Despite remarkable advances in large Vision-Language Models (VLMs), spatial reasoning remains a persistent challenge. In this work, we investigate how attention heads within VLMs contribute to spatial reasoning by analyzing their functional roles through a mechanistic interpretability lens. We introduce CogVSR, a dataset that decomposes complex spatial reasoning questions into step-by-step subquestions designed to simulate human-like reasoning via a chain-of-thought paradigm, with each subquestion linked to specific cognitive functions such as spatial perception or relational reasoning. Building on CogVSR, we develop a probing framework to identify and characterize attention heads specialized for these functions. Our analysis across diverse VLM families reveals that these functional heads are universally sparse, vary in number and distribution across functions. Notably, spatially specialized heads are fewer than those for other cognitive functions, highlighting their scarcity. We propose methods to activate latent spatial heads, improving spatial understanding. Intervention experiments further demonstrate their critical role in spatial reasoning: removing functional heads leads to performance degradation, while emphasizing them enhances accuracy. This study provides new interpretability driven insights into how VLMs attend to space and paves the way for enhancing complex spatial reasoning in multimodal models.

Local Intrinsic Dimensionality (LID) is a measure of data complexity in the vicinity of a query point. In this work, we propose a novel Bayesian framework for LID estimation that improves robustness and accuracy, especially in scenarios with small neighborhood sizes (k <= 10), where maintaining locality is critical. Our framework allows the incorporation of both informative and non-informative priors, enabling the integration of prior knowledge to enhance the estimation process. Using this framework, we derive new LID estimators and provide insights into transitional ones. Furthermore, we propose aggregation methods using linear and logarithmic pooling to combine multiple LID posteriors. These methods allow for principled integration of LID estimates across different training states of Deep Neural Networks (DNNs), such as epochs, thereby improving estimation stability and performance. We also derive a posterior predictive distribution (PPD) for modeling and synthesizing nearest-neighbor distances. Experiments demonstrate that the LID Bayesian estimation methods achieves a balanced approach between reducing variance and lowering squared bias, often leading to lower mean squared error (MSE) values for small number of nearest neighbors.

Muon orthogonalizes the momentum buffer before each update, replacing its singular values with ones via Newton-Schulz iterations. This simple change lets Muon tolerate far larger learning rates and converge faster than other optimizers, but why? We show that the mechanism is spectral flattening, and develop two results around it. First, we prove that Muon's maximal stable step size scales with the average singular value of the gradient rather than the largest, which bottlenecks standard gradient descent. Second, we recast Muon as a preconditioned gradient method and show, under a Kronecker-factored curvature model, that it improves the effective convergence factor, with the improvement controlled by the spectrum of the gradient covariance. Extensive experiments validate both results: Muon remains stable at learning rates that cause SGD to diverge within the first few iterations, and reaches accuracy milestones several epochs earlier even at identical step sizes. Taken together, our results offer a principled, geometric explanation for Muon's empirical success.

Embodied Artificial Intelligence (Embodied AI) integrates perception, cognition, planning, and interaction into agents that operate in open-world, safety-critical environments. As these systems gain autonomy and enter domains such as transportation, healthcare, and industrial or assistive robotics, ensuring their safety becomes both technically challenging and socially indispensable. Unlike digital AI systems, embodied agents must act under uncertain sensing, incomplete knowledge, and dynamic human-robot interactions, where failures can directly lead to physical harm. This survey provides a comprehensive and structured review of safety research in embodied AI, examining attacks and defenses across the full embodied pipeline, from perception and cognition to planning, action and interaction, and agentic system. We introduce a multi-level taxonomy that unifies fragmented lines of work and connects embodied-specific safety findings with broader advances in vision, language, and multimodal foundation models. Our review synthesizes insights from over 500 papers spanning adversarial, backdoor, jailbreak, and hardware-level attacks; attack detection, safe training and robust inference; and risk-aware human-agent interaction. This analysis reveals several overlooked challenges, including the fragility of multimodal perception fusion, the instability of planning under jailbreak attacks, and the trustworthiness of human-agent interaction in open-ended scenarios. By organizing the field into a coherent framework and identifying critical research gaps, this survey provides a roadmap for building embodied agents that are not only capable and autonomous but also safe, robust, and reliable in real-world deployment.

Self-supervised speech models (S3Ms) achieve strong downstream performance, yet their learned representations remain poorly understood under natural and adversarial perturbations. Prior studies rely on representation similarity or global dimensionality, offering limited visibility into local geometric changes. We ask: how do perturbations deform local geometry, and do these shifts track downstream automatic speech recognition (ASR) degradation? To address this, we present GRIDS, a framework using Local Intrinsic Dimensionality (LID) across layer-wise representations in WavLM and wav2vec 2.0. We find that LID increases for all low signal-to noise ratio (SNR) perturbations and diverges at high SNR: benign noise converges toward the clean profile, while adversarial inputs retain early-layer LID elevation. We show LID elevation co-occurs with increased WER, and that layer-wise LID features enable anomaly detection (AUROC 0.78-1.00), opening the door to transcript-free monitoring in S3Ms.

While prior work has explored emotion control in hybrid text-to-speech systems, the geometric properties of these modules, and their implications for steerability, remain poorly understood. We present the first comparative study of speech language model (SLM) and conditional flow-matching (CFM) modules as activation steering sites for mixed emotion speech synthesis. We first characterize emotion representations using linear probing and local intrinsic dimensionality (LID), and then evaluate single-site and joint steering for mixed-emotion synthesis. Our results show that SLM offers a clean, low-dimensional emotion-specific subspace with strong speaker–emotion disentanglement, while CFM exhibitspoor cross-speaker generalization due to speaker–emotion entanglement. Joint steering increases emotion intensity but degrades proportional control and speech quality on in-distribution data. These findings provide practical guidance for multi-site activation steering in hybrid TTS systems and highlight the importance of representation geometry in controllable speech generation.

Vision-language pre-trained models (VLPs) are widely used in real-world applications. However, they remain vulnerable to adversarial attacks. Although adversarial detection methods have demonstrated success in single-modality settings (either vision or language), their effectiveness and reliability in multimodal models such as VLPs remain largely unexplored. In this work, we study the geometry of VLP embedding spaces and observe structured anisotropy that differs from unimodal vision models. Our theoretical analysis shows that under this anisotropic structure, adversarial attacks increase the expected geometric separation between clean and adversarial examples (AEs). Specifically, we demonstrate that AEs consistently exhibit greater expected distances to randomly sampled points than their clean counterparts, indicating that AEs tend to push representations out of manifold regions. Building on these insights, we propose GeoDetect, which leverages these off-manifold deviations via geometric scores to identify AEs. Through comprehensive evaluations, we show that our approach reliably detects AEs across diverse VLP architectures and threat settings, covering unimodal and multimodal attacks as well as adaptive attacks, thereby providing a robust and practical approach to improving the safety and reliability of these models.

Zero-shot skeleton-based action recognition (ZS-SAR) is fundamentally constrained by prevailing approaches that rely on aligning skeleton features with static, class-level semantics. This coarse-grained alignment fails to bridge the domain shift between seen, unseen classes, thereby impeding the effective transfer of fine-grained visual knowledge. To address these limitations, we introduce DynaPURLS, a unified framework that establishes robust, multi-scale visual-semantic correspondences, dynamically refines them at inference time to enhance generalization. Our framework leverages a large language model to generate hierarchical textual descriptions that encompass both global movements, local body-part dynamics. Concurrently, an adaptive partitioning module produces fine-grained visual representations by semantically grouping skeleton joints. To fortify this fine-grained alignment against the train-test domain shift, DynaPURLS incorporates a dynamic refinement module. During inference, this module adapts textual features to the incoming visual stream via a lightweight learnable projection. This refinement process is stabilized by a confidence-aware, class-balanced memory bank, which mitigates error propagation from noisy pseudo-labels. Extensive experiments on three large-scale benchmark datasets, including NTU RGB+D 60/120, PKU-MMD, demonstrate that DynaPURLS significantly outperforms prior art, setting new state-of-the-art records.

Emotional expression in human speech is nuanced and compositional, often involving multiple, sometimes conflicting, affective cues that may diverge from linguistic content. In contrast, most expressive text-to-speech (TTS) systems enforce a single utterance-level emotion, collapsing affective diversity and suppressing mixed or text–emotion–misaligned expression. While activation steering via latent direction vectors offers a promising solution, it remains unclear whether emotion representations are linearly steerable in TTS, where steering should be applied within hybrid TTS architectures, and how such complex emotion behaviors should be evaluated. This paper presents the first systematic analysis of activation steering for emotional control in hybrid TTS models, introducing a quantitative, controllable steering framework, and multi-rater evaluation protocols that enable composable mixed-emotion synthesis and reliable text–emotion mismatch synthesis. Our results demonstrate, for the first time, that emotional prosody and expressive variability are primarily synthesized by the TTS language module instead of the flow-matching module, and also provide a lightweight steering approach for generating natural, human-like emotional speech.

The size of sea ice floes in the marginal ice zone (MIZ) is a key factor influencing ice coverage, albedo, wave propagation, and ocean-atmosphere energy exchanges. Floe size can be observed by processing visual-range imagery from ships, aircraft, or satellites. However, autonomously capturing floe boundaries remains challenging, particularly due to sea ice heterogeneity, which impairs boundary definition and reduces image clarity. This study evaluates the accuracy of sea ice floe segmentation using the gradient vector flow (GVF) active contour method, the deep learning-based Segment Anything Model (SAM), and a hybrid approach combining GVF and SAM. Methods are evaluated on a representative subset of a large data set of close-range, high-resolution imagery collected from cameras aboard an icebreaker during an Antarctic winter expedition. Spanning a wide range of ice conditions and image clarity in the MIZ, the subset provides a rigorous segmentation test bed. Performance is assessed in terms of floe detection accuracy, size distribution, and ice concentration, with results compared against a manually segmented benchmark. Results indicate SAM, in prompt-driven mode, offers the best balance between accuracy and computational efficiency. Its strong performance in estimating sea ice concentration and detecting floes, while maintaining close agreement with benchmark floe size distributions, makes it suitable for real-time applications and scalable analyses of large imagery data sets. Compared with SAM, the combined SAM-GVF method provides more accurate floe boundary delineation, although at much higher computational cost, and is therefore better suited for analyses requiring precise floe shapes.

Local Intrinsic Dimensionality (LID) has shown strong potential for anomaly detection in high-dimensional data, including landslide failure detection in granular media, where early and accurate identification of failure zones is crucial for effective geohazard mitigation. However, this task is still challenging due to the spatial correlations and temporal dynamics that are inherently present in surface displacement data. To address this gap, we propose a novel unsupervised framework called spatiotemporal LID (st-LID) that generalizes the LID for robust failure detection in landslide monitoring networks. Our approach introduces three key innovations: (1) Kinematic enhancement, incorporating velocity into the LID computation to capture instantaneous deformation rates and short-term temporal dynamics; (2) Bayesian spatial fusion, which aggregates LID values across spatial neighborhoods via Bayesian estimation, to embed spatial correlations and account for localized noise; and (3) Temporal modeling (t-LID), a new variant that characterizes long-term dynamics of displacement data, providing a robust temporal representation of displacement behavior. By unifying these components, st-LID identifies complex, multi-stage failure zones often overlooked by existing methods. Extensive experiments show that st-LID consistently outperforms state-of-the-art unsupervised baselines in detection precision and lead-time, providing a robust foundation for landslide early warning systems and targeted risk intervention to enhance community resilience and preparedness strategies.

Humans process rich auditory environments through tightly integrated cognitive capabilities such as audio perception, audio reasoning, and memory. Despite recent progress in large audio-language models (LALMs) across speech understanding and multimodal audio reasoning, current evaluation paradigms remain largely task- or modality-centric, focusing on end performance while overlooking underlying auditory cognitive behaviours. This reveals a fundamental gap between how auditory cognition is understood in humans and how it is evaluated in LALMs, particularly in the lack of frameworks that operationalise cognitive principles beyond task-level metrics to systematically capture model behaviour. In this work, we introduce RAIL, a human-centric evaluation paradigm grounded in the Cattell-Horn-Carroll (CHC) cognitive framework. RAIL formalises auditory cognition into five core capabilities and develop them into structured evaluation tasks that probe how models process, retain, and integrate auditory information. We further construct a cognitively grounded benchmark with principled data curation and human-aligned evaluation protocols. Evaluating 26 state-of-the-art LALMs, we find that current models exhibit highly uneven performance across cognitive abilities. RAIL establishes a new evaluation paradigm that moves beyond task-centric benchmarking toward cognitively grounded assessment of auditory intelligence.

The theory of Local Intrinsic Dimensionality (LID) has become a valuable tool for characterizing local complexity within and across data manifolds, supporting a range of data mining and machine learning tasks. Accurate LID estimation requires samples drawn from small neighborhoods around each query to avoid biases from nonlocal effects and potential manifold mixing, yet limited data within such neighborhoods tends to cause high estimation variance. As a variance reduction strategy, we propose an ensemble approach that uses subbagging to preserve the local distribution of nearest neighbor (NN) distances. The main challenge is that the uniform reduction in total sample size within each subsample increases the proximity threshold for finding a fixed number k of NNs around the query. As a result, in the specific context of LID estimation, the sampling rate has an additional, complex interplay with the neighborhood size, where both combined determine the sample size as well as the locality and resolution considered for estimation. We analyze both theoretically and experimentally how the choice of the sampling rate and the k-NN size used for LID estimation, alongside the ensemble size, affects performance, enabling informed prior selection of these hyper-parameters depending on application-based preferences. Our results indicate that within broad and well-characterized regions of the hyper-parameters space, using a bagged estimator will most often significantly reduce variance as well as the mean squared error when compared to the corresponding non-bagged baseline, with controllable impact on bias. We additionally propose and evaluate different ways of combining bagging with neighborhood smoothing for substantial further improvements on LID estimation performance.

Modern speech systems increasingly use discretized self-supervised speech representations for compression and integration with token-based models, yet their impact on emotional information remains unclear. We study how residual vector quantization (RVQ) reshapes emotional information in discrete speech representations from both representation- and task-level perspectives. Our analysis shows that aggressive compression disproportionately degrades emotion, with uneven loss across emotion classes and model architectures. To address this, we introduce emotion-aware quantization using emotion-specific and emotion-biased codebooks, improving the preservation of both hard and soft emotion perception. We further propose Emo-Q, a lightweight routed quantization method that selects emotion-specialized codebooks, improving emotion recognition performance at lower bitrates. These results highlight the importance of emotion-aware discretization for robust affective speech processing.

Developing open-set classification methods capable of classifying in-distribution (ID) data while detecting out-of-distribution (OOD) samples is essential for deploying graph neural networks (GNNs) in open-world scenarios. Existing methods typically treat all OOD samples as a single class, despite real-world applications—especially high-stake settings like fraud detection and medical diagnosis—demanding deeper insights into OOD samples, including their probable labels. This raises a critical question: Can OOD detection be extended to OOD classification without true label information? To answer this question, we introduce a Coarse-to-Fine open-set Classification (CFC) method that leverages large language models (LLMs) for text-attributed graphs. CFC consists of three key components: (1) A coarse classifier that utilizes LLM prompts for OOD detection and outlier label generation; (2) A GNN-based fine classifier trained with OOD samples from (1) for enhanced OOD detection and ID classification; and (3) Refined OOD classification achieved through LLM prompts and post-processed OOD labels. Unlike methods relying on synthetic or auxiliary OOD samples, CFC employs semantic OOD data-instances that are genuinely out-of-distribution based on their inherent meaning, thus improving interpretability and practical utility. CFC enhances OOD detection by 10% compared to state-of-the-art approaches on text-attributed graphs and in the text domain, while achieving up to 70% accuracy in OOD classification on graph datasets.

Deep learning has been a critical part of designing inverse design methods that are computationally efficient and accurate. An example of this is the design of photonic metasurfaces by using their photoluminescent spectrum as the input data to predict their topology. One fundamental challenge of these systems is their ability to represent nonlinear relationships between sets of data that have different dimensionalities. Existing design methods often implement a conditional generative adversarial network in order to solve this problem, but in many cases, the solution is unable to generate structures that provide multiple peaks when validated. It is demonstrated that in response to the target spectrum, the bidirectional adversarial autoencoder is able to generate structures that provide multiple peaks on several occasions. As a result, the proposed model represents an important advance toward the generation of nonlinear photonic metasurfaces that can be used in advanced metasurface design. (c) 2025 Author(s). All article content, except where otherwise noted, is licensed under a Creative Commons Attribution-NonCommercial-NoDerivs 4.0 International (CC BY-NC-ND) license (https://creativecommons.org/licenses/by-nc-nd/4.0/), https://doi.org/10.1063/50284338

Modern data analysis platforms, such as Tableau, Microsoft Power BI, Google Looker Studio, Kibana, and Splunk, have democratized data exploration by enabling users to interact with data through intuitive visual interfaces, eliminating the need for proficiency in query languages like SQL. These platforms allow both experts and non-experts to perform high-level operations and incrementally construct complex analysis workflows. As the volume and complexity of data grow, assisting users in navigating these workflows becomes increasingly important. One promising direction is to provide intelligent next-action recommendations that guide users through meaningful and efficient exploration paths. In this paper, we present ExplorAct, a context-aware next-action recommendation framework that leverages historical session logs to predict and suggest relevant next steps during data exploration. Unlike existing approaches that suffer from scalability issues due to log-size-dependent retrieval, ExplorAct achieves constant-time inference by employing a deep learning architecture that models both the structural and sequential aspects of exploration sessions. Through extensive experiments on four real-world datasets, we show that ExplorAct consistently outperforms state-of-the-art (SOTA) baselines across three core recommendation tasks, while maintaining stable and low-latency inference regardless of log size.

Contrastive language-image pretraining (CLIP) has been found to be vulnerable to poisoning backdoor attacks where the adversary can achieve an almost perfect attack success rate on CLIP models by poisoning only 0.01\% of the training dataset. This raises security concerns on the current practice of pretraining large-scale models on unscrutinized web data using CLIP. In this work, we analyze the representations of backdoor-poisoned samples learned by CLIP models and find that they exhibit unique characteristics in their local subspace, i.e., their local neighborhoods are far more sparse than that of clean samples. Based on this finding, we conduct a systematic study on detecting CLIP backdoor attacks and show that these attacks can be easily and efficiently detected by traditional density ratio-based local outlier detectors, whereas existing backdoor sample detection methods fail. Our experiments also reveal that an unintentional backdoor already exists in the original CC3M dataset and has been trained into a popular open-source model released by OpenCLIP. Based on our detector, one can clean up a million-scale web dataset (e.g., CC3M) efficiently within 15 minutes using 4 Nvidia A100 GPUs.

Large language models (LLMs) have achieved state-of-the-art performance in a variety of tasks, but remain largely opaque in terms of their internal mechanisms. Understanding these mechanisms is crucial to improve their reasoning abilities. Drawing inspiration from the interplay between neural processes and human cognition, we propose a novel interpretability framework to systematically analyze the roles and behaviors of attention heads, which are key components of LLMs. We introduce CogQA, a dataset that decomposes complex questions into step-by-step subquestions with a chain-of-thought design, each associated with specific cognitive functions such as retrieval or logical reasoning. By applying a multi-label probing method, we identify the attention heads responsible for these functions. Our analysis across multiple LLM families reveals that attention heads exhibit functional specialization, characterized as cognitive heads. These cognitive heads exhibit several key properties: they are universally sparse, and vary in number and distribution across different cognitive functions, and they display interactive and hierarchical structures. We further show that cognitive heads play a vital role in reasoning tasks—removing them leads to performance degradation, while augmenting them enhances reasoning accuracy. These insights offer a deeper understanding of LLM reasoning and suggest important implications for model design, training and fine-tuning strategies.

Recent successes in diffusion probabilistic models have demonstrated their strength in modeling and generating different types of data, paving the way for their application in generative time series forecasting. However, most existing diffusion based approaches rely on sequential models and unimodal latent variables to capture global dependencies and model entire observable data, resulting in difficulties when it comes to highly stochastic time series data. In this paper, we propose a novel Stochastic Diffusion (StochDiff) model that integrates the diffusion process into time series modeling stage and utilizes the representational power of the stochastic latent spaces to capture the variability of the stochastic time series data. Specifically, the model applies diffusion module at each time step within the sequential framework and learns a step-wise, data-driven prior for generative diffusion process. These features enable the model to effectively capture complex temporal dynamics and the multi-modal nature of the highly stochastic time series data. Through extensive experiments on real-world datasets, we demonstrate the effectiveness of our proposed model for probabilistic time series forecasting, particularly in scenarios with high stochasticity. Additionally, with a real-world surgical use case, we highlight the model's potential in a medical application.