AAAI 2025 论文列表

Label skews, characterized by disparities in local label distribution across clients, pose a significant challenge in federated learning. As minority classes suffer from worse accuracy due to overfitting on local imbalanced data, prior methods often incorporate class-balanced learning techniques during local training. Although these methods improve the mean accuracy across all classes, we observe that vacant classes—referring to categories absent from a client's data distribution—remain poorly recognized. Besides, there is still a gap in the accuracy of local models on minority classes compared to the global model. This paper introduces FedVLS, a novel approach to label-skewed federated learning that integrates both vacant-class distillation and logit suppression simultaneously. Specifically, vacant-class distillation leverages knowledge distillation during local training on each client to retain essential information related to vacant classes from the global model. Moreover, logit suppression directly penalizes network logits for non-label classes, effectively addressing misclassifications in minority classes that may be biased toward majority classes. Extensive experiments validate the efficacy of FedVLS, demonstrating superior performance compared to previous state-of-the-art (SOTA) methods across diverse datasets with varying degrees of label skews.

The generalization problem is broadly recognized as a critical challenge in detecting deepfakes. Most previous work believes that the generalization gap is caused by the differences among various forgery methods. However, our investigation reveals that the generalization issue can still occur when forgery-irrelevant factors shift. In this work, we identify two biases that detectors may also be prone to overfitting: position bias and content bias, as depicted in Fig. 1. For the position bias, we observe that detectors are prone to “lazily” depending on the specific positions within an image (e.g., central regions even no forgery). As for content bias, we argue that detectors may potentially and mistakenly utilize forgery-unrelated information for detection (e.g., background, and hair). To intervene in these biases, we propose two branches for shuffling and mixing with tokens in the latent space of transformers. For the shuffling branch, we rearrange the tokens and corresponding position embedding for each image while maintaining the local correlation. For the mixing branch, we randomly select and mix the tokens in the latent space between two images with the same label within the mini-batch to recombine the content information. During the learning process, we align the outputs of detectors from different branches in both feature space and logit space. Contrastive losses for features and divergence losses for logits are applied to obtain unbiased feature representation and classifiers. We demonstrate and verify the effectiveness of our method through extensive experiments on widely used evaluation datasets.

Biological neural systems often represent information on low-dimensional manifolds that reflect the topology of their encoded variables. This suggests that neural activity can be naturally organized in geometrically meaningful ways, as seen in rodent head direction cells forming circular manifolds. This proposal examines whether artificial neural networks (ANNs) trained on tasks with well-defined topologies—such as planar or spherical coordinates from autonomous driving datasets like Apolloscape, cyclic temporal variables, or graph-structured road networks—develop similar low-dimensional representations aligned with the variables' inherent topology. We consider convolutional and vision transformer models for image data, graph neural networks for road network graphs, and 3D or point-based models for LIDAR point clouds, analyzing their internal activations with dimensionality reduction and topological data analysis. If successful, this approach not only elucidates the nature of internal representations in ANNs but also offers insights into the computational principles that bridge artificial systems and biological cognition.

Dense video captioning aims to detect and describe all events in untrimmed videos. This paper presents a dense video captioning network called Multi-Concept Cyclic Learning (MCCL), which aims to: (1) detect multiple concepts at the frame level and leverage these concepts to provide temporal event cues; and (2) establish cyclic co-learning between the generator and the localizer within the captioning network to promote semantic perception and event localization. Specifically, weakly supervised concept detection is performed for each frame, and the detected concept embeddings are integrated into the video features to provide event cues. Additionally, video-level concept contrastive learning is introduced to produce more discriminative concept embeddings. In the captioning network, a cyclic co-learning strategy is proposed, where the generator guides the localizer for event localization through semantic matching, while the localizer enhances the generator’s event semantic perception through location matching, making semantic perception and event localization mutually beneficial. MCCL achieves state-of-the-art performance on the ActivityNet Captions and YouCook2 datasets. Extensive experiments demonstrate its effectiveness and interpretability.

With the development of Vision Foundation Models (VFMs) in recent years, Visual In-Context Learning (VICL) has become a better choice compared to modifying models in most scenarios. Different from retraining or fine-tuning models, VICL does not require modifications to the model's weights and architecture, and only needs a prompt with demonstrations to teach VFM how to solve tasks. Currently, significant computational cost for finding optimal prompts for every test sample hinders the deployment of VICL, as determining which demonstrations to use for constructing the prompt is very costly. In this paper, however, we find a counterintuitive phenomenon that most test samples actually achieve optimal performance under the same prompts, and searching for sample-level prompts only costs much time but results in completely identical prompts actually. Therefore, we propose task-level prompting to reduce the cost of searching for prompts during the inference stage and introduce two time-saving yet effective task-level prompt search strategies accordingly. Extensive experimental results show that our proposed method can identify near-optimal prompts and reach the best VICL performance with a minimal cost that prior work has never achieved.

Domain Generalized Semantic Segmentation (DGSS) seeks to utilize source domain data exclusively to enhance the generalization of semantic segmentation across unknown target domains. Prevailing studies predominantly concentrate on feature normalization and domain randomization, these approaches exhibit significant limitations. Feature normalization-based methods tend to confuse semantic features in the process of constraining the feature space distribution, resulting in classification misjudgment. Domain randomization-based methods frequently incorporate domain-irrelevant noise due to the uncontrollability of style transformations, resulting in segmentation ambiguity. To address these challenges, we introduce a novel framework, named SCSD for Semantic Consistency prediction and Style Diversity generalization. It comprises three pivotal components: Firstly, a Semantic Query Booster is designed to enhance the semantic awareness and discrimination capabilities of object queries in the mask decoder, enabling cross-domain semantic consistency prediction. Secondly, we develop a Text-Driven Style Transform module that utilizes domain difference text embeddings to controllably guide the style transformation of image features, thereby increasing inter-domain style diversity. Lastly, to prevent the collapse of similar domain feature spaces, we introduce a Style Synergy Optimization mechanism that fortifies the separation of inter-domain features and the aggregation of intra-domain features by synergistically weighting style contrastive loss and style aggregation loss. Extensive experiments demonstrate that the proposed SCSD significantly outperforms existing state-of-theart methods. Notably, SCSD trained on GTAV achieved an average of 49.11 mIoU on the four unseen domain datasets, surpassing the state-of-the-art method by +4.08 mIoU.

In this paper, we explore how to develop salient object detection models using adder neural networks (ANNs), which are more energy efficient than convolutional neural networks (CNNs), especially for real-world applications. Based on our empirical studies, we show that directly replacing the convolutions in CNN-based models with adder layers leads to a substantial loss of activations in the decoder part. This makes the feature maps learned in the decoder lack pattern diversity and hence results in a significant performance drop. To alleviate this issue, by investigating the statistics of the feature maps produced by adder layers, we introduce a simple yet effective differential merging strategy to augment the feature representations learned by adder layers and present a simple baseline for SOD using ANNs. Experiments on popular salient object detection benchmarks demonstrate that our proposed method with a simple feature pyramid network (FPN) architecture achieves comparable performance to previous state-of-theart CNN-based models and consumes much less energy. We hope this work could facilitate the development of ANNs in binary segmentation tasks.

Catastrophic forgetting poses a significant challenge for graph neural networks in continuously updating their knowledge base with data streams. To address this issue, much of the research has focused on node-level continual learning using parameter regularization or rehearsal-based strategies, while little attention given to graph-level tasks. Furthermore, current paradigms for continual graph learning may inadvertently capture spurious correlations for specific tasks through shortcuts, thereby exacerbating the forgetting of previous knowledge when new tasks are introduced. To tackle these challenges, we propose a novel paradigm, Rationale Learning GNN (RL-GNN), for graph-level continual graph learning. Specifically, we harness the invariant learning principle to incorporate environmental interventions into both the current and historical distributions, aiming to uncover rationales by minimizing empirical risk across all environments. The rationale serves as the sole factor guiding the learning process. Therefore, continual graph learning is redefined as capturing these invariant rationales within task sequences, alleviating catastrophic forgetting caused by spurious features. Extensive experiments on real-world datasets with varying task lengths demonstrate the effectiveness of our RL-GNN in continuous knowledge assimilation and reduction of catastrophic forgetting.

Data-free model stealing involves replicating the functionality of a target model into a substitute model without accessing the target model's structure, parameters, or training data. Instead, the adversary can only access the target model's predictions for generated samples. Once the substitute model closely approximates the behavior of the target model, attackers can exploit its white-box characteristics for subsequent malicious activities, such as adversarial attacks. Existing methods within cooperative game frameworks often produce samples with high confidence for the prediction of the substitute model, which makes it difficult for the substitute model to replicate the behavior of the target model. This paper presents a new data-free model stealing approach called Query Efficient Data Generation (QEDG). We introduce two distinct loss functions to ensure the generation of sufficient samples that closely and uniformly align with the target model's decision boundary across multiple classes. Building on the limitation of current methods, which typically yield only one piece of supervised information per query, we propose the query-free sample augmentation that enables the acquisition of additional supervised information without increasing the number of queries. Motivated by theoretical analysis, we adopt the consistency rate metric, which more accurately evaluates the similarity between the substitute and target models. We conducted extensive experiments to verify the effectiveness of our proposed method, which achieved better performance with fewer queries compared to the state-of-the-art methods on the real MLaaS scenario and five datasets.

The gait, as a kind of soft biometric characteristic, can reflect the distinct walking patterns of individuals at a distance, exhibiting a promising technique for unrestrained human identification. With largely excluding gait-unrelated cues hidden in RGB videos, the silhouette and skeleton, though visually compact, have acted as two of the most prevailing gait modalities for a long time. Recently, several attempts have been made to introduce more informative data forms like human parsing and optical flow images to capture gait characteristics, along with multi-branch architectures. However, due to the inconsistency within model designs and experiment settings, we argue that a comprehensive and fair comparative study among these popular gait modalities, involving the representational capacity and fusion strategy exploration, is still lacking. From the perspectives of fine vs. coarse-grained shape and whole vs. pixel-wise motion modeling, this work presents an in-depth investigation of three popular gait representations, i.e., silhouette, human parsing, and optical flow, with various fusion evaluations, and experimentally exposes their similarities and differences. Based on the obtained insights, we further develop a C²Fusion strategy, consequently building our new framework MultiGait++. C²Fusion preserves commonalities while highlighting differences to enrich the learning of gait features. To verify our findings and conclusions, extensive experiments on Gait3D, GREW, CCPG, and SUSTech1K are conducted.

Model editing aims at selectively updating a small subset of a neural model's parameters with an interpretable strategy to achieve desired modifications. It can significantly reduce computational costs to adapt to large language models(LLMs). Given its ability to precisely target critical components within LLMs, model editing shows great potential for efficient fine-tuning applications. In this work, we investigate model editing to serve as an efficient method for adapting LLMs to solve aspect-based sentiment classification. Through causal interventions, we trace and determine which neuron hidden states are essential for the model’s predictions. By performing interventions and restorations on each component of an LLM, we identify the importance of these components for aspect-based sentiment classification. Our findings reveal that a distinct set of mid-layer representations is essential for detecting the sentiment polarity of given aspect words. Leveraging these insights, we develop a model editing approach that focuses exclusively on these critical parts of the LLM, leading to a more efficient method for adapting LLMs. Our in and out of domain experiments demonstrate that this approach achieves competitive results compared to the currently strongest methods with significantly fewer trainable parameters, highlighting a more efficient and interpretable fine-tuning strategy.

Large language models (LLMs) have demonstrated strong capabilities in language understanding and generation, and their potential in educational contexts is increasingly being explored. One promising area is learnersourcing, where students engage in creating their own educational content, such as multiple-choice questions. A critical step in this process is generating effective explanations for the solutions to these questions, as such explanations aid in peer understanding and promote deeper conceptual learning. However, students often find it difficult to craft high-quality explanations due to limited understanding or gaps in their subject knowledge. To support this task, we introduce ``ILearner-LLM,'' a framework that uses iterative enhancement with LLMs to improve generated explanations. The framework combines an explanation generation model and an explanation evaluation model fine-tuned using student preferences for quality, where feedback from the evaluation model is fed back into the generation model to refine the output. Our experiments with LLaMA2-13B and GPT-4 using five large datasets from the PeerWise MCQ platform show that ILearner-LLM produces explanations of higher quality that closely align with those written by students. Our findings represent a promising approach for enriching the learnersourcing experience for students and for leveraging the capabilities of large language models for educational applications.

Recent years have witnessed rapid advancements in the safety alignments of large language models (LLMs). Methods such as supervised instruction fine-tuning (SFT) and reinforcement learning with human feedback (RLHF) have thus emerged as vital components in constructing LLMs. While these methods achieve robust and fine-grained alignment to human values, their practical application is still hindered by high annotation costs and incomplete human alignments. Besides, the intrinsic human values within training corpora have not been fully exploited. To address these issues, we propose ISAAC (Intrinsically Supervised Alignments by Assessing Corpus), a primary and coarse-grained safety alignment strategy for LLMs. ISAAC only relies on a prior assumption about the text corpus, and does not require preferences in RLHF or human responses selection in SFT. Specifically, it assumes a long-tail distribution of text corpus and employs a specialized sampling strategy to automatically sample high-quality responses. Theoretically, we prove that this strategy can improve the safety of LLMs under our assumptions. Empirically, our evaluations on mainstream LLMs show that ISAAC achieves a safety score comparable to current SFT solutions. Moreover, we conduct experiments on ISAAC for some RLHF-based LLMs, where we find that ISAAC can even improve the safety of these models under specific safety domains. These findings demonstrate that ISAAC can provide preliminary alignment to LLMs, thereby reducing the construction costs of existing human-feedback-based methods.

Contextual information at the video level has become increasingly crucial for visual object tracking. However, existing methods typically use only a few tokens to convey this information, which can lead to information loss and limit their ability to fully capture the context. To address this issue, we propose a new video-level visual object tracking framework called MCITrack. It leverages Mamba's hidden states to continuously record and transmit extensive contextual information throughout the video stream, resulting in more robust object tracking. The core component of MCITrack is the Contextual Information Fusion module, which consists of the mamba layer and the cross-attention layer. The mamba layer stores historical contextual information, while the cross-attention layer integrates this information into the current visual features of each backbone block. This module enhances the model's ability to capture and utilize contextual information at multiple levels through deep integration with the backbone. Experiments demonstrate that MCITrack achieves competitive performance across numerous benchmarks. For instance, it gets 76.6% AUC on LaSOT and 80.0% AO on GOT-10k, establishing a new state-of-the-art performance.

This paper presents a practical problem in dialogue systems: the capability to adapt to changing user intentions and resolve inconsistencies in conversation histories. It is crucial in scenarios like train ticket booking, where travel plans often change dynamically. Notwithstanding the advancements in NLP and large language models (LLMs), these systems struggle with real-time information updates during conversations. We introduce a specialized dataset to evaluate LLM-based chatbots on such conversational adaptability by asking a broad range of open-domain questions, focusing on scenarios where users modify their requests mid-conversation. Additionally, as LLMs are susceptible to generating superfluous sentences, we propose a novel, Chain-of-Thought-free evaluation framework to distill the user intention from their responses. Through extensive investigations on four LLMs, we observe that these contemporary LLMs are not well-aligned with the latest user intent in long-term conversations; they often fail to capture the nuances of natural conversations in a zero-shot setting. Interestingly, the results demonstrate that GPT-4, widely recognized as having the most advanced reasoning capabilities to date, is bested by GPT-3.5 in this task. This work aims to improve the practicality of LLM-based chatbots, bridging the gap between the current capabilities of dialogue systems and the fluidity of human interactions.

Most work treats large language models as black boxes without an in-depth understanding of their internal working mechanism. To explain the internal representations of LLMs, we utilize a gradient-based metric to assess the activation level of model parameters. Based on this metric, we obtain three preliminary findings. (1) When the inputs are in the same domain, parameters in the shallow layers will be activated densely, which means a larger portion of parameters will have great impacts on the outputs. In contrast, parameters in the deep layers are activated sparsely. (2) When the inputs are across different domains, parameters in shallow layers exhibit higher similarity in the activation behavior than in deep layers. (3) In deep layers, the similarity of the distributions of activated parameters is positively correlated to the empirical data relevance. Further, we develop three validation experiments to solidify these findings. (1) Firstly, starting from the first finding, we attempt to configure different sparsities for different layers and find this method can benefit model pruning. (2) Secondly, we find that a pruned model based on one calibration set can better handle tasks related to the calibration task than those not related, which validates the second finding. (3) Thirdly, Based on the STS-B and SICK benchmarks, we find that two sentences with consistent semantics tend to share similar parameter activation patterns in deep layers, which aligns with our third finding. Our work sheds light on the behavior of parameter activation in LLMs, and we hope these findings will have the potential to inspire more practical applications.

Complex question answering (QA) is a challenging task in artificial intelligence research which requires reasoning based on related knowledge. The retrieval-augmented generation (RAG) based on large language models (LLMs) have become one promising solution in QA. To facilitate RAG more effectively, the LLM needs to precisely evaluate knowledge required in QA. That is, first, the LLM needs to examine its knowledge boundary (what the LLM does not know) to retrieve external knowledge as supplement. Second, the LLM needs to evaluate the utility of the retrieved knowledge (whether it helps in reasoning) for robust RAG. To this end, in this paper, we propose a novel Question Answering with Knowledge Evaluation (KEQA) framework to promote the effectiveness and efficiency of RAG in QA. First, inspired by quizzes in classroom, we propose a quiz-based method to precisely examine the knowledge state of the uninterpretable LLM for QA. We ask indicative quizzes on each required knowledge, and inspect whether the LLM can consistently answer the quiz to examine its knowledge boundary. Second, we retrieve the unknown knowledge from external source, and evaluate its utility to pick the helpful ones for reasoning. We design a reasoning-based metric to evaluate utility, and construct a demonstration set in training data for reference to guide knowledge picking in inference. We conduct extensive experiments on four widely-used QA datasets, and the results demonstrate the effectiveness of the proposed method.

In-context segmentation has drawn increasing attention with the advent of vision foundation models. Its goal is to segment objects using given reference images. Most existing approaches adopt metric learning or masked image modeling to build the correlation between visual prompts and input image queries. This work approaches the problem from a fresh perspective - unlocking the capability of the latent diffusion model (LDM) for in-context segmentation and investigating different design choices. Specifically, we examine the problem from three angles: instruction extraction, output alignment, and meta-architectures. We design a two-stage masking strategy to prevent interfering information from leaking into the instructions. In addition, we propose an augmented pseudo-masking target to ensure the model predicts without forgetting the original images. Moreover, we build a new and fair in-context segmentation benchmark that covers both image and video datasets. Experiments validate the effectiveness of our approach, demonstrating comparable or even stronger results than previous specialist or visual foundation models. We hope our work inspires others to rethink the unification of segmentation and generation.

In eXplainable Constraint Solving (XCS), it is common to extract a Minimal Unsatisfiable Subset (MUS) from a set of unsatisfiable constraints. This helps explain to a user why a constraint specification does not admit a solution. Finding MUSes can be computationally expensive for highly symmetric problems, as many combinations of constraints need to be considered. In the traditional context of solving satisfaction problems, symmetry has been well studied, and effective ways to detect and exploit symmetries during the search exist. However, in the setting of finding MUSes of unsatisfiable constraint programs, symmetries are understudied. In this paper, we take inspiration from existing symmetry-handling techniques and adapt well-known MUS-computation methods to exploit symmetries in the specification, speeding-up overall computation time. Our results display a significant reduction of runtime for our adapted algorithms compared to the baseline on symmetric problems.