AAAI 2025 Papers

In this paper, we augment online algorithms for the knapsack problem using the total weight information. The conventional optimal online algorithm achieves the ln(U/L)+1 competitive ratio where L and U are the upper and lower bounds of the value-to-weight ratio. However, it does not consider that decision makers can know the total weight information or obtain it through machine-learned predictions. To fill this gap, we first propose the Known Weight Algorithm (KWA) which uses the exact total weight information to achieve a competitive ratio of W((U-L)/(eL))+1, where W denotes the Lambert-W function. We prove that it is optimal and tight. After that, we extend KWA to the Predicted Weight Algorithm (PWA), a learning-augmented online algorithm that uses predicted total weight. We show the consistency and robustness of PWA, and prove that its competitive ratio degrades gracefully as the prediction error grows. Finally, we introduce the Limited Volume Algorithm (LWA), which achieves a better competitive ratio than ln(U/L)+1 when the total weight is less than twice the capacity.

Despite the advancements in large language models (LLMs) for mathematical reasoning, solving competition-level math problems remains a significant challenge, especially for open-source LLMs without external tools. We introduce the MMIQC dataset, comprising a mixture of processed web data and synthetic question-response pairs, aimed at enhancing the mathematical reasoning capabilities of base language models. Models fine-tuned on MMIQC consistently surpass their counterparts in performance on the MATH benchmark across various model sizes. Notably, Qwen-72B-MMIQC achieves a 45.0% accuracy, exceeding the previous open-source state-of-the-art by 8.2% and outperforming the initial version GPT-4 released in 2023. Extensive evaluation results on Hungarian high school finals suggest that such improvement can generalize to unseen data. Our ablation study on MMIQC reveals that a large part of the improvement can be attributed to our novel augmentation method, Iterative Question Composing (IQC), which involves iteratively composing new questions from seed problems using an LLM and applying rejection sampling through another LLM.

We applied Risk-averse Reinforcement Learning (RL) to optimize investment portfolios while incorporating risk constraints. Given that portfolios must adhere to risk constraints set by investors and regulators, enforcing hard constraints is essential for practical portfolio optimization. Traditional techniques often lack the flexibility to model the complexities of dynamic financial markets. To address this, we used the Augmented Lagrangian Multiplier (ALM) to impose constraints on the agent, reducing risk during decision-making. Our risk-constrained RL algorithm demonstrated no constraint violations during testing and outperformed other Risk-averse RL methods, indicating its potential for optimizing portfolios for risk-averse investors.

3D visual grounding (3DVG), which aims to correlate a natural language description with the target object within a 3D scene, is a significant yet challenging task. Despite recent advancements in this domain, existing approaches commonly encounter a shortage: a limited amount and diversity of text-3D pairs available for training. Moreover, they fall short in effectively leveraging different contextual clues (e.g., rich spatial relations within the 3D visual space) for grounding. To address these limitations, we propose AugRefer, a novel approach for advancing 3D visual grounding. AugRefer introduces cross-modal augmentation designed to extensively generate diverse text-3D pairs by placing objects into 3D scenes and creating accurate and semantically rich descriptions using foundation models. Notably, the resulting pairs can be utilized by any existing 3DVG methods for enriching their training data. Besides, AugRefer presents a language-spatial adaptive decoder that effectively adapts the potential referring objects based on the language description and various 3D spatial relations. Extensive experiments on three benchmark datasets clearly validate the effectiveness of AugRefer.

Conditional demographic parity (CDP) is a measure of the demographic parity of a predictive model or decision process when conditioning on an additional feature or set of features. Many algorithmic fairness techniques exist to target demographic parity, but CDP is much harder to achieve, particularly when the conditioning variable has many levels and/or when the model outputs are continuous. The problem of auditing and enforcing CDP is understudied in the literature. In light of this, we propose novel measures of conditional demographic disparity (CDD) which rely on statistical distances borrowed from the optimal transport literature. We further design and evaluate regularization-based approaches based on these CDD measures. Our methods, FairBiT and FairLeap, allow us to target conditional demographic parity even when the conditioning variable has many levels. When model outputs are continuous, our methods target full equality of the conditional distributions, unlike other methods that only consider first moments or related proxy quantities. We validate our approaches on real-world datasets.

Text-to-audio generation models (TAG) have achieved significant advances in generating audio conditioned on text descriptions. However, a critical challenge lies in the lack of transparency regarding how each textual input impacts the generated audio. To address this issue, we introduce AudioGenX, an Explainable AI (XAI) method that provides explanations for text-to-audio generation models by highlighting the importance of input tokens. AudioGenX optimizes an Explainer by leveraging factual and counterfactual objective functions to provide faithful explanations at the audio token level. This method offers a detailed and comprehensive understanding of the relationship between text inputs and audio outputs, enhancing both the explainability and trustworthiness of TAG models. Extensive experiments demonstrate the effectiveness of AudioGenX in producing faithful explanations, benchmarked against existing methods using novel evaluation metrics specifically designed for audio generation tasks.

The Audio-Visual Question Answering (AVQA) task involves extracting question-related audio-visual clues from both temporal and spatial perspectives to answer questions accurately. Despite the promising performance of existing multi-modal AVQA models, thanks to large-scale pre-trained models, challenges remain in the field. Firstly, aligning audio-visual information across temporal and spatial dimensions is difficult. Secondly, the fusion of audio-visual information is often weighted inadequately, limiting model performance. To address the above issues, we design the Audio-Visual Adaptive Fusion Network (AVAF-Net), which uses contrastive learning to align audio-visual information temporally and spatially and adaptively adjusts fusion weights based on the question. Specifically, we initially align visual and audio information temporally through a temporal-alignment contrastive loss. This is followed by an audio-visual clue-mining module that highlights question-related cues, aligning them with the vocal region spatially using spatial alignment contrastive loss. Additionally, a question-oriented adaptive fusion module assigns different weights to audio and visual modalities based on the question content and then fuses them. The fused audio-visual cues are finally used to predict the answer. Extensive experiments on the MUSIC-AVQA dataset show that AVAF-Net surpasses all baseline models, with a maximum improvement of 15.90% in average accuracy and an average improvement of 9.80%.

Recent literature uses language to build foundation models for audio. These Audio-Language Models (ALMs) are trained on a vast number of audio-text pairs and show remarkable performance in tasks including Text-to-Audio Retrieval, Captioning, and Question Answering. However, their ability to engage in more complex open-ended tasks, like Interactive Question-Answering, requires proficiency in logical reasoning- a skill not yet benchmarked. We introduce the novel task of Audio Entailment to evaluate an ALM's deductive reasoning ability. This task assesses whether a text description (hypothesis) of audio content can be deduced from an audio recording (premise), with potential conclusions being entailment, neutral, or contradiction, depending on the sufficiency of the evidence. We create two datasets for this task with audio recordings sourced from two audio captioning datasets-AudioCaps and Clotho-and hypotheses generated using Large Language Models (LLMs). We benchmark state-of-the-art ALMs and find deficiencies in logical reasoning with both zero-shot and linear probe evaluations. Finally, we propose "caption-before-reason", an intermediate step of captioning that improves the Zero-Shot and linear-probe performance of ALMs by an absolute 6% and 3%, respectively.

This study investigates user engagement and political polarization on YouTube Shorts, a special category of YouTube videos with a duration of 15-60 seconds. Via a substantial corpus of 38,838 videos gleaned from 100 YouTube channels focusing on political content, we contrast YouTube Shorts with long-form content in terms of user engagement, content toxicity, and polarization. Our analyses reveal that (1) YouTube Shorts receive more likes and views and fewer comments as compared to their long-form video counterparts; (2) YouTube Shorts are more toxic; and (3) considerably more polarized than long-form YouTube videos.

In recent years, large language models (LLMs) have demonstrated outstanding capabilities in various tasks. However, LLMs also have various drawbacks, especially hallucination. Hallucination refers to the generation of content that does not align with the user input, contradicts previously generated content or world knowledge. Current research on hallucination mainly include knowledge retrieval, prompt engineering, training data improvement, reinforcement learning, etc. However, these methods do not involve different categories of hallucinations which is important on hallucination analysis, and make detailed investigation for the internal state of LLMs which indicates the direction on hallucination occurrence. Therefore, in our research, we introduce an attribution framework to trace the origins of hallucinations based on the internal signals of LLMs. To support this framework, we develop a new benchmark named RelQA-Cate, which includes eight categories of hallucinations for the answers generated by LLMs. After that, we present a novel Differential Penalty Decoding (DPD) strategy for reducing hallucinations through adjusting post-probabilities of each answer. We conduct a series of experiments and the performance on answer reliability has significant improvement, achieving 28.25% at most, which demonstrates the effectiveness of our proposed DPD and its generalization in mitigating hallucination in LLMs.

Model editing aims to correct outdated or erroneous knowledge in large models without costly retraining. Recent research discovered that the mid-layer representation of the subject's final token in a prompt has a strong influence on factual predictions, and developed Large Language Model (LLM) editing techniques based on this observation. However, for Vision-LLMs (VLLMs), how visual representations impact the predictions from a decoder-only language model remains largely unexplored. To the best of our knowledge, model editing for VLLMs has not been extensively studied in the literature. In this work, we employ the contribution allocation and noise perturbation methods to measure the contributions of visual representations for token predictions. Our attribution analysis shows that visual representations in mid-to-later layers that are highly relevant to the prompt contribute significantly to predictions. Based on these insights, we propose *VisEdit*, a novel model editor for VLLMs that effectively corrects knowledge by editing intermediate visual representations in regions important to the edit prompt. We evaluated *VisEdit* using multiple VLLM backbones and public VLLM editing benchmark datasets. The results show the superiority of *VisEdit* over the strong baselines adapted from existing state-of-the-art editors for LLMs.

Federated Learning (FL) enables multiple clients, such as mobile phones and IoT devices, to collaboratively train a global machine learning model while keeping their data localized. However, recent studies have revealed that the training phase of FL is vulnerable to reconstruction attacks, such as attribute inference attacks (AIA), where adversaries exploit exchanged messages and auxiliary public information to uncover sensitive attributes of targeted clients. While these attacks have been extensively studied in the context of classification tasks, their impact on regression tasks remains largely unexplored. In this paper, we address this gap by proposing novel model-based AIAs specifically designed for regression tasks in FL environments. Our approach considers scenarios where adversaries can either eavesdrop on exchanged messages or directly interfere with the training process. We benchmark our proposed attacks against state-of-the-art methods using real-world datasets. The results demonstrate a significant increase in reconstruction accuracy, particularly in heterogeneous client datasets, a common scenario in FL. The efficacy of our model-based AIAs makes them better candidates for empirically quantifying privacy leakage for federated regression tasks.

Recently, diffusion models have emerged as promising newcomers in the field of generative models, shining brightly in image generation. However, when employed for object removal tasks, they still encounter issues such as generating random artifacts and the incapacity to repaint foreground object areas with appropriate content after removal. To tackle these problems, we propose Attentive Eraser, a tuning-free method to empower pre-trained diffusion models for stable and effective object removal. Firstly, in light of the observation that the self-attention maps influence the structure and shape details of the generated images, we propose Attention Activation and Suppression (ASS), which re-engineers the self-attention mechanism within the pre-trained diffusion models based on the given mask, thereby prioritizing the background over the foreground object during the reverse generation process. Moreover, we introduce Self-Attention Redirection Guidance (SARG), which utilizes the self-attention redirected by ASS to guide the generation process, effectively removing foreground objects within the mask while simultaneously generating content that is both plausible and coherent. Experiments demonstrate the stability and effectiveness of Attentive Eraser in object removal across a variety of pre-trained diffusion models, outperforming even training-based methods. Furthermore, Attentive Eraser can be implemented in various diffusion model architectures and checkpoints, enabling excellent scalability.

With the successful transition of Transformers from natural language processing (NLP) to computer vision (CV) domains, Vision Transformers (ViTs) have achieved state-of-the-art performance in many CV tasks. However, backdoor attacks, a significant threat in deep learning, also pose a risk to the security of ViT models. Recently, several backdoor attack methods targeting the patch-level self-attention mechanism in ViTs have been proposed, but they are relatively naive in terms of stealthiness and robustness against defensive measures, lacking in-depth investigation. In this paper, we explore the crucial role of attention-level imperceptibility in backdoor attacks for ViTs and propose an Attention-Imperceptible Backdoor Attacks on Vision Transformers (AIBA). In AIBA, a constrained adversarial perturbation is used as the trigger to achieve visual imperceptibility. Additionally, the trigger is designed to seamlessly implant into the focal areas of the image, ensuring that the trigger receives enough attention from the model without causing anomalies at the attention level. During the backdoor learning process, we designed an efficient constrained bi-level optimization training strategy at the mini-batch level to implant an effective backdoor in the victim model using the imperceptible trigger. We evaluated the effectiveness of the proposed AIBA across multiple datasets and ViT benchmarks and explored the robustness of AIBA against current ViT-specific defense methods. The experimental results demonstrate that our backdoor attack method can successfully implant a powerful and stealthy backdoor into ViTs.

Recent advancements in Multimodal Large Language Models (MLLMs) have generated significant interest in their ability to autonomously interact with and interpret Graphical User Interfaces (GUIs). A major challenge in these systems is grounding—accurately identifying critical GUI components such as text or icons based on a GUI image and a corresponding text query. Traditionally, this task has relied on fine-tuning MLLMs with specialized training data to predict component locations directly. However, in this paper, we propose a novel Tuning-free Attention-driven Grounding (TAG) method that leverages the inherent attention patterns in pretrained MLLMs to accomplish this task without the need for additional fine-tuning. Our method involves identifying and aggregating attention maps from specific tokens within a carefully constructed query prompt. Applied to MiniCPM-Llama3-V 2.5, a state-of-the-art MLLM, our tuning-free approach achieves performance comparable to tuning-based methods, with notable success in text localization. Additionally, we demonstrate that our attention map-based grounding technique significantly outperforms direct localization predictions from MiniCPM-Llama3-V 2.5, highlighting the potential of using attention maps from pretrained MLLMs and paving the way for future innovations in this domain.

Test-time adaptation aims to adapt a well-trained model to potential distribution shifts at test time using only unlabeled test data, without access to the original training data. While previous efforts mainly focus on a single modality, test-time distribution shift in the multi-modal setting is more complex and calls for new solutions. This paper tackles the problem of multi-modal test-time adaptation by proposing a novel method named Attention Bootstrapping with Principal Entropy Minimization (ABPEM). We observe that test-time distribution shift causes misalignment across modalities, leading to a large gap between intra-modality discrepancies (measured by self-attention) and inter-modality discrepancies (measured by cross-attention). We name this the attention gap. This attention gap widens with more severe distribution shifts, hindering effective modality fusion. To mitigate this attention gap and encourage better modality fusion, we propose attention bootstrapping that promotes cross-attention with the guidance of self-attention. Moreover, to reduce the gradient noise in the commonly-used entropy minimization, we adopt principal entropy minimization, a refinement of entropy minimization that reduces gradient noise by focusing on the principal parts of entropy, excluding less reliable gradient information. Extensive experiments on the benchmarks validate the effectiveness of the proposed ABPEM in comparison with competing baselines.

Zero-shot learning (ZSL) endeavors to transfer knowledge from the seen categories to recognize unseen categories, which mostly relies on the semantic-visual interactions between image and attribute tokens. Recently, the prompt learning has emerged in ZSL and demonstrated significant potential as it allows the zero-shot transfer of diverse visual concepts to downstream tasks. However, current methods explore the fixed adaptation of the learnable prompt on the seen domains, which make them over-emphasize the primary visual features observed during training, limiting their generalization capabilities to the unseen domains. In this work, we propose AENet, which endows semantic information into the visual prompt to distill semantic-enhanced prompt for visual representation enrichment, enabling effective knowledge transfer for ZSL. AENet comprises two key steps: 1) exploring the concept-harmonized tokens for the visual and attribute modalities, grounded on the modal-sharing token that represents consistent visual-semantic concepts; and 2) yielding the semantic-enhanced prompt via the visual residual refinement unit with attribute consistency supervision. It is further integrated with primary visual features to attend to semantic-related information for visual enhancement, thus strengthening transferable ability. Experimental results on three benchmarks show that our AENet outperforms existing state-of-the-art ZSL methods.

While novel gradient-based attacks are continuously proposed to improve the optimization of adversarial examples, each is shown to outperform its predecessors using different experimental setups, implementations, and computational budgets, leading to biased and unfair comparisons. In this work, we overcome this issue by proposing AttackBench, i.e., an attack evaluation framework that evaluates the effectiveness of each attack (along with its different library implementations) under the same maximum available computational budget. To this end, we (i) define a novel optimality metric that quantifies how close each attack is to the optimal solution (empirically estimated by ensembling all attacks), and (ii) limit the maximum number of forward and backward queries that each attack can execute on the target model. Our extensive experimental analysis compares more than 100 attack implementations over 800 different configurations, considering both CIFAR-10 and ImageNet models, and shows that only a few attack implementations outperform all the remaining approaches. These findings suggest that novel defenses should be evaluated against different attacks than those normally used in the literature to avoid overly-optimistic robustness evaluations. We release AttackBench as a publicly-available benchmark that will be continuously updated with new attack implementations to maintain an up-to-date ranking of the best gradient-based attacks. We release AttackBench as a publicly available benchmark, including a continuously updated leaderboard and source code to maintain an up-to-date ranking of the best gradient-based attacks.

Deep neural networks (DNNs) have substantially achieved high predictive accuracy in many vision tasks. However, we find that they are poorly calibrated for crack recognition tasks, as these DNNs tend to produce both under-confident and over-confident predictions in such safety-critical applications, thereby limiting their practical use in real-world scenarios. To address this issue, we propose a novel attack-inspired calibration loss (AICL) that explicitly regularizes class probabilities to be better confidence estimation. Specifically, we first propose the attack-inspired correctness estimation method (ACE) that aims to estimate the correctness degree of each sample via adversarial attacks. Then, we propose Correctness-aware Distribution Guidance, which starts from a distribution perspective that enforces the ordinal ranking of the predicted confidence referring to the estimated correctness degree. The proposed method can be conveniently implemented on top of any DNNs-based crack recognition model by serving as a plug-and-play loss function. To address the limited availability of related benchmarks, we collect a fully annotated dataset, namely, Bridge2024, which involves inconsistent cracks and noisy backgrounds in real-world bridges. Our AICL outperforms the state-of-art calibration methods on various benchmark datasets including CRACK2019, SDNET2018, and our BRIDGE2024.