AAAI 2025 论文列表

Recent advances in text-to-image diffusion models have shown an outstanding ability in zero-shot style transfer. However, existing methods often struggle to balance preserving the semantic content of the input image and faithfully transferring the target style in line with the edit prompt. Especially when applied to complex traffic scenes with diverse objects, layouts, and stylistic variations, current diffusion models tend to exhibit Style Neglection, i.e., failing to generate the required style in the prompt. To address this issue, we propose Style Nursing, which directs the model to focus on style subject tokens in the text prompt and excites their corresponding visual activations. Moreover, we introduce Spatial and Semantic Guidance to guide the preservation of content after editing, which utilizes spatial features from the DDIM sampling process together with attention maps from the semantic reconstruction. To evaluate the performance of zero-shot style transfer methods in traffic scenes, we present STREET-6K, a new benchmark dataset comprising 6000 images showcasing diverse traffic scenes and style transfer variations, accompanied by comprehensive annotations and evaluation metrics. Our approach beats state-of-the-art image translation methods in comprehensive quantitative metrics and human evaluations on traffic scene image synthesis while seamlessly generalizing to various other types of images without training or fine-tuning. Further experiments on detection and segmentation tasks show that fine-tuning perception models on our synthesized images improves Recall and mean Intersection over Union (mIoU) by over 10% and 3% respectively in rarely-seen traffic scenes.

This paper focuses on face stylization with a single artistic target. Existing works for this task often fail to retain the source content while achieving geometry variation. Here, we present a novel StyO model, i.e., Stylize the face in only One-shot, to solve the above problem. In particular, StyO exploits a disentanglement and recombination strategy. It first disentangles the content and style of source and target images into identifiers, which are then recombined in a cross manner to derive the stylized face image. In this way, StyO decomposes complex images into independent and specific attributes, and simplifies one-shot face stylization as the combination of different attributes from input images, thus producing results better matching face geometry of target image and content of source one. StyO is implemented with latent diffusion models (LDM) and composed of two key modules: 1) Identifier Disentanglement Learner (IDL) for disentanglement phase. It represents identifiers as contrastive text prompts, i.e. positive and negative descriptions. And it introduces a novel triple reconstruction loss to fine-tune the pre-trained LDM for encoding style and content into corresponding identifiers; 2) Fine-graind Content Controller (FCC) for recombination phase. It recombines disentangled identifiers from IDL to form an augmented text prompt for generating stylized faces. In addition, FCC also constrains the cross-attention maps of latent and text features to preserve source face details in results. The extensive evaluation shows that StyO produces high-quality images on numerous paintings of various styles and outperforms the current state-of-the-art.

Recent advancements in long-context language modeling have attracted significant attention, yet their practical applications often suffer from suboptimal context utilization. To efficiently address this issue, we introduce the Structured Packing for Long Context, SPLiCe, a method that uses retrieval to collate mutually relevant documents into long training samples. We demonstrate that SPLiCe improves performance on long-context tasks, particularly by achieving perfect accuracy on the synthetic Needle in the Haystack benchmark, and effectively mitigating the ‘lost-in-the-middle’ phenomenon often observed in large language models. Notably, these long-context capabilities also extend to realistic downstream tasks, such as Qasper, across multiple model sizes—3B, 7B, and 13B—and are achieved with only brief fine-tuning on 2-6 billion tokens. We supplement these results with a detailed analysis of SPLiCe, examining the impact of hyperparameter choices, the different mixtures and proportions of SPLiCe-generated training data, and the choice of the retriever. We also study the transfer of long-context utilization skills between the modalities. An intriguing finding from our analysis is that training on a corpus of code can enhance performance on natural language tasks.

The Information Bottleneck (IB) principle has emerged as a promising approach for enhancing the generalization, robustness, and interpretability of deep neural networks, demonstrating efficacy across image segmentation, document clustering, and semantic communication. Among IB implementations, the IB Lagrangian method, employing Lagrangian multipliers, is widely adopted. While numerous methods for the optimizations of IB Lagrangian based on variational bounds and neural estimators are feasible, their performance is highly dependent on the quality of their design, which is inherently prone to errors. To address this limitation, we introduce Structured IB, a framework for investigating potential structured features. By incorporating auxiliary encoders to extract missing informative features, we generate more informative representations. Our experiments demonstrate superior prediction accuracy and task-relevant information preservation compared to the original IB Lagrangian method, even with reduced network size.

We describe an application that uses large language models to generate structured documents related to industrial equipment, specifically focusing on Failure Modes and Effects Analysis (FMEAs). Our novel application uses techniques in structured document generation, in-context learning, and ensembling to create high-quality structured content that subject matter experts supervise through a user-centric interface that presents FMEA entities as UI elements. Novel evaluation metrics for structured document generation are also proposed. Our empirical results, based on 71 asset evaluations, demonstrate the individual and combined contributions of these techniques, with an overall effectiveness that varies between a recall of 0.669 and a precision of 0.91. Qualitative feedback from target users validates the practicality of the described approach to seamlessly integrate expert supervision with generative AI in a labour-saving workflow.

Multi-view clustering (MVC) for remote sensing data is a critical and challenging task in Earth observation. Although recent advances in graph neural network (GNN)-based MVC have shown remarkable success, the most prevalent approaches have two major limitations: 1) heavily relying on a predefined yet fixed graph, which limits the performance of clustering because the large number of indistinguishable background samples contained in remote sensing data would introduce noise information and increase structure heterogeneity; 2) ignoring the effect of confusing samples on cluster structure compactness, which leads to fluffy cluster structure and decrease feature discriminability. To address these issues, we propose a Structure-Adaptive Multi-View Graph Clustering method named SAMVGC on remote sensing data which boosts the structure homogeneity and cluster compactness by adaptively learning the graph and cluster structures, respectively. Concretely, we use the geometric structure within the feature embedding space to refine adjacency matrices. The adjacency matrices are dynamically fused with the previous ones to improve the homogeneity and stability of structure information. Additionally, the samples are separated into two categories, including the central (intra-cluster center samples) and the confusing (inter-cluster boundary samples). On the basis, we deploy the contrastive learning paradigm on the central samples within views and the consistent learning paradigm on the confusing samples between views, improving the cluster compactness and consistency. Finally, we conduct extensive experiments on four benchmarks and achieve promising results, well demonstrating the effectiveness and superiority of the proposed method.

Training graph neural networks (GNNs) for graph representation has received increasing concerns due to its outstanding performance in the link prediction and node classification tasks, but it incurs much time and storage for tackling large-scale graphs. To alleviate this issue, graph condensation has been emerged to condense the large graph into a small but highly-informative graph, while achieving comparable performance of GNNs trained on the small graph and large graph. However, existing works mainly focus on the gradient or distribution matching under GNN training trajectories to condense simple link structures, while overlooking the structure matching for condensing signed graph that exists conflict links and structural balance among nodes. To bridge this gap, we propose a novel Structure Balance and Gradient Matching-Based Signed Graph Condensation (SGSGC) method for condensing signed graph with node attributes, conflict links and structural balance into informative smaller ones. Specifically, we first propose a structure-balanced matching to match the structural balance between the original and condensed signed graph, and then combine it with the gradient matching to condense signed graph for the link sign prediction task, while preserving both conflicting link structures and node attributes. Moreover, we use the feature smoothing and the graph sparsification technique to improve the robustness for the GNN training, respectively. Finally, a bi-level optimization technique is proposed to simultaneously find the optimal node attributes and conflict structure of the condensed graph. Experiments on six datasets demonstrate that SGSGC achieves excellent performance. On Epinions, 94% test accuracy of training on the original signed graph, while reducing their graph size by 99.95% - 99.99%, and there exist 2.24% – 6.26% accuracy improvements for link sign prediction compared to the state-of-the-arts.

In recent years, semantic segmentation has flourished in various applications. However, the high computational cost remains a significant challenge that hinders its further adoption. The filter pruning method for structured network slimming offers a direct and effective solution for the reduction of segmentation networks. Nevertheless, we argue that most existing pruning methods, originally designed for image classification, overlook the fact that segmentation is a location-sensitive task, which consequently leads to their suboptimal performance when applied to segmentation networks. To address this issue, this paper proposes a novel approach, denoted as Spatial-aware Information Redundancy Filter Pruning (SIRFP), which aims to reduce feature redundancy between channels. First, we formulate the pruning process as a maximum edge weight clique problem (MEWCP) in graph theory, thereby minimizing the redundancy among the remaining features after pruning. Within this framework, we introduce a spatial-aware redundancy metric based on feature maps, thus endowing the pruning process with location sensitivity to better adapt to pruning segmentation networks. Additionally, based on the MEWCP, we propose a low computational complexity greedy strategy to solve this NP-hard problem, making it feasible and efficient for structured pruning. To validate the effectiveness of our method, we conducted extensive comparative experiments on various challenging datasets. The results demonstrate the superior performance of SIRFP for semantic segmentation tasks.

With the emerging of huge amount of unlabeled data, unsupervised out-of-distribution (OOD) detection is vital for ensuring the reliability of graph neural networks (GNNs) by identifying OOD samples from in-distribution (ID) ones during testing, where encountering novel or unknown data is inevitable. Existing methods often suffer from compromised performance due to redundant information in graph structures, which impairs their ability to effectively differentiate between ID and OOD data. To address this challenge, we propose SEGO, an unsupervised framework that integrates structural entropy into OOD detection regarding graph classification. Specifically, within the architecture of contrastive learning, SEGO introduces an anchor view in the form of coding tree by minimizing structural entropy. The obtained coding tree effectively removes redundant information from graphs while preserving essential structural information, enabling the capture of distinct graph patterns between ID and OOD samples. Furthermore, we present a multi-grained contrastive learning scheme at local, global, and tree levels using triplet views, where coding trees with essential information serve as the anchor view. Extensive experiments on real-world datasets validate the effectiveness of SEGO, demonstrating superior performance over state-of-the-art baselines in OOD detection. Specifically, our method achieves the best performance on 9 out of 10 dataset pairs, with an average improvement of 3.7% on OOD detection datasets, significantly surpassing the best competitor by 10.8% on the FreeSolv/ToxCast dataset pair.

Probabilistic embeddings have several advantages over deterministic embeddings as they map each data point to a distribution, which better describes the uncertainty and complexity of data. Many works focus on adjusting the distribution constraint under the Information Bottleneck (IB) principle to enhance representation learning. However, these proposed regularization terms only consider the constraint of each latent variable, omitting the structural information between latent variables. In this paper, we propose a novel structural entropy-guided probabilistic coding model, named SEPC. Specifically, we incorporate the relationship between latent variables into the optimization by proposing a structural entropy regularization loss. Besides, as traditional structural information theory is not well-suited for regression tasks, we propose a probabilistic encoding tree, transferring regression tasks to classification tasks while diminishing the influence of the transformation. Experimental results across 12 natural language understanding tasks, including both classification and regression tasks, demonstrate the superior performance of SEPC compared to other state-of-the-art models in terms of effectiveness, generalization capability, and robustness to label noise.

Diffusion-based models have shown great promise in real-world image super-resolution (Real-ISR), but often generate content with structural errors and spurious texture details due to the empirical priors and illusions of these models. To address this issue, we introduce StructSR, a simple, effective, and plug-and-play method that enhances structural fidelity and suppresses spurious details for diffusion-based Real-ISR. StructSR operates without the need for additional fine-tuning, external model priors, or high-level semantic knowledge. At its core is the Structure-Aware Screening (SAS) mechanism, which identifies the image with the highest structural similarity to the low-resolution (LR) input in the early inference stage, allowing us to leverage it as a historical structure knowledge to suppress the generation of spurious details. By intervening in the diffusion inference process, StructSR seamlessly integrates with existing diffusion-based Real-ISR models. Our experimental results demonstrate that StructSR significantly improves the fidelity of structure and texture, improving the PSNR and SSIM metrics by an average of 5.27% and 9.36% on a synthetic dataset (DIV2K-Val) and 4.13% and 8.64% on two real-world datasets (RealSR and DRealSR) when integrated with four state-of-the-art diffusion-based Real-ISR methods.

The super-alignment problem of how humans can effectively supervise super-human AI has garnered increasing attention. Recent research has focused on investigating the weak-to-strong generalization (W2SG) scenario as an analogy for super-alignment. This scenario examines how a pre-trained strong model, supervised by an aligned weak model, can outperform its weak supervisor. Despite good progress, current W2SG methods face two main issues: 1) The annotation quality is limited by the knowledge scope of the weak model; 2) It is risky to position the strong model as the final corrector. To tackle these issues, we propose a ``Strong Empowered and Aligned Weak Mastered'' (SEAM) framework for weak annotations in W2SG. This framework can leverage the vast intrinsic knowledge of the pre-trained strong model to empower the annotation and position the aligned weak model as the annotation master. Specifically, the pre-trained strong model first generates principle fast-and-frugal trees for samples to be annotated, encapsulating rich sample-related knowledge. Then, the aligned weak model picks informative nodes based on the tree's information distribution for final annotations. Experiments on six datasets for preference tasks in W2SG scenarios validate the effectiveness of our proposed method.

Human beings often experience stress, which can significantly influence their performance. This study explores whether Large Language Models (LLMs) exhibit stress responses similar to those of humans and whether their performance fluctuates under different stress-inducing prompts. To investigate this, we developed a novel set of prompts, termed StressPrompt, designed to induce varying levels of stress. These prompts were derived from established psychological frameworks and carefully calibrated based on ratings from human participants. We then applied these prompts to several LLMs to assess their responses across a range of tasks, including instruction-following, complex reasoning, and emotional intelligence. The findings suggest that LLMs, like humans, perform optimally under moderate stress, consistent with the Yerkes-Dodson law. Notably, their performance declines under both low and high-stress conditions. Our analysis further revealed that these StressPrompts significantly alter the internal states of LLMs, leading to changes in their neural representations that mirror human responses to stress. This research provides critical insights into the operational robustness and flexibility of LLMs, demonstrating the importance of designing AI systems capable of maintaining high performance in real-world scenarios where stress is prevalent, such as in customer service, healthcare, and emergency response contexts. Moreover, this study contributes to the broader AI research community by offering a new perspective on how LLMs handle different scenarios and their similarities to human cognition.

Multimodal models, namely vision-language models, present unique possibilities through the seamless integration of different information mediums for data generation. These models mostly act as a black-box, making them lack transparency and explicability. Reliable results require accountable and trustworthy Artificial Intelligence (AI), namely when in use for critical tasks, such as the automatic generation of medical imaging reports for healthcare diagnosis. By exploring stress-testing techniques, multimodal generative models can become more transparent by disclosing their shortcomings, further supporting their responsible usage in the medical field.

The rapid advancement of large language models (LLMs) has led to significant improvements in their capabilities, but also to increased concerns about their alignment with human values and intentions. Current alignment strategies, including adaptive training and inference-time methods, have demonstrated potential in this area. However, these approaches still struggle to balance deployment complexity and capability across various tasks and difficulties. In this work, we introduce the Streaming Distribution Induce Aligner (Stream Aligner), a novel alignment paradigm that combines efficiency with enhanced performance in various tasks throughout the generation process. Stream Aligner achieves dynamic sentence-level correction by using a small model to learn the preferences of the suffix sentence, iteratively correcting the suffix sentence output by the upstream model, and then using the corrected sentence to replace the suffix sentence in subsequent generations. Compared to Aligner, our experiments demonstrate that Stream Aligner reduces reliance on the capabilities of additional models, enhances the reasoning abilities of LLMs, and decreases latency during user interaction. Specifically, Stream Aligner-2B model has achieved an improvement of 76.1% in helpfulness, 36.0% in harmlessness on the tested Llama2-70B-chat model, and Stream Aligner-8B has achieved an improvement of 3.5% on the math ability of the tested Llama3-70B-Instruct model.

We initiate the study of matching roommates and rooms wherein the preferences of agents over other agents and rooms are complementary and represented by Leontief utilities. In this setting, 2n agents must be paired up and assigned to n rooms. Each agent has cardinal valuations over the rooms as well as compatibility values over all other agents. Under Leontief preferences, an agent’s utility for a matching is the minimum of the two values. We focus on the tradeoff between maximizing utilitarian social welfare and strategyproofness. Our main result shows that—in a stark contrast to the additive case— under binary Leontief utilities, there exist strategyproof mechanisms that maximize the social welfare. We further devise a strategyproof mechanism that implements such a welfare maximizing algorithm and is parameterized by the number of agents. Along the way, we highlight several possibility and impossibility results, and give upper bounds and lower bounds for welfare with or without strategyproofness.

Today we rely on networks that are created and maintained by smart devices. For such networks, there is no governing central authority but instead the network structure is shaped by the decisions of selfish intelligent agents. A key property of such communication networks is that they should be easy to navigate for routing data. For this, a common approach is greedy routing, where every device simply routes data to a neighbor that is closer to the respective destination. Networks of intelligent agents can be analyzed via a game-theoretic approach and in the last decades many variants of network creation games have been proposed and analyzed. In this paper we present the first game-theoretic network creation model that incorporates greedy routing, i.e., the strategic agents in our model are embedded in some metric space and strive for creating a network among themselves where all-pairs greedy routing is enabled. Besides this, the agents optimize their connection quality within the created network by aiming for greedy routing paths with low stretch. For our model, we analyze the existence of (approximate)-equilibria and the computational hardness in different underlying metric spaces. E.g., we characterize the set of equilibria in 1-2-metrics and tree metrics and show that Nash equilibria always exist. For Euclidean space, the setting which is most relevant in practice, we prove that equilibria are not guaranteed to exist but that the well-known Θ-graph construction yields networks having a low stretch that are game-theoretically almost stable. For general metric spaces, we show that approximate equilibria exist where the approximation factor depends on the cost of maintaining any link.

We study a model of sequential decision-making where voters have dynamic preferences over a set of candidates that are undesirable. This models scenarios such as the implementation of projects that are overall beneficial to society, but impose individual costs on certain affected individuals. We show that while minimizing the sum of agents' disutilities can be done in polynomial time, minimizing the maximum disutility obtained by any agent is computationally intractable, even in restricted cases. We then examine the potential for agents to engage in strategic manipulation in response to these welfare objectives, offering insights into possible misconduct within such decision-making environments.

Recent advances in text-to-image diffusion models have spurred significant interest in continuous story image generation. In this paper, we introduce Storynizor, a model capable of generating coherent stories with strong inter-frame character consistency, effective foreground-background separation, and diverse pose variation. The core innovation of Storynizor lies in its key modules: ID-Synchronizer and ID-Injector. The ID-Synchronizer employs an auto-mask self-attention module and a mask perceptual loss across inter-frame images to improve the consistency of character generation, vividly representing their postures and backgrounds. The ID-Injector utilize a Shuffling Reference Strategy (SRS) to integrate ID features into specific locations, enhancing ID-based consistent character generation. Additionally, to facilitate the training of Storynizor, we have curated a novel dataset called StoryDB comprising 100, 000 images. This dataset contains single and multiple-character sets in diverse environments, layouts, and gestures with detailed descriptions. Experimental results indicate that Storynizor demonstrates superior coherent story generation with high-fidelity character consistency, flexible postures, and vivid backgrounds compared to other character-specific methods.