Hui Huang

Bing Wang, Xinnian Liang, Jian Yang et al.

Large Language Models (LLMs) are constrained by their inability to process lengthy inputs, resulting in the loss of critical historical information. To address this limitation, in this paper, we propose the Self-Controlled Memory (SCM) framework to enhance the ability of LLMs to maintain long-term memory and recall relevant information. Our SCM framework comprises three key components: an LLM-based agent serving as the backbone of the framework, a memory stream storing agent memories, and a memory controller updating memories and determining when and how to utilize memories from memory stream. Additionally, the proposed SCM is able to process ultra-long texts without any modification or fine-tuning, which can integrate with any instruction following LLMs in a plug-and-play paradigm. Furthermore, we annotate a dataset to evaluate the effectiveness of SCM for handling lengthy inputs. The annotated dataset covers three tasks: long-term dialogues, book summarization, and meeting summarization. Experimental results demonstrate that our method achieves better retrieval recall and generates more informative responses compared to competitive baselines in long-term dialogues. (https://github.com/wbbeyourself/SCM4LLMs)

Ruizhen Hu, Xiangyu Su, Xiangkai Chen et al.

We introduce a method for assigning photorealistic relightable materials to 3D shapes in an automatic manner. Our method takes as input a photo exemplar of a real object and a 3D object with segmentation, and uses the exemplar to guide the assignment of materials to the parts of the shape, so that the appearance of the resulting shape is as similar as possible to the exemplar. To accomplish this goal, our method combines an image translation neural network with a material assignment neural network. The image translation network translates the color from the exemplar to a projection of the 3D shape and the part segmentation from the projection to the exemplar. Then, the material prediction network assigns materials from a collection of realistic materials to the projected parts, based on the translated images and perceptual similarity of the materials. One key idea of our method is to use the translation network to establish a correspondence between the exemplar and shape projection, which allows us to transfer materials between objects with diverse structures. Another key idea of our method is to use the two pairs of (color, segmentation) images provided by the image translation to guide the material assignment, which enables us to ensure the consistency in the assignment. We demonstrate that our method allows us to assign materials to shapes so that their appearances better resemble the input exemplars, improving the quality of the results over the state-of-the-art method, and allowing us to automatically create thousands of shapes with high-quality photorealistic materials. Code and data for this paper are available at https://github.com/XiangyuSu611/TMT.

Kaijie Zheng, Weiqin Wang, Yile Wang et al.

Calculating semantic textual similarity is a foundational task in natural language processing. Current large language models (LLMs) based methods typically rely on extracting last-layer hidden states with fixed dimensions to compute similarity for every text pairs. We argue that this paradigm is suffer from two limitations: (i) The last hidden layer encodes more general knowledge rather than just semantic knowledge, making it suboptimal for semantic similarity computation; (ii) The hidden layer dimensions of LLMs are generally very large, which introduces some redundancy and noise for representing semantics. In this work, we propose DySem, a novel training-free framework that investigates more semantic-related internal components of LLMs via multilingual consensus, and shifts away from static representation spaces in favor of dynamic, sample-specific semantic dimensions by constructing text-dependent joint semantic set and computes similarity over this shared dimensional subset. Extensive experiments across various LLMs show that our method consistently outperforms recent baselines while maintaining lower dimensions for similarity calculation. The code is released at https://github.com/szu-tera/DySem.

Jingyuan Yang, Qirui Huang, Tingting Ding et al.

Visual Emotion Analysis (VEA) aims at predicting people's emotional responses to visual stimuli. This is a promising, yet challenging, task in affective computing, which has drawn increasing attention in recent years. Most of the existing work in this area focuses on feature design, while little attention has been paid to dataset construction. In this work, we introduce EmoSet, the first large-scale visual emotion dataset annotated with rich attributes, which is superior to existing datasets in four aspects: scale, annotation richness, diversity, and data balance. EmoSet comprises 3.3 million images in total, with 118,102 of these images carefully labeled by human annotators, making it five times larger than the largest existing dataset. EmoSet includes images from social networks, as well as artistic images, and it is well balanced between different emotion categories. Motivated by psychological studies, in addition to emotion category, each image is also annotated with a set of describable emotion attributes: brightness, colorfulness, scene type, object class, facial expression, and human action, which can help understand visual emotions in a precise and interpretable way. The relevance of these emotion attributes is validated by analyzing the correlations between them and visual emotion, as well as by designing an attribute module to help visual emotion recognition. We believe EmoSet will bring some key insights and encourage further research in visual emotion analysis and understanding. Project page: https://vcc.tech/EmoSet.

Mengke Li, Zhikai Hu, Yang Lu et al.

The imbalanced distribution of long-tailed data presents a considerable challenge for deep learning models, as it causes them to prioritize the accurate classification of head classes but largely disregard tail classes. The biased decision boundary caused by inadequate semantic information in tail classes is one of the key factors contributing to their low recognition accuracy. To rectify this issue, we propose to augment tail classes by grafting the diverse semantic information from head classes, referred to as head-to-tail fusion (H2T). We replace a portion of feature maps from tail classes with those belonging to head classes. These fused features substantially enhance the diversity of tail classes. Both theoretical analysis and practical experimentation demonstrate that H2T can contribute to a more optimized solution for the decision boundary. We seamlessly integrate H2T in the classifier adjustment stage, making it a plug-and-play module. Its simplicity and ease of implementation allow for smooth integration with existing long-tailed recognition methods, facilitating a further performance boost. Extensive experiments on various long-tailed benchmarks demonstrate the effectiveness of the proposed H2T. The source code is available at https://github.com/Keke921/H2T.

Zhanyu Shen, Sijie Cheng, Zhicheng Guo et al. · tsinghua

While large language models have achieved remarkable performance in complex tasks, they still need a memory system to utilize historical experience in long-term interactions. Existing memory methods (e.g., A-Mem, Mem0) place excessive emphasis on organizing interactions by frequently rewriting them, however, this heavy reliance on summarization risks diluting essential contextual nuances and obscuring key retrieval features. To bridge this gap, we introduce AnchorMem, a novel memory framework inspired by the Proust Phenomenon in cognitive science, where a specific anchor triggers a holistic recollection. We propose a method that decouples the retrieval unit from the generation context. AnchorMem extracts atomic facts from interaction history to serve as retrieval anchors, while preserving the original context as the immutable context. To reveal implicit narrative cues, we construct an associative event graph that uses higher-order event links that bind sets of related facts into shared event representations, strengthening cross-memory integration without relying on generic entities as bridges. During retrieval, the system anchors queries to specific facts and events to locate relevant memories, but reconstructs the context using the associated raw chunks and events. Our method reconciles fine-grained retrieval with the contextual integrity of interactions. Experiments across three closed-source and open-source models on the LoCoMo benchmark demonstrate that AnchorMem significantly outperforms baselines. Code is available at https://github.com/RayNeo-AI-2025/AnchorMem.

Dongsheng Wang, Dawei Su, Hui Huang

Recently, zero-shot 3D scene understanding via 2D Vision-Language Models (VLMs) has gained increasing research interest due to their promising spatial reasoning capabilities. Typically, multiple 2D views are sampled from a 3D point cloud and fed into pre-trained VLMs to answer a given question. This paradigm highlights the critical role of input context quality and raises the challenge of retaining as many task-relevant 3D details as possible under a limited input budget. We propose \texttt{KeyVT}, a hierarchical approach for input context collection at both the view and token levels. Specifically, we combine pixel features with camera parameters and assess view importance based on both semantic content and geometric position, resulting in spatially consistent and task-relevant views. Furthermore, we address redundancy among patches across selected views by identifying representative tokens under the optimal transport (OT) framework, where view tokens and key tokens are formulated as two discrete distributions in the embedding space. These key tokens are expected to cover all view features by minimizing the OT distance. We evaluate our framework on three widely used benchmarks, demonstrating significant improvements over existing tuning-free methods and performance comparable to training-based approaches.

Xinnian Liang, Zefan Zhou, Hui Huang et al.

Pretrained language models (PLMs) have shown marvelous improvements across various NLP tasks. Most Chinese PLMs simply treat an input text as a sequence of characters, and completely ignore word information. Although Whole Word Masking can alleviate this, the semantics in words is still not well represented. In this paper, we revisit the segmentation granularity of Chinese PLMs. We propose a mixed-granularity Chinese BERT (MigBERT) by considering both characters and words. To achieve this, we design objective functions for learning both character and word-level representations. We conduct extensive experiments on various Chinese NLP tasks to evaluate existing PLMs as well as the proposed MigBERT. Experimental results show that MigBERT achieves new SOTA performance on all these tasks. Further analysis demonstrates that words are semantically richer than characters. More interestingly, we show that MigBERT also works with Japanese. Our code and model have been released here~\footnote{https://github.com/xnliang98/MigBERT}.

Hongli Zhou, Hui Huang, Yunfei Long et al.

Recently, there has been a trend of evaluating the Large Language Model (LLM) quality in the flavor of LLM-as-a-Judge, namely leveraging another LLM to evaluate the current output quality. However, existing judges are proven to be biased, namely they would favor answers which present better superficial quality (such as verbosity, fluency) while ignoring the instruction following ability. In this work, we propose systematic research about the bias of LLM-as-a-Judge. Specifically, for closed-source judge models, we apply calibration to mitigate the significance of superficial quality, both on probability level and prompt level. For open-source judge models, we propose to mitigate the bias by contrastive training, with curated negative samples that deviate from instruction but present better superficial quality. We apply our methods on the bias evaluation benchmark, and experiment results show our methods mitigate the bias by a large margin while maintaining a satisfactory evaluation accuracy.

Yilin Liu, Liqiang Lin, Yue Hu et al.

We introduce the first learning-based reconstructability predictor to improve view and path planning for large-scale 3D urban scene acquisition using unmanned drones. In contrast to previous heuristic approaches, our method learns a model that explicitly predicts how well a 3D urban scene will be reconstructed from a set of viewpoints. To make such a model trainable and simultaneously applicable to drone path planning, we simulate the proxy-based 3D scene reconstruction during training to set up the prediction. Specifically, the neural network we design is trained to predict the scene reconstructability as a function of the proxy geometry, a set of viewpoints, and optionally a series of scene images acquired in flight. To reconstruct a new urban scene, we first build the 3D scene proxy, then rely on the predicted reconstruction quality and uncertainty measures by our network, based off of the proxy geometry, to guide the drone path planning. We demonstrate that our data-driven reconstructability predictions are more closely correlated to the true reconstruction quality than prior heuristic measures. Further, our learned predictor can be easily integrated into existing path planners to yield improvements. Finally, we devise a new iterative view planning framework, based on the learned reconstructability, and show superior performance of the new planner when reconstructing both synthetic and real scenes.

Mengke Li, Haiquan Ling, Lihao Chen et al.

Learning from real-world data is frequently hindered by the compound challenge of long-tailed class distributions and noisy annotations. Existing methods partially address these issues but typically ignore the non-uniform impact of label noise across classes, resulting in ineffective correction for tail classes and over-regularization for head classes. To address this issue, we propose Class-Adaptive Rectification with Experts (CARE), a parameter-efficient framework that leverages three complementary supervision sources from vision-language models (VLM): observed noisy labels, VLM text embeddings, and visual features. CARE introduces a class-adaptive expert consensus mechanism that enforces stricter agreement for tail classes and more permissive agreement for head classes based on class frequency. By aggregating high-confidence predictions across these sources, CARE filters unreliable signals and recalibrates class distributions, yielding more reliable rectification under long-tailed distributions. Extensive experiments on both synthetic and real-world benchmarks demonstrate that CARE consistently outperforms state-of-the-art methods, achieving up to 3.0\% performance gains. The source code is available at https://github.com/qwq123-study/CARE.

Yichuan Mo, Hui Huang, Mingjie Li et al.

Diffusion models have achieved notable success in image generation, but they remain highly vulnerable to backdoor attacks, which compromise their integrity by producing specific undesirable outputs when presented with a pre-defined trigger. In this paper, we investigate how to protect diffusion models from this dangerous threat. Specifically, we propose TERD, a backdoor defense framework that builds unified modeling for current attacks, which enables us to derive an accessible reversed loss. A trigger reversion strategy is further employed: an initial approximation of the trigger through noise sampled from a prior distribution, followed by refinement through differential multi-step samplers. Additionally, with the reversed trigger, we propose backdoor detection from the noise space, introducing the first backdoor input detection approach for diffusion models and a novel model detection algorithm that calculates the KL divergence between reversed and benign distributions. Extensive evaluations demonstrate that TERD secures a 100% True Positive Rate (TPR) and True Negative Rate (TNR) across datasets of varying resolutions. TERD also demonstrates nice adaptability to other Stochastic Differential Equation (SDE)-based models. Our code is available at https://github.com/PKU-ML/TERD.

Qijin She, Ruizhen Hu, Juzhan Xu et al.

We approach the problem of high-DOF reaching-and-grasping via learning joint planning of grasp and motion with deep reinforcement learning. To resolve the sample efficiency issue in learning the high-dimensional and complex control of dexterous grasping, we propose an effective representation of grasping state characterizing the spatial interaction between the gripper and the target object. To represent gripper-object interaction, we adopt Interaction Bisector Surface (IBS) which is the Voronoi diagram between two close by 3D geometric objects and has been successfully applied in characterizing spatial relations between 3D objects. We found that IBS is surprisingly effective as a state representation since it well informs the fine-grained control of each finger with spatial relation against the target object. This novel grasp representation, together with several technical contributions including a fast IBS approximation, a novel vector-based reward and an effective training strategy, facilitate learning a strong control model of high-DOF grasping with good sample efficiency, dynamic adaptability, and cross-category generality. Experiments show that it generates high-quality dexterous grasp for complex shapes with smooth grasping motions.

Gengxin Liu, Oliver van Kaick, Hui Huang et al.

Since the preparation of labeled data for training semantic segmentation networks of point clouds is a time-consuming process, weakly supervised approaches have been introduced to learn from only a small fraction of data. These methods are typically based on learning with contrastive losses while automatically deriving per-point pseudo-labels from a sparse set of user-annotated labels. In this paper, our key observation is that the selection of what samples to annotate is as important as how these samples are used for training. Thus, we introduce a method for weakly supervised segmentation of 3D scenes that combines self-training with active learning. The active learning selects points for annotation that likely result in performance improvements to the trained model, while the self-training makes efficient use of the user-provided labels for learning the model. We demonstrate that our approach leads to an effective method that provides improvements in scene segmentation over previous works and baselines, while requiring only a small number of user annotations.

Jie Liu, Zhiba Su, Hui Huang et al.

Thanks to recent advancements in end-to-end speech modeling technology, it has become increasingly feasible to imitate and clone a user`s voice. This leads to a significant challenge in differentiating between authentic and fabricated audio segments. To address the issue of user voice abuse and misuse, the second Audio Deepfake Detection Challenge (ADD 2023) aims to detect and analyze deepfake speech utterances. Specifically, Track 2, named the Manipulation Region Location (RL), aims to pinpoint the location of manipulated regions in audio, which can be present in both real and generated audio segments. We propose our novel TranssionADD system as a solution to the challenging problem of model robustness and audio segment outliers in the trace competition. Our system provides three unique contributions: 1) we adapt sequence tagging task for audio deepfake detection; 2) we improve model generalization by various data augmentation techniques; 3) we incorporate multi-frame detection (MFD) module to overcome limited representation provided by a single frame and use isolated-frame penalty (IFP) loss to handle outliers in segments. Our best submission achieved 2nd place in Track 2, demonstrating the effectiveness and robustness of our proposed system.

Hui Huang, Yanping Chen, Ruizhang Huang et al.

Generative LLMs typically improve Named Entity Recognition (NER) performance through instruction tuning. They excel at generating entities by semantic pattern matching but lack an explicit, verifiable reasoning mechanism. This "cognitive shortcutting" leads to suboptimal performance and brittle generalization, especially in zero-shot and lowresource scenarios where reasoning from limited contextual cues is crucial. To address this issue, a reasoning framework is proposed for NER, which shifts the extraction paradigm from implicit pattern matching to explicit reasoning. This framework consists of three stages: Chain of Thought (CoT) generation, CoT tuning, and reasoning enhancement. First, a dataset annotated with NER-oriented CoTs is generated, which contain task-relevant reasoning chains. Then, they are used to tune the NER model to generate coherent rationales before deriving the final answer. Finally, a reasoning enhancement stage is implemented to optimize the reasoning process using a comprehensive reward signal. This stage ensures explicit and verifiable extractions. Experiments show that ReasoningNER demonstrates impressive cognitive ability in the NER task, achieving competitive performance. In zero-shot settings, it achieves state-of-the-art (SOTA) performance, outperforming GPT-4 by 12.3 percentage points on the F1 score. Analytical results also demonstrate its great potential to advance research in reasoningoriented information extraction. Our codes are available at https://github.com/HuiResearch/ReasoningIE.

Xinnian Liang, Shuangzhi Wu, Hui Huang et al.

Retrieval augmented methods have shown promising results in various classification tasks. However, existing methods focus on retrieving extra context to enrich the input, which is noise sensitive and non-expandable. In this paper, following this line, we propose a $k$-nearest-neighbor (KNN) -based method for retrieval augmented classifications, which interpolates the predicted label distribution with retrieved instances' label distributions. Different from the standard KNN process, we propose a decoupling mechanism as we find that shared representation for classification and retrieval hurts performance and leads to training instability. We evaluate our method on a wide range of classification datasets. Experimental results demonstrate the effectiveness and robustness of our proposed method. We also conduct extra experiments to analyze the contributions of different components in our model.\footnote{\url{https://github.com/xnliang98/knn-cls-w-decoupling}}

Gengxin Liu, Qian Sun, Haibin Huang et al.

Given a 3D object, kinematic motion prediction aims to identify the mobile parts as well as the corresponding motion parameters. Due to the large variations in both topological structure and geometric details of 3D objects, this remains a challenging task and the lack of large scale labeled data also constrain the performance of deep learning based approaches. In this paper, we tackle the task of object kinematic motion prediction problem in a semi-weakly supervised manner. Our key observations are two-fold. First, although 3D dataset with fully annotated motion labels is limited, there are existing datasets and methods for object part semantic segmentation at large scale. Second, semantic part segmentation and mobile part segmentation is not always consistent but it is possible to detect the mobile parts from the underlying 3D structure. Towards this end, we propose a graph neural network to learn the map between hierarchical part-level segmentation and mobile parts parameters, which are further refined based on geometric alignment. This network can be first trained on PartNet-Mobility dataset with fully labeled mobility information and then applied on PartNet dataset with fine-grained and hierarchical part-level segmentation. The network predictions yield a large scale of 3D objects with pseudo labeled mobility information and can further be used for weakly-supervised learning with pre-existing segmentation. Our experiments show there are significant performance boosts with the augmented data for previous method designed for kinematic motion prediction on 3D partial scans.

Yizhi Wang, Zeyu Huang, Ariel Shamir et al.

We introduce anchored radial observations (ARO), a novel shape encoding for learning implicit field representation of 3D shapes that is category-agnostic and generalizable amid significant shape variations. The main idea behind our work is to reason about shapes through partial observations from a set of viewpoints, called anchors. We develop a general and unified shape representation by employing a fixed set of anchors, via Fibonacci sampling, and designing a coordinate-based deep neural network to predict the occupancy value of a query point in space. Differently from prior neural implicit models that use global shape feature, our shape encoder operates on contextual, query-specific features. To predict point occupancy, locally observed shape information from the perspective of the anchors surrounding the input query point are encoded and aggregated through an attention module, before implicit decoding is performed. We demonstrate the quality and generality of our network, coined ARO-Net, on surface reconstruction from sparse point clouds, with tests on novel and unseen object categories, "one-shape" training, and comparisons to state-of-the-art neural and classical methods for reconstruction and tessellation.

Junyu Lu, Qi Wei, Peishuo Zheng et al.

Legal Judgment Prediction (LJP) has become a core benchmark for evaluating AI in the criminal legal domain, but it only sees criminal cases that have already passed prosecutorial review and been formally indicted. As a result, LJP leaves a substantial blind spot in assessing criminal liability, overlooking cases involving insufficient evidence, no criminal liability, or guilt exempted from punishment. To fill this gap, we propose \textbf{Prosecution Decision Prediction (PDP)}, the first Legal AI task built around prosecutorial review, which classifies each case into prosecution or one of three non-prosecution decisions and reflects legal AI's capabilities in evidence evaluation, legal subsumption, and value-based discretion. We further construct \textbf{PDP-Bench}, a benchmark of 4{,}630 real Chinese prosecutorial decisions spanning 190 charges. Extensive experiments show that state-of-the-art LLMs perform substantially worse on PDP than on LJP and that mainstream enhancement routes fail to close the gap. Moreover, controlled RLVR interventions show that simple outcome rewards fail to produce generalizable PDP discrimination.

Jinzuomu Zhong, Yang Li, Hui Huang et al.

In expressive and controllable Text-to-Speech (TTS), explicit prosodic features significantly improve the naturalness and controllability of synthesised speech. However, manual prosody annotation is labor-intensive and inconsistent. To address this issue, a two-stage automatic annotation pipeline is novelly proposed in this paper. In the first stage, we use contrastive pretraining of Speech-Silence and Word-Punctuation (SSWP) pairs to enhance prosodic information in latent representations. In the second stage, we build a multi-modal prosody annotator, comprising pretrained encoders, a text-speech fusing scheme, and a sequence classifier. Experiments on English prosodic boundaries demonstrate that our method achieves state-of-the-art (SOTA) performance with 0.72 and 0.93 f1 score for Prosodic Word and Prosodic Phrase boundary respectively, while bearing remarkable robustness to data scarcity.

Juzhan Xu, Minglun Gong, Hao Zhang et al.

We present a novel learning framework to solve the transport-and-packing (TAP) problem in 3D. It constitutes a full solution pipeline from partial observations of input objects via RGBD sensing and recognition to final box placement, via robotic motion planning, to arrive at a compact packing in a target container. The technical core of our method is a neural network for TAP, trained via reinforcement learning (RL), to solve the NP-hard combinatorial optimization problem. Our network simultaneously selects an object to pack and determines the final packing location, based on a judicious encoding of the continuously evolving states of partially observed source objects and available spaces in the target container, using separate encoders both enabled with attention mechanisms. The encoded feature vectors are employed to compute the matching scores and feasibility masks of different pairings of box selection and available space configuration for packing strategy optimization. Extensive experiments, including ablation studies and physical packing execution by a real robot (Universal Robot UR5e), are conducted to evaluate our method in terms of its design choices, scalability, generalizability, and comparisons to baselines, including the most recent RL-based TAP solution. We also contribute the first benchmark for TAP which covers a variety of input settings and difficulty levels.

Zeyu Huang, Juzhan Xu, Sisi Dai et al.

We introduce NIFT, Neural Interaction Field and Template, a descriptive and robust interaction representation of object manipulations to facilitate imitation learning. Given a few object manipulation demos, NIFT guides the generation of the interaction imitation for a new object instance by matching the Neural Interaction Template (NIT) extracted from the demos in the target Neural Interaction Field (NIF) defined for the new object. Specifically, the NIF is a neural field that encodes the relationship between each spatial point and a given object, where the relative position is defined by a spherical distance function rather than occupancies or signed distances, which are commonly adopted by conventional neural fields but less informative. For a given demo interaction, the corresponding NIT is defined by a set of spatial points sampled in the demo NIF with associated neural features. To better capture the interaction, the points are sampled on the Interaction Bisector Surface (IBS), which consists of points that are equidistant to the two interacting objects and has been used extensively for interaction representation. With both point selection and pointwise features defined for better interaction encoding, NIT effectively guides the feature matching in the NIFs of the new object instances such that the relative poses are optimized to realize the manipulation while imitating the demo interactions. Experiments show that our NIFT solution outperforms state-of-the-art imitation learning methods for object manipulation and generalizes better to objects from new categories.

Hongxiao Zhang, Siyu Lai, Songming Zhang et al.

Translation suggestion (TS) models are used to automatically provide alternative suggestions for incorrect spans in sentences generated by machine translation. This paper introduces the system used in our submission to the WMT'22 Translation Suggestion shared task. Our system is based on the ensemble of different translation architectures, including Transformer, SA-Transformer, and DynamicConv. We use three strategies to construct synthetic data from parallel corpora to compensate for the lack of supervised data. In addition, we introduce a multi-phase pre-training strategy, adding an additional pre-training phase with in-domain data. We rank second and third on the English-German and English-Chinese bidirectional tasks, respectively.

Qi Zhang, Daijie Chen, Yunfei Gong et al.

Existing multi-view crowd counting and localization methods are evaluated under relatively small scenes with limited crowd numbers, camera views, and frames. This makes the evaluation and comparison of existing methods impractical, as small datasets are easily overfit by these methods. To avoid these issues, 3DROM proposes a data augmentation method. Instead, in this paper, we propose a large synthetic benchmark, SynMVCrowd, for more practical evaluation and comparison of multi-view crowd counting and localization tasks. The SynMVCrowd benchmark consists of 50 synthetic scenes with a large number of multi-view frames and camera views and a much larger crowd number (up to 1000), which is more suitable for large-scene multi-view crowd vision tasks. Besides, we propose strong multi-view crowd localization and counting baselines that outperform all comparison methods on the new SynMVCrowd benchmark. Moreover, we prove that better domain transferring multi-view and single-image counting performance could be achieved with the aid of the benchmark on novel new real scenes. As a result, the proposed benchmark could advance the research for multi-view and single-image crowd counting and localization to more practical applications. The codes and datasets are here: https://github.com/zqyq/SynMVCrowd.

Bingxue Zhao, Qi Zhang, Hui Huang

Modeling realistic pedestrian trajectories requires accounting for both social interactions and environmental context, yet most existing approaches largely emphasize social dynamics. We propose \textbf{EnvSocial-Diff}: a diffusion-based crowd simulation model informed by social physics and augmented with environmental conditioning and individual--group interaction. Our structured environmental conditioning module explicitly encodes obstacles, objects of interest, and lighting levels, providing interpretable signals that capture scene constraints and attractors. In parallel, the individual--group interaction module goes beyond individual-level modeling by capturing both fine-grained interpersonal relations and group-level conformity through a graph-based design. Experiments on multiple benchmark datasets demonstrate that EnvSocial-Diff outperforms the latest state-of-the-art methods, underscoring the importance of explicit environmental conditioning and multi-level social interaction for realistic crowd simulation. Code is here: https://github.com/zqyq/EnvSocial-Diff.

Zihao Yan, Fubao Su, Mingyang Wang et al.

We introduce an active 3D reconstruction method which integrates visual perception, robot-object interaction, and 3D scanning to recover both the exterior and interior, i.e., unexposed, geometries of a target 3D object. Unlike other works in active vision which focus on optimizing camera viewpoints to better investigate the environment, the primary feature of our reconstruction is an analysis of the interactability of various parts of the target object and the ensuing part manipulation by a robot to enable scanning of occluded regions. As a result, an understanding of part articulations of the target object is obtained on top of complete geometry acquisition. Our method operates fully automatically by a Fetch robot with built-in RGBD sensors. It iterates between interaction analysis and interaction-driven reconstruction, scanning and reconstructing detected moveable parts one at a time, where both the articulated part detection and mesh reconstruction are carried out by neural networks. In the final step, all the remaining, non-articulated parts, including all the interior structures that had been exposed by prior part manipulations and subsequently scanned, are reconstructed to complete the acquisition. We demonstrate the performance of our method via qualitative and quantitative evaluation, ablation studies, comparisons to alternatives, as well as experiments in a real environment.

Hui Huang, Xingyuan Bu, Hongli Zhou et al.

Recently, there has been a growing trend of utilizing Large Language Model (LLM) to evaluate the quality of other LLMs. Many studies have fine-tuned judge models based on open-source LLMs for evaluation. While the fine-tuned judge models are claimed to achieve comparable evaluation capability with GPT-4, in this work, we conduct an empirical study of LLM-as-a-Judge. Our findings indicate that although the fine-tuned judge models achieve high performance on in-domain test sets, even surpassing GPT-4, they underperform GPT-4 across several dimensions, including generalizability, fairness and adaptability. We also reveal that the fine-tuned judge model inherently operates as a task-specific classifier, consequently imposing the limitations.

Mengke Li, Da Li, Guoqing Yang et al.

Pre-trained large-scale models have exhibited remarkable efficacy in computer vision, particularly for 2D image analysis. However, when it comes to 3D point clouds, the constrained accessibility of data, in contrast to the vast repositories of images, poses a challenge for the development of 3D pre-trained models. This paper therefore attempts to directly leverage pre-trained models with 2D prior knowledge to accomplish the tasks for 3D point cloud analysis. Accordingly, we propose the Adaptive PointFormer (APF), which fine-tunes pre-trained 2D models with only a modest number of parameters to directly process point clouds, obviating the need for mapping to images. Specifically, we convert raw point clouds into point embeddings for aligning dimensions with image tokens. Given the inherent disorder in point clouds, in contrast to the structured nature of images, we then sequence the point embeddings to optimize the utilization of 2D attention priors. To calibrate attention across 3D and 2D domains and reduce computational overhead, a trainable PointFormer with a limited number of parameters is subsequently concatenated to a frozen pre-trained image model. Extensive experiments on various benchmarks demonstrate the effectiveness of the proposed APF. The source code and more details are available at https://vcc.tech/research/2024/PointFormer.

Chaoqi Chen, Qile Xu, Wenjun Zhou et al.

Understanding 3D point clouds through language remains a fundamental challenge in computer graphics and visual computing, due to the irregular structure of point cloud data and the lack of explicit reasoning in existing 3D multimodal models. While Chain-of-Thought (CoT) reasoning has shown strong effectiveness in LLMs and image-based MLLMs, its extension to 3D understanding remains largely underexplored. In this paper, we propose a data-centric framework for constructing large-scale CoT supervision tailored to 3D point cloud understanding. Our framework consists of a two-stage pipeline that first refines point-text instruction data via vision-language-model-based quality evaluation and reference-guided refinement, and then synthesizes high-quality reasoning paths through Human-in-the-Loop Prompt Optimization (HiLPO). Using this approach, we build PoCoTI, a CoT-enhanced point-text instruction-following dataset containing 55K samples with explicit reasoning paths. Fine-tuning PointLLM on PoCoTI yields PointLLM-R, a reasoning-capable 3D multimodal language model. Extensive experiments on generative 3D classification and captioning demonstrate that PointLLM-R achieves state-of-the-art performance and generalizes robustly to real-world scanned point clouds and multi-turn dialogue scenarios.

Xin Hu, Pengfei Xu, Jin Zhou et al.

Structured layouts are preferable in many 2D visual contents (\eg, GUIs, webpages) since the structural information allows convenient layout editing. Computational frameworks can help create structured layouts but require heavy labor input. Existing data-driven approaches are effective in automatically generating fixed layouts but fail to produce layout structures. We present StructLayoutFormer, a novel Transformer-based approach for conditional structured layout generation. We use a structure serialization scheme to represent structured layouts as sequences. To better control the structures of generated layouts, we disentangle the structural information from the element placements. Our approach is the first data-driven approach that achieves conditional structured layout generation and produces realistic layout structures explicitly. We compare our approach with existing data-driven layout generation approaches by including post-processing for structure extraction. Extensive experiments have shown that our approach exceeds these baselines in conditional structured layout generation. We also demonstrate that our approach is effective in extracting and transferring layout structures. The code is publicly available at %\href{https://github.com/Teagrus/StructLayoutFormer} {https://github.com/Teagrus/StructLayoutFormer}.

Yang Zhou, Xu Gao, Zichong Chen et al.

Recent advances in generative diffusion models have shown a notable inherent understanding of image style and semantics. In this paper, we leverage the self-attention features from pretrained diffusion networks to transfer the visual characteristics from a reference to generated images. Unlike previous work that uses these features as plug-and-play attributes, we propose a novel attention distillation loss calculated between the ideal and current stylization results, based on which we optimize the synthesized image via backpropagation in latent space. Next, we propose an improved Classifier Guidance that integrates attention distillation loss into the denoising sampling process, further accelerating the synthesis and enabling a broad range of image generation applications. Extensive experiments have demonstrated the extraordinary performance of our approach in transferring the examples' style, appearance, and texture to new images in synthesis. Code is available at https://github.com/xugao97/AttentionDistillation.

Mengke Li, Ye Liu, Yang Lu et al.

Long-tail learning has garnered widespread attention and achieved significant progress in recent times. However, even with pre-trained prior knowledge, models still exhibit weaker generalization performance on tail classes. The promising Sharpness-Aware Minimization (SAM) can effectively improve the generalization capability of models by seeking out flat minima in the loss landscape, which, however, comes at the cost of doubling the computational time. Since the update rule of SAM necessitates two consecutive (non-parallelizable) forward and backpropagation at each step. To address this issue, we propose a novel method called Random SAM prompt tuning (RSAM-PT) to improve the model generalization, requiring only one-step gradient computation at each step. Specifically, we search for the gradient descent direction within a random neighborhood of the parameters during each gradient update. To amplify the impact of tail-class samples and avoid overfitting, we employ the deferred re-weight scheme to increase the significance of tail-class samples. The classification accuracy of long-tailed data can be significantly improved by the proposed RSAM-PT, particularly for tail classes. RSAM-PT achieves the state-of-the-art performance of 90.3\%, 76.5\%, and 50.1\% on benchmark datasets CIFAR100-LT (IF 100), iNaturalist 2018, and Places-LT, respectively. The source code is temporarily available at https://github.com/Keke921/GNM-PT.

Honghao Xu, Juzhan Xu, Zeyu Huang et al.

In this paper, we introduce a novel method called FRI-Net for 2D floorplan reconstruction from 3D point cloud. Existing methods typically rely on corner regression or box regression, which lack consideration for the global shapes of rooms. To address these issues, we propose a novel approach using a room-wise implicit representation with structural regularization to characterize the shapes of rooms in floorplans. By incorporating geometric priors of room layouts in floorplans into our training strategy, the generated room polygons are more geometrically regular. We have conducted experiments on two challenging datasets, Structured3D and SceneCAD. Our method demonstrates improved performance compared to state-of-the-art methods, validating the effectiveness of our proposed representation for floorplan reconstruction.

Weidan Xiong, Hongqian Zhang, Botao Peng et al.

Coarse architectural models are often generated at scales ranging from individual buildings to scenes for downstream applications such as Digital Twin City, Metaverse, LODs, etc. Such piece-wise planar models can be abstracted as twins from 3D dense reconstructions. However, these models typically lack realistic texture relative to the real building or scene, making them unsuitable for vivid display or direct reference. In this paper, we present TwinTex, the first automatic texture mapping framework to generate a photo-realistic texture for a piece-wise planar proxy. Our method addresses most challenges occurring in such twin texture generation. Specifically, for each primitive plane, we first select a small set of photos with greedy heuristics considering photometric quality, perspective quality and facade texture completeness. Then, different levels of line features (LoLs) are extracted from the set of selected photos to generate guidance for later steps. With LoLs, we employ optimization algorithms to align texture with geometry from local to global. Finally, we fine-tune a diffusion model with a multi-mask initialization component and a new dataset to inpaint the missing region. Experimental results on many buildings, indoor scenes and man-made objects of varying complexity demonstrate the generalization ability of our algorithm. Our approach surpasses state-of-the-art texture mapping methods in terms of high-fidelity quality and reaches a human-expert production level with much less effort. Project page: https://vcc.tech/research/2023/TwinTex.

Yang Zhou, Zichong Chen, Hui Huang

This paper addresses the complex issue of one-shot face stylization, focusing on the simultaneous consideration of appearance and structure, where previous methods have fallen short. We explore deformation-aware face stylization that diverges from traditional single-image style reference, opting for a real-style image pair instead. The cornerstone of our method is the utilization of a self-supervised vision transformer, specifically DINO-ViT, to establish a robust and consistent facial structure representation across both real and style domains. Our stylization process begins by adapting the StyleGAN generator to be deformation-aware through the integration of spatial transformers (STN). We then introduce two innovative constraints for generator fine-tuning under the guidance of DINO semantics: i) a directional deformation loss that regulates directional vectors in DINO space, and ii) a relative structural consistency constraint based on DINO token self-similarities, ensuring diverse generation. Additionally, style-mixing is employed to align the color generation with the reference, minimizing inconsistent correspondences. This framework delivers enhanced deformability for general one-shot face stylization, achieving notable efficiency with a fine-tuning duration of approximately 10 minutes. Extensive qualitative and quantitative comparisons demonstrate our superiority over state-of-the-art one-shot face stylization methods. Code is available at https://github.com/zichongc/DoesFS

Yang Zhou, Rongjun Xiao, Dani Lischinski et al.

This paper addresses the challenge of example-based non-stationary texture synthesis. We introduce a novel twostep approach wherein users first modify a reference texture using standard image editing tools, yielding an initial rough target for the synthesis. Subsequently, our proposed method, termed "self-rectification", automatically refines this target into a coherent, seamless texture, while faithfully preserving the distinct visual characteristics of the reference exemplar. Our method leverages a pre-trained diffusion network, and uses self-attention mechanisms, to gradually align the synthesized texture with the reference, ensuring the retention of the structures in the provided target. Through experimental validation, our approach exhibits exceptional proficiency in handling non-stationary textures, demonstrating significant advancements in texture synthesis when compared to existing state-of-the-art techniques. Code is available at https://github.com/xiaorongjun000/Self-Rectification

Mengke Li, Haiquan Ling, Yiqun Zhang et al.

Real-world data often exhibit long-tailed distributions with numerous noisy labels, substantially degrading the performance of deep models. While prior research has made progress in addressing this combined challenge, it overlooks the severe label-image mismatch inherent to high-noise settings, thereby limiting their effectiveness. Given that observed labels, though mismatched with images, still retain category information, we propose employing auxiliary text information from labels to address label-image inconsistencies in long-tailed noisy data. Specifically, we leverage the intrinsic cross-modal alignment in pre-trained visual-language models to correct the label-image inconsistencies. This supervisory signal, referred to as Weak Teacher Supervision (WTS), is unaffected by label noise and data distribution biases, albeit exhibits limited accuracy. Therefore, the activation of WTS is determined by evaluating the discrepancy between text-predicted labels and observed labels. Extensive experiments demonstrate the superior performance of WTS across synthetic and real-world datasets, particularly under high-noise conditions. The source code is available at https://anonymous.4open.science/r/WTS-0F3C.

Gheorghe Comanici, Eric Bieber, Mike Schaekermann et al. · amazon-science, baidu

In this report, we introduce the Gemini 2.X model family: Gemini 2.5 Pro and Gemini 2.5 Flash, as well as our earlier Gemini 2.0 Flash and Flash-Lite models. Gemini 2.5 Pro is our most capable model yet, achieving SoTA performance on frontier coding and reasoning benchmarks. In addition to its incredible coding and reasoning skills, Gemini 2.5 Pro is a thinking model that excels at multimodal understanding and it is now able to process up to 3 hours of video content. Its unique combination of long context, multimodal and reasoning capabilities can be combined to unlock new agentic workflows. Gemini 2.5 Flash provides excellent reasoning abilities at a fraction of the compute and latency requirements and Gemini 2.0 Flash and Flash-Lite provide high performance at low latency and cost. Taken together, the Gemini 2.X model generation spans the full Pareto frontier of model capability vs cost, allowing users to explore the boundaries of what is possible with complex agentic problem solving.

Hongli Zhou, Hui Huang, Wei Liu et al.

Reward models (RMs) play a pivotal role in aligning large language models (LLMs) with human preferences. Due to the difficulty of obtaining high-quality human preference annotations, distilling preferences from generative LLMs has emerged as a standard practice. However, existing approaches predominantly treat teacher models as simple binary annotators, failing to fully exploit the rich knowledge and capabilities for RM distillation. To address this, we propose RM-Distiller, a framework designed to systematically exploit the multifaceted capabilities of teacher LLMs: (1) Refinement capability, which synthesizes highly correlated response pairs to create fine-grained and contrastive signals. (2) Scoring capability, which guides the RM in capturing precise preference strength via a margin-aware optimization objective. (3) Generation capability, which incorporates the teacher's generative distribution to regularize the RM to preserve its fundamental linguistic knowledge. Extensive experiments demonstrate that RM-Distiller significantly outperforms traditional distillation methods both on RM benchmarks and reinforcement learning-based alignment, proving that exploiting multifaceted teacher capabilities is critical for effective reward modeling. To the best of our knowledge, this is the first systematic research on RM distillation from generative LLMs.

Qi Zhang, Kaiyi Zhang, Antoni B. Chan et al.

Multi-view crowd localization predicts the ground locations of all people in the scene. Typical methods usually estimate the crowd density maps on the ground plane first, and then obtain the crowd locations. However, the performance of existing methods is limited by the ambiguity of the density maps in crowded areas, where local peaks can be smoothed away. To mitigate the weakness of density map supervision, optimal transport-based point supervision methods have been proposed in the single-image crowd localization tasks, but have not been explored for multi-view crowd localization yet. Thus, in this paper, we propose a novel Mahalanobis distance-based multi-view optimal transport (M-MVOT) loss specifically designed for multi-view crowd localization. First, we replace the Euclidean-based transport cost with the Mahalanobis distance, which defines elliptical iso-contours in the cost function whose long-axis and short-axis directions are guided by the view ray direction. Second, the object-to-camera distance in each view is used to adjust the optimal transport cost of each location further, where the wrong predictions far away from the camera are more heavily penalized. Finally, we propose a strategy to consider all the input camera views in the model loss (M-MVOT) by computing the optimal transport cost for each ground-truth point based on its closest camera. Experiments demonstrate the advantage of the proposed method over density map-based or common Euclidean distance-based optimal transport loss on several multi-view crowd localization datasets. Project page: https://vcc.tech/research/2024/MVOT.

Junkai Liu, Yujie Tong, Hui Huang et al.

Legal judgment prediction (LJP), which enables litigants and their lawyers to forecast judgment outcomes and refine litigation strategies, has emerged as a crucial legal NLP task. Existing studies typically utilize legal facts, i.e., facts that have been established by evidence and determined by the judge, to predict the judgment. However, legal facts are often difficult to obtain in the early stages of litigation, significantly limiting the practical applicability of fact-based LJP. To address this limitation, we propose a novel legal NLP task: legal fact prediction (LFP), which takes the evidence submitted by litigants for trial as input to predict legal facts, thereby empowering fact-based LJP technologies to make predictions in the absence of ground-truth legal facts. We also propose the first benchmark dataset, LFPBench, for evaluating the LFP task. Our extensive experiments on LFPBench demonstrate the effectiveness of LFP-empowered LJP and highlight promising research directions for LFP.

Xueqi Ma, Xingguang Yan, Congyue Zhang et al.

We present MeshTailor, the first mesh-native generative framework for synthesizing edge-aligned seams on 3D surfaces. Unlike prior optimization-based or extrinsic learning-based methods, MeshTailor operates directly on the mesh graph, eliminating projection artifacts and fragile snapping heuristics. We introduce ChainingSeams, a hierarchical serialization of the seam graph that prioritizes global structural cuts before local details in a coarse-to-fine manner, and a dual-stream encoder that fuses topological and geometric context. Leveraging this hierarchical representation and enriched vertex embeddings, our MeshTailor Transformer utilizes an autoregressive pointer layer to trace seams vertex-by-vertex within local neighborhoods, ensuring projection-free, edge-aligned seams. Extensive evaluations show that MeshTailor produces more coherent, professional-quality seam layouts compared to recent optimization-based and learning-based baselines.

Qi Zhang, Jixuan Chen, Kaiyi Zhang et al.

Multi-view crowd tracking estimates each person's tracking trajectories on the ground of the scene. Recent research works mainly rely on CNNs-based multi-view crowd tracking architectures, and most of them are evaluated and compared on relatively small datasets, such as Wildtrack and MultiviewX. Since these two datasets are collected in small scenes and only contain tens of frames in the evaluation stage, it is difficult for the current methods to be applied to real-world applications where scene size and occlusion are more complicated. In this paper, we propose a Transformer-based multi-view crowd tracking model, \textit{MVTrackTrans}, which adopts interactions between camera views and the ground plane for enhanced multi-view tracking performance. Besides, for better evaluation, we collect and label two large real-world multi-view tracking datasets, MVCrowdTrack and CityTrack, which contain a much larger scene size over a longer time period. Compared with existing methods on the two large and new datasets, the proposed MVTrackTrans model achieves better performance, demonstrating the advantages of the model design in dealing with large scenes. We believe the proposed datasets and model will push the frontiers of the task to more practical scenarios, and the datasets and code are available at: https://github.com/zqyq/MVTrackTrans.

Hui Huang, Jiaheng Liu, Yancheng He et al.

Complex instruction-following with elaborate constraints is imperative for Large Language Models (LLMs). While existing methods have constructed data for complex instruction alignment, they all rely on a more advanced model, especially GPT-4, limiting their application. In this paper, we propose a Multi-granularity Self-Contrastive Training (MuSC) framework, to improve the complex instruction alignment without relying on a stronger model. Our method is conducted on both coarse and fine granularity. On coarse-granularity, we construct constraint-aware preference data based on instruction decomposition and recombination. On fine-granularity, we perform token-aware preference optimization with dynamic token-level supervision. Our method is evaluated on open-sourced models, and experiment results show our method achieves significant improvement on both complex and general instruction-following benchmarks, surpassing previous self-alignment methods.

Yan Gao, Yazheng Yang, Zhibin Lan et al.

Code-switching (CS) speech translation (ST) refers to translating speech that alternates between two or more languages into a target language text, which poses significant challenges due to the complexity of semantic modeling and the scarcity of CS data. Previous studies tend to rely on the model itself to implicitly learn semantic modeling during training, and resort to inefficient and costly manual annotations for these two challenges. To mitigate these limitations, we propose enhancing Large Language Models (LLMs) with a Mixture of Experts (MoE) speech projector, where each expert specializes in the semantic subspace of a specific language, enabling fine-grained modeling of speech features. Additionally, we introduce a multi-stage training paradigm that utilizes readily available monolingual automatic speech recognition (ASR) and monolingual ST data, facilitating speech-text alignment and improving translation capabilities. During training, we leverage a combination of language-specific loss and intra-group load balancing loss to guide the MoE speech projector in efficiently allocating tokens to the appropriate experts, across expert groups and within each group, respectively. To bridge the data gap across different training stages and improve adaptation to the CS scenario, we further employ a transition loss, enabling smooth transitions of data between stages, to effectively address the scarcity of high-quality CS speech translation data. Extensive experiments on widely used datasets demonstrate the effectiveness and generality of our approach.

Hui Huang, Yingqi Qu, Jing Liu et al.

The proliferation of open-source Large Language Models (LLMs) underscores the pressing need for evaluation methods. Existing works primarily rely on external evaluators, focusing on training and prompting strategies. However, a crucial aspect, model-aware glass-box features, is overlooked. In this study, we explore the utility of glass-box features under the scenario of self-evaluation, namely applying an LLM to evaluate its own output. We investigate various glass-box feature groups and discovered that the softmax distribution serves as a reliable quality indicator for self-evaluation. Experimental results on public benchmarks validate the feasibility of self-evaluation of LLMs using glass-box features.

Mingrong Gong, Chaoqi Chen, Qingqiang Sun et al.

Out-of-distribution (OOD) detection is a crucial task for deploying deep learning models in the wild. One of the major challenges is that well-trained deep models tend to perform over-confidence on unseen test data. Recent research attempts to leverage real or synthetic outliers to mitigate the issue, which may significantly increase computational costs and be biased toward specific outlier characteristics. In this paper, we propose a simple yet effective framework, Prototypical Outlier Proxy (POP), which introduces virtual OOD prototypes to reshape the decision boundaries between ID and OOD data. Specifically, we transform the learnable classifier into a fixed one and augment it with a set of prototypical weight vectors. Then, we introduce a hierarchical similarity boundary loss to impose adaptive penalties depending on the degree of misclassification. Extensive experiments across various benchmarks demonstrate the effectiveness of POP. Notably, POP achieves average FPR95 reductions of 7.70%, 6.30%, and 5.42% over the second-best methods on CIFAR-10, CIFAR-100, and ImageNet-200, respectively. Moreover, compared to the recent method NPOS, which relies on outlier synthesis, POP trains 7.2X faster and performs inference 19.5X faster. The source code is available at: https://github.com/gmr523/pop.

Jianyu Liu, Hangyu Guo, Ranjie Duan et al.

Multimodal Large Language Models (MLLMs) pose unique safety challenges due to their integration of visual and textual data, thereby introducing new dimensions of potential attacks and complex risk combinations. In this paper, we begin with a detailed analysis aimed at disentangling risks through step-by-step reasoning within multimodal inputs. We find that systematic multimodal risk disentanglement substantially enhances the risk awareness of MLLMs. Via leveraging the strong discriminative abilities of multimodal risk disentanglement, we further introduce \textbf{DREAM} (\textit{\textbf{D}isentangling \textbf{R}isks to \textbf{E}nhance Safety \textbf{A}lignment in \textbf{M}LLMs}), a novel approach that enhances safety alignment in MLLMs through supervised fine-tuning and iterative Reinforcement Learning from AI Feedback (RLAIF). Experimental results show that DREAM significantly boosts safety during both inference and training phases without compromising performance on normal tasks (namely oversafety), achieving a 16.17\% improvement in the SIUO safe\&effective score compared to GPT-4V. The data and code are available at https://github.com/Kizna1ver/DREAM.

Jingyuan Yang, Rucong Chen, Hui Huang

Story generation aims to produce image sequences that depict coherent narratives while maintaining subject consistency across frames. Although existing methods have excelled in producing coherent and expressive stories, they remain largely emotion-neutral, focusing on what subject appears in a story while overlooking how emotions shape narrative interpretation and visual presentation. As stories are intended to engage audiences emotionally, we introduce emotion-aware story generation, a new task that aims to generate subject-consistent visual stories with explicit emotional directions. This task is challenging due to the abstract nature of emotions, which must be grounded in concrete visual elements and consistently expressed across a narrative through visual composition. To address these challenges, we propose EmoStory, a two-stage framework that integrates agent-based story planning and region-aware story generation. The planning stage transforms target emotions into coherent story prompts with emotion agent and writer agent, while the generation stage preserves subject consistency and injects emotion-related elements through region-aware composition. We evaluate EmoStory on a newly constructed dataset covering 25 subjects and 600 emotional stories. Extensive quantitative and qualitative results, along with user studies, show that EmoStory outperforms state-of-the-art story generation methods in emotion accuracy, prompt alignment, and subject consistency.