AAAI 2025 Papers

In recent years, the distributed training of foundation models (FMs) has seen a surge in popularity. In particular, federated learning enables collaborative model training among edge clients while safeguarding the privacy of their data. However, federated training of FMs across resource-constrained and highly heterogeneous edge devices encounter several challenges. These include the difficulty of deploying FMs on clients with limited computational resources and the high computation and communication costs associated with fine-tuning and collaborative training. To address these challenges, we propose FedCKMS, a Cluster-Aware Framework with Knowledge-Aware Model Search. Specifically, FedCKMS incorporates three key components. The first component is multi-factor heterogeneity-aware clustering, which groups clients based on both data distribution and resource limitations and selects an appropriate model for each cluster. The second component is knowledge-aware model architecture search, which enables each client to identify the optimal sub-model from the cluster model, facilitating adaptive deployment that accommodates highly heterogeneous computational resources across clients. The final component is cluster-aware knowledge transfer, which facilitates knowledge sharing between clusters and the server, addressing model heterogeneity, and reducing communication overhead. Extensive experiments demonstrate that FedCKMS outperforms state-of-the-art baselines by 3-10% in accuracy.

Every pharmaceutical product must be accompanied by a comprehensive label that delineates its indications, usage, dosages, and side effects, essential for safe medication practices. Traditionally, creating drug labels is labor-intensive and dependent on manual quality checks. Recent advancements in Large Language Models (LLMs) offer a promising avenue to streamline this process. In this paper we introduce ClinicalRAG, an automated labeling quality control pipeline that integrates LLM with hierarchical Retrieval Augmented Generation that allows to cross-check every statement in the drug label document. ClinicalRAG enhances the reliability of automated drug labeling by systematically reducing hallucination risks, achieving an accuracy of 96.1% in internal validation. With user-friendly interface, our pipeline aims to support pharmaceutical company in drug approval and expedite patients' access to new treatments.

Recent advancements in generative models have revolutionized image generation and editing, making these tasks accessible to non-experts. This paper focuses on local image editing, particularly the task of adding new content to a loosely specified area. Existing methods often require a precise mask or a detailed description of the location, which can be cumbersome and prone to errors. We propose Click2Mask, a novel approach that simplifies the local editing process by requiring only a single point of reference (in addition to the content description). A mask is dynamically grown around this point during a Blended Latent Diffusion (BLD) process, guided by a masked CLIP-based semantic loss. Click2Mask surpasses the limitations of segmentation-based and fine-tuning dependent methods, offering a more user-friendly and contextually accurate solution. Our experiments demonstrate that Click2Mask not only minimizes user effort but also enables competitive or superior local image manipulations compared to SoTA methods, according to both human judgement and automatic metrics. Key contributions include the simplification of user input, the ability to freely add objects unconstrained by existing segments, and the integration potential of our dynamic mask approach within other editing methods.

Graph neural networks (GNNs) have achieved impressive results in various graph learning tasks. Backdoor attacks pose a significant threat to GNNs, with a focus on dirty-label attacks. However, these attacks often necessitate the inclusion of blatantly incorrect inputs into the training set, rendering them easily detectable through simple filtering. In response to this challenge, we introduce Clean-Label Graph Backdoor Attack (CGBA). The majority of features in the generated poisoned nodes align with their true labels, significantly enhancing the difficulty of detecting the attack. Firstly, leveraging the uncertainty inherent in the GNNs, we develop a low-budget strategy for selecting poisoned nodes. This approach focuses on nodes in the target class with uncertain and low-degree classifications, allowing for efficient attacks within a limited budget while mitigating the impact on other clean nodes. Secondly, we present an innovative strategy for generating feature triggers. By boosting the confidence of poisoned samples in the target class, this tactic establishes a robust association between the trigger and the target class, even without modifying the labels of poisoned nodes. Additionally, we incorporate two constraints to reduce disruption to the graph structure. In conclusion, comprehensive experimental results unequivocally showcase CGBA's exceptional attack performance across three benchmark datasets and four GNNs models. Notably, the attack targeting the GraphSAGE model attains a 100% success rate, accompanied by a marginal benign accuracy drop of no more than 0.5%.

Compared to conventional long-tail learning, which focuses on addressing class-wise imbalances, generalized long-tail (GLT) learning considers that samples within each class still conform to long-tailed distributions due to varying attributes, known as attribute imbalance. In the presence of such imbalance, the assumption of equivalence between the class-conditional probability densities of the training and testing sets is no longer tenable. Existing GLT approaches typically employ regularization techniques to avoid directly modeling the class-conditional probability density (CCPD) ratio between training and test data, leading to suboptimal performance. This study aims to directly estimate this ratio, for which a novel class-attribute aware logit-adjusted (CALA) loss incorporating both the CCPD ratio and the class priors is presented. Two new GLT learning methods, named Heuristic-CALA and Meta-CALA, are then proposed, which estimate the CCPD ratio in the CALA loss by leveraging the neighborhood information of samples. Extensive experiments across diverse scenarios susceptible to class and attribute imbalances showcase the state-of-the-art performance of Meta-CALA. Furthermore, while Heuristic-CALA exhibits inferior performance compared to Meta-CALA, it incurs only negligible additional training time compared to the Cross-Entropy loss, yet surpasses existing methods by a significant margin.

Deep learning has achieved remarkable success in supervised image classification tasks, which relies on a large number of labeled samples for each class. Recently, zero-shot learning has garnered significant attention, which aims to recognize unseen classes using only training samples from seen classes. To bridge the gap between images and classes, class semantic attributes are introduced, making the alignment between image and class semantic attributes critical to zero-shot learning. However, existing methods often struggle to accurately focus on the image regions corresponding to individual class semantic attributes and tend to overlook the relations between different regions of an image, leading to poor alignment. To address these challenges, we propose a class semantic attribute perception guided zero-shot learning method. Specifically, we achieve coarse-grained perception of class semantic attributes across the entire image through contrastive semantic learning. Additionally, we attain fine-grained perception of individual class semantic attributes within image regions via region partitioning-based attribute alignment, which fully considers the relations between different regions of an image. By integrating these two processes into a unified network, we achieve multi-grained class semantic attribute perception, thereby enhancing the alignment between images and class semantic attributes. We validate the effectiveness of the proposed method on zero-shot learning benchmark data sets.

Question answering systems are rapidly advancing, but their opaque nature may impact user trust. We explored trust through an anti-monitoring framework, where trust is predicted to be correlated with presence of citations and inversely related to checking citations. We tested this hypothesis with a live question-answering experiment that presented text responses generated using a commercial Chatbot along with varying citations (zero, one, or five), both relevant and random, and recorded if participants checked the citations and their self-reported trust in the generated responses. We found a significant increase in trust when citations were present, a result that held true even when the citations were random; we also found a significant decrease in trust when participants checked the citations. These results highlight the importance of citations in enhancing trust in AI-generated content.

We present an open-source tool for manipulating Boolean circuits. It implements efficient algorithms, both existing and novel, for a rich variety of frequently used circuit tasks such as satisfiability, synthesis, and minimization. We tested the tool on a wide range of practically relevant circuits (computing, in particular, symmetric and arithmetic functions) that have been optimized intensively by the community for the last three years. The tool helped us to win the IWLS 2024 Programming Contest. In 2023, it was Google DeepMind who took the first place in the competition. We were able to reduce the size of the best circuits from 2023 by 12% on average, whereas for some individual circuits, our size reduction was as large as 83%.

Transfer learning for bio-signals has recently become an important technique to improve prediction performance on downstream tasks with small bio-signal datasets. Recent works have shown that pre-training a neural network model on a large dataset (e.g. EEG) with a self-supervised task, replacing the self-supervised head with a linear classification head, and fine-tuning the model on different downstream bio-signal datasets (e.g., EMG or ECG) can dramatically improve the performance on those datasets. In this paper, we propose a new convolution-transformer hybrid model architecture with masked auto-encoding for low-data bio-signal transfer learning, introduce a frequency-based masked auto-encoding task, employ a more comprehensive evaluation framework, and evaluate how much and when (multimodal) pre-training improves fine-tuning performance. We also introduce a dramatically more performant method of aligning a downstream dataset with a different temporal length and sampling rate to the original pre-training dataset. Our findings indicate that the convolution-only part of our hybrid model can achieve state-of-the-art performance on some low-data downstream tasks. The performance is often improved even further with our full model. In the case of transformer-based models we find that pre-training especially improves performance on downstream datasets, multimodal pre-training often increases those gains further, and our frequency-based pre-training performs the best on average for the lowest and highest data regimes.

The importance of time series forecasting drives continuous research and the development of new approaches to tackle this problem. Typically, these methods are introduced through empirical studies that frequently claim superior accuracy for the proposed approaches. Nevertheless, concerns are rising about the reliability and generalizability of these results due to limitations in experimental setups. This paper addresses a critical limitation: the number and representativeness of the datasets used. We investigate the impact of dataset selection bias, particularly the practice of cherry-picking datasets, on the performance evaluation of forecasting methods. Through empirical analysis with a diverse set of benchmark datasets, our findings reveal that cherry-picking datasets can significantly distort the perceived performance of methods, often exaggerating their effectiveness. Furthermore, our results demonstrate that by selectively choosing just four datasets — what most studies report — 46% of methods could be deemed best in class, and 77% could rank within the top three. Additionally, recent deep learning-based approaches show high sensitivity to dataset selection, whereas classical methods exhibit greater robustness. Finally, our results indicate that, when empirically validating forecasting algorithms on a subset of the benchmarks, increasing the number of datasets tested from 3 to 6 reduces the risk of incorrectly identifying an algorithm as the best one by approximately 40%. Our study highlights the critical need for comprehensive evaluation frameworks that more accurately reflect real-world scenarios. Adopting such frameworks will ensure the development of robust and reliable forecasting methods.

Large Language Models (LLMs) have achieved remarkable success and have been applied across various scientific fields, including chemistry. However, many chemical tasks require the processing of visual information, which cannot be successfully handled by existing chemical LLMs. This brings a growing need for models capable of integrating multimodal information in the chemical domain. In this paper, we introduce ChemVLM, an open-source chemical multimodal large language model specifically designed for chemical applications. ChemVLM is trained on a carefully curated bilingual multimodal dataset that enhances its ability to understand both textual and visual chemical information, including molecular structures, reactions, and chemistry examination questions. We develop three datasets for comprehensive evaluation, tailored to Chemical Optical Character Recognition (OCR), Multimodal Chemical Reasoning (MMCR), and Multimodal Molecule Understanding tasks. We benchmark ChemVLM against a range of open-source and proprietary multimodal large language models on various tasks. Experimental results demonstrate that ChemVLM achieves competitive performance across all evaluated tasks.

We investigate the problem of checking the consistency of qualitative preferences expressed in CP-theory. This problem is PSPACE-Complete even when the preferences are locally consistent or the preference variables have binary domain. We present a new sufficient condition for consistency of preferences and show that the condition can be checked in polynomial time in settings of practical relevance (locally consistent or binary domain preference variables). We further show how the resulting sufficient condition can be used to efficiently identify a subset of outcomes that are non-dominated with respect to a set of qualitative preferences.

In this study, we establish a benchmark and a baseline approach for Multimodal referring and grounding with Chain-of-Questions (MCQ), opening up a promising direction for ‘logical’ multimodal dialogues. The newly collected dataset, named CB-300K, spans challenges including probing dialogues with spatial relationship among multiple objects, consistent reasoning, and complex question chains. The baseline approach, termed ChatterBox, involves a modularized design and a referent feedback mechanism to ensure logical coherence in continuous referring and grounding tasks. This design reduces the risk of referential confusion, simplifies the training process, and presents validity in retaining the language model’s generation ability. Experiments show that ChatterBox demonstrates superiority in MCQ both quantitatively and qualitatively, paving a new path towards multimodal dialogue scenarios with logical interactions.

Human experts typically integrate numerical and textual multimodal information to analyze time series. However, most traditional deep learning predictors rely solely on unimodal numerical data, using a fixed-length window for training and prediction on a single dataset, and cannot adapt to different scenarios. The powered pre-trained large language model has introduced new opportunities for time series analysis. Yet, existing methods are either inefficient in training, incapable of handling textual information, or lack zero-shot forecasting capability. In this paper, we innovatively model time series as a foreign language and construct ChatTime, a unified framework for time series and text processing. As an out-of-the-box multimodal time series foundation model, ChatTime provides zero-shot forecasting capability and supports bimodal input/output for both time series and text. We design a series of experiments to verify the superior performance of ChatTime across multiple tasks and scenarios, and create four multimodal datasets to address data gaps. The experimental results demonstrate the potential and utility of ChatTime.

Large language models (LLMs) are expected to follow instructions from users and engage in conversations. Techniques to enhance LLMs' instruction-following capabilities typically fine-tune them using data structured according to a predefined chat template. Although chat templates are shown to be effective in optimizing LLM performance, their impact on safety alignment of LLMs has been less understood, which is crucial for deploying LLMs safely at scale. In this paper, we investigate how chat templates affect safety alignment of LLMs. We identify a common vulnerability, named ChatBug, that is introduced by chat templates. Our key insight to identify ChatBug is that the chat templates provide a rigid format that need to be followed by LLMs, but not by users. Hence, a malicious user may not necessarily follow the chat template when prompting LLMs. Instead, malicious users could leverage their knowledge of the chat template and accordingly craft their prompts to bypass safety alignments of LLMs. We study two attacks to exploit the ChatBug vulnerability. Additionally, we demonstrate that the success of multiple existing attacks can be attributed to the ChatBug vulnerability. We show that a malicious user can exploit the ChatBug vulnerability of eight state-of-the-art (SOTA) LLMs and effectively elicit unintended responses from these models. Moreover, we show that ChatBug can be exploited by existing jailbreak attacks to enhance their attack success rates. We investigate potential countermeasures to ChatBug. Our results show that while adversarial training effectively mitigates the ChatBug vulnerability, the victim model incurs significant performance degradation. These results highlight the trade-off between safety alignment and helpfulness. Developing new methods for instruction tuning to balance this trade-off is an open and critical direction for future research.

In Tennenholtz’s program equilibrium, players of a game submit programs to play on their behalf. Each program receives the other programs’ source code and outputs an action. This can model interactions involving AI agents, mutually transparent institutions, or commitments. Tennenholtz 2004 (https://doi.org/10.1016/j.geb.2004.02.002) proves a folk theorem for program games, but the equilibria constructed are very brittle. We therefore consider simulation-based programs – i.e., programs that work by running opponents’ programs. These are relatively robust (in particular, two programs that act the same are treated the same) and are more practical than proof-based approaches. Oesterheld’s (2019, https://doi.org/10.1007/s11238-018-9679-3) epsilon-Grounded-pi-Bot is such an approach. Unfortunately, it is not generally applicable to games of three or more players, and only allows for a limited range of equilibria in two player games. In this paper, we propose a generalisation to Oesterheld’s (2019) epsilon-Grounded-pi-Bot. We prove a folk theorem for our programs in a setting with access to a shared source of randomness. We then characterise their equilibria in a setting without shared randomness. Both with and without shared randomness, we achieve a much wider range of equilibria than Oesterheld’s (2019) epsilon-Grounded-pi-Bot. Finally, we explore the limits of simulation-based program equilibrium, showing that the Tennenholtz folk theorem cannot be attained by simulation-based programs without access to shared randomness.

Character-based dialogue (aka role-playing) enables users to freely customize characters for interaction, which often relies on LLMs, raising the need to evaluate LLMs’ character customization capability. However, existing benchmarks fail to ensure a robust evaluation as they often only involve a single character category or evaluate limited dimensions. Moreover, the sparsity of character features in responses makes feature-focused generative evaluation both ineffective and inefficient. To address these issues, we propose CharacterBench, the largest bilingual generative benchmark, with 22,859 human-annotated samples covering 3,956 characters from 25 detailed character categories. We define 11 dimensions of 6 aspects, classified as sparse and dense dimensions based on whether character features evaluated by specific dimensions manifest in each response. We enable effective and efficient evaluation by crafting tailored queries for each dimension to induce characters’ responses related to specific dimensions. Further, we develop CharacterJudge model for cost-effective and stable evaluations. Experiments show its superiority over SOTA automatic judges (e.g., GPT-4) and our benchmark’s potential to optimize LLMs’ character customization.

Lately, the practice of utilizing task-specific fine-tuning has been implemented to improve the performance of large language models (LLM) in subsequent tasks. Through the integration of diverse LLMs, the overall competency of LLMs is significantly boosted. Nevertheless, traditional ensemble methods are notably memory-intensive, necessitating the simultaneous loading of all specialized models into GPU memory. To address the inefficiency, model merging strategies have emerged, merging all LLMs into one model to reduce the memory footprint during inference. Despite these advances, model merging often leads to parameter conflicts and performance decline as the number of experts increases. Previous methods to mitigate these conflicts include post-pruning and partial merging. However, both approaches have limitations, particularly in terms of performance and storage efficiency when merged experts increase. To address these challenges, we introduce Channel Merging, a novel strategy designed to minimize parameter conflicts while enhancing storage efficiency. This method initially clusters and merges channel parameters based on their similarity to form several groups offline. By ensuring that only highly similar parameters are merged within each group, it significantly reduces parameter conflicts. During inference, we can instantly look up the expert parameters from the merged groups, preserving specialized knowledge. Our experiments demonstrate that Channel Merging consistently delivers high performance, matching unmerged models in tasks like English and Chinese reasoning, mathematical reasoning, and code generation. Moreover, it obtains results comparable to model ensemble with just 53% parameters when used with a task-specific router.

Data-driven deep learning models have enabled tremendous progress in change detection (CD) with the support of pixel-level annotations. However, collecting diverse data and manually annotating them is costly, laborious, and knowledge-intensive. Existing generative methods for CD data synthesis show competitive potential in addressing this issue but still face the following limitations: 1) difficulty in flexibly controlling change events, 2) dependence on additional data to train the data generators, 3) focus on specific change detection tasks. To this end, this paper focuses on the semantic CD (SCD) task and develops a multi-temporal SCD data generator ChangeDiff by exploring powerful diffusion models. ChangeDiff innovatively generates change data in two steps: first, it uses text prompts and a text-to-layout (T2L) model to create continuous layouts, and then it employs layout-to-image (L2I) to convert these layouts into images. Specifically, we propose multi-class distribution-guided text prompts (MCDG-TP), allowing for layouts to be generated flexibly through controllable classes and their corresponding ratios. Subsequently, to generalize the T2L model to the proposed MCDG-TP, a class distribution refinement loss is further designed as training supervision. Our generated data shows significant progress in temporal continuity, spatial diversity, and quality realism, empowering change detectors with accuracy and transferability.