This system paper presents the DeMeVa team's approaches to the third edition of the Learning with Disagreements shared task (LeWiDi 2025; Leonardelli et al., 2025). We explore two directions: in-context learning (ICL) with large language models, where we compare example sampling strategies; and label distribution learning (LDL) methods with RoBERTa (Liu et al., 2019b), where we evaluate several fine-tuning methods. Our contributions are twofold: (1) we show that ICL can effectively predict annotator-specific annotations (perspectivist annotations), and that aggregating these predictions into soft labels yields competitive performance; and (2) we argue that LDL methods are promising for soft label predictions and merit further exploration by the perspectivist community.

This technical report analyzes the challenge of "hallucinations" (false information) in LLMs applied to law. It examines their causes, manifestations, and the effectiveness of the RAG mitigation strategy, highlighting its limitations and proposing holistic optimizations. The paper explores the ethical and regulatory implications, emphasizing human oversight as an irreplaceable role. It concludes that the solution lies not in incrementally improving generative models, but in adopting a "consultative" AI paradigm that prioritizes veracity and traceability, acting as a tool to amplify, not replace, professional judgment. -- Este informe técnico analiza el desafío de las "alucinaciones" (información falsa) en los LLMs aplicados al derecho. Se examinan sus causas, manifestaciones y la efectividad de la estrategia de mitigación RAG, exponiendo sus limitaciones y proponiendo optimizaciones holísticas. Se exploran las implicaciones éticas y regulatorias, enfatizando la supervisión humana como un rol insustituible. El documento concluye que la solución no reside en mejorar incrementalmente los modelos generativos, sino en adoptar un paradigma de IA "consultiva" que priorice la veracidad y la trazabilidad, actuando como una herramienta para amplificar, y no sustituir, el juicio profesional.

Assessing the reliability of Large Language Models (LLMs) by confidence elicitation is a prominent approach to AI safety in high-stakes applications, such as healthcare and finance. Existing methods either require expensive computational overhead or suffer from poor calibration, making them impractical and unreliable for real-world deployment. In this work, we propose GrACE, a Generative Approach to Confidence Elicitation that enables scalable and reliable confidence elicitation for LLMs. GrACE adopts a novel mechanism in which the model expresses confidence by the similarity between the last hidden state and the embedding of a special token appended to the vocabulary, in real-time. We fine-tune the model for calibrating the confidence with calibration targets associated with accuracy. Experiments with three LLMs and two benchmark datasets show that the confidence produced by GrACE achieves the best discriminative capacity and calibration on open-ended generation tasks, outperforming six competing methods without resorting to additional sampling or an auxiliary model. Moreover, we propose two strategies for improving test-time scaling based on confidence induced by GrACE. Experimental results show that using GrACE not only improves the accuracy of the final decision but also significantly reduces the number of required samples in the test-time scaling scheme, indicating the potential of GrACE as a practical solution for deploying LLMs with scalable, reliable, and real-time confidence estimation.

Secure inference enables privacy-preserving machine learning by leveraging cryptographic protocols that support computations on sensitive user data without exposing it. However, integrating cryptographic protocols with large language models (LLMs) presents significant challenges, as the inherent complexity of these protocols, together with LLMs' massive parameter scale and sophisticated architectures, severely limits practical usability. In this work, we propose ENSI, a novel non-interactive secure inference framework for LLMs, based on the principle of co-designing the cryptographic protocols and LLM architecture. ENSI employs an optimized encoding strategy that seamlessly integrates CKKS scheme with a lightweight LLM variant, BitNet, significantly reducing the computational complexity of encrypted matrix multiplications. In response to the prohibitive computational demands of softmax under homomorphic encryption (HE), we pioneer the integration of the sigmoid attention mechanism with HE as a seamless, retraining-free alternative. Furthermore, by embedding the Bootstrapping operation within the RMSNorm process, we efficiently refresh ciphertexts while markedly decreasing the frequency of costly bootstrapping invocations. Experimental evaluations demonstrate that ENSI achieves approximately an 8x acceleration in matrix multiplications and a 2.6x speedup in softmax inference on CPU compared to state-of-the-art method, with the proportion of bootstrapping is reduced to just 1%.

Recent advances in Large Language Models (LLMs) have demonstrated their remarkable capacity to process and reason over structured and unstructured data modalities beyond natural language. In this work, we explore the applications of Vision Language Models (VLMs), specifically a fine-tuned variant of LLaMa 3.2, to the task of identifying neutrino interactions in pixelated detector data from high-energy physics (HEP) experiments. We benchmark this model against a state-of-the-art convolutional neural network (CNN) architecture, similar to those used in the NOvA and DUNE experiments, which have achieved high efficiency and purity in classifying electron and muon neutrino events. Our evaluation considers both the classification performance and interpretability of the model predictions. We find that VLMs can outperform CNNs, while also providing greater flexibility in integrating auxiliary textual or semantic information and offering more interpretable, reasoning-based predictions. This work highlights the potential of VLMs as a general-purpose backbone for physics event classification, due to their high performance, interpretability, and generalizability, which opens new avenues for integrating multimodal reasoning in experimental neutrino physics.

Large Audio Language Models (LALMs) are rapidly advancing, but evaluating them remains challenging due to inefficient toolkits that limit fair comparison and systematic assessment. Current frameworks suffer from three critical issues: slow processing that bottlenecks large-scale studies, inconsistent prompting that hurts reproducibility, and narrow task coverage that misses important audio reasoning capabilities. We introduce AU-Harness, an efficient and comprehensive evaluation framework for LALMs. Our system achieves a speedup of up to 127% over existing toolkits through optimized batch processing and parallel execution, enabling large-scale evaluations previously impractical. We provide standardized prompting protocols and flexible configurations for fair model comparison across diverse scenarios. Additionally, we introduce two new evaluation categories: LLM-Adaptive Diarization for temporal audio understanding and Spoken Language Reasoning for complex audio-based cognitive tasks. Through evaluation across 380+ tasks, we reveal significant gaps in current LALMs, particularly in temporal understanding and complex spoken language reasoning tasks. Our findings also highlight a lack of standardization in instruction modality existent across audio benchmarks, which can lead up performance differences up to 9.5 absolute points on the challenging complex instruction following downstream tasks. AU-Harness provides both practical evaluation tools and insights into model limitations, advancing systematic LALM development.

Vision-Language Models (VLMs) have demonstrated remarkable success across diverse visual tasks, yet their performance degrades in complex visual environments. While existing enhancement approaches require additional training, rely on external segmentation tools, or operate at coarse-grained levels, they overlook the innate ability within VLMs. To bridge this gap, we investigate VLMs' attention patterns and discover that: (1) visual complexity strongly correlates with attention entropy, negatively impacting reasoning performance; (2) attention progressively refines from global scanning in shallow layers to focused convergence in deeper layers, with convergence degree determined by visual complexity. Building on these insights, we propose Contrastive Attention Refinement for Visual Enhancement (CARVE), a training-free method that extracts task-relevant visual signals through attention contrasting at the pixel level. Extensive experiments demonstrate that CARVE consistently enhances performance, achieving up to 75% improvement on open-source models. Our work provides critical insights into the interplay between visual complexity and attention mechanisms, offering an efficient pathway for improving visual reasoning with contrasting attention. Vision-Language Models (VLMs) have achieved remarkable success across diverse tasks (Radford et al., 2021; Jia et al., 2021; Alayrac et al., 2022).

The increasing availability of personal data has enabled significant advances in fields such as machine learning, healthcare, and cybersecurity. However, this data abundance also raises serious privacy concerns, especially in light of powerful re-identification attacks and growing legal and ethical demands for responsible data use. Differential privacy (DP) has emerged as a principled, mathematically grounded framework for mitigating these risks. This review provides a comprehensive survey of DP, covering its theoretical foundations, practical mechanisms, and real-world applications. It explores key algorithmic tools and domain-specific challenges - particularly in privacy-preserving machine learning and synthetic data generation. The report also highlights usability issues and the need for improved communication and transparency in DP systems. Overall, the goal is to support informed adoption of DP by researchers and practitioners navigating the evolving landscape of data privacy.

This paper presents MagicGUI, a foundational mobile GUI agent designed to address critical challenges in perception, grounding, and reasoning within real-world mobile GUI environments. The framework is underpinned by following six key components: (1) a comprehensive and accurate dataset, constructed via the scalable GUI Data Pipeline, which aggregates the largest and most diverse GUI-centric multimodal data to date from open-source repositories, automated crawling, and targeted manual annotation; (2) enhanced perception and grounding capabilities, facilitating fine-grained multimodal alignment for UI element referencing, grounding, and screen comprehension; (3) a comprehensive and unified action space, encompassing both fundamental UI operations and complex interactive intents to support human-agent interactions; (4) planning-oriented reasoning mechanisms that enable the model to decompose complex user instructions into sequential actions with explicit intermediate meta-paln reasoning; (5) an iterative two-stage training procedure, combining large-scale continue pre-training on 7.8M samples with reinforcement fine-tuning utilizing a spatially enhanced composite reward and dual filtering strategy; and (6) competitive performance on both the proprietary Magic-RICH benchmark and over a dozen public benchmarks, achieving superior performance across GUI perception and agent tasks, while demonstrating robust generalization and real-world deployment potential in practical mobile GUI scenarios, as detailed in Figure 1.

Ontology Matching (OM) is a cornerstone task of semantic interoperability, yet existing systems often rely on handcrafted rules or specialized models with limited adaptability. We present KROMA, a novel OM framework that harnesses Large Language Models (LLMs) within a Retrieval-Augmented Generation (RAG) pipeline, to dynamically enrich the semantic context of OM tasks with structural, lexical, and definitional knowledge. To optimize both performance and efficiency, KROMA integrates a bisimilarity-based concept matching and a lightweight ontology refinement step, which prune candidate concepts and substantially reduce the communication overhead from invoking LLMs. Through experiments on multiple benchmark datasets, we show that integrating knowledge retrieval with context-augmented LLMs significantly enhances ontology matching--outperforming both classic OM systems and cutting-edge LLM-based approaches--while keeping communication overhead comparable. Our study highlights the feasibility and benefit of the proposed optimization techniques (targeted knowledge retrieval, prompt enrichment, and ontology refinement) for ontology matching at scale. Our code and experimental dataset has been made available at: https://github.com/lamng3/kroma