AAAI 2025 Papers

As AI agents are increasingly adopted to collaborate on complex objectives, ensuring the security of autonomous multi-agent systems becomes crucial. We develop simulations of agents collaborating on shared objectives to study these security risks and security trade-offs. We focus on scenarios where an attacker compromises one agent, using it to steer the entire system toward misaligned outcomes by corrupting other agents. In this context, we observe infectious malicious prompts - the multi-hop spreading of malicious instructions. To mitigate this risk, we evaluated several strategies: two "vaccination" approaches that insert false memories of safely handling malicious input into the agents' memory stream, and two versions of a generic safety instruction strategy. While these defenses reduce the spread and fulfillment of malicious instructions in our experiments, they tend to decrease collaboration capability in the agent network. Our findings illustrate potential trade-off between security and collaborative efficiency in multi-agent systems, providing insights for designing more secure yet effective AI collaborations.

Multi-Agent Motion Planning (MAMP) finds various applications in fields such as traffic management, airport operations, and warehouse automation. In many of these environments, differential drive robots are commonly used. These robots have a kinodynamic model that allows only in-place rotation and movement along their current orientation, subject to speed and acceleration limits. However, existing Multi-Agent Path Finding (MAPF)-based methods often use simplified models for robot kinodynamics, which limits their practicality and realism. In this paper, we introduce a three-level framework called MASS to address these challenges. MASS combines MAPF-based methods with our proposed stationary state search planner to generate high-quality kinodynamically-feasible plans. We further extend MASS using an adaptive window mechanism to address the lifelong MAMP problem. Empirically, we tested our methods on the single-shot grid map domain and the lifelong warehouse domain. Our method shows up to 400% improvements in terms of throughput compared to existing methods.

Assignment problems are a classic combinatorial optimization problem in which a group of agents must be assigned to a group of tasks such that maximum utility is achieved while satisfying assignment constraints. Given the utility of each agent completing each task, polynomial-time algorithms exist to solve a single assignment problem in its simplest form. However, in many modern-day applications such as satellite constellations, power grids, and mobile robot scheduling, assignment problems unfold over time, with the utility for a given assignment depending heavily on the state of the system. We apply multi-agent reinforcement learning to this problem, learning the value of assignments by bootstrapping from the known polynomial-time greedy solver and then learning from further experience. We then choose assignments using a distributed optimal assignment mechanism rather than by selecting them directly. We demonstrate that this algorithm is theoretically justified and avoids pitfalls experienced by other RL algorithms in this setting. Finally, we show that our algorithm significantly outperforms other methods in the literature, even while scaling to realistic scenarios with hundreds of agents and tasks.

Understanding people's social interactions in complex real-world scenarios often relies on intricate mental reasoning. To truly understand how and why people interact with one another, we must infer the underlying mental states that give rise to the social interactions, i.e., Theory of Mind reasoning in multi-agent interactions. Additionally, social interactions are often multi-modal -- we can watch people's actions, hear their conversations, and/or read about their past behaviors. For AI systems to successfully and safely interact with people in real-world environments, they also need to understand people's mental states as well as their inferences about each other's mental states based on multi-modal information about their interactions. For this, we introduce MuMA-ToM, a Multi-modal Multi-Agent Theory of Mind benchmark. MuMA-ToM is the first multi-modal Theory of Mind benchmark that evaluates mental reasoning in embodied multi-agent interactions. In MuMA-ToM, we provide video and text descriptions of people's multi-modal behavior in realistic household environments. Based on the context, we then ask questions about people's goals, beliefs, and beliefs about others' goals. We validated MuMA-ToM in a human experiment and provided a human baseline. We also proposed a novel multi-modal, multi-agent ToM model, LIMP (Language model-based Inverse Multi-agent Planning). Our experimental results show that LIMP significantly outperforms state-of-the-art methods, including large multi-modal models (e.g., GPT-4o, Gemini-1.5 Pro) and a recent multi-modal ToM model, BIP-ALM.

Current methods commonly utilize three-branch structures of inversion, reconstruction, and editing, to tackle consistent image editing task. However, these methods lack control over the generation position of the edited object and have issues with background preservation. To overcome these limitations, we propose a tuning-free method with only two branches: inversion and editing. This approach allows users to simultaneously edit the object's action and control the generation position of the edited object. Additionally, it achieves improved background preservation. Specifically, we transfer the edited object information to the target area and repair or preserve the background of other areas during the inversion process at a specific time step. In the editing stage, we use the image features in self-attention to query the key and value of the corresponding time step in the inversion to achieve consistent image editing. Impressive image editing results and quantitative evaluation demonstrate the effectiveness of our method.

Whole-body multimodal motion generation, controlled by text, speech, or music, has numerous applications including video generation and character animation. However, employing a unified model to process different condition modalities presents two main challenges: motion distribution drifts across different tasks (e.g., co-speech gestures and text-driven daily actions) and the complex optimization of mixed conditions with varying granularities (e.g., text and audio). In this paper, we propose MotionCraft, a unified diffusion transformer that crafts whole-body motion with plug-and-play multimodal control. Our framework employs a coarse-to-fine training strategy, starting with the text-to-motion semantic pre-training, followed by the multimodal low-level control adaptation. To effectively learn and transfer motion knowledge across different distributions, we design MC-Attn for parallel modeling of static and dynamic human topology graphs. To overcome the motion format inconsistency of existing benchmarks, we introduce MC-Bench, the first available multimodal whole-body motion generation benchmark based on the unified SMPL-X format. Extensive experiments show that MotionCraft achieves state-of-the-art performance on various standard motion generation tasks.

To equip artificial intelligence with a comprehensive understanding towards a temporal world, video and 4D panoptic scene graph generation abstracts visual data into nodes to represent entities and edges to capture temporal relations. Existing methods encode entity masks tracked across temporal dimensions (mask tubes), then predict their relations with temporal pooling operation, which does not fully utilize the motion indicative of the entities' relation. To overcome this limitation, we introduce a contrastive representation learning framework that focuses on motion pattern for temporal scene graph generation. Firstly, our framework encourages the model to learn close representations for mask tubes of similar subject-relation-object triplets. Secondly, we seek to push apart mask tubes from their temporally shuffled versions. Moreover, we also learn distant representations for mask tubes belonging to the same video but different triplets. Extensive experiments show that our motion-aware contrastive framework significantly improves state-of-the-art methods on both video and 4D datasets.

Event cameras, which capture pixel-level brightness changes asynchronously, provide rich motion information that is often missed during traditional frame-based camera exposures, thereby offering fresh perspectives for motion deblurring. Although current approaches incorporate event intensity, they neglect essential spatial motion information. Unlike their CNN architectures, Transformers excel in modeling long-range dependencies but struggle with establishing relevant non-local connections in sparse events and fail to highlight significant interactions in dense images. To address these limitations, we introduce a Motion-Adaptive Transformer network (MAT) that utilizes spatial motion information to forge robust global connections. The core design is an Adaptive Motion Mask Predictor (AMMP) that identifies key motion regions, guiding the Motion-Sparse Attention (MSA) to eliminate irrelevant event tokens and enabling the Motion-Aware Attention (MAA) to focus on relevant ones, thereby enhancing long-range dependency modeling. Additionally, we elaborately design a Cross-Modal Intensity Gating mechanism that efficiently merges intensity data across modalities while minimizing parameter use. The learnable Expansion-Controlled Spatial Gating further optimizes the transmission of event features. Comprehensive testing confirms that our approach sets a new benchmark in image deblurring, surpassing previous methods by up to 0.60dB on the GoPro dataset, 1.04dB on the HS-ERGB dataset, and achieving an average improvement of 0.52dB across two real-world datasets.

Recent large-scale pre-trained diffusion models have demonstrated a powerful generative ability to produce high-quality videos from detailed text descriptions. However, exerting control over the motion of objects in videos generated by any video diffusion model remains a challenging problem. In this paper, we propose a novel zero-shot moving object trajectory control framework, Motion-Zero, to enable arbitrary single-object-trajectory control for the text-to-video diffusion model. To this end, an initial noise prior module is designed to provide a position-based prior to improve the stability of the appearance of the moving object and the accuracy of position. In addition, based on the attention map of the U-Net, spatial constraints are directly applied to the denoising process of diffusion models, which further ensures the positional consistency of moving objects during the inference. Furthermore, temporal consistency is guaranteed with a proposed shift temporal attention mechanism. Our method can be flexibly applied to various state-of-the-art video diffusion models without any training process. Extensive experiments demonstrate our proposed method can control the motion trajectories of arbitrary objects while preserving the original ability to generate high-quality videos.

Different from traditional object detection, pure vision is not enough to infrared small target detection, due to small target size and weak background contrast. For promoting detection performance, more target representations are needed. Currently, motion representations have been proved to be one of the most potential feature kinds for infrared small target detection. Existing methods have an obvious weakness, that besides vision features, they could only capture coarse motion representations from temporal domain. With vision features, fine motion representations could be more effective to enhance detection performance. To overcome this weakness, inspired by prevalent vision-language models, we propose the first vision-language framework with motion prior knowledge learning (MoPKL). Breaking through traditional pure-vision modality, it utilizes homogeneous language descriptions, formatted for moving targets, to directionally guide vision channel learning motion prior knowledge. With the facilitation of motion-vision alignment and motion-relation mining, the motion of infrared small targets is further refined by graph attention, to generate more fine motion representations. The extensive experiments on datasets ITSDT-15K and IRDST show that our framework is effective. It could often obviously outperform other methods.

Dynamic object modeling is a critical challenge in 3D scene reconstruction. Previous methods typically maintain a canonical space to represent the object model, and a deformation field to express the object motion. However, this approach fails when the object undergoes large motions. The position variation caused by significant motion not only complicates the establishment of a canonical space, but also misleads the interpretation of the deformation field. To overcome the above weaknesses, we propose Motion Decoupled Dynamic 3D Gaussian Splatting (M5D-GS), the first 3D-GS model that separates motion and deformation modeling for dynamic object representation with large motion from a monocular camera. M5D-GS increases the practicality of 3D-GS, as it is common for objects to move, rotate, and deform simultaneously. Current datasets only contain object deformations with slight motions. We introduce a pipeline to reuse current benchmarks by adding large motions into the scene. We also introduce a new benchmark featuring several new scenes with complex motions, scenes augmented from previous datasets, and some real world recorded testcases, to fully demonstrate our improvements. Our M5D-GS significantly increases the accuracy under large motion scenarios while maintaining high rendering speed, which makes it suitable for dynamic object representation tasks including 4D novel view synthesis and real-time rendering.

Motion artifacts present in magnetic resonance imaging (MRI) can seriously interfere with clinical diagnosis. Removing motion artifacts is a straightforward solution and has been extensively studied. However, paired data are still heavily relied on in recent works and the perturbations in k-space (frequency domain) are not well considered, which limits their applications in the clinical field. To address these issues, we propose a novel unsupervised purification method which leverages pixel-frequency information of noisy MRI images to guide a pre-trained diffusion model to recover clean MRI images. Specifically, considering that motion artifacts are mainly concentrated in high-frequency components in k-space, we utilize the low-frequency components as the guide to ensure correct tissue textures. Additionally, given that high-frequency and pixel information are helpful for recovering shape and detail textures, we design alternate complementary masks to simultaneously destroy the artifact structure and exploit useful information. Quantitative experiments are performed on datasets from different tissues and show that our method achieves superior performance on several metrics. Qualitative evaluations with radiologists also show that our method provides better clinical feedback.

Social segregation in cities, spanning racial, residential, and income dimensions, is becoming increasingly diverse and severe. As urban spaces and social dynamics grow more complex, residents experience varying levels of segregation, which, if left unaddressed, could exacerbate crime rates, fuel social tensions, and lead to other societal challenges. Effectively addressing these issues requires a comprehensive analysis of the underlying structures of urban spaces and resident interactions. While previous studies have primarily focused on surface-level indicators of segregation, they often fail to explore the complexity of urban structure and mobility dynamics, leaving gaps in understanding modern segregation patterns. To fill this gap, we propose the Motif-Enhanced Graph Prototype Learning (MotifGPL) framework, offering a novel approach to studying urban segregation. The framework consists of three key modules: prototype-based graph structure extraction, motif distribution discovery, and urban graph reconstruction. Specifically, we use prototype-based learning to extract key urban graph prototypes from both spatial and origin-destination graphs, incorporating attributes such as points of interest, street images, and flow indices. The motif distribution discovery module enhances interpretability by matching each prototype to similar motifs, which represent simplified graph structures that reflect local patterns. These motifs are then used to guide the reconstruction of urban graphs, enabling a more detailed exploration of spatial structures and mobility patterns. By identifying critical motifs influencing urban segregation, MotifGPL offers insights to guide the design of urban environments that can help reduce segregation. Experimental results demonstrate that MotifGPL effectively uncovers these key motifs and provides actionable insights for mitigating segregation.

Networked time series are time series on a graph, one for each node, with applications in traffic and weather monitoring. Graph neural networks are natural candidates for networked time series imputation and have recently outperformed existing alternatives such as recurrent and generative models for time series imputation as they utilize a relational inductive bias for imputation. However, existing GNN-based approaches fail to capture the higher-order topological structure between sensors, which are shaped by recurring substructures in the graph, referred to as temporal motifs. In addition, it remains uncertain which motifs are the most pivotal motifs guiding the imputation task in networked time series. In this paper, we fill in this gap by proposing a graph neural network designed to leverage motif structures within the network by employing weighted motif adjacency matrices to capture higher-order neighborhood information. In particular, (1) we design a motif-wise multi-view attention module that explicitly captures various higher-order structures along with an attention mechanism that automatically assigns high weights to informative ones in order to maximize the use of higher-order information. (2) We introduce a gated fusion module by merging gated recurrent networks and graph convolutional networks to capture the spatial and temporal dependency in order to reflect the intricate impacts of temporal and spatial influence. Experimental results demonstrate that when compared to state-of-the-art models for time-series imputation tasks, our proposed model can reduce the error by around 19%.

Drug-Drug Interaction (DDI) prediction has attracted considerable attention in designing multi-drug combination strategies and avoiding adverse reactions. Notably, Artificial Intelligence (AI)-driven DDI prediction methods have emerged as a pivotal research paradigm. However, most AI-driven DDI prediction methods fall short in exploring intra-molecular motifs, and heavily rely on the overly idealized assumption of the complete inter-molecular topology, limiting their expressive capacities. To this end, we propose a Motif-Oriented representation learning with TOpology Refinement for DDI prediction, namely MOTOR, to exploit both the multi-granularity motif information and the topological structure of DDI networks. Specifically, MOTOR effectively captures motif internal structures, motif local contexts, and motif global semantics. Furthermore, MOTOR employs an iterative learning strategy to continuously refine the DDI topology and optimize the corresponding drug representations. Extensive experimental results demonstrate that MOTOR exhibits superior performance with interpretable insights in DDI prediction tasks across three real-world datasets, thereby opening up new avenues in AI-driven DDI prediction.

Person re-identification (re-ID) via 3D skeleton data is a challenging task with significant value in many scenarios. Existing skeleton-based methods typically assume virtual motion relations between all joints, and adopt average joint or sequence representations for learning. However, they rarely explore key body structure and motion such as gait to focus on more important body joints or limbs, while lacking the ability to fully mine valuable spatial-temporal sub-patterns of skeletons to enhance model learning. This paper presents a generic Motif guided graph transformer with Combinatorial skeleton prototype learning (MoCos) that exploits structure-specific and gait-related body relations as well as combinatorial features of skeleton graphs to learn effective skeleton representations for person re-ID. In particular, motivated by the locality within joints' structure and the body-component collaboration in gait, we first propose the motif guided graph transformer (MGT) that incorporates hierarchical structural motifs and gait collaborative motifs, which simultaneously focuses on multi-order local joint correlations and key cooperative body parts to enhance skeleton relation learning. Then, we devise the combinatorial skeleton prototype learning (CSP) that leverages random spatial-temporal combinations of joint nodes and skeleton graphs to generate diverse sub-skeleton and sub-tracklet representations, which are contrasted with the most representative features (prototypes) of each identity to learn class-related semantics and discriminative skeleton representations. Extensive experiments validate the superior performance of MoCos over existing state-of-the-art models. We further show its generality under RGB-estimated skeletons, different graph modeling, and unsupervised scenarios.

Enabling Large Language Models (LLMs) to comprehend the 3D physical world remains a significant challenge. Due to the lack of large-scale 3D-text pair datasets, the success of LLMs has yet to be replicated in 3D understanding. In this paper, we rethink this issue and propose a new task: 3D Data-Efficient Point-Language Understanding. The goal is to enable LLMs to achieve robust 3D object understanding with minimal 3D point cloud and text data pairs. To address this task, we introduce GreenPLM, which leverages more text data to compensate for the lack of 3D data. First, inspired by using CLIP to align images and text, we utilize a pre-trained point cloud-text encoder to map the 3D point cloud space to the text space. This mapping leaves us to seamlessly connect the text space with LLMs. Once the point-text-LLM connection is established, we further enhance text-LLM alignment by expanding the intermediate text space, thereby reducing the reliance on 3D point cloud data. Specifically, we generate 6M free-text descriptions of 3D objects, and design a three-stage training strategy to help LLMs better explore the intrinsic connections between different modalities. To achieve efficient modality alignment, we design a zero-parameter cross-attention module for token pooling. Extensive experimental results show that GreenPLM requires only 12% of the 3D training data used by existing state-of-the-art models to achieve superior 3D understanding. Remarkably, GreenPLM also achieves competitive performance using text-only data.

Procedural Content Generation via Machine Learning (PCGML) has enhanced game content creation, yet challenges in controllability and limited training data persist. This study addresses these issues by distilling a constructive PCG algorithm into a controllable PCGML model. We first generate a large amount of content with a constructive algorithm and label it using a Large Language Model (LLM). We use these synthetic labels to condition two PCGML models for content-specific generation, a diffusion model and the five-dollar model. This neural network distillation process ensures that the generation aligns with the original algorithm while introducing controllability through plain text. We define this text-conditioned PCGML as a Text-to-game-Map (T2M) task, offering an alternative to prevalent text-to-image multi-modal tasks. We compare our distilled models with the baseline constructive algorithm. Our analysis of the variety, accuracy, and quality of our generation demonstrates the efficacy of distilling constructive methods into controllable text-conditioned PCGML models.

We propose MonoBox, an innovative box-supervised segmentation method constrained by monotonicity to liberate its training from the user-unfriendly box-tightness assumption. In contrast to conventional box-supervised segmentation, where the box edges must precisely touch the target boundaries, MonoBox leverages imprecisely-annotated boxes to achieve robust pixel-wise segmentation. The 'linchpin' is that, within the noisy zones around box edges, MonoBox discards the traditional misguiding multiple-instance learning loss, and instead optimizes a carefully-designed objective, termed monotonicity constraint. Along directions transitioning from the foreground to background, this new constraint steers responses to adhere to a trend of monotonically decreasing values. Consequently, the originally unreliable learning within the noisy zones is transformed into a correct and effective monotonicity optimization. Moreover, an adaptive label correction is introduced, enabling MonoBox to enhance the tightness of box annotations using predicted masks from the previous epoch and dynamically shrink the noisy zones as training progresses. We verify MonoBox in the box-supervised segmentation task of polyps, where satisfying box-tightness is challenging due to the vague boundaries between the polyp and normal tissues. Experiments on both public synthetic and in-house real noisy datasets demonstrate that MonoBox exceeds other anti-noise state-of-the-arts by improving Dice by at least 5.5% and 3.3%, respectively.