AAAI 2025 论文列表

Knowledge distillation (KD) is a model compression technique that transfers knowledge from a large teacher model to a smaller student model to enhance its performance. Existing methods often assume that the student model is inherently inferior to the teacher model. However, we identify that the fundamental issue affecting student performance is the bias transferred by the teacher. Current KD frameworks transmit both right and wrong knowledge, introducing bias that misleads the student model. To address this issue, we propose a novel strategy to rectify bias and greatly improve the student model's performance. Our strategy involves three steps: First, we differentiate knowledge and design a bias elimination method to filter out biases, retaining only the right knowledge for the student model to learn. Next, we propose a bias rectification method to rectify the teacher model's wrong predictions, fundamentally addressing bias interference. The student model learns from both the right knowledge and the rectified biases, greatly improving its prediction accuracy. Additionally, we introduce a dynamic learning approach with a loss function that updates weights dynamically, allowing the student model to quickly learn right knowledge-based easy tasks initially and tackle hard tasks corresponding to biases later, greatly enhancing the student model's learning efficiency. To the best of our knowledge, this is the first strategy enabling the student model to surpass the teacher model. Experiments demonstrate that our strategy, as a plug-and-play module, is versatile across various mainstream KD frameworks.

We show that current SOTA methods for privately and fairly training models are unreliable in many practical scenarios. Specifically, we (1) introduce a new type of adversarial attack that seeks to introduce unfairness into private model training, and (2) demonstrate that the use of methods for training on private data that are robust to adversarial attacks often leads to unfair models, regardless of the use of fairness-enhancing training methods. This leads to a dilemma when attempting to train fair models on private data: either (A) we use a robust training method which may introduce unfairness to the model itself, or (B) we train models which are vulnerable to adversarial attacks that introduce unfairness. This paper highlights flaws in robust learning methods when training fair models, yielding a new perspective for the design of robust and private learning systems.

As one of the key technologies leading to Artificial General Intelligence (AGI), Large Language Models (LLMs) have achieved remarkable accomplishments. Exploring the capabilities of LLMs is crucial for scientific research, and many studies propose new challenges from various aspects to explore the boundaries of capabilities in LLMs. This paper attempts to push the challenges of information understanding, synthesizing and reasoning to the extreme, in order to explore the boundaries of more advanced dimensional cognitive capabilities in LLMs. It is defined as the task of High-Level Cognition (HLC), which involves obtaining high-level conclusions from low-level and fragmented foundational information. To evaluate HLC, we construct a dataset based on soccer matches. Experiments and analysis on this dataset show that current state-of-the-art LLMs lack the ability to effectively solve the task of HLC, because their performance is equivalent to random-level. However, by fine-tuning Llama3-8B-Instruct, there are improvements of 14.4%, 48.1%, and 19.4% over random-level in three types of evaluation tasks. This indicates that LLMs have great potential to solve the task of HLC.

Large Vision-Language Models (LVLMs) have recently garnered significant attention, with many efforts aimed at harnessing their general knowledge to enhance the interpretability and robustness of autonomous driving models. However, LVLMs typically rely on large, general-purpose datasets and lack the specialized expertise required for professional and safe driving. Existing vision-language driving datasets focus primarily on scene understanding and decision-making, without providing explicit guidance on traffic rules and driving skills, which are critical aspects directly related to driving safety. To bridge this gap, we propose IDKB, a large-scale dataset containing over one million data items collected from various countries, including driving handbooks, theory test data, and simulated road test data. Much like the process of obtaining a driver's license, IDKB encompasses nearly all the explicit knowledge needed for driving from theory to practice. In particular, we conducted comprehensive tests on 15 LVLMs using IDKB to assess their reliability in the context of autonomous driving and provided extensive analysis. We also fine-tuned popular models, achieving notable performance improvements, which further validate the significance of our dataset.

Simulating learner actions helps stress-test open-ended interactive learning environments and prototype new adaptations before deployment. While recent studies show the promise of using large language models (LLMs) for simulating human behavior, such approaches have not gone beyond rudimentary proof-of-concept stages due to key limitations. First, LLMs are highly sensitive to minor prompt variations, raising doubts about their ability to generalize to new scenarios without extensive prompt engineering. Moreover, apparently successful outcomes can often be unreliable, either because domain experts unintentionally guide LLMs to produce expected results, leading to self-fulfilling prophecies; or because the LLM has encountered highly similar scenarios in its training data, meaning that models may not be simulating behavior so much as regurgitating memorized content. To address these challenges, we propose Hyp-Mix, a simulation authoring framework that allows experts to develop and evaluate simulations by combining testable hypotheses about learner behavior. Testing this framework in a physics learning environment, we found that GPT-4 Turbo maintains calibrated behavior even as the underlying learner model changes, providing the first evidence that LLMs can be used to simulate realistic behaviors in open-ended interactive learning environments, a necessary prerequisite for useful LLM behavioral simulation.

Malware authors often employ code obfuscations to make their malware harder to detect. Existing tools for generating obfuscated code often require access to the original source code (e.g., C++ or Java), and adding new obfuscations is a non-trivial, labor-intensive process. In this study, we ask the following question: Can Large Language Models (LLMs) potentially generate a new obfuscated assembly code? If so, this poses a risk to anti-virus engines and potentially increases the flexibility of attackers to create new obfuscation patterns. We answer this in the affirmative by developing the MetamorphASM benchmark comprising MetamorphASM Dataset (MAD) along with three code obfuscation techniques: dead code, register substitution, and control flow change. The MetamorphASM systematically evaluates the ability of LLMs to generate and analyze obfuscated code using MAD, which contains 328,200 obfuscated assembly code samples. We release this dataset and analyze the success rate of various LLMs (e.g., GPT-3.5/4, GPT-4o-mini, Starcoder, CodeGemma, CodeLlama, CodeT5, and LLaMA 3.1) in generating obfuscated assembly code. The evaluation was performed using established information-theoretic metrics and manual human review to ensure correctness and provide the foundation for researchers to study and develop remediations to this risk.

Prior work found that superhuman Go AIs like KataGo can be defeated by simple adversarial strategies. In this paper, we study if defenses can improve KataGo's worst-case performance. We test three natural defenses: adversarial training on hand-constructed positions, iterated adversarial training, and changing the network architecture. We find that though some of these defenses protect against previously discovered attacks, none withstand adaptive attacks. In particular, we are able to train new adversaries that reliably defeat our defended agents by causing them to blunder in ways humans would not. Our results suggest that building robust AI systems is challenging even for superhuman systems in narrow domains like Go.

The rapid advancement in self-supervised representation learning has highlighted its potential to leverage unlabeled data for learning rich visual representations. However, the existing techniques, particularly those employing different augmentations of the same image, often rely on a limited set of simple transformations that cannot fully capture variations in the real world. This constrains the diversity and quality of samples, which leads to sub-optimal representations. In this paper, we introduce a framework that enriches the self-supervised learning (SSL) paradigm by utilizing generative models to produce semantically consistent image augmentations. By directly conditioning generative models on a source image, our method enables the generation of diverse augmentations while maintaining the semantics of the source image, thus offering a richer set of data for SSL. Our extensive experimental results on various joint-embedding SSL techniques demonstrate that our framework significantly enhances the quality of learned visual representations by up to 10% Top-1 accuracy in downstream tasks. This research demonstrates that incorporating generative models into the joint-embedding SSL workflow opens new avenues for exploring the potential of synthetic data. This development paves the way for more robust and versatile representation learning techniques.

Accurately gauging the confidence level of Large Language Models' (LLMs) predictions is pivotal for their reliable application. However, LLMs are often uncalibrated inherently and elude conventional calibration techniques due to their proprietary nature and massive scale. In this work, we derive model confidence from the distribution of multiple randomly sampled generations, using three measures of consistency. We extensively evaluate eleven open and closed-source models on nine reasoning datasets. Results show that consistency-based calibration methods outperform existing post-hoc approaches in terms of calibration error. Meanwhile, we find that factors such as intermediate explanations, model scaling, and larger sample sizes enhance calibration, while instruction-tuning makes calibration more difficult. Moreover, confidence scores obtained from consistency can potentially enhance model performance. Finally, we offer guidance on choosing suitable consistency metrics for calibration, tailored to model characteristics such as the exposure to instruction-tuning and RLHF.

Partial multi-label learning (PML) aims to train a classifier on dataset whose instances are over-annotated with not only relevant labels but also irrelevant labels, which is common when datasets are collected from crowd-sourcing platform. Existing works primarily approach it from a curriculum learning perspective, leveraging the memorization effect to disambiguate noisy labels and produce robust predictions. However, these methods are based on non-adaptive weighting functions and lack theoretical guidance for optimal weighting. To overcome these issues, a calibrated disambiguation model named PML-CD is proposed. We firstly formulate the optimal weighting function for curriculum-based disambiguation, which is equivalent to the calibration of the model's predicted confidences, thus provide a guidance for curriculum designing. To obtain the optimal weighting function from PML dataset during the training, a transferable calibrator is designed, which takes the histogram of positive samples' confidences as input, and outputs the optimal curriculum weighting for training. Prototype alignment regularization is also proposed to promote the model's performance. Experiments conducted on Pascal VOC, MS-COCO, NUS-WIDE and CUB have verified that our method outperforms existing state-of-the-art PML methods.

Video saliency prediction aims to identify the regions in a video that attract human attention and gaze, driven by bottom-up features from the video and top-down processes like memory and cognition. Among these top-down influences, language plays a crucial role in guiding attention by shaping how visual information is interpreted. Existing methods primarily focus on modeling perceptual information while neglecting the reasoning process facilitated by language, where ranking cues are crucial outcomes of this process and practical guidance for saliency prediction. In this paper, we propose CaRDiff (Caption, Rank, and generate with Diffusion), a framework that imitates the process by integrating multimodal large language model (MLLM), a grounding module, and a diffusion model, to enhance video saliency prediction. Specifically, we introduce a novel prompting method VSOR-CoT (Video Slient Object Ranking Chain of Thought), which utilizes an MLLM with a grounding module to caption video content and infer salient objects along with their rankings and positions. This process derives ranking maps that can be sufficiently leveraged by the diffusion model to accurately decode the saliency maps for the given video. Extensive experiments showcase the effectiveness of VSOR-CoT in improving the performance of video saliency prediction. The proposed CaRDiff performs better than state-of-the-art models on the MVS dataset and demonstrates cross-dataset capabilities on the DHF1k dataset through zero-shot evaluation.

Despite the rapid development of Chinese vision-language models (VLMs), most existing Chinese vision-language (VL) datasets are constructed on Western-centric images from existing English VL datasets. The cultural bias in the images makes these datasets unsuitable for evaluating VLMs in Chinese culture. To remedy this issue, we present a new Chinese Vision-Language Understanding Evaluation (CVLUE) benchmark dataset, where the selection of object categories and images is entirely driven by Chinese native speakers, ensuring that the source images are representative of Chinese culture. The benchmark contains four distinct VL tasks ranging from image-text retrieval to visual question answering, visual grounding and visual dialogue. We present a detailed statistical analysis of CVLUE and provide a baseline performance analysis with several open-source multilingual VLMs on CVLUE and its English counterparts to reveal their performance gap between English and Chinese. Our in-depth category-level analysis reveals a lack of Chinese cultural knowledge in existing VLMs. We also find that fine-tuning on Chinese culture-related VL datasets effectively enhances VLMs' understanding of Chinese culture.

The National Vulnerability Database (NVD) publishes over a thousand new vulnerabilities monthly, with a projected 25 percent increase in 2024, highlighting the crucial need for rapid vulnerability identification to mitigate cybersecurity attacks and save costs and resources. In this work, we propose using large language models (LLMs) to learn vulnerability evaluation from historical assessments of medical device vulnerabilities in a single manufacturer's portfolio. We highlight the effectiveness and challenges of using LLMs for automatic vulnerability evaluation and introduce a method to enrich historical data with cybersecurity ontologies, enabling the system to understand new vulnerabilities without retraining the LLM. Our LLM system integrates with the in-house application - Cybersecurity Management System (CSMS) - to help Siemens Healthineers (SHS) product cybersecurity experts efficiently assess the vulnerabilities in our products. Also, we present guidelines for efficient integration of LLMs into the cybersecurity tool.

Trajectory prediction models that can infer both future trajectories and their associated uncertainties of the target vehicles is crucial for safe and robust navigation and path planning of autonomous vehicles. However, the majority of existing trajectory prediction models have neither considered reducing the uncertainty as one objective during the training stage nor provided reliable uncertainty quantification during inference stage, especially under potential distribution shift. Therefore, in this paper, we propose the Conformal Uncertainty Quantification under Distribution Shift framework, CUQDS, to quantify the uncertainty of the predicted trajectories of existing trajectory prediction models under potential data distribution shift, while improving the prediction accuracy of the models and reducing the estimated uncertainty during the training stage. Specifically, CUQDS includes 1) a learning-based Gaussian process regression module that models the output distribution of the base model (any existing trajectory prediction neural networks) and reduces the estimated uncertainty by an additional loss term, and 2) a statistical-based Conformal P control module to calibrate the estimated uncertainty from the Gaussian process regression module in an online setting under potential distribution shift between training and testing data. Experimental results on various state-of-the-art methods using benchmark motion forecasting datasets demonstrate the effectiveness of our proposed design.

Medical benchmark datasets significantly contribute to developing Large Language Models (LLMs) for medical knowledge extraction, diagnosis, summarization, and other uses. Yet, current benchmarks are mainly derived from exam questions given to medical students or cases described in the medical literature, lacking the complexity of real-world patient cases that deviate from classic textbook abstractions. These include rare diseases, uncommon presentations of common diseases, and unexpected treatment responses. Here, we construct Clinically Uncommon Patient Cases and Diagnosis Dataset (CUPCase) based on 3,563 real-world case reports from BMC, which we formulate into diagnoses in open-ended textual format and as multiple-choice options with distractors. Using this dataset, we evaluate the ability of state-of-the-art LLMs, including both general-purpose and Clinical LLMs, to identify and correctly diagnose a patient case, and test models' performance when only partial information about cases is available. Our findings show that general-purpose GPT-4o attains the best performance in both the multiple-choice task (average accuracy of 87.9%) and the open-ended task (BERTScore F1 of 0.764), outperforming several LLMs with a focus on the medical domain such as Meditron-70B and MedLM-Large. Moreover, GPT-4o was able to maintain 87% and 88% of its performance with only the first 20% of tokens of the case presentation in multiple-choice and free text, respectively, highlighting the potential of LLMs to aid in early diagnosis in real-world cases. An error analysis demonstrates the complexity of the task, and attempts to hypothesise about the models' reasoning. CUPCase expands our ability to evaluate LLMs for clinical decision support in an open and reproducible manner.

Recommendation systems (RS) play a crucial role in assisting decision-making but often suffer from either a lack of credibility or unfairness problems. A few recommendation models have endeavored to address the problem from only one aspect, and approaches to solving both problems remain to be explored. This paper aims to construct a generalized fairness-based recommendation framework that can also provide the credibility of recommendation models. Generally, we propose a reliable and fair recommendation framework called Conformalized User Group Fairness (CUGF) based on the inspiration of conformal prediction. Specifically, we construct dynamic prediction sets that are guaranteed to cover the true item with a user pre-specified probability to ensure credibility while designing novel fairness metrics based on empirical risks to guarantee the fairness of users across different groups. Furthermore, we design a novel CUGF Algorithm to optimize the parameter γ that dominates the prediction sets and also the fairness. Besides, we conduct extensive experiments by applying CUGF on top of various recommendation models and representative datasets to validate its effectiveness with respect to recommendation performance (in terms of average set size) and fairness (in terms of the two defined fairness metrics), the results of which demonstrate the validity of the proposed framework.

Categorical Distributional Reinforcement Learning (CDRL) has demonstrated superior sample efficiency in learning complex tasks compared to conventional Reinforcement Learning (RL) approaches. However, the practical application of CDRL is encumbered by challenging projection steps, detailed parameter tuning, and domain knowledge. This paper addresses these challenges by introducing a pioneering Continuous Distributional Model-Free RL algorithm tailored for continuous action spaces. The proposed algorithm simplifies the implementation of distributional RL, adopting an actor-critic architecture wherein the critic outputs a continuous probability distribution. Additionally, we propose an ensemble of multiple critics fused through a Kalman fusion mechanism to mitigate overestimation bias. Through a series of experiments, we validate that our proposed method provides a sample-efficient solution for executing complex continuous-control tasks.

In the domain of multivariate time series analysis, the concept of channel independence has been increasingly adopted, demonstrating excellent performance due to its ability to eliminate noise and the influence of irrelevant variables. However, such a concept often simplifies the complex interactions among channels, potentially leading to information loss. To address this challenge, we propose a strategy of channel independence followed by mixing. Based on this strategy, we introduce CSformer, a novel framework featuring a two-stage multiheaded self-attention mechanism. This mechanism is designed to extract and integrate both channel-specific and sequence-specific information. Distinctively, CSformer employs parameter sharing to enhance the cooperative effects between these two types of information. Moreover, our framework effectively incorporates sequence and channel adapters, significantly improving the model's ability to identify important information across various dimensions. Extensive experiments on several real-world datasets demonstrate that CSformer achieves state-of-the-art results in terms of overall performance.

Singing Voice Synthesis (SVS) aims to generate singing voices of high fidelity and expressiveness. Conventional SVS systems usually utilize an acoustic model to transform a music score into acoustic features, followed by a vocoder to reconstruct the singing voice. It was recently shown that end-to-end modeling is effective in the fields of SVS and Text to Speech (TTS). In this work, we thus present a fully end-to-end SVS method together with a chunkwise streaming inference to address the latency issue for practical usages. Note that this is the first attempt to fully implement end-to-end streaming audio synthesis using latent representations in VAE. We have made specific improvements to enhance the performance of streaming SVS using latent representations. Experimental results demonstrate that the proposed method achieves synthesized audio with high expressiveness and pitch accuracy in both streaming SVS and TTS tasks.