Erratum to the paper (Zhang et al., 2025): corrections to Table IV and the data in Page 3, Section A. In the post-pandemic era, a high proportion of civil aviation passengers wear masks during security checks, posing significant challenges to traditional face recognition models. The backbone network serves as the core component of face recognition models. In standard tests, r100 series models excelled (98%+ accuracy at 0.01% FAR in face comparison, high top1/top5 in search). r50 ranked second, r34_mask_v1 lagged. In masked tests, r100_mask_v2 led (90.07% accuracy), r50_mask_v3 performed best among r50 but trailed r100. Vit-Small/Tiny showed strong masked performance with gains in effectiveness. Through extensive comparative experiments, this paper conducts a comprehensive evaluation of several core backbone networks, aiming to reveal the impacts of different models on face recognition with and without masks, and provide specific deployment recommendations.

Existing face-swapping methods often deliver competitive results in constrained settings but exhibit substantial quality degradation when handling extreme facial poses. To improve facial pose robustness, explicit geometric features are applied, but this approach remains problematic since it introduces additional dependencies and increases computational cost. Diffusion-based methods have achieved remarkable results; however, they are impractical for real-time processing. We introduce AlphaFace, which leverages an open-source vision-language model and CLIP image and text embeddings to apply novel visual and textual semantic contrastive losses. AlphaFace enables stronger identity representation and more precise attribute preservation, all while maintaining real-time performance. Comprehensive experiments across FF++, MPIE, and LPFF demonstrate that AlphaFace surpasses state-of-the-art methods in pose-challenging cases. The project is publicly available on `https://github.com/andrewyu90/Alphaface_Official.git'.

Infrared small target detection (IRSTD) is critical for applications like remote sensing and surveillance, which aims to identify small, low-contrast targets against complex backgrounds. However, existing methods often struggle with inadequate joint modeling of local-global features (harming target-background discrimination) or feature redundancy and semantic dilution (degrading target representation quality). To tackle these issues, we propose DCCS-Det (Directional Context and Cross-Scale Aware Detector for Infrared Small Target), a novel detector that incorporates a Dual-stream Saliency Enhancement (DSE) block and a Latent-aware Semantic Extraction and Aggregation (LaSEA) module. The DSE block integrates localized perception with direction-aware context aggregation to help capture long-range spatial dependencies and local details. On this basis, the LaSEA module mitigates feature degradation via cross-scale feature extraction and random pooling sampling strategies, enhancing discriminative features and suppressing noise. Extensive experiments show that DCCS-Det achieves state-of-the-art detection accuracy with competitive efficiency across multiple datasets. Ablation studies further validate the contributions of DSE and LaSEA in improving target perception and feature representation under complex scenarios. \href{https://huggingface.co/InPeerReview/InfraredSmallTargetDetection-IRSTD.DCCS}{DCCS-Det Official Code is Available Here!}

Designing experiments that systematically gather data from complex physical systems is central to accelerating scientific discovery. While Bayesian experimental design (BED) provides a principled, information-based framework that integrates experimental planning with probabilistic inference, the selection of utility functions in BED is a long-standing and active topic, where different criteria emphasize different notions of information. Although Kullback--Leibler (KL) divergence has been one of the most common choices, recent studies have proposed Wasserstein distance as an alternative. In this work, we first employ a toy example to illustrate an issue of Wasserstein distance - the value of Wasserstein distance of a fixed-shape posterior depends on the relative position of its main mass within the support and can exhibit false rewards unrelated to information gain, especially with a non-informative prior (e.g., uniform distribution). We then further provide a systematic comparison between these two criteria through a classical source inversion problem in the BED literature, revealing that the KL divergence tends to lead to faster convergence in the absence of model discrepancy, while Wasserstein metrics provide more robust sequential BED results if model discrepancy is non-negligible. These findings clarify the trade-offs between KL divergence and Wasserstein metrics for the utility function and provide guidelines for selecting suitable criteria in practical BED applications.

We present RENEW, a global path planner for Autonomous Surface Vehicle (ASV) in dynamic environments with external disturbances (e.g., water currents). RENEW introduces a unified risk- and energy-aware strategy that ensures safety by dynamically identifying non-navigable regions and enforcing adaptive safety constraints. Inspired by maritime contingency planning, it employs a best-effort strategy to maintain control under adverse conditions. The hierarchical architecture combines high-level constrained triangulation for topological diversity with low-level trajectory optimization within safe corridors. Validated with real-world ocean data, RENEW is the first framework to jointly address adaptive non-navigability and topological path diversity for robust maritime navigation.

Deep convolutional neural networks can use hierarchical information to progressively extract structural information to recover high-quality images. However, preserving the effectiveness of the obtained structural information is important in image super-resolution. In this paper, we propose a cosine network for image super-resolution (CSRNet) by improving a network architecture and optimizing the training strategy. To extract complementary homologous structural information, odd and even heterogeneous blocks are designed to enlarge the architectural differences and improve the performance of image super-resolution. Combining linear and non-linear structural information can overcome the drawback of homologous information and enhance the robustness of the obtained structural information in image super-resolution. Taking into account the local minimum of gradient descent, a cosine annealing mechanism is used to optimize the training procedure by performing warm restarts and adjusting the learning rate. Experimental results illustrate that the proposed CSRNet is competitive with state-of-the-art methods in image super-resolution.

Vapnik--Chervonenkis' theorem is a seminal result in machine learning. It establishes sufficient conditions for empirical probabilities to converge to theoretical probabilities, uniformly over families of events. It also provides an estimate for the rate of such uniform convergence. We revisit the probabilistic component of the classical argument. Instead of applying Hoeffding's inequality at the final step, we use a normal approximation with explicit Berry--Esseen error control. This yields a moderate-deviation sharpening of the usual VC estimate, with an additional factor of order $(\varepsilon\sqrt{n})^{-1}$ in the leading exponential term when $\varepsilon\sqrt{n}$ is large.

Rare-event prediction is critical in domains such as healthcare, finance, reliability engineering, customer support, aviation safety, where positive outcomes are infrequent yet potentially catastrophic. Extreme class imbalance biases conventional models toward majority-class predictions, limiting recall, calibration, and operational usefulness. We propose LPCORP (Low-Prevalence CORrector for Prediction)*, a two-stage framework that combines reasoningenhanced prediction with confidence-based outcome correction. A reasoning model first produces enriched predictions from narrative inputs, after which a lightweight logistic-regression classifier evaluates and selectively corrects these outputs to mitigate prevalence-driven bias. We evaluate LPCORP on real-world datasets from medical and consumer service domains. The results show that this method transforms a highly imbalanced setting into a well-balanced one while preserving the original number of samples and without applying any resampling strategies. Test-set evaluation demonstrates substantially improved performance, particularly in precision, which is a known weakness in low-prevalence data. We further provide a costreduction analysis comparing the expenses associated with rare-event damage control without preventive measures to those incurred when low-cost, prediction-based preventive interventions are applied that showed more than 50% reduction in some cases. * Patent pending: U.S. Provisional 63/933,518, filed 8 December 2025.

Coverage Path Planning (CPP) is a fundamental capability for agricultural robots; however, existing solutions often overlook energy constraints, resulting in incomplete operations in large-scale or resource-limited environments. This paper proposes an energy-aware CPP framework grounded in Soft Actor-Critic (SAC) reinforcement learning, designed for grid-based environments with obstacles and charging stations. To enable robust and adaptive decision-making under energy limitations, the framework integrates Convolutional Neural Networks (CNNs) for spatial feature extraction and Long Short-Term Memory (LSTM) networks for temporal dynamics. A dedicated reward function is designed to jointly optimize coverage efficiency, energy consumption, and return-to-base constraints. Experimental results demonstrate that the proposed approach consistently achieves over 90% coverage while ensuring energy safety, outperforming traditional heuristic algorithms such as Rapidly-exploring Random Tree (RRT), Particle Swarm Optimization (PSO), and Ant Colony Optimization (ACO) baselines by 13.4-19.5% in coverage and reducing constraint violations by 59.9-88.3%. These findings validate the proposed SAC-based framework as an effective and scalable solution for energy-constrained CPP in agricultural robotics.

Reinforcement Learning from Human Feedback (RLHF) and its variants have emerged as the dominant approaches for aligning Large Language Models with human intent. While empirically effective, the theoretical generalization properties of these methods in high-dimensional settings remain to be explored. To this end, we build the generalization theory on RLHF of LLMs under the linear reward model, through the framework of algorithmic stability. In contrast to the existing works built upon the consistency of maximum likelihood estimations on reward model, our analysis is presented under an end-to-end learning framework, which is consistent with practice. Concretely, we prove that under a key \textbf{feature coverage} condition, the empirical optima of policy model have a generalization bound of order $\mathcal{O}(n^{-\frac{1}{2}})$. Moreover, the results can be extrapolated to parameters obtained by gradient-based learning algorithms, i.e., Gradient Ascent (GA) and Stochastic Gradient Ascent (SGA). Thus, we argue that our results provide new theoretical evidence for the empirically observed generalization of LLMs after RLHF.

Real-world visual question answering (VQA) is often context-dependent: an image-question pair may be under-specified, such that the correct answer depends on external information that is not observable in the image. In such cases, directly answering can lead to confident but incorrect predictions. We propose CoA(Clarify-or-Answer), an ask-or-answer agent that separately models the decision to ask or answer, and what to ask if needed. CoA first determines whether clarification is necessary; if so, it asks a single focused question and then incorporates the response to produce the final answer. We introduce CONTEXTCLARIFY with a set of ambiguous VQA questions and the contrast set that is non-ambiguous. We further introduce GRPO-CR (Clarification Reasoning), a reinforcement learning approach that optimizes clarification question generation with multiple reward signals encouraging well-formed, focused, non-trivial questions that resolve ambiguity. Across three VLLMs and three datasets, CoA achieves consistent improvements at both the module and system levels, improving end-to-end VQA accuracy by an average of +15.3 points (83%) over prompting-based baselines

Referring Expression Segmentation (RES) is a core vision-language segmentation task that enables pixel-level understanding of targets via free-form linguistic expressions, supporting critical applications such as human-robot interaction and augmented reality. Despite the progress of Multimodal Large Language Model (MLLM)-based approaches, existing RES methods still suffer from two key limitations: first, the coarse bounding boxes from MLLMs lead to redundant or non-discriminative point prompts; second, the prevalent reliance on textual coordinate reasoning is unreliable, as it fails to distinguish targets from visually similar distractors. To address these issues, we propose \textbf{\model}, a novel RES framework integrating \textbf{E}ntropy-\textbf{B}ased Point \textbf{D}iscovery (\textbf{EBD}) and \textbf{V}ision-\textbf{B}ased \textbf{R}easoning (\textbf{VBR}). Specifically, EBD identifies high-information candidate points by modeling spatial uncertainty within coarse bounding boxes, treating point selection as an information maximization process. VBR verifies point correctness through joint visual-semantic alignment, abandoning text-only coordinate inference for more robust validation. Built on these components, \model implements a coarse-to-fine workflow: bounding box initialization, entropy-guided point discovery, vision-based validation, and mask decoding. Extensive evaluations on four benchmark datasets (RefCOCO, RefCOCO+, RefCOCOg, and ReasonSeg) demonstrate that \model achieves new state-of-the-art performance across all four benchmarks, highlighting its effectiveness in generating accurate and semantically grounded segmentation masks with minimal prompts.

Large language models exhibit strong multilingual capabilities, yet significant performance gaps persist between dominant and non-dominant languages. Prior work attributes this gap to imbalances between shared and language-specific neurons in multilingual representations. We propose Cross-Lingual Activation Steering (CLAS), a training-free inference-time intervention that selectively modulates neuron activations. We evaluate CLAS on classification and generation benchmarks, achieving average improvements of 2.3% (Acc.) and 3.4% (F1) respectively, while maintaining high-resource language performance. We discover that effective transfer operates through functional divergence rather than strict alignment; performance gains correlate with increased language cluster separation. Our results demonstrate that targeted activation steering can unlock latent multilingual capacity in existing models without modification to model weights.

Few-Shot Anomaly Detection (FSAD) has emerged as a critical paradigm for identifying irregularities using scarce normal references. While recent methods have integrated textual semantics to complement visual data, they predominantly rely on features pre-trained on natural scenes, thereby neglecting the granular, domain-specific semantics essential for industrial inspection. Furthermore, prevalent fusion strategies often resort to superficial concatenation, failing to address the inherent semantic misalignment between visual and textual modalities, which compromises robustness against cross-modal interference. To bridge these gaps, this study proposes VTFusion, a vision-text multimodal fusion framework tailored for FSAD. The framework rests on two core designs. First, adaptive feature extractors for both image and text modalities are introduced to learn task-specific representations, bridging the domain gap between pre-trained models and industrial data; this is further augmented by generating diverse synthetic anomalies to enhance feature discriminability. Second, a dedicated multimodal prediction fusion module is developed, comprising a fusion block that facilitates rich cross-modal information exchange and a segmentation network that generates refined pixel-level anomaly maps under multimodal guidance. VTFusion significantly advances FSAD performance, achieving image-level AUROCs of 96.8% and 86.2% in the 2-shot scenario on the MVTec AD and VisA datasets, respectively. Furthermore, VTFusion achieves an AUPRO of 93.5% on a real-world dataset of industrial automotive plastic parts introduced in this paper, further demonstrating its practical applicability in demanding industrial scenarios.

Multimodal language models (MLMs) perform well on semantic vision-language tasks but fail at spatial reasoning that requires adopting another agent's visual perspective. These errors reflect a persistent egocentric bias and raise questions about whether current models support allocentric reasoning. Inspired by human spatial cognition, we introduce perspective tokens, specialized embeddings that encode orientation through either (1) embodied body-keypoint cues or (2) abstract representations supporting mental rotation. Integrating these tokens into LLaVA-1.5-13B yields performance on level-2 visual perspective-taking tasks. Across synthetic and naturalistic benchmarks (Isle Bricks V2, COCO, 3DSRBench), perspective tokens improve accuracy, with rotation-based tokens generalizing to non-human reference agents. Representational analyses reveal that fine-tuning enhances latent orientation sensitivity already present in the base model, suggesting that MLMs contain precursors of allocentric reasoning but lack appropriate internal structure. Overall, embedding cognitively grounded spatial structure directly into token space provides a lightweight, model-agnostic mechanism for perspective-taking and more human-like spatial reasoning.

This study investigates regional bias in large language models (LLMs), an emerging concern in AI fairness and global representation. We evaluate ten prominent LLMs: GPT-3.5, GPT-4o, Gemini 1.5 Flash, Gemini 1.0 Pro, Claude 3 Opus, Claude 3.5 Sonnet, Llama 3, Gemma 7B, Mistral 7B, and Vicuna-13B using a dataset of 100 carefully designed prompts that probe forced-choice decisions between regions under contextually neutral scenarios. We introduce FAZE, a prompt-based evaluation framework that measures regional bias on a 10-point scale, where higher scores indicate a stronger tendency to favor specific regions. Experimental results reveal substantial variation in bias levels across models, with GPT-3.5 exhibiting the highest bias score (9.5) and Claude 3.5 Sonnet scoring the lowest (2.5). These findings indicate that regional bias can meaningfully undermine the reliability, fairness, and inclusivity of LLM outputs in real-world, cross-cultural applications. This work contributes to AI fairness research by highlighting the importance of inclusive evaluation frameworks and systematic approaches for identifying and mitigating geographic biases in language models.

Art technological investigations of historical panel paintings rely on acquiring multi-modal image data, including visual light photography, infrared reflectography, ultraviolet fluorescence photography, x-radiography, and macro photography. For a comprehensive analysis, the multi-modal images require pixel-wise alignment, which is still often performed manually. Multi-modal image registration can reduce this laborious manual work, is substantially faster, and enables higher precision. Due to varying image resolutions, huge image sizes, non-rigid distortions, and modality-dependent image content, registration is challenging. Therefore, we propose a coarse-to-fine non-rigid multi-modal registration method efficiently relying on sparse keypoints and thin-plate-splines. Historical paintings exhibit a fine crack pattern, called craquelure, on the paint layer, which is captured by all image systems and is well-suited as a feature for registration. In our one-stage non-rigid registration approach, we employ a convolutional neural network for joint keypoint detection and description based on the craquelure and a graph neural network for descriptor matching in a patch-based manner, and filter matches based on homography reprojection errors in local areas. For coarse-to-fine registration, we introduce a novel multi-level keypoint refinement approach to register mixed-resolution images up to the highest resolution. We created a multi-modal dataset of panel paintings with a high number of keypoint annotations, and a large test set comprising five multi-modal domains and varying image resolutions. The ablation study demonstrates the effectiveness of all modules of our refinement method. Our proposed approaches achieve the best registration results compared to competing keypoint and dense matching methods and refinement methods.

Data science agents promise to accelerate discovery and insight-generation by turning data into executable analyses and findings. Yet existing data science benchmarks fall short due to fragmented evaluation interfaces that make cross-benchmark comparison difficult, narrow task coverage and a lack of rigorous data grounding. In particular, we show that a substantial portion of tasks in current benchmarks can be solved without using the actual data. To address these limitations, we introduce DSGym, a standardized framework for evaluating and training data science agents in self-contained execution environments. Unlike static benchmarks, DSGym provides a modular architecture that makes it easy to add tasks, agent scaffolds, and tools, positioning it as a live, extensible testbed. We curate DSGym-Tasks, a holistic task suite that standardizes and refines existing benchmarks via quality and shortcut solvability filtering. We further expand coverage with (1) DSBio: expert-derived bioinformatics tasks grounded in literature and (2) DSPredict: challenging prediction tasks spanning domains such as computer vision, molecular prediction, and single-cell perturbation. Beyond evaluation, DSGym enables agent training via execution-verified data synthesis pipeline. As a case study, we build a 2,000-example training set and trained a 4B model in DSGym that outperforms GPT-4o on standardized analysis benchmarks. Overall, DSGym enables rigorous end-to-end measurement of whether agents can plan, implement, and validate data analyses in realistic scientific context.

College students experience many stressors, resulting in high levels of anxiety and depression. Wearable technology provides unobtrusive sensor data that can be used for the early detection of mental illness. However, current research is limited concerning the variety of psychological instruments administered, physiological modalities, and time series parameters. In this research, we collect the Student Mental and Environmental Health (StudentMEH) Fitbit dataset from students at our institution during the pandemic. We provide a comprehensive assessment of the ability of predictive machine learning models to screen for depression, anxiety, and stress using different Fitbit modalities. Our findings indicate potential in physiological modalities such as heart rate and sleep to screen for mental illness with the F1 scores as high as 0.79 for anxiety, the former modality reaching 0.77 for stress screening, and the latter modality achieving 0.78 for depression. This research highlights the potential of wearable devices to support continuous mental health monitoring, the importance of identifying best data aggregation levels and appropriate modalities for screening for different mental ailments.

Large language models (LLMs) enable rapid and consistent automated evaluation of open-ended exam responses, including dimensions of content and argumentation that have traditionally required human judgment. This is particularly important in cases where a large amount of exams need to be graded in a limited time frame, such as nation-wide graduation exams in various countries. Here, we examine the applicability of automated scoring on two large datasets of trial exam essays of two full national cohorts from Estonia. We operationalize the official curriculum-based rubric and compare LLM and statistical natural language processing (NLP) based assessments with human panel scores. The results show that automated scoring can achieve performance comparable to that of human raters and tends to fall within the human scoring range. We also evaluate bias, prompt injection risks, and LLMs as essay writers. These findings demonstrate that a principled, rubric-driven, human-in-the-loop scoring pipeline is viable for high-stakes writing assessment, particularly relevant for digitally advanced societies like Estonia, which is about to adapt a fully electronic examination system. Furthermore, the system produces fine-grained subscore profiles that can be used to generate systematic, personalized feedback for instruction and exam preparation. The study provides evidence that LLM-assisted assessment can be implemented at a national scale, even in a small-language context, while maintaining human oversight and compliance with emerging educational and regulatory standards.

Federated learning enables collaborative model training across geographically distributed medical centers while preserving data privacy. However, domain shifts and heterogeneity in data often lead to a degradation in model performance. Medical imaging applications are particularly affected by variations in acquisition protocols, scanner types, and patient populations. To address these issues, we introduce Federated Template and Task Learning (FeTTL), a novel framework designed to harmonize multi-institutional medical imaging data in federated environments. FeTTL learns a global template together with a task model to align data distributions among clients. We evaluated FeTTL on two challenging and diverse multi-institutional medical imaging tasks: retinal fundus optical disc segmentation and histopathological metastasis classification. Experimental results show that FeTTL significantly outperforms the state-of-the-art federated learning baselines (p-values <0.002) for optical disc segmentation and classification of metastases from multi-institutional data. Our experiments further highlight the importance of jointly learning the template and the task. These findings suggest that FeTTL offers a principled and extensible solution for mitigating distribution shifts in federated learning, supporting robust model deployment in real-world, multi-institutional environments.

Multi-agent systems powered by large language models (LLMs) are transforming enterprise automation, yet systematic evaluation methodologies for assessing tool-use reliability remain underdeveloped. We introduce a comprehensive diagnostic framework that leverages big data analytics to evaluate procedural reliability in intelligent agent systems, addressing critical needs for SME-centric deployment in privacy-sensitive environments. Our approach features a 12-category error taxonomy capturing failure modes across tool initialization, parameter handling, execution, and result interpretation. Through systematic evaluation of 1,980 deterministic test instances spanning both open-weight models (Qwen2.5 series, Functionary) and proprietary alternatives (GPT-4, Claude 3.5/3.7) across diverse edge hardware configurations, we identify actionable reliability thresholds for production deployment. Our analysis reveals that procedural reliability, particularly tool initialization failures, constitutes the primary bottleneck for smaller models, while qwen2.5:32b achieves flawless performance matching GPT-4.1. The framework demonstrates that mid-sized models (qwen2.5:14b) offer practical accuracy-efficiency trade-offs on commodity hardware (96.6\% success rate, 7.3 s latency), enabling cost-effective intelligent agent deployment for resource-constrained organizations. This work establishes foundational infrastructure for systematic reliability evaluation of tool-augmented multi-agent AI systems.

Large Language Models (LLMs) have demonstrated remarkable multilingual capabilities, making them promising tools in both high- and low-resource languages. One particularly valuable use case is generating synthetic samples that can be used to train smaller models in low-resource scenarios where human-labelled data is scarce. In this work, we investigate whether these synthetic data generation capabilities can serve as a form of distillation, producing smaller models that perform on par with or even better than massive LLMs across languages and tasks. To this end, we use a state-of-the-art multilingual LLM to generate synthetic datasets covering 11 languages and 4 classification tasks. These datasets are then used to train smaller models via fine-tuning or instruction tuning, or as synthetic in-context examples for compact LLMs. Our experiments show that even small amounts of synthetic data enable smaller models to outperform the large generator itself, particularly in low-resource languages. Overall, the results suggest that LLMs are best utilised as generators (teachers) rather than classifiers, producing data that empowers smaller and more efficient multilingual models.

Strategic decision-making in multi-agent settings is a key challenge for large language models (LLMs), particularly when coordination and negotiation must unfold over extended conversations. While recent work has explored the use of LLMs in isolated decision tasks, little attention has been given to optimizing long-term objectives through dialogue. We introduce \textbf{GameTalk}, a framework for training LLMs to make strategic decisions via multi-turn interactions. Unlike prior work that focuses on single-turn objectives or static action prediction, we train LLMs to optimize a global objective across full conversations. We achieve this by adapting fine-tuning methods like GRPO, DPO, and STaR to incorporate reward signals that depend on the entire interaction. We evaluate this approach on a suite of increasingly complex games, designed to stress different aspects of reasoning, coordination, and opponent modeling. Our results show that GameTalk significantly outperforms untrained models, especially under reward shaping, with DPO consistently yielding the strongest gains. These findings position conversational fine-tuning as a promising path for LLMs to reason, negotiate, and act in interactive environments.

This technical report describes our submission to the ICME 2025 audio encoder challenge. Our submitted system is built on BEATs, a masked speech token prediction based audio encoder. We extend the BEATs model using 74,000 hours of data derived from various speech, music, and sound corpora and scale its architecture upto 300 million parameters. We experiment with speech-heavy and balanced pre-training mixtures to study the impact of different domains on final performance. Our submitted system consists of an ensemble of the Dasheng 1.2 billion model with two custom scaled-up BEATs models trained on the aforementioned pre-training data mixtures. We also propose a simple ensembling technique that retains the best capabilities of constituent models and surpasses both the baseline and Dasheng 1.2B. For open science, we publicly release our trained checkpoints via huggingface at https://huggingface.co/shikhar7ssu/OpenBEATs-ICME-SOUND and https://huggingface.co/shikhar7ssu/OpenBEATs-ICME.

Personalized speech enhancement (PSE) has shown convincing results when it comes to extracting a known target voice among interfering ones. The corresponding systems usually incorporate a representation of the target voice within the enhancement system, which is extracted from an enrollment clip of the target voice with upstream models. Those models are generally heavy as the speaker embedding's quality directly affects PSE performances. Yet, embeddings generated beforehand cannot account for the variations of the target voice during inference time. In this paper, we propose to perform on-thefly refinement of the speaker embedding using a tiny speaker encoder. We first introduce a novel contrastive knowledge distillation methodology in order to train a 150k-parameter encoder from complex embeddings. We then use this encoder within the enhancement system during inference and show that the proposed method greatly improves PSE performances while maintaining a low computational load.

Multimodal foundation models that integrate audio, vision, and language achieve strong performance on reasoning and generation tasks, yet their robustness to adversarial manipulation remains poorly understood. We study a realistic and underexplored threat model: untargeted, audio-only adversarial attacks on trimodal audio-video-language models. We analyze six complementary attack objectives that target different stages of multimodal processing, including audio encoder representations, cross-modal attention, hidden states, and output likelihoods. Across three state-of-the-art models and multiple benchmarks, we show that audio-only perturbations can induce severe multimodal failures, achieving up to 96% attack success rate. We further show that attacks can be successful at low perceptual distortions (LPIPS <= 0.08, SI-SNR >= 0) and benefit more from extended optimization than increased data scale. Transferability across models and encoders remains limited, while speech recognition systems such as Whisper primarily respond to perturbation magnitude, achieving >97% attack success under severe distortion. These results expose a previously overlooked single-modality attack surface in multimodal systems and motivate defenses that enforce cross-modal consistency.

Large language models (LLMs) are typically personalized via prompt engineering or parameter-efficient fine-tuning such as LoRA. However, writing style can be difficult to distill into a single prompt, and LoRA fine-tuning requires computationally intensive training and infrastructure. We investigate a possible lightweight alternative: steering a frozen LLM with n-gram style priors injected in logit space at decoding time. We train an n-gram model on stylistically distinct corpora -- including Don Quixote, CNN/DailyMail news headlines, and arXiv abstracts -- constructing an interpolated 1-to-3-gram prior over next-token probabilities. During generation we modify the LLM's logits by adding a weighted sum of style log-probabilities from each n-gram order that matches the current context, scaled by a control parameter lambda in [0, 1]. We sweep lambda and style corpora and report style perplexity under the n-gram model, base-model perplexity as a proxy for fluency, Jensen-Shannon (JS) divergence between the original and steered token distributions, and token-overlap statistics. On TinyLlama-1.1B we identify a single narrow regime (for the Don Quixote corpus at lambda=0.1) where style perplexity improves by 24.7% and base-model perplexity improves by 51.4% relative to the frozen model. Outside this regime, and for multi-author corpora such as CNN/DailyMail and arXiv abstracts, even small nonzero lambda values generally result in worse style and fluency, and larger lambda values lead to collapse with extreme perplexities and incoherent text. Logit-space injection of n-gram style priors provides lightweight, tunable style control, but it is fragile: it operates effectively only within a narrow range of low lambda values and is consistently outperformed by prompting and LoRA.

Language diffusion models aim to improve sampling speed and coherence over autoregressive LLMs. We introduce Neural Flow Diffusion Models for language generation, an extension of NFDM that enables the straightforward application of continuous diffusion models to discrete state spaces. NFDM learns a multivariate forward process from the data, ensuring that the forward process and generative trajectory are a good fit for language modeling. Our model substantially reduces the likelihood gap with autoregressive models of the same size, while achieving sample quality comparable to that of previous latent diffusion models.

Large Language Models (LLMs) excel in general tasks but often struggle with hallucinations when handling domain-specific or institutional knowledge absent from their pre-training. We present an offline response-based knowledge distillation method that develops high-accuracy specialized assistants under constrained hardware resources. We evaluate three distinct data strategies: general domain adaptation (15,000 lines), unstructured knowledge injection (2,000 lines), and a context-aware synthetic dataset (500 lines) generated by a teacher model. To minimize computational costs, we utilize the Unsloth library to optimize the Qwen-2.5-7B student model, reducing NVIDIA A100 GPU memory requirements from 40 GB to 16 GB. Experimental results demonstrate that while larger unstructured datasets suffer from persistent hallucinations, the 500-line context-aware dataset achieves a 96.7% accuracy rate and robust rejection capability. These findings validate the LIMA hypothesis, showing that data quality and structural alignment are more critical than quantity for domain adaptation in low-resource settings.