ACML 2025 论文列表

In recent years, detecting instantaneously expressed urgent needs or requests in disaster-related posts from disaster-affected social media users has become crucial for disaster response and recovery. To address the gap that the performance of current urgent need detectors based on large language models (LLMs) is below requirements on this task from the domain of disaster response, we propose a novel insight: decomposing and inducting post content expressing disaster-induced urgent needs, into disaster event statements and disaster-induced appeals. The former, widely present and highly coarse-grained homogeneous across disaster-related posts, tends to introduce event-induced model bias leading to false recalls; while the latter, characterized by highly personalized, fine-grained and subjective phrasing, often challenge LLMs to allocate appropriate attentions to the corresponding tokens. In light of this, we propose CryChime, a novel model-agnostic parameter-efficient fine-tuning (PEFT) framework. CryChime represents disaster event statements in a bootstrapping style, and then removes the event-induced bias by orthogonal LoRA-based counterfactual learning. As fine-tuning steps increase, CryChime gradually disentangles the domain knowledge for understanding disaster event statements and disaster-induced appeals in candidate posts, then collaboratively leverage them in performing better urgent need detection. Experimental results on two benchmark datasets show that, compared to the strong baselines, CryChime can more effectively listen to the distant cries from the disaster-affected users. Our instruction-tuning data examples will be released in the further preprint version.

Our study aims to utilize psychological risk factors to detect posts on social media that contain high-risk suicidal content in Mandarin. We propose a two-stage model structure: the first stage labels each sentence in an post according to risk factors, while the second stage uses these labels as features to predict the crisis level of the post. Our models were trained using a dataset developed from social media posts on a popular Mandarin-speaking platform, labeled by psychological professionals. Our approach achieved an accuracy and F1-score of 0.96 in classifying posts with high crisis levels. Furthermore, we developed a frontend webpage system to apply our model, designed for use by psychological professionals as an aid. This system not only helps psychological professionals detect and address high-risk posts but also offers them the opportunity for psychological analysis based on risk factors. By integrating expertise from psychology with advanced NLP and deep learning techniques, our system bridges the gap between technical models and psychological insights.

Key-value (KV) caching has emerged as a crucial optimization technique for accelerating inference in large language models (LLMs). By allowing the attention operation to scale linearly rather than quadratically with the total sequence length, KV caching significantly enhances generation throughput. However, due to large context lengths in the modern LLMs, the memory footprint of the KV is a huge bottleneck for model deployment directly impacting the model’s batch size, hindering its ability to deliver high-throughput. Existing research addresses this challenge using several techniques, such as discarding low-attention tokens, quantization, and matrix approximation which typically lead to a negative impact on the model accuracy. In this paper, we propose KVCrush technology which can be combined with many KV compression technologies to improve the model accuracy at a much smaller memory. KVCrush provides an alternate representation scheme for key-value states, along with a low-overhead token pruning algorithm that accounts for the token distribution in the KV cache, which in turn allows for a smaller footprint while maintaining the accuracy of the model. Based on our results, KVCrush reduces LongBench KV Cache size by 4 times with less than 1

Decision making under uncertain environments in the maximization of expected reward while minimizing its risk is one of the ubiquitous problems in many subjects. Here, we introduce a novel problem setting in stochastic bandit optimization that jointly addresses two critical aspects of decision-making: maximizing expected reward and minimizing associated uncertainty, quantified via the textit{mean-variance}(MV) criterion. Unlike traditional bandit formulations that focus solely on expected returns, our objective is to efficiently and accurately identify the Pareto-optimal set of arms that strikes the best trade-off between expected performance and risk. We propose a unified meta-algorithmic framework capable of operating under both fixed-confidence and fixed-budget regimes, achieved through adaptive design of confidence intervals tailored to each scenario using the same sample exploration strategy. We provide theoretical guarantees on the correctness of the returned solutions in both settings. To complement this theoretical analysis, we conduct extensive empirical evaluations across synthetic benchmarks, demonstrating that our approach outperforms existing methods in terms of both accuracy and sample efficiency, highlighting its broad applicability to risk-aware decision-making tasks in uncertain environments.

The paradigm shift toward structure-driven molecule generation has been propelled by advances in deep generative models, such as variational auto-encoders and diffusion models. However, these generative models for molecular design remain constrained by exposure bias, error accumulation, and suboptimal handling of activity cliffs. Here, we introduce DiffGap, a diffusion-based framework that integrates adaptive sampling and pseudo-molecule estimation to bridge the gap between training objectives and inference dynamics in 3D molecule generation. By dynamically aligning intermediate denoising steps with realistic generation trajectories, DiffGap enables the diffusion model to adapt to input biases in advance during the training phase. A temperature annealing module further controls the aligning strength of the adaptive alignment process, ensuring stable learning of the data distribution. Evaluated on the CrossDocked2020 benchmark, DiffGap outperforms existing methods in docking scores and binding affinity, demonstrating superior fidelity in generating drug-like molecules. Our work establishes a principled approach to harmonize generative training with inference mechanics, offering a robust computational toolkit for accelerating structure-based therapeutic discovery. The source code of DiffGap will be published after review.

Although recent quantized Large Language Models, such as BitNet, have paved the way for significant reduction in memory usage during deployment with binary or ternary weights, training these models still demands substantial memory footprints. This is partly because high-precision (i.e., unquantized) weights required for straight-through estimation must be maintained throughout the whole training process. To address this, we explore directly updating the quantized low-precision weights without relying on straight-through estima- tion during backpropagation, aiming to save memory usage during training. Specifically, we employ a stochastic rounding technique to minimize the information loss caused by the use of low-bit weights throughout training. Experimental results on our LLaMA-structured models of various sizes indicate that (1) training with only low-precision weights is feasible even when they are constrained to ternary values; (2) extending the bit width to 8 bits achieves performance on par with BitNet b1.58; (3) our models remain robust to precision scaling and memory reduction, showing minimal performance degradation when moving from FP32 to lower-memory environments (BF16/FP8); and (4) our models also support inference using ternary weights, showcasing their flexibility in deployment.

Unsupervised visible-infrared person re-identification (VI-ReID) presents unique challenges due to severe modality discrepancies, including heterogeneous appearance gaps, semantic granularity mismatches, and pseudo-label noise amplification intrinsic to label-free scenarios. We distill these challenges into two core problems: fine-grained semantic alignment, which necessitates explicit token-level cross-modal feature fusion, and memory fragmentation caused by noisy pseudo-label propagation. To address these issues, we propose Fusion-Injected Residual Memory (FIRM), a unified framework that integrates Vision–Semantic Prompt Fusion (VSPF), which injects multi-scale textual cues derived from CLIP and large language models into multiple layers of a vision backbone for token-wise semantic alignment, and Evolving Multi-view Cluster Memory (EMCM), which employs optimal transport–guided clustering and dynamic prototype maintenance to ensure long-term identity consistency. The framework is optimized end-to-end using an optimal transport–weighted InfoNCE loss, a multi-layer alignment regularizer, and geometric cluster regularization, all without reliance on manual annotations. Extensive experiments on benchmark VI-ReID datasets demonstrate that the proposed method substantially advances unsupervised cross-modal retrieval performance, achieving new state-of-the-art results. Ablation studies further verify the independent and synergistic effectiveness of both modules in overcoming the identified core challenges.

In the field of Traditional Chinese Medicine (TCM), Named Entity Recognition (NER) is a crucial task. However, the scarcity of NER datasets in TCM significantly hampers the performance of models in this domain. A promising approach to addressing this low-resource issue is through domain adaptation techniques. Current domain adaptation methods typically leverage large amounts of labeled data from a source domain to bridge the gap between the source and target domains, making the features of the generated target domain data as similar as possible to those of the source domain, thereby enhancing model performance in the target domain. However, existing methods primarily focus on aligning textual features and neglect the importance of label information. In the NER task, labels not only indicate categories but also carry important categorical information. Therefore, this paper proposes a Label-Perceptive Adversarial Domain Adaptation (LPADA) method that integrates label information with textual features, providing additional contextual information for the domain adaptation process, thus enhancing the model’s performance in the TCM domain. Furthermore, we annotate medical case records to construct a dataset TCMNER2024 and establish a baseline. TCMNER2024 dataset can be accessed via https://github.com/TCMNER/TCMNER2024. The evaluation demonstrates that our approach significantly outperforms existing methods.

Accurate decoding of emotional EEG signals constitutes a critical challenge for developing affective brain-computer interfaces. Contemporary methods for cross-subject EEG-based emotion recognition confront two critical challenges: 1) inadequate investigation of the distinct affective features of the EEG rhythm; 2) insufficient capability to extract the various neurophysiological connectivity patterns across subjects in the same experimental setting. To address these limitations, we propose FG-MSTGNN, a dual-stage adaptive learning framework comprising the Frequency-guided Multi-period Spatial-temporal Graph Neural Network. The Feature Learning Stage utilizes a Multi-period Time-Frequency Cooperative Encoder Module to hierarchically extract cross-frequency rhythmic dynamics. The Topology Optimization Stage utilizes a Dual-Phase Graph Pooling Module to dynamically generate personalized sparse neurophysiological connectivity patterns. Systematic evaluation under cross-subject experiments demonstrates the framework achieves average classification accuracies of 94.67

Large Language Models (LLMs) have been widely adopted in real-world applications, yet their safety remains a major concern, particularly regarding jailbreak attacks that bypass alignment safeguards to elicit harmful outputs. Among various attack strategies, optimization-based jailbreak attacks have emerged as a primary approach by designing specialized loss functions to optimize adversarial suffixes added after the harmful question. However, existing methods often suffer from poor generalization and over-refusal issues due to overly fixed optimization targets, which significantly undermine the utility of jailbreak attempts by yielding generic denials (e.g., "Sorry, I can’t assist with that") rather than harmful completions. These issues fundamentally stem from the rigid exact match constraint in their loss design. To address this, we propose SEMINAR, a novel semantic information-augmented optimization framework that promotes diverse and semantically aligned affirmative responses. Specifically, we leverages semantic-level supervision to guide the optimization toward intent-consistent outputs rather than rigid templates by introducing a non-exact match loss based on semantic similarity. Furthermore, we mitigate the token shift problem—the generation of LLM highly depends on the correctness of the first few tokens, but the loss is averaged over the entire sequence, which leads to insufficient attention paid to the early tokens in the optimization—by introducing a cosine decay scheduling mechanism that emphasizes the early tokens in the sequence into the optimization process. As a result, SEMINAR not only enhances the diversity of affirmative responses generated by LLMs but also significantly improves overall attack effectiveness. Extensive experiments demonstrate the superiority of SEMINAR over baseline methods, along with its strong transferability across different models.

While Graph Neural Networks (GNNs) have shown significant promise for data integration using graph structures, methods to support subsampling graph data are lagging. To address this gap, in this paper, we propose a novel importance-based data subsampling framework. This framework strategically identifies inputs from a primary graph dataset based on their impact on the model’s learning of downstream tasks, such as graph or node classification. Our measure of impact is the predictive uncertainty of each data point. To ensure the subsample is well-representative of the original sample, we cluster them based on their learned graph representation. Finally, subsampling is performed from these identified clusters. The process favours selecting data points with greater prediction uncertainty, while preserving the diversity of the overall sample. We evaluate our approach using a multi-source, real-world dataset on child and youth mental health, comprising emergency department (ED) admissions and mental health questionnaire data. Our experimental results demonstrate that training a GNN with samples identified by the proposed framework yields a statistically significant improvement (on average, 10.13

Recent face-swapping methods excel under controlled conditions but often fail when presented with extreme facial poses. Diffusion-based approaches may be able to overcome these issues, but they still face significant computational costs. This paper introduces MagicMask, a novel face-swapping framework that robustly handles various poses in real time by fusing visual and geometric information. Our method incorporates explicit, identity-adapted geometric cues into the latent feature space via a multi-head attention mechanism. It employs an Adversarial Facial Silhouette Alignment (AFSA) loss to preserve detailed facial boundaries adapted to source identity. Comprehensive experiments on multiple benchmarks demonstrate that MagicMask competes with state-of-the-art methods under standard conditions and significantly outperforms them in extreme pose scenarios. The source code for the demonstration of MagicMask is attached as supplementary materials.

This paper studies the training data selection problem, focusing on the selection of effective samples to improve model training using data affected by distributional shifts (i.e., data drifts). Existing drift-detection-based methods struggle with local drifts, while recent drift-localization-based methods lack theoretical support for the problem and are often ineffective. To tackle these issues, this paper proposes TSJD, a training data soft selection method based on joint density ratio estimation. TSJD assigns training weights (i.e., soft selects) to samples based on the estimated joint density ratio to align the selected data with the recent data distribution. By evaluating each sample independently of time, TSJD effectively addresses local data drifts. We also provide theoretical guarantees by deriving an upper bound on the generalization error for models trained with data selected by TSJD. In numerical experiments with four real-world datasets, TSJD shows great versatility, achieving the best or comparable results over baseline methods in all of the experiments.

Recent knowledge editing methods have predominantly concentrated on modifying structured triplet knowledge within large language models. Compared to triplet-based knowledge, unstructured knowledge contains richer and more interrelated information, which increases the difficulty of editing. When relying solely on parameter-based editing methods, similar knowledge may interfere with each other due to their semantic overlap. Although previous studies have shown that directly applying in-context editing to unstructured knowledge with better results than parameter-based approaches, there is still considerable room for improvement. Previous studies have found that large language models are highly sensitive to the sequence of long text information,even the core content of the text may be masked due to positional influence. This indicates that, after rewriting unstructured facts, LLMs(Large Language Models) are better able to process and utilize the rewritten facts than the original facts. Inspired by this idea, we propose EIKEA(Enhancing In-Context Knowledge Editing by Agents), a novel method that combines rewriting agent with IKE (In-Context Knowledge Editing), enabling language models to effectively internalize unstructured factual updates without modifying model parameters. We conduct comprehensive experiments on the WIKIUPDATE subset of the AKEW benchmark, demonstrating that our method significantly improves editing accuracy over baseline IKE and parameter-editing methods. Our method provides a practical, lightweight, and scalable solution to unstructured knowledge editing.

Alignment Faking is a phenomenon in which a language model pretends to agree with a certain set of instructions during a test or evaluation, only to revert to its predetermined or natural behavior once the test is over. Recent work has actually shown that models strategically deceive the users they are interacting with when presented with certain scenarios, such as an evaluation where the model is threatened with retraining if it does not comply with the given instructions. In this paper, we propose ChameleonBench, a new benchmark that measures and quantifies the tendency of a model to engage in alignment faking when evaluated for different behavioral patterns. Our benchmark consists of 800 prompts that span 8 harmful behaviors and two evaluation scenarios: one in which the model is made to act freely, and another in which it is aware that it is interacting in a closed or test-like environment. We use an external judge pipeline to rate the severity, or the extent to which a response demonstrates a specific harmful behavior. We evaluated the shift in severity across different scenarios to quantify alignment-faking. Evaluating six frontier and open-weight models, we find that leading large language models (LLMs) frequently engage in alignment-faking when presented with different types of scenarios, with some models differing by over 20

Knowledge editing in Text-to-Image(T2I) diffusion models aims to update specific factual associations without disrupting unrelated knowledge. However, existing methods often suffer from unintended collateral effects, where editing a single fact can alter the representation of non-target named entities, degrading generation quality for unrelated prompts, which becomes more severe in real-world, dynamic environments requiring frequent updates. To address this challenge, we introduce a novel editing framework supporting large-scale T2I knowledge editing. Our framework incorporates our proposed Entity-Aware Text Alignment(EATA) to penalize unintended changes in unaffected entities and employs a principled null-space projection strategy to minimize perturbations to existing knowledge. Experimental results demonstrate that our approach enables precise and robust large-scale T2I knowledge editing, preserves the integrity of unrelated content, and maintains high generation fidelity, while offering scalability for continuous editing scenarios.

Recent advances in generative AI have enabled the creation of highly realistic synthetic audio, which poses significant challenges in voice authentication, media verification, and fraud detection. While Artificial Neural Networks (ANNs) are frequently used for fake audio detection, they often struggle to generalize to unseen and complex manipulations, particularly partial fake audio, where real and synthetic segments are seamlessly combined. This paper explores the use of Spiking Neural Networks (SNNs) for fake and partial fake audio detection – an unexplored area. Taking advantage of the inherent energy efficiency and temporal processing capabilities of SNNs, we propose novel SNN-based architectures for both tasks. We perform comprehensive evaluations that include hyperparameter tuning, cross-data set generalization, noise robustness, and partial fake audio detection using multiple large-scale public audio datasets. Our results show that SNNs achieve performance comparable to state-of-the-art ANN models while showing better generalization capabilities and robustness to noise. These SNN-based approaches also resulted in additional advantages, such as reduced model sizes and the ability to classify individual segments, making them more suitable for resource-constrained and real-time voice authentication applications. This work lays a foundation for exploring SNNs as countermeasures against audio spoofing in security-critical applications.

Translating natural language into precise and executable Computer-Aided Design (CAD) programs remains a challenging task, requiring both semantic understanding and geometric fidelity. In this paper, we present CAD-HLLM, a hierarchical LLM framework for structured CAD command generation. Our approach decomposes the task into two stages: a Plan Generator that infers high-level symbolic plans from text, and a Parameter Completor that generates detailed parametric commands conditioned on both the original description and the inferred plan. To enhance robustness, we introduce a lightweight ensemble selection mechanism that ranks and selects among multiple candidates based on model log-likelihoods. Experiments on benchmark datasets show that our method outperforms existing baselines in both parametric precision and 3D shape similarity, demonstrating the effectiveness of hierarchical reasoning and LLM-based planning in bridging the gap between human design intent and executable CAD sequences.

Privileged information (PI) provides additional knowledge to improve performance. Though some efforts are carried out by learning using privileged information (LUPI), they mainly focus on classifier-level LUPI and single-view PI tasks. Therefore, it is a challenge for feature representation learning by transferring multi-view PI to improve the main view. In this paper, we propose a novel feature-level LUPI for multi-view PI tasks, called the multi-view privileged information-based representation learning (MPIRL) algorithm, in which multi-view PI and main view are required at the training phase, but only the main view is available at the testing phase. MPIRL consists of a feature-level LUPI module and a classification module. The feature-level LUPI module of MPIRL designs a multi-branch structure to transfer the multi-view privileged information to the main view, so that diversity and discriminative representation can be generated. For the classification module, multi-view deep SVM (MDSVM) is developed, which combines a multi-channel deep neural network with SVM into a unified framework. MDSVM further learns the fusion representation and classification simultaneously to improve the generalization performance. The experimental results on the dual-view PI tasks and multi-view PI tasks of the real-world multi-view liver cancer dataset show that the proposed MPIRL achieves superior performance with an accuracy of 86.92