ACML 2025 Papers

Personality detection from text aims to infer an individual’s personality traits based on linguistic patterns. However, existing machine learning approaches often struggle to capture contextual information spanning multiple posts and tend to fall short in extracting representative and robust features in semantically sparse environments. This paper presents HIPPD, a brain-inspired framework for personality detection that emulates the hierarchical information processing of the human brain. HIPPD utilises a large language model to simulate the cerebral cortex, enabling global semantic reasoning and deep feature abstraction. A dynamic memory module, modelled after the prefrontal cortex, performs adaptive gating and selective retention of critical features, with all adjustments driven by dopaminergic prediction error feedback. Subsequently, a set of specialised lightweight models, emulating the basal ganglia, are dynamically routed via a strict winner-takes-all mechanism to capture the personality-related patterns they are most proficient at recognising. Extensive experiments on the Kaggle and Pandora datasets demonstrate that HIPPD consistently outperforms state-of-the-art baselines.

Accurate skin lesion segmentation is essential for the early diagnosis of dermatological conditions, including the timely detection of malignant skin cancers. Enabling such analysis on personal devices—such as smartphones—offers greater accessibility but introduces critical challenges related to computational constraints and privacy preservation. Performing segmentation directly on mobile edge devices avoids the need to transmit sensitive data to the cloud but requires models that are both lightweight and highly accurate. To this end, we propose SparseSegNet, an organically efficient segmentation framework that combines architectural simplicity with training-time innovations to enable real-time, on-device inference. SparseSegNet is built upon a Deep Layer Aggregation (DLA)-inspired encoder–decoder backbone, which effectively captures multi-scale lesion features while maintaining a compact model size. To further enhance boundary precision and generalization, we introduce a novel dual-teacher distillation strategy, termed Agreement-Guided Orthogonal Projection (AG-OP). This method transfers complementary spatial cues from two powerful vision foundation models— Segment Anything Model (SAM) based on Vision Transformer-Huge (ViT-H), and Segment Everything Everywhere Model (SEEM). Unlike traditional single-teacher distillation approaches, AG-OP encourages alignment between hard and soft pseudo-labels through orthogonal subspace projection, improving the robustness of the student model. We validate SparseSegNet across five public skin lesion segmentation benchmarks—ISIC 2017, ISIC 2018, PH^2, HAM10000, and Derm7pt —under a unified preprocessing and training pipeline. SparseSegNet achieves up to 0.91 Dice coefficient, 0.85 Intersection-over-Union (IoU), and 38ms latency with only 7 million parameters, outperforming recent compact models such as MobileSAM, CMUNeXt, and YOLOv8n-seg. Paired t-tests (p < 0.01) confirm the statistical significance of our improvements. SparseSegNet thus presents a privacy-preserving, boundary-aware solution for real-time skin lesion analysis on edge devices.

Predictive maintenance is essential for commercial vehicle fleets to reduce unexpected downtime and emergency repair costs. While standardized fault codes (SPN/FMI, representing Suspect Parameter Numbers and Failure Mode Indicators) assist in diagnosis, their temporal and spatial inconsistency limits the effectiveness of conventional time-series models in identifying high-cost failures. We propose a Hybrid Node-level Relationship-based Graph Convolutional Network with Random Forest (NRP-GCN-RF), which encodes fault interactions as graphs to capture non-temporal dependencies. Built on a real-world dataset from Dongfeng Motor Corporation, our study follows a dual-task design. (1) We construct a predictive model using graph neural networks (GCN) and random forests (RF) to forecast emergency repair costs and fault categories based on chassis-level fault sequences. (2) In parallel, we apply the Apriori algorithm to mine frequent co-occurring SPN-FMI pairs, revealing interpretable fault patterns and subsystem-level dependencies. This interpretable analysis complements the graph-based model by supporting feature design and failure diagnostics. Experiments show that our approach achieves 98.93

This paper introduces a novel oversampling technique designed to improve classification performance on imbalanced datasets. The proposed method, Kernel-Weighted SMOTE (KWSMOTE), enhances the traditional SMOTE algorithm by employing a kernel-based weighting scheme to prioritize closer neighbors, which guides a convex combination that ensures the generated samples are geometrically bounded. This dual-mechanism approach generates synthetic samples that better represent the minority class. Through experiments on multiple real-world datasets, we demonstrate that KWSMOTE outperforms existing methods in terms of F1-score, G-mean, and AUC, providing a robust solution for handling imbalanced datasets in classification tasks.

Large language models (LLMs) have demonstrated remarkable performance in code generation and evaluation tasks, particularly for Python, which dominates the pre-training corpora. However, the evaluation of code in low-resource programming languages remains challenging due to limited data and suboptimal model alignment. In this paper, we propose CrossPyEval, a novel cross-language code evaluation framework that uses an LLM to translate code from other languages into Python, verifies consistency with an SMT solver, and then analyzes the translated code via abstract syntax trees before performing the final evaluation. Experiments on public benchmarks and our custom low-resource datasets demonstrate that CrossPyEval substantially boosts evaluation accuracy for non-Python languages, achieving up to an 8.83

Vision-Language Models (VLMs) have achieved remarkable performance across various tasks. Unfortunately, due to their multimodal nature, a common jailbreak strategy transforms harmful instructions into visual formats like stylized typography or AI-generated images to bypass safety alignment. Despite numerous heuristic defenses, little research has investigated the underlying rationale behind the jailbreak. In this paper, we introduce an information-theoretic framework to explore the fundamental trade-off between attack effectiveness and stealthiness. Leveraging Fano’s inequality, we show that an attacker’s success probability intrinsically relates to the stealthiness of the generated prompts. We further propose an efficient algorithm to detect non-stealthy jailbreak attacks. Experimental results highlight the inherent tension between strong attacks and detectability, offering a formal lower bound on adversarial strategies and potential defense mechanisms.

Generalized eigenvalues serve as a foundational tool for extracting insights from data and constructing robust statistical learning models, while differential privacy ensures the protection of individual information within these models by minimizing the impact of any single data point. In this work, we propose an ( epsilon, delta)-differential privacy algorithm to solve the generalized eigenvalue problem (GEP). Our algorithm gives better classification accuracy over existing methods and has the nearly optimal ell_2-norm error bounds in both low and high dimensions. Furthermore, our algorithm guarantees convergence to the solution regardless of the initial vector and this improves a previous method that requires a specific procedure to find a proper starting vector. Our experiments confirm the effectiveness of our algorithm in safeguarding privacy while simultaneously boosting classification accuracy.

In the field of underwater image enhancement, existing methods generally rely heavily on the RGB color space and ignore the potential advantages of the perceptually uniform XYZ color space in color correction. Additionally, CNN-based methods are prone to losing long-distance dependency relationships during local feature extraction process, thus affecting the image restoration quality. To address the above issues, we propose DSCNet, an underwater image enhancement framework based on dual color spaces. The framework aims to break through the limitations of traditional methods, by innovatively introducing a parallel processing mechanism for both RGB and XYZ color spaces. Upon fully taking advantages of the XYZ space in terms of perceptual linearity, the model can improve the color correction and brightness enhancement processes. Furthermore, we design a hybrid computing architecture which combines convolutional operations with a novel lightweight Transformer module. Through channel splitting and dimensionality reduction strategies, the computational complexity is reduced significantly while maintaining the ability to effectively model global contextual information. Experimental results show that DCSNet exhibits excellent enhancement performance in various underwater scenarios and delivers superior visual effects. Moreover, with its small number of model parameters, DCSNet can be deployed on embedded or edge devices for practical underwater visualization applications.

To enable parameter-efficient fine-tuning of large language models (LLMs), Low-Rank Adaptation (LoRA) reduces parameters by freezing pretrained weights W_{0} and approximating updates via low-rank matrices Delta W = BA. However, standard LoRA neglects the differential impact of low-rank matrix components on model performance and suffers from slow convergence due to random initialization. To address this, we propose a dual-module architecture: The shared module inherits pretrained weights’s core semantic representations through principal component initialization, retaining residuals in the original model.The expert module incorporates a selection mechanism guided by importance screening, with orthogonality constraints imposed through loss regularization to ensure independence in parameter update directions. The shared module accelerates convergence by updating world knowledge, while the expert module dynamically screens domain knowledge to achieve efficient allocation of updated budgets. Extensive experiments under identical configurations show our method achieves 76.8

Stereo image compression (SIC) aims to enhance compression performance and efficiency by exploiting cross-view redundancy in overlapping fields between stereo images. However, current SIC methods faces practical limitations in adequately exploiting inter-view correlations and contextual information due to occlusions, disparity variations, and computational overhead. To effectively extract contextual information and efficiently model cross-view dependencies in stereo images, we propose a novel distributed stereo image compression framework, Reparameterized Swin Transformer Stereo Image Compression (RSTSIC) integrating Reparameterized Swin Block (RSB) and Cross Feature Enhancement Modules (CFEMs) in the joint decoder. CFEMs progressively aggregate cross-view dependencies and enhance cross feature interaction efficiency. RSB integrates window-based self-attention with convolutional operations to effectively leverage non-local contextual information, while maintaining inference efficiency through structural reparameterization. RSTSIC outperforms traditional codecs and deep stereo compression methods on both Cityscapes and InStereo2K datasets, with at least 58.57

End-Edge-Cloud Computing (EECC) has emerged as the mainstream computing paradigm, integrating edge computing to overcome the limitations of traditional federated learning in communication efficiency and resource scheduling. However, existing studies reveal that most frameworks still struggle with challenges such as computing resource allocation and high end-to-end latency in EECC. To address these issues, we propose Mega-CE^2, a novel multi-modal heterogeneous aggregation framework. Mega-CE^2 establishes a closed-loop feedback mechanism from the bottom-up to the top down through end-device data serialization, edge-server model personalization, and cloud-based optimization. Notably Mega-CE^2 incorporates lightweight adapters for fine-tuning, enabling efficient deployment while preserving local model personalization. These adapters, with fewer parameters than the global model, optimize model parameters during edge-to-cloud aggregation, thereby achieving both lightweight and personalized capabilities. In experiments on three open-source standard datasets, we show that the performance of Mega-CE^2 improves by 3

Domain shifts present significant challenges for data-driven modeling of dynamical systems, as they may reduce state prediction accuracy and degrade model-based control performance. Transfer learning is a promising way to mitigate the effect of changes in dynamics. In this study, we investigate a domain adaptation framework based on hybrid modeling with fine-tuning. Hybrid models integrate physics-based components derived from prior knowledge of system dynamics into neural ordinary differential equations (neural ODEs). They are expected to facilitate efficient and enhanced fine-tuning because the structures of physics parts often remain invariant under domain shifts. We evaluated the hybrid neural ODE approach through experiments on multicopters undergoing concept shifts and found that introducing physics models significantly enhanced the domain adaptation capabilities, even when the physics-based components included unidentified parameters. Moreover, the results demonstrated that the hybrid modeling strategy reduced the amount of data required in the target domain, enabling efficient domain adaptation.

We address two-player general-sum stochastic Stackelberg games (SSGs), where the leader’s policy is optimized considering the best-response follower whose policy is optimal for its reward under the leader. Existing policy gradient and value iteration approaches for SSGs do not guarantee monotone improvement in the leader’s policy under the best-response follower. Consequently, their performance is not guaranteed when their limits are not stationary Stackelberg equilibria (SSEs), which do not necessarily exist. In this paper, we derive a policy improvement theorem for SSGs under the best-response follower and propose a novel policy iteration algorithm that guarantees monotone improvement in the leader’s performance. Additionally, we introduce Pareto-optimality as an extended optimality of the SSE and prove that our method converges to the Pareto front when the leader is myopic.

We present R^2-FFNO, a novel neural operator architecture designed to address the overparameterization common in Factorized Fourier Neural Operators (FFNO) through reduced-rank factorization of spectral components. While neural operators are effective for learning solutions to partial differential equations (PDE), their architectures often contain an excessive number of parameters, which can lead to overfitting and diminished generalization capabilities. Inspired from reduced-rank learning techniques, the R^2-FFNO approach decomposes spectral kernels into lower-rank representations, enabling systematic control over the model’s capacity. This low-rank factorization facilitates a balance between the model’s expressiveness and capability of generalization. Empirical analysis reveals that performance saturates once an optimal rank is achieved and degrades if the rank is increased beyond this point. This observation highlights an optimal trade-off between model complexity and accuracy, underscoring the importance of principled rank selection in designing neural operators. To further enhance performance, a targeted data augmentation strategy is utilized. This strategy introduces high-frequency variations during training to address spectral bias, thereby enhancing the model’s capacity to resolve fine-scale PDE dynamics. A comprehensive evaluation on benchmark datasets confirms the efficacy of the R^2-FFNO. Compared to the original FFNO architecture, R^2-FFNO demonstrates significant error reductions: 46.5

Negative sampling algorithms are critical for training dense retrievers, which in turn impact retrieval performance in information systems. Among these, hard negative sampling is of great value, and the denoised negative sampling methods in particular. Strategically selecting relevant negative samples, these methods effectively enhance the effectiveness of model training. However, they are either restricted to single-stage retrieval, failing to fully explore potential effective negatives, or demand additional training for a filter, which compromises sampling efficiency. To address this issue, the paper introduces a two-stage Retrieval-Sampling Algorithm(ReSa2). It integrates document vector-based retrieval to refine candidate selection progressively while preserving semantic relevance. In Stage 1, ReSa2 uses query vectors for broad retrieval, generating a candidate subset from the corpus to narrow the search space. In Stage 2, it reuses the retriever to perform positive-centric retrieval within this subset, leveraging positive sample vectors to re-rank candidates and enrich hard negatives with semantic similarity to the query. During the whole process, the effect is further enhanced by conducting probability-weighted sampling on the candidate subset. Insight experiments on 40,000 query-sample pairs show ReSa2 suppresses false negatives by 69.1

Synthetic tabular data generation has gained significant attention for its potential in data augmentation and privacy-preserving data sharing. While recent methods like diffusion and auto-regressive models (i.e., transformer) have advanced the field, generative adversarial networks (GANs) remain highly competitive due to their training efficiency and strong data generation capabilities. In this paper, we introduce Tabular Auto-Encoder Generative Adversarial Network (TAEGAN), a novel GAN-based framework that leverages a masked auto-encoder as the generator. TAEGAN is the first to incorporate self-supervised warmup training of generator into tabular GANs. It enhances GAN stability and exposes the generator to richer information beyond the discriminator’s feedback. Additionally, we propose a novel sampling method tailored for imbalanced or skewed data and an improved loss function to better capture data distribution and correlations. We evaluate TAEGAN against seven state-of-the-art synthetic tabular data generation algorithms. Results from eight datasets show that TAEGAN outperforms all baselines on five datasets, achieving a 27

This paper develops a new framework to understand generalization for pairwise learning problems, which covers many popular machine learning problems as specific examples. By integrating robust optimization principles with pairwise loss structures, we establish data-dependent generalization bounds that significantly improve over existing approaches. Our method overcomes key limitations of prior work by leveraging observable training data properties rather than restrictive theoretical assumptions. This results in tighter performance guarantees that better reflect real-world learning behavior, particularly for complex datasets with dependent training pairs.

High-Definition (HD) maps are essential for both Advanced Driver-Assistance Systems (ADAS) and autonomous driving. However, offline HD map construction remains costly and challenging to maintain due to the dynamic nature of real-world environments. Consequently, online HD map generation using onboard sensors has become a key area of research. Despite recent advancements, existing deep learning-based methods often provide inaccurate output even using computationally heavy architectures, limiting their practicality for real-world applications. We introduce TLSD, an efficient end-to-end neural network that generates HD maps, incorporating both topological and geometric road information. To enhance both accuracy and efficiency, we introduce four key innovations: (1) an iterative refinement scheme within the decoder to progressively improve map predictions, (2) a group-wise one-to-many assignment strategy that accelerates training convergence, (3) a graph neural network (GNN) module that integrates lane segment coordinates for improved spatial reasoning, and (4) a distance-aware topological post-processing method that enhances the quality of connectivity outputs. We performed extensive experiments

Large language models (LLMs) demonstrate remarkable capabilities across various tasks. However, their deployment introduces significant risks related to intellectual property. In this context, we focus on model stealing attacks, where adversaries replicate the behaviors of these models to steal services. These attacks are highly relevant to proprietary LLMs and pose serious threats to revenue and financial stability. To mitigate these risks, the watermarking solution embeds imperceptible patterns in LLM outputs, enabling model traceability and intellectual property verification. In this paper, we study the vulnerability of LLM service providers by introducing delta-Steal, a novel model stealing attack that bypasses the service provider’s watermark detectors while preserving the adversary’s model utility. delta-Steal injects noise into the token embeddings of the adversary’s model during fine-tuning in a way that satisfies local differential privacy (LDP) guarantees. The adversary queries the service provider’s model to collect outputs and form input-output training pairs. By applying LDP-preserving noise to these pairs, delta- Steal obfuscates watermark signals, making it difficult for the service provider to determine whether its outputs were used, thereby preventing claims of model theft. Our experiments show that delta-Steal with lightweight modifications achieves attack success rates of up to 96.95