arXiv Research Digest

Large Language Models (LLMs) face significant inference latency challenges stemming from their autoregressive design and large size . To address this, speculative decoding emerges as a solution, enabling the simultaneous generation and validation of multiple tokens .

Knowledge distillation (KD) demonstrates promising results in transferring knowledge from larger to smaller VLMs . applying KD in multilingualism is an underexplored area . We study five distillation formulations across CLIP and SigLIP2 .

Human Activity Recognition (HAR) via Wi-Fi Channel State Information (CSI) presents a privacy-preserving, contactless sensing approach suitable for smart homes, healthcare monitoring, and mobile IoT systems .

Knowledge distillation (KD) is an effective method for model compression and transferring knowledge between models . However, its effect on model's robustness against spurious correlations that degrade performance on out-of-distribution data remains underexplored . This study investigates the effect of knowledge distillation on the transferability of ``debiasing'' capabilities from teacher models to student models .

We present a novel framework for Industry 5.0 that simplifies the deployment of AI models on edge devices in various industrial settings . The design reduces latency and avoids external data transfer by enabling local inference and real-time processing .

Autonomous driving is an emerging technology that is expected to bring significant social, economic, and environmental benefits . However, these benefits come with rising energy consumption by computation engines limiting the driving range of vehicles, especially electric ones . Perception computing is typically the most power-intensive component, as it relies on deep learning models to extract environmental features . To address these challenges, we propose an energy-efficient autonomous driving framework, called EneAD .

While machine learning has transformed numerous application domains, its growing computational demands increasingly constrain scalability and efficiency . In practice, neural networks must not only operate efficiently but also provide reliable predictions under distributional shifts or unseen data . This work advances resource-efficient and robust inference for both conventional and Bayesian neural networks .

Knowledge distillation (KD) is an effective model compression technique that transfers knowledge from a high-performance teacher to a lightweight student, reducing cost while maintaining accuracy . In visual applications, where large-scale image models are widely used, KD enables efficient deployment .

Vision-Language-Action models (VLAs) represent a significant frontier in embodied intelligence, aiming to bridge digital knowledge with physical-world interaction . While these models have demonstrated remarkable generalist capabilities, deployment is severely hampered by the substantial computational and data requirements .

Edge intelligent applications like VR/AR and language model based chatbots have become widespread with the rapid expansion of IoT and mobile devices . But constrained edge devices often cannot serve the increasingly large and complex deep learning (DL) models . Research aims to balance accuracy, computation delay, transmission delay, and privacy concerns .

Over-the-air (OTA) federated learning (FL) has been well recognized as a scalable paradigm that exploits the waveform superposition of the wireless multiple-access channel . We develop novel OTA-FL SGD updates that allow a structured, time-invariant model bias while facilitating reduced variance updates .

We present a simple perturbation mechanism for the release of $d-dimensional covariance matrices under pure differential privacy . For large datasets with at least $n\geq d^2/\varepsilon elements, our mechanism recovers the provably optimal Frobenius norm error guarantees of \cite{nikolov2023private}.

We revisit the task of computing the span of the top $r$ singular vectors $u_1, \ldots, u_r$ of a matrix under differential privacy . We show that a simple and efficient algorithm -- based on singular value decomposition and standard perturbation mechanisms -- returns a private rank-$r$ approximation whose error depends only on the coherence of the input .

Federated learning (FL) has emerged as a key paradigm for collaborative model training across multiple clients without sharing raw data . We propose UnifiedFL, a dynamic federated learning framework that represents heterogeneous local networks as nodes and edges in a directed model graph optimized by a shared graph neural network .

Local Differential Privacy (LDP) is a widely adopted privacy-protection model in the Internet of Things . However, existing defenses either incur prohibitive resource overheads or rely on domain-specific prior knowledge . We propose PEEL, a Poisoning-Exposing Encoding theoretical framework for LDP, which departs from resource- or prior-dependent countermeasures . PEEL amplifies stealthy poisoning effects by re-encoding LDP-perturbed data via sparsification, normalization, and low-rank projection .

Contribution-Guided Asymmetric Learning (CAL) aims to enhance the contribution of high-contribution modalities while compressing weak modalities to increase their contribution . CAL has shown outstanding performance in imbalanced fusion tasks and noise robustness tests . CAL is based on a modality contribution metric W^m combining the information quantity I(m) and confidence D(m).

Recent advances in speech foundation models (SFMs) have enabled the direct processing of spoken language from raw audio . This capability allows SFMs to be exposed to rich paralinguistic variations embedded in the input speech signal . One under-explored dimension of this variation is voice quality, encompassing phonation types such as creaky and breathy voice .

Advances in spoken language processing have driven the development of spoken language models . We evaluate four SLMs on the task of speech emotion recognition using a dataset of emotionally incongruent speech samples . Results indicate that SLMs rely predominantly on textual semantics rather than speech emotion to perform the task .

Multimodal emotion recognition is crucial for future human-computer interaction . However accurate emotion recognition still faces significant challenges due to differences between different modalities and the difficulty of characterizing unimodal emotional information . A hybrid network model based on multipath cross-modal interaction (MCIHN) is proposed .

Emotional Rationale Verifier (ERV) and an Explanation Reward are novel approaches to predicting emotions . Authors propose a novel approach: the ERV and an explanation reward . Their method significantly improves explanation-prediction consistency and explanation emotion accuracy .

Negative emotions are linked to the onset of neurodegenerative diseases and dementia . Physiological signals from wearable devices offer a promising noninvasive method for continuous emotion monitoring . The method is designed for deployment in resource-constrained systems, such as Internet of Things .

LUNA (Latent Unified Network Architecture) is a self-supervised foundation model that reconciles disparate electrode geometries while scaling linearly -- not quadratically -- with channel count . LUNA compresses multi-channel EEG into a fixed-size, topology-agnostic latent space via learned queries and cross-attention . It demonstrates highly competitive performance across several benchmarks .

The method outperforms state-of-the-art large-scale EEG models by an average of 4.57% in AUROC for few-shot emotion recognition and 11.92% in accuracy for zero-shot generalization to a new dataset .

Maithili is an Indo-Aryan language spoken by more than 13 million people in the Purvanchal region of India . It is valued for its rich linguistic structure and cultural significance .

Foundation models have transformed AI by reducing reliance on task-specific data through large-scale pretraining . While successful in language and vision, their adoption in EEG has lagged due to the heterogeneity of public datasets . Existing EEG foundation models struggle to generalize across these variations, often restricting pretraining to a single setup . We present REVE (Representation for EEG with Versatile Embeddings) a pretrained model .

Recommender systems often struggle with data sparsity and cold-start scenarios . This paper presents a Graph Attention Network (GAT) based Collaborative Filtering (CF) framework enhanced with context aware embeddings .

In this article, we give completely new examples of embedded complex manifolds the germ of neighborhood of which is holomorphically equivalent to the zero section in its normal bundle . The first set of examples is composed of connected abelian complex Lie groups, embedded in some complex manifold $M$. The second set is $n$-dimensional Hopf manifolds, embedded as hypersurfaces .

The prohibitive cost of evaluating large language models (LLMs) on comprehensive benchmarks necessitates the creation of small yet representative data subsets that enable efficient assessment while retaining predictive fidelity . Current methods for this task operate under a model-centric paradigm, selecting benchmarking items based on the collective performance of existing models . Such approaches are limited by large upfront costs, an inability to immediately handle new benchmarks (`cold-start'), and the fragile assumption that future models will share the failure patterns of their predecessors .

This paper studies how to achieve accurate modeling and effective control in stochastic nonlinear dynamics with multiple interacting objects . Non-uniform interactions and random topologies make this task challenging .

In this paper, we derive a new $p$-Logarithmic Sobolev inequality and optimal continuous and compact embeddings into Orlicz-type spaces of the function space associated with the logarathmic $p$.-Laplacian . By employing the method of the Nehari manifold, we prove the existence of a nontrivial weak solution . We conduct an asymptotic analysis of a weighted nonlocal, nonlinear problem governed by the fractional

Stellar-mass compact objects (COs) embedded in active galactic nucleus (AGN) discs are commonly assumed to accrete via Bondi or Bondi-Hoyle-Lyttleton prescriptions . We show that differential rotation in AGN discs can impart non-negligible angular momentum, in which case accretion proceeds through a viscous disc rather than Bondi/BHL flow .

We show that the heat kernel measures based at the north pole of the spheres converges to a Gaussian measure in $R^\infty$ We also find an explicit formula for this measure .

This is an updated version of a paper which appeared in the proceedings of the 1979 Berlin Colloquium on Global Differential Geometry . The main result of this paper is that every compact isoparametric hypersurface $M^n \subset S^{n+1} is tight .

We show that by embedding liquid-filled acoustofluidic cavities in a metamaterial, the quality factor of the cavity at selected acoustic resonance modes can be enhanced by 2 to 3 orders of magnitude .

Modeling complex spatiotemporal dynamics, particularly in far-from-equilibrium systems, remains a challenge in science . The governing partial differential equations (PDEs) for these systems are often intractable to derive from first principles .

Honeypots are decoy systems used for gathering valuable threat intelligence . Maximising attacker engagement is essential to their utility . Research has highlighted that context-awareness is necessary to increase engagement . Large Language Models (LLMs) have been shown as one approach to increase context awareness .

Large Language Models offer potential for improving RE efficiency through automated comprehension and commenting . Cloud-hosted, closed-weight models pose privacy and security risks and cannot be used in closed-network facilities . REx86 reduces test-set cross-entropy loss by 64.2% and improves semantic cosine similarity against ground truth by 20.3\% over its base model .

Mobile agents rely on Large Language Models (LLMs) to plan and execute tasks on smartphone user interfaces . While cloud-based LLMs achieve high task accuracy, they require uploading the full UI state at every step . In contrast, local LLMs avoid UI uploads but suffer from limited capacity, resulting in lower task success rates . CORE comprises three key components: (1) layout-aware block partitioning, (2) Co-planning) and Co-decision-making .

Modern IR systems are increasingly tasked with answering complex, multi-faceted queries that require deep reasoning . We introduce LATTICE, a hierarchical retrieval framework that enables an LLM to reason over and navigate large corpora with logarithmic search complexity .

COSTAR-A is a novel prompt engineering framework that enhances the existing COSTAR method . COSTAR stands for Context, Objective, Style, Tone, Audience, and Response, by adding the 'Answer' component at the end .

Software vulnerabilities remain a persistent risk, yet static and dynamic analyses often overlook structural dependencies that shape insecure behaviors . Viewing programs as heterogeneous graphs, we capture control- and data-flow relations as complex interaction networks . Our hybrid framework combines these graph representations with light-weight (<4B) local LLMs .

The toolkit is designed to be open (open-source and local), extensible ( plugin support and users can interact with multiple models) The extensibility is enabled through a two-pronged plugin architecture and a community platform for sharing, importing, and adapting extensions .

Large Language Models (LLMs) consistently underperform compared to frontier models in tool calling scenarios . We propose "decoupled fine-tuning" to create dedicated LoRA adapters for tool selection and tool-specific argument generation using separate loss masking for each of the subtasks . DualTune is an inference framework that leverages the LRA adapters created using decoupled fines-tune to perform efficient agent orchestration with the help of local models .

Static analysis tools like Bandit are effective at vulnerability detection but suffer from high false positives and lack repair capabilities . Large Language Models (LLMs) can suggest fixes but often hallucinate changes and lack self-validation . We present SecureFixAgent, a hybrid repair framework integrating Bandit with lightweight local LLMs in an iterative detect-repair-validate loop .

PrivWeb is a trusted add-on on web agents that anonymizes private information on interfaces according to user preferences . It features privacy categorization and adaptive notifications that selectively pauses tasks for user control over information collection for highly sensitive information . PrivWeb reduces perceived privacy risks with no associated increase in cognitive effort, and resulted in higher overall satisfaction .