This work addresses the challenge of cross-modal bird species retrieval in the absence of audio-image paired data, where representation alignment is inherently difficult. The authors propose leveraging textual semantics as an intermediary by distilling the text embedding space of a pretrained image-text model (BioCLIP-2) into an audio-text model (BioLingual), followed by contrastive learning-based fine-tuning of the audio encoder. This approach achieves implicit alignment between audio and image embeddings without any direct audio-image supervision. By circumventing the conventional reliance on paired multimodal data, the method significantly outperforms zero-shot model ensembles and text-mapping baselines on bioacoustic benchmarks such as SSW60, while preserving audio discriminability and enhancing audio-text alignment.

Monocular video-based dynamic scene reconstruction often struggles to balance global structural consistency with fine local details due to sparse multi-view cues. This work proposes WebSpline, a novel framework that introduces structure-aware, learnable cubic Hermite splines to model dynamic 3D Gaussian trajectories and explicitly organizes motion structures via a Structure Proxy Graph (SPG). By integrating spatio-temporal neighborhood constraints, point trajectory initialization, and rigidity regularization, WebSpline employs a two-stage optimization strategy that ensures structural coherence while enabling efficient inference. Evaluated on both iPhone-captured and NVIDIA dynamic scene benchmarks, WebSpline substantially outperforms existing methods, achieving state-of-the-art rendering quality and inference speeds over ten times faster than WorldTree.

Identifying intrinsic modal coordinates for nonlinear dynamical systems remains challenging due to the absence of closed-form models and inherent nonlinearity. To address this, we propose a physics-informed, data-driven method that directly learns the nonlinear normal mode (NNM) transformation from vibration response data—enabling modal decomposition, reconstruction, and prediction without explicit governing equations. This work introduces, for the first time, physics-informed neural networks (PINNs) into NNM identification, embedding fundamental physical constraints such as energy conservation and symmetry to overcome reliance on analytical system models. The approach successfully isolates the first two NNMs of a nonlinear beam in both numerical simulations and free-vibration experiments, accurately capturing hallmark nonlinear phenomena—including amplitude-dependent frequency hardening and distorted mode shapes. Reconstruction and prediction errors remain below 5%, demonstrating high fidelity and generalizability.

Verifying behavioral compliance of AI systems in safety-critical scenarios remains challenging due to the limited expressiveness of formal methods and the high false-positive/negative rates in natural-language-constraint evaluation. To address this, we propose RepV, a neuro-symbolic verifier that constructs a safety-separable latent space—enabling linear separation of safe versus unsafe plans—and provides position-based probabilistic verification guarantees. RepV integrates model-checking–generated initial labels, a lightweight projector, a frozen linear classifier, and LLM-generated reasoning traces in a synergistic inference framework. It achieves efficient, single-forward-pass verification with <0.2M parameters and zero human annotation. Experiments across multiple tasks demonstrate that RepV improves compliance prediction accuracy by up to 15% over state-of-the-art baselines, significantly outperforming conventional fine-tuning approaches.

This work addresses the reference-free evaluation challenge in multimodal presentation generation. Methodologically, it introduces (1) the first reference-free evaluation paradigm based on negative-sample perturbation, generating discriminative signals via controlled semantic, structural, and visual perturbations; (2) RefSlides, the first high-quality, human-annotated benchmark for presentation evaluation; and (3) a decoupled multi-metric modeling framework coupled with LLM instruction tuning to enable fine-grained scoring of intrinsic attributes—such as summarization capability and concept conveyance—and generation of interpretable natural-language feedback. Experiments demonstrate that our approach significantly outperforms heuristic baselines and state-of-the-art LLM-based evaluators in both automated and human evaluations: scoring consistency improves by 21.3%, and feedback usefulness increases by 34.7%.

In RGB-to-thermal infrared (IR) unsupervised domain adaptation (UDA) object detection, poor pseudo-label quality and high false-positive rates arise from the absence of color and texture cues in IR imagery. To address this, we propose Semantic-Aware Grayscale Augmentation (SAGA): it extracts object-level IR-specific semantic features and employs semantic-segmentation-guided grayscale mapping to mitigate inter-modal color bias. Additionally, we introduce IndraEye—the first multi-view, multi-temporal RGB-IR paired dataset tailored for airborne-ground collaborative perception. Integrated into the UDA-Faster R-CNN framework with high-precision multi-sensor calibration, SAGA achieves mAP improvements of 0.4–7.6% on both autonomous driving benchmarks and IndraEye, significantly enhancing pseudo-label reliability and cross-domain detection robustness. The code and dataset are publicly released.

This work addresses the challenge in zero-shot skeleton-based action recognition where the absence of contextual cues—such as interacting objects—hinders effective alignment between skeletal and semantic representations, leading to difficulties in distinguishing visually similar actions. To overcome this limitation, the authors propose SkeletonContext, a novel framework that introduces a language model–guided contextual prompt reconstruction mechanism. By integrating cross-modal prompt learning with disentangled motion features of key joints, the method enables fine-grained action discrimination without requiring explicit object information. Extensive experiments demonstrate that SkeletonContext significantly outperforms existing approaches across multiple benchmark datasets, achieving state-of-the-art performance under both conventional and generalized zero-shot settings, thereby enhancing both semantic alignment accuracy and action recognition robustness.

To address the insufficient perception, comprehension, and decision-making capabilities of vision-language models (VLMs) in safety-critical autonomous driving scenarios, this paper proposes SafeDriveRAG: a knowledge-augmented retrieval-augmented generation framework. It introduces SafeDrive228K—the first large-scale multimodal traffic safety QA benchmark covering 18 traffic-related question categories—and constructs a domain-specific knowledge graph integrating traffic regulations and safety commonsense. A multi-scale subgraph retrieval algorithm is designed to enable fine-grained, context-aware knowledge grounding. Evaluated on five mainstream VLMs, SafeDriveRAG achieves average improvements of +4.73% on accident analysis, +8.79% on edge-case identification, and +14.57% on safety commonsense reasoning. This work establishes the first systematic multimodal safety evaluation benchmark for autonomous driving and pioneers a knowledge-driven RAG paradigm for traffic safety understanding, thereby bridging a critical research gap in safety-enhanced VLMs.

This paper challenges the conventional wisdom that bandwidth is negligible and latency dominates web performance, systematically investigating the actual impact of mobile network bandwidth and latency on modern complex web pages. Method: We propose a novel metric—Critical Path of Improvement (CPI)—to quantify websites’ sensitivity distributions to bandwidth versus latency. Leveraging real-world network measurements across four operators and 57 cities, we construct an empirically grounded CPI model and perform critical-path decomposition and correlation analysis. Contribution/Results: Our findings reveal that 18% of websites are predominantly bandwidth-limited, while over half exhibit significant bandwidth sensitivity. The study uncovers the underappreciated role of bandwidth in mobile web performance and delivers a practical, deployable framework for diagnosing bandwidth sensitivity—providing empirical guidance for developers optimizing resource loading strategies and operators designing differentiated QoS policies.

This work addresses the limitation of existing reinforcement learning approaches for large language models, which apply uniform optimization across all tokens and struggle to dynamically balance exploration and exploitation. To overcome this, the authors propose HTPO, a novel algorithm that introduces a hierarchical token-level objective control mechanism for the first time. HTPO categorizes response tokens into distinct groups based on prompt difficulty, answer correctness, and token entropy, and assigns each group a tailored optimization objective to enable fine-grained exploration-exploitation trade-offs. Integrated with a verifiable reward framework, HTPO outperforms the DAPO baseline by 8.6% and 6.7% on the challenging AIME'24 and AIME'25 reasoning benchmarks, respectively, with performance gains further amplified as test-time compute resources increase.