ACML 2025 Papers

Real-time object detection on resource-constrained edge devices presents a significant challenge in balancing performance and efficiency. This paper introduces a novel knowledge distillation framework designed to enhance the capabilities of lightweight student models for object detection tasks. Our approach, Multi-Scale Frequency-Aware Distillation (MSFAD), integrates three key components: multi-scale distillation, frequency domain mask distillation, and feature alignment distillation. Multi-scale distillation enables the student to learn feature representations at various levels of granularity. Frequency domain mask distillation improves the student’s ability to focus on relevant regions. Feature alignment distillation facilitates the transfer of channel-wise knowledge from teacher to student. We combine these techniques with a traditional detection loss to form a comprehensive loss function, balanced by a hyperparameter alpha. Experimental results across various scenarios demonstrate that MSFAD significantly improves detection accuracy while reducing computational and storage costs. Our approach clearly presents significant performance gains and faster inference speeds.

Counterfactual explanations in reinforcement learning (RL) aim to answer what-if questions by demonstrating sparse and minimal changes to states, resulting in the probability mass moving from one action to another. Although these explanations are effective in classification tasks that look for the presence of concepts, RL brings new challenges that counterfactual methods need to solve. These challenges include defining state similarity, avoiding out-of-distribution states, and improving discriminative power of explanations. Given a state of interest called the query state, we solve these problems by asking how long the agent can execute the query state action without incurring a negative outcome regarding the expected return. We coin this outcome-based semifactual (OSF) explanation and find the OSF state by simulating trajectories from the query state. The last state in a subtrajectory where we can take the same action as in the query state without incurring a negative outcome is the OSF state. This state is discriminative, plausible, and similar to the query state. It abstracts away unimportant action switching with little explanatory value and shows the boundary between positive and negative outcomes. Qualitatively, we show that our method explains when an agent must switch actions. As a result, it is easier to understand the agent’s behavior. Quantitatively, we demonstrate that our method can increase policy performance and, at the same time, reduce how often the agent switches its action across six environments. The code and trained models are made open source.

Imitation learning (IL) enables agents to learn policies by mimicking expert demonstrations. While online IL methods require interaction with the environment, which is costly, risky, or impractical, offline IL allows agents to learn solely from expert datasets without any interaction with the environment.In this paper, we propose Preference Optimization for Imitation Learning (POIL), a novel approach inspired by preference optimization techniques in large language model alignment. POIL eliminates the need for adversarial training and reference models by directly comparing the agent’s actions to expert actions using a preference-based loss function. We evaluate POIL on MuJoCo control tasks and Adroit manipulation tasks.Our experiments show that POIL consistently delivers superior or competitive performance against state-of-the-art methods in the past, including Behavioral Cloning (BC), IQ-Learn, MCNN, and O-DICE, especially in data-scarce scenarios, such as using single trajectory.These results demonstrate that POIL enhances data efficiency and stability in offline imitation learning, making it a promising solution for applications where environment interaction is infeasible and expert data is limited, even in high-dimensional and complex control tasks.

The recent advances in Legal Large Language Models (LLMs) have transformed the landscape of legal research and practice by automating tasks, enhancing research precision, and supporting complex decision-making processes. However, effectively adapting LLMs to the legal domain remains challenging due to the complexity of legal reasoning, the need for precise interpretation of specialized language, and the potential for hallucinations. This paper examines the efficacy of Domain-Adaptive Continual Pre-Training (DACP) in improving the legal reasoning capabilities of LLMs. Through a series of experiments on legal reasoning tasks within the Taiwanese legal framework, we demonstrate that while DACP enhances domain-specific knowledge, it does not uniformly improve performance across all legal tasks. We discuss the trade-offs involved in DACP, particularly its impact on model generalization and performance in prompt-based tasks, and propose directions for future research to optimize domain adaptation strategies in legal AI.

In causal effect estimation, determining the appropriate sampling size is critical for ensuring reliability and validity in both experimental and observational studies, a challenge closely tied to robust model generalization under limited data conditions in machine learning. This paper tackles these challenges by leveraging the Probably Approximately Correct (PAC) theory to establish a theoretically grounded framework for determining sampling boundaries. We utilize Hoeffding’s inequality and Vapnik–Chervonenkis (VC) dimension to set upper boundaries for dataset adequacy in diverse scenarios: no confounders, confounders with a finite hypothesis space, and confounders with an infinite hypothesis space. Our work ensures that if the dataset size exceeds the upper boundary, the error probability for the estimated causal effect stays within a specified threshold at the given confidence level. Additionally, we demonstrate that when the dataset size is inadequate, the error of the estimated average treatment effects is bounded by the estimation of the outcome variable, which forms the theoretical basis for data augmentation strategies to improve the accuracy of causal effect estimation. Extensive experiments on synthetic and semi-synthetic datasets validate the correctness of our presented sampling upper limitations under different error and confidence level constraints. Our findings not only offer a systematic and reliable method for determining sample size in causal effect estimation but also provide actionable guidance for developing causal inference models in data-scarce environments, enhancing their applicability and robustness across fields such as healthcare, social sciences, and policy evaluation.

Continuous time-event sequence (CTES) forecasting is essential across diverse domains, from healthcare to finance, requiring accurate prediction of both future event types and their timestamps. Traditionally, CTES forecasting has been driven by Temporal Point Processes (TPPs), which rely on intensity function-based priors. However, these methods often fail to generalize effectively to real-world scenarios as well as perform poorly over longer horizons. While recent diffusion-based approaches are promising, they are limited by the need to define a fixed prior while training, such as number of events to forecast or time horizon, which requires training multiple models for different horizon lengths. We present Kairos, a novel model that reformulates CTES forecasting as a language modeling task. Our model employs a decoder-only transformer architecture with a unified tokenization approach that represents time and events in a shared embedding space. By structuring the input as alternating event and time tokens, the model learns to capture the inherent temporal relationships between events. Through comprehensive experiments on multiple large-scale datasets, we demonstrate that Kairos consistently outperforms state-of-the-art baselines, achieving average improvements of 4.5

Graph Neural Networks (GNNs) have achieved remarkable success in modeling structured data through local message passing. However, their effectiveness diminishes on graphs with low homophily or irregular structures, where long-range dependencies are hard to capture and features tend to suffer from over-smoothing and noise amplification. To address these limitations, we propose GMN, a novel dual-path Graph Mediator Network that explicitly enhances both global information propagation and spectral stability. In the spatial path, GMN introduces a lightweight Mediator node connected to all graph nodes, allowing long-range communication to occur in a single hop without increasing network depth. In parallel, the spectral path leverages multi-scale Chebyshev filtering along with a spectral energy regularization term that suppresses high-frequency noise, leading to smoother and more stable node embeddings. These two complementary pathways are adaptively integrated via a gated fusion mechanism, which dynamically balances their contributions based on structural context. Final graph-level representations are obtained through task-specific pooling strategies, enabling GMN to generalize effectively across different tasks. Extensive experiments on benchmark datasets with varying homophily levels and structural perturbations demonstrate that GMN consistently achieves state-of-the-art performance in terms of accuracy, robustness, and generalization. Code is available at: https://github.com/sun2017bupt/GMN.

Emotion recognition using Electroencephalography (EEG) is challenging due to its low signal-to-noise ratios and high-dimensional sparsity. We propose FeelNet, a novel Fast Fourier Transform (FFT)-based architecture that simultaneously extracts global and local features across joint frequency-time domains. FeelNet incorporates an adaptive Rhythm Spectral Block (RSB) for capturing key frequency patterns and filtering task-irrelevant noise through power spectral thresholding. Additionally, the Multi-scale Temporal Conv Block (MTCB) enhances the model’s ability to decode complex temporal dynamics. Extensive evaluations on the DEAP and DREAMER datasets demonstrate that FeelNet outperforms existing state-of-the-art methods in accuracy and flexibility, even under noise-contaminated conditions. Owing to its computational efficiency and noise resilience, FeelNet provides an alternative perspective for EEG-based affective computing.

Cross-domain facial expression recognition (CD-FER) remains difficult due to severe domain shift between training and deployment data. We propose Graph-Attention Network with Adversarial Domain Alignment (GAT-ADA), a hybrid framework that couples a ResNet-50 as backbone with a batch-level Graph Attention Network (GAT) to model inter-sample relations under shift. Each mini-batch is cast as a sparse ring graph so that attention aggregates cross-sample cues that are informative for adaptation. To align distributions, GAT-ADA combines adversarial learning via a Gradient Reversal Layer (GRL) with statistical alignment using CORAL and MMD. GAT-ADA is evaluated under a standard unsupervised domain adaptation protocol: training on one labeled source (RAF-DB) and adapting to multiple unlabeled targets (CK+, JAFFE, SFEW 2.0, FER2013, and ExpW). GAT-ADA attains 74.39

This paper introduces Data-Centric Graph Condensation (named DCGC), a task- and model-agnostic method for condensing a large graph into a smaller one by matching the distribution between two graphs. DCGC defines the distribution of a graph as the trajectories of its node signals (such as node features and node labels) induced by a diffusion process over the geometric structure, which accommodates multi-order structural information. Built upon this, DCGC compresses the topological knowledge of the original graph into the orders-of-magnitude smaller synthetic one by aligning their distributions in input space. Compared with existing methods that stick to particular GNN architectures and require solving complicated optimization, DCGC can be flexibly applied to arbitrary off-the-shelf GNNs and achieve graph condensation with a much faster speed. Apart from the cross-architecture generalization ability and training efficiency, experiments demonstrate that DCGC yields consistently superior performance than existing methods on datasets with varying scales and condensation ratios.

Vision Transformers (ViTs) in human activity recognition tasks suffer from inadequate spatial modeling through conventional position embeddings, leading to over-reliance on fixed positional information. This paper proposes Shuffled Positional Embedding (SPE), a mechanism that randomly disrupts the order of positional encoding during each forward propagation, reducing model dependence on position embedding and encouraging exploration of intrinsic spatial relationships. While SPE enhances general spatial awareness, it lacks targeted guidance for human-centric modeling. To address this limitation, Local Shuffled Skeleton Position Embedding (LSSPE) is developed, which leverages 2D skeleton data to provide human body structure-aware spatial representation. LSSPE computes attention weights based on spatial distances between image patches and skeleton keypoints, incorporating joint motion amplitudes for enhanced modeling. To further utilize skeleton data, a dual-stream architecture is designed combining TimeSFormer with LSSPE (LSSPE-TimeSFormer) for RGB processing and SkateFormer for skeleton processing. The proposed dual-stream model achieves outstanding performance of 95.8

Dataset distillation aims to give models trained on synthetic datasets the same performance as models trained with complete real datasets. Trajectory matching distillation, as an efficient dataset distillation method, achieves this goal gradually by accurately matching the dynamic trajectories of the target dataset and the synthetic dataset during the training process. Where the training trajectory is composed of the time series parameters of the agent model, and each time series contains the network parameters of all the layers in the agent model, i.e., trajectory matching distillation achieves its goal by matching the network parameters between the target dataset and the synthetic dataset. However, the variability of the training datasets used by the teacher and student networks can lead to the problem of difficult alignment of network parameters during the distillation process, so this paper proposes Difference-Driven Pruning Distillation (DPD), an innovative approach to pruning the difficult-to-align parameters according to the magnitude of the difference in parameter comparisons to alleviate the above problem. Comparative experimental results show that DPD achieves a significant performance improvement, with a greatly reduced memory footprint and superior performance in several benchmarks.

With the rapid expansion of the Internet, people commonly maintain multiple account identities across different online platforms, creating latent cross-network associations. User Identity Linkage (UIL), which seeks to identify and associate multiple accounts belonging to the same individual across platforms, has emerged as a vital research direction with broad applications in cross-platform recommendation, unified user profiling, and so on. However, existing methods face two major challenges in real-world environments: cross-platform feature heterogeneity and attribute-structure representation fusion. To address these challenges, this paper propose a Multi-View Feature High-Order Consistency-Guided User Identity Linkage method UIL-HC-MV. Our approach mitigates cross-network heterogeneity by deeply integrating multi-view features and mining consistency in shared thematic information among users and their relational networks. We decompose cross-platform feature heterogeneity into two subproblems: attribute heterogeneity and structural heterogeneity. We first fuse attribute and structural views by coupling nodes’ random-walk sequences with neighborhood sampling to jointly extract node attributes and topological context. We then employ a Large Language Model to capture deep semantic information and contextual relationships across multiple text segments, distilling unified themes or high-order community features from the combined attribute-structure representation. Finally, we fine-tune a BERT model on the extracted high-order information to reinforce feature consistency and enable transfer learning for improved generalization. Extensive comparative experiments on real-world datasets demonstrate significant performance improvements over existing mainstream methods, validating the effectiveness of high-order information in alleviating cross-network heterogeneity and confirm the contribution of each component within our deeply integrated multi-view feature learning framework.

Chinese-Kazakh cross-lingual information retrieval (CLIR) aims to search relevant content from a collection of Kazakh documents using Chinese query statements. The intrinsic differences in grammar, vocabulary, and semantic expression between languages pose significant challenges for semantic alignment in CLIR. Existing CLIR methods that incorporate multilingual knowledge graph (MLKG) typically use simple vector stacking approaches to integrate entity information, failing to leverage deeper entity relationships and semantic connections. To address these challenges, we propose GIIM, a graph information integration method for Chinese-Kazakh CLIR that leverages the rich multilingual entity information embedded in MLKG as semantic bridges to narrow the linguistic gap during query-document matching process. Unlike previous methods, GIIM unifies query-document pairs and entity information into a graph structure and employs Graph Convolutional Network to aggregate both direct and multi-hop relations among entities, effectively modeling complex semantic paths and hierarchical knowledge propagation. To comprehensively evaluate GIIM, we construct CKIRD, a Chinese-Kazakh information retrieval dataset containing approximately 11,820 annotated query-paragraph pairs, and conduct experiments on both CKIRD and the public CLIRMatrix datasets. Experimental results show that GIIM outperforms existing baseline models across multiple ranking metrics, demonstrating its effectiveness on the Chinese-Kazakh CLIR task.

Deep learning models today are indispensable tools for image compression and restoration. However, despite recent progress, many existing models often lack generalization upon facing with different types and coding strength designs of image restoration, thus limiting their practical application. In this paper, a novel approach called { em dual-Channel Transformers and Multi-scale attention Prompt learning (CTMP)} is introduced to bridge the gap on blind image restoration. The prompt-based learning approach is employed in the model to address two key image restoration tasks: 1) compressed image artifact removal, and 2) image denoising. By utilizing adaptive prompts to accommodate varying quantization parameter (QP) values and noise conditions, and enhancing adaptability through the integration of multi-scale attention mechanisms, an advanced Transformer architecture in our model can tackle diverse image degradations in blind image restoration. That is, our Transformer module is improved through merging and harnessing the strengths of both channel attention and self-attention. The design is adept at extracting both high-frequency details and low-frequency structures, thereby significantly enhancing overall restoration performance. Using the Kodak dataset in experiments, our model outperforms conventional deep learning techniques with a 2.44

Jailbreak attacks reveal a persistent gap between the intended alignment of language models and their actual behavior during inference. To address this, we investigate how such attacks succeed at the internal level of model computation, focusing on attention heads. Unlike previous studies that primarily analyzed why jailbreaks work, our approach aims to develop a defense mechanism. We identify attention heads that influence whether a model produces a harmful or safe response by comparing activation patterns between a harmful prompt that is rejected and its adversarial variant that elicits a harmful response. By interpolating the internal representations of these heads between the two scenarios, we suppress harmful outputs while maintaining appropriate responses to benign prompts. Experiments with representative jailbreak methods, including GCG and AutoDAN, show that our method significantly reduces attack success rates without degrading response quality. For instance, with Llama-2-7b-chat, the average success rate drops from 39.3

Knowledge editing has emerged as an efficient approach for updating factual knowledge in large language models (LLMs), typically achieved by first locating key knowledge-storage modules and then modifying their parameters. However, most existing methods focus exclusively on updating the weights of Multi-Layer Perceptron (MLP) modules, which are commonly identified as the primary repositories of factual information. Other important components, such as attention (Attn) modules—one of the core modules in LLMs—are often ignored during editing. This biased allocation of updates can leave residual outdated knowledge in the model and limit the effectiveness of knowledge editing. In this paper, we conduct comprehensive and systematic knowledge localization experiments on advanced LLMs, revealing that Attn modules play a substantial role in factual knowledge storage and retrieval, especially in earlier layers. Building on these insights, we propose textit{IntAttn-Edit}, a novel method that extends the associative memory paradigm to jointly update both MLP and Attn modules. Our approach employs a knowledge balancing strategy that proportionally allocates update magnitudes based on each module’s measured contribution to knowledge storage. Extensive experiments on popular benchmarks demonstrate that textit{IntAttn-Edit} consistently achieves superior results over existing methods, delivering higher edit success, improved generalization, and robust knowledge preservation. Further empirical analysis shows that our knowledge balancing strategy enables the editing performance to remain within the optimal range across different settings.

We study the non-contextual multi-armed bandit problem in a transfer learning setting: before any pulls, the learner is given N’_k i.i.d. samples from each source distribution nu’_k, and the true target distributions nu_k lie within a known distance bound d_k( nu_k, nu’_k) le L_k. In this framework, we first derive a problem-dependent asymptotic lower bound on cumulative regret that extends the classical Lai–Robbins result to incorporate the transfer parameters (d_k,L_k,N’_k). We then propose textsc{KL-UCB-Transfer}, a simple index policy that matches this new bound in the Gaussian case. Finally, we validate our approach via simulations, showing that textsc{KL-UCB-Transfer} significantly outperforms the no-prior baseline when source and target distributions are sufficiently close.

In time-series unsupervised domain adaptation (UDA), the adaptation between temporal and frequency domain features has been relatively underexplored. To address this gap, we conduct a comprehensive series of experiments to revisit the roles of these domains in UDA. Our findings reveal that the temporal domain contains more diverse features, offering higher discriminability, while the frequency domain is more domain-invariant, providing better transferability. Combining the strengths of both domains, we propose TF-DAN, a UDA framework that synergistically integrates temporal and frequency domain features. TF-DAN enhances feature extraction and captures subtle, class-specific features without relying on traditional alignment strategies. By utilizing simple hyperparameter adjustments and using frequency embeddings from the source domain as reference points for domain adaptation, TF-DAN achieves nearly a 10