This survey explores LiDAR applications in rehabilitation, post-injury care, and hospital environments from 2019 to 2025. LiDAR offers advantages over camera-based and wearable sensors, including privacy preservation and comfort. Key applications include 3D body scanning, gait analysis, robotic rehabilitation, real-time monitoring, and activity recognition. The paper also analyzes AI-based processing techniques and identifies research gaps and future directions.

Infrared small target detection (IRSTD) is critical for applications like maritime surveillance and military search and rescue, but faces challenges due to low signal-to-clutter ratios and small target sizes. This paper proposes SANet, a selective attention network built on U-Net, featuring a Dual-path Semantic-aware Module (DSM) and a Selective Attention Fusion Module (SAFM) to enhance target perception and reduce false alarms.

Existing single image dehazing methods perform well on homogeneous thin haze but struggle with non-homogeneous haze. CPIFNet decomposes the problem into homogeneous sub-problems using a two-stage network: IENet for enhancing regions with different haze densities and IFNet for fusing them into a clear image. A comprehensive loss function supervises both stages.

Researchers present LiteVLA-H, a compact 256M-parameter vision-language-action (VLA) system designed for drone deployment. It operates on an NVIDIA Jetson AGX Orin with dual-rate scheduling: a fast guidance mode (50.65ms/19.74Hz) for short action tokens and a slower semantic mode (149.90-164.57ms) for scene understanding and hazard descriptions. Key insight is that latency is dominated by multimodal pre-fill, not token decoding. Using knowledge-preserving fine-tuning, the system achieves high edge inference rates while retaining descriptive competence, outperforming recent architectures like AnywhereVLA, FutureVLA, and ReMem-VLA.

This paper presents a comprehensive survey and benchmark for face swapping. It organizes existing methods into five major paradigms and introduces CASIA FaceSwapping, a high-quality benchmark with balanced demographics and explicit attribute variations, along with standardized evaluation protocols. Extensive experiments provide new insights into current techniques' performance and limitations.

Remote Photoplethysmography (rPPG) enables convenient non-contact physiological measurement. Existing Self-Supervised Learning (SSL) methods commonly fall into a correlation trap: they tend to learn the most dominant periodic signals in the data, such as high-energy motion or illumination noise, rather than the faint, true rPPG signal, leading to poor model generalization. To address this, we propose a new SSL paradigm, Physiological Causal Probing (PCP), which treats the latent rPPG signal as the underlying physical source and the resulting pixel chrominance variations as its visual manifestation. Its core idea is to shift from passive correlation learning to active, precise intervention: it intervenes on the video based on a proposed rPPG hypothesis, and verifies whether the post-interv…

A new gray-box attack framework, Adversarial Flow Matching (AFM), targets Transformer vulnerabilities in end-to-end autonomous driving models, enabling efficient one-step generation of imperceptible perturbations that degrade performance across both VLA and modular architectures while maintaining high visual imperceptibility and cross-model transferability.

A hybrid defogging model integrating a fourth-order nonlinear PDE with a physical haze formation model. Uses Dark Channel Prior for parameter estimation and a fourth-order telegraph PDE evolution with edge-adaptive diffusion and fidelity term. Results show comparable visual quality and structural detail preservation.

This paper systematically compares different visual representations for cryptocurrency regime prediction, evaluating three image encoding methods, five chart component configurations, four neural network architectures, and the impact of transfer learning. Results show a simple 4-layer CNN on raw candlestick charts achieves 0.892 AUC-ROC, outperforming larger pretrained models, and simpler representations consistently outperform complex alternatives. Transfer learning improves performance by 4–16% despite domain gap.

This research proposes a latent space probing framework that intercepts denoised latent representations from the CogVideoX video diffusion model during inference, attaching lightweight classifiers for real-time adult content detection. Built on a dataset of 11,039 ten-second clips (5,086 violating, 5,953 non-violating), it achieves 97.29% F1 on a held-out test set with only 4-6ms overhead.

Current synthetic data pipelines generate images without diagnosing downstream model needs. This paper proposes Synthetic Designed Experiments for Representational Sufficiency (SDRS), leveraging Design of Experiments (DoE) theory. Using fractional factorial designs and ANOVA, SDRS audits factor sensitivity, classifies failures into Type I (coverage gaps) and Type II (spurious dependencies), and prescribes targeted synthetic data. Experiments show accuracy improvement from 49.9% to 79.0% on dSprites, mIoU from 0.948 to 0.998 on segmentation, and cross-factor contamination detection.

This paper studies the local geometry of embedding clouds induced by controlled local classes of semantically close sentences. It introduces affine, quadratic, and cubic fitted models along with a surface-based latent probing procedure. Experiments show nonlinear models outperform affine models, but geometric validity does not automatically improve classification performance. The CoPaGE-300K dataset is introduced.

This study systematically evaluates text decomposition and privacy budget distribution techniques in differentially private text obfuscation, finding that design choices significantly impact privacy-utility trade-offs and providing evidence for optimizing DP obfuscation procedures.

A new paper argues that counterfactual prompting evaluations often fail to account for incidental surface-form variations, leading to false attributions of model sensitivity. The authors propose a framework that compares target interventions to paraphrasing baselines using statistical testing. Revisiting the MedPerturb dataset, they find that only 5 out of 120 tests show significance, while a directional gender bias in occupational biography classification remains detectable. Per-sample and regression metrics are more powerful than aggregate metrics.

A new arXiv preprint formalizes adversarial attacks and defenses in large language models as a theoretical game, deriving provably optimal strategies and demonstrating superior empirical performance against existing methods.

A new benchmark, XHS-SCoRE, reveals that while LLMs can fluently generate social comparison triggers in Xiaohongshu posts, their prompt-based classifiers fail to reliably detect these signals. The study finds stable failure modes like neutralization and directional skew, and shows that LLM-generated posts can actually affect readers' perceptions despite detection fragility.

The CLEAR framework systematically perturbs the number of plausible answer options, the presence of a ground truth or abstention option, and the semantic framing of answer options to evaluate 17 LLMs on three medical benchmarks. Findings show that increasing plausible answers degrades correctness and abstention; abstention framing matters, with 'I don't know' (IDK) increasing errors; and a 'humility deficit' worsens with model scale, indicating scaling alone does not resolve LLM reliability issues.

Large language models can improve cross-partisan trust by reframing liberal news headlines, but only when targeting ideological framing, not surface-level language. The models also overestimate their effectiveness and lack psychological accuracy.

This paper introduces a simple perplexity-based method to surface finetuning objectives from model organisms by comparing the perplexity gap between finetuned and reference models on random prompts. Tested on 76 model organisms (0.5B to 70B parameters), the method effectively reveals backdoors, false fact internalization, adversarial behaviors, and emergent misalignment, and can work without exact pre-finetuning checkpoints using other reference models.

Diagram question answering requires reasoning-level attribution that links each QA pair to all visual regions needed to derive the answer, not just the region with the final answer. Existing annotation tools are tightly coupled to dataset-specific formats. DIAGRAMS is a lightweight, schema-driven review framework that decouples interface logic from dataset-specific JSON structures via an internal meta-schema and dataset adapters. It performs QA-conditioned evidence selection and proposes required regions, generating missing QA pairs or candidate regions when needed, with human verification. Across six datasets, model-suggested evidence achieves 85.39% precision and 75.30% recall against reviewer-final selections, indicating reduced manual effort with high agreement.

Researchers propose H-Probes, linear probes that extract hierarchical depth and pairwise distance from LLM latent representations. Experiments show hierarchy subspaces are low-dimensional, causally important, and generalize across domains. Analogous structures found in real-world mathematical reasoning.

PhaseNet++ introduces a frequency-domain autoencoder that leverages phase spectra from STFT for anomaly detection in ICS. Using a Phase Coherence Index to guide graph attention networks, it achieves competitive results on the SWaT benchmark, marking the first systematic exploration of phase-domain methods in this domain.

This paper reviews the Receiver Operating Characteristic (ROC) curve for binary classifiers, formally proves that the area under the curve (AUC) equals the probability that a classifier ranks a random positive instance higher than a random negative one, and provides an error bound when the required hypothesis is not met.

Haiku is a tri-modal contrastive learning model that integrates multiplexed immunofluorescence, H&E histology, and clinical metadata, enabling cross-modal retrieval, improved clinical predictions, zero-shot biomarker inference, and counterfactual analysis revealing niche-specific molecular shifts in cancer progression.

StyleShield is a novel flow matching framework that effectively evades AI-generated content detectors through conditional text style transfer in continuous token embedding space. It achieves a 94.6% evasion rate against a trained detector and over 99% against three unseen detectors on a multi-domain Chinese benchmark, while maintaining 0.928 semantic similarity. The work also introduces RateAudit, a scheduling algorithm that exposes the arbitrariness of detection scores.

GAZE enables medical VLMs to iteratively analyze images using viewer tools and retrieve literature from PubMed/Open-i. On the NOVA benchmark (906 rare brain MRI cases), it achieves 58.2 [email protected] for lesion localization and 34.9% Top-1 diagnostic accuracy, with disproportionate benefits for rare pathologies. Ablations reveal a model-dependent trade-off between diagnosis and localization.

FastSinkhorn is a lightweight native CUDA implementation of the log-domain Sinkhorn algorithm for entropic regularized optimal transport. It achieves 12x speedup over POT and 5.9x over PyTorch on dense OT problems with n=m=8192, while using only 256 MB of GPU memory.

A paper derives and benchmarks four classical ODE solvers (Euler, Midpoint, RK4, Dormand-Prince) for Flow Matching generative models. It finds RK4 at 80 function evaluations matches Euler at 200, and observes stiffness near t=1 and greater solver gap for undertrained models.

This study benchmarks six classical and Bayesian sparse regression methods under realistic conditions of correlated features, weak signals, and growing dimensionality. Over 2,600 experiments show Bayesian methods excel in prediction error (MSE 72 vs 108-267) and coverage, while Lasso matches Spike-and-Slab in variable selection (F1 ~0.47).

This matters because reduces group selection from exponential to polynomial time with complexity O(d^2M^2 + d^3).