment, and stylistic integrity across diverse languages and formats. We introduce a structured framework that combines automated components with human oversight to address the complexities of advertisement localization. To the best of our knowledge, this is the first work to integrate scene text detection, inpainting, machine translation (MT), and text reimposition specifically for accelerating ad localization evaluation workflows. Qualitative results across six locales demonstrate that our approach produces semantically accurate and visually coherent localized advertisements, suitable for deployment in real-world workflows.

cs.AI↗

arXiv2025

From Intent to Execution: Multimodal Chain-of-Thought Reinforcement Learning for Precise CAD Code Generation

Computer-Aided Design (CAD) plays a vital role in engineering and manufacturing, yet current CAD workflows require extensive domain expertise and manual modeling effort. Recent advances in large language models (LLMs) have made it possible to generate code from natural language, opening new opportunities for automating parametric 3D modeling. However, directly translating human design intent into executable CAD code remains highly challenging, due to the need for logical reasoning, syntactic correctness, and numerical precision. In this work, we propose CAD-RL, a multimodal Chain-of-Thought (CoT) guided reinforcement learning post training framework for CAD modeling code generation. Our method combines CoT-based Cold Start with goal-driven reinforcement learning post training using three task-specific rewards: executability reward, geometric accuracy reward, and external evaluation reward. To ensure stable policy learning under sparse and high-variance reward conditions, we introduce three targeted optimization strategies: Trust Region Stretch for improved exploration, Precision Token Loss for enhanced dimensions parameter accuracy, and Overlong Filtering to reduce noisy supervision. To support training and benchmarking, we release ExeCAD, a noval dataset comprising 16,540 real-world CAD examples with paired natural language and structured design language descriptions, executable CADQuery scripts, and rendered 3D models. Experiments demonstrate that CAD-RL achieves significant improvements in reasoning quality, output precision, and code executability over existing VLMs.

cs.LG↗

arXiv2025

Interpretable Oracle Bone Script Decipherment through Radical and Pictographic Analysis with LVLMs

As the oldest mature writing system, Oracle Bone Script (OBS) has long posed significant challenges for archaeological decipherment due to its rarity, abstractness, and pictographic diversity. Current deep learning-based methods have made exciting progress on the OBS decipherment task, but existing approaches often ignore the intricate connections between glyphs and the semantics of OBS. This results in limited generalization and interpretability, especially when addressing zero-shot settings and undeciphered OBS. To this end, we propose an interpretable OBS decipherment method based on Large Vision-Language Models, which synergistically combines radical analysis and pictograph-semantic understanding to bridge the gap between glyphs and meanings of OBS. Specifically, we propose a progressive training strategy that guides the model from radical recognition and analysis to pictographic analysis and mutual analysis, thus enabling reasoning from glyph to meaning. We also design a Radical-Pictographic Dual Matching mechanism informed by the analysis results, significantly enhancing the model's zero-shot decipherment performance. To facilitate model training, we propose the Pictographic Decipherment OBS Dataset, which comprises 47,157 Chinese characters annotated with OBS images and pictographic analysis texts. Experimental results on public benchmarks demonstrate that our approach achieves state-of-the-art Top-10 accuracy and superior zero-shot decipherment capabilities. More importantly, our model delivers logical analysis processes, possibly providing archaeologically valuable reference results for undeciphered OBS, and thus has potential applications in digital humanities and historical research. The dataset and code will be released in https://github.com/PKXX1943/PD-OBS.

cs.CV↗

arXiv2025

IADGPT: Unified LVLM for Few-Shot Industrial Anomaly Detection, Localization, and Reasoning via In-Context Learning

Few-Shot Industrial Anomaly Detection (FS-IAD) has important applications in automating industrial quality inspection. Recently, some FS-IAD methods based on Large Vision-Language Models (LVLMs) have been proposed with some achievements through prompt learning or fine-tuning. However, existing LVLMs focus on general tasks but lack basic industrial knowledge and reasoning capabilities related to FS-IAD, making these methods far from specialized human quality inspectors. To address these challenges, we propose a unified framework, IADGPT, designed to perform FS-IAD in a human-like manner, while also handling associated localization and reasoning tasks, even for diverse and novel industrial products. To this end, we introduce a three-stage progressive training strategy inspired by humans. Specifically, the first two stages gradually guide IADGPT in acquiring fundamental industrial knowledge and discrepancy awareness. In the third stage, we design an in-context learning-based training paradigm, enabling IADGPT to leverage a few-shot image as the exemplars for improved generalization to novel products. In addition, we design a strategy that enables IADGPT to output image-level and pixel-level anomaly scores using the logits output and the attention map, respectively, in conjunction with the language output to accomplish anomaly reasoning. To support our training, we present a new dataset comprising 100K images across 400 diverse industrial product categories with extensive attribute-level textual annotations. Experiments indicate IADGPT achieves considerable performance gains in anomaly detection and demonstrates competitiveness in anomaly localization and reasoning. We will release our dataset in camera-ready.

cs.CV↗

arXiv2025

CrowdTrack: A Benchmark for Difficult Multiple Pedestrian Tracking in Real Scenarios

Multi-object tracking is a classic field in computer vision. Among them, pedestrian tracking has extremely high application value and has become the most popular research category. Existing methods mainly use motion or appearance information for tracking, which is often difficult in complex scenarios. For the motion information, mutual occlusions between objects often prevent updating of the motion state; for the appearance information, non-robust results are often obtained due to reasons such as only partial visibility of the object or blurred images. Although learning how to perform tracking in these situations from the annotated data is the simplest solution, the existing MOT dataset fails to satisfy this solution. Existing methods mainly have two drawbacks: relatively simple scene composition and non-realistic scenarios. Although some of the video sequences in existing dataset do not have the above-mentioned drawbacks, the number is far from adequate for research purposes. To this end, we propose a difficult large-scale dataset for multi-pedestrian tracking, shot mainly from the first-person view and all from real-life complex scenarios. We name it ``CrowdTrack'' because there are numerous objects in most of the sequences. Our dataset consists of 33 videos, containing a total of 5,185 trajectories. Each object is annotated with a complete bounding box and a unique object ID. The dataset will provide a platform to facilitate the development of algorithms that remain effective in complex situations. We analyzed the dataset comprehensively and tested multiple SOTA models on our dataset. Besides, we analyzed the performance of the foundation models on our dataset. The dataset and project code is released at: https://github.com/loseevaya/CrowdTrack .

cs.CV↗

arXiv2025

ChatReID: Open-ended Interactive Person Retrieval via Hierarchical Progressive Tuning for Vision Language Models

Person re-identification (Re-ID) is a crucial task in computer vision, aiming to recognize individuals across non-overlapping camera views. While recent advanced vision-language models (VLMs) excel in logical reasoning and multi-task generalization, their applications in Re-ID tasks remain limited. They either struggle to perform accurate matching based on identity-relevant features or assist image-dominated branches as auxiliary semantics. In this paper, we propose a novel framework ChatReID, that shifts the focus towards a text-side-dominated retrieval paradigm, enabling flexible and interactive re-identification. To integrate the reasoning abilities of language models into Re-ID pipelines, We first present a large-scale instruction dataset, which contains more than 8 million prompts to promote the model fine-tuning. Next. we introduce a hierarchical progressive tuning strategy, which endows Re-ID ability through three stages of tuning, i.e., from person attribute understanding to fine-grained image retrieval and to multi-modal task reasoning. Extensive experiments across ten popular benchmarks demonstrate that ChatReID outperforms existing methods, achieving state-of-the-art performance in all Re-ID tasks. More experiments demonstrate that ChatReID not only has the ability to recognize fine-grained details but also to integrate them into a coherent reasoning process.

cs.CV↗

arXiv2024

STNMamba: Mamba-based Spatial-Temporal Normality Learning for Video Anomaly Detection

Video anomaly detection (VAD) has been extensively researched due to its potential for intelligent video systems. However, most existing methods based on CNNs and transformers still suffer from substantial computational burdens and have room for improvement in learning spatial-temporal normality. Recently, Mamba has shown great potential for modeling long-range dependencies with linear complexity, providing an effective solution to the above dilemma. To this end, we propose a lightweight and effective Mamba-based network named STNMamba, which incorporates carefully designed Mamba modules to enhance the learning of spatial-temporal normality. Firstly, we develop a dual-encoder architecture, where the spatial encoder equipped with Multi-Scale Vision Space State Blocks (MS-VSSB) extracts multi-scale appearance features, and the temporal encoder employs Channel-Aware Vision Space State Blocks (CA-VSSB) to capture significant motion patterns. Secondly, a Spatial-Temporal Interaction Module (STIM) is introduced to integrate spatial and temporal information across multiple levels, enabling effective modeling of intrinsic spatial-temporal consistency. Within this module, the Spatial-Temporal Fusion Block (STFB) is proposed to fuse the spatial and temporal features into a unified feature space, and the memory bank is utilized to store spatial-temporal prototypes of normal patterns, restricting the model's ability to represent anomalies. Extensive experiments on three benchmark datasets demonstrate that our STNMamba achieves competitive performance with fewer parameters and lower computational costs than existing methods.

cs.CV↗

arXiv2023

Learning Causality-inspired Representation Consistency for Video Anomaly Detection

Video anomaly detection is an essential yet challenging task in the multimedia community, with promising applications in smart cities and secure communities. Existing methods attempt to learn abstract representations of regular events with statistical dependence to model the endogenous normality, which discriminates anomalies by measuring the deviations to the learned distribution. However, conventional representation learning is only a crude description of video normality and lacks an exploration of its underlying causality. The learned statistical dependence is unreliable for diverse regular events in the real world and may cause high false alarms due to overgeneralization. Inspired by causal representation learning, we think that there exists a causal variable capable of adequately representing the general patterns of regular events in which anomalies will present significant variations. Therefore, we design a causality-inspired representation consistency (CRC) framework to implicitly learn the unobservable causal variables of normality directly from available normal videos and detect abnormal events with the learned representation consistency. Extensive experiments show that the causality-inspired normality is robust to regular events with label-independent shifts, and the proposed CRC framework can quickly and accurately detect various complicated anomalies from real-world surveillance videos.

cs.MM↗

arXiv2023

LGN-Net: Local-Global Normality Network for Video Anomaly Detection

Video anomaly detection (VAD) has been intensively studied for years because of its potential applications in intelligent video systems. Existing unsupervised VAD methods tend to learn normality from training sets consisting of only normal videos and regard instances deviating from such normality as anomalies. However, they often consider only local or global normality in the temporal dimension. Some of them focus on learning local spatiotemporal representations from consecutive frames to enhance the representation for normal events. But powerful representation allows these methods to represent some anomalies and causes miss detection. In contrast, the other methods are devoted to memorizing prototypical normal patterns of whole training videos to weaken the generalization for anomalies, which also restricts them from representing diverse normal patterns and causes false alarm. To this end, we propose a two-branch model, Local-Global Normality Network (LGN-Net), to simultaneously learn local and global normality. Specifically, one branch learns the evolution regularities of appearance and motion from consecutive frames as local normality utilizing a spatiotemporal prediction network, while the other branch memorizes prototype features of the whole videos as global normality by a memory module. LGN-Net achieves a balance of representing normal and abnormal instances by fusing local and global normality. In addition, the fused normality enables LGN-Net to generalize to various scenes more than exploiting single normality. Experiments demonstrate the effectiveness and superior performance of our method. The code is available online: https://github.com/Myzhao1999/LGN-Net.

cs.CV↗