User behavior on online platforms is evolving, reflecting real-world changes in how people post, whether it's helpful messages or hate speech. Models that learn to capture this content can experience a decrease in performance over time due to data drift, which can lead to ineffective behavioral analytics systems. However, fine-tuning such a model over time with new data can be detrimental due to catastrophic forgetting. Replay-based approaches in continual learning offer a simple yet efficient method to update such models, minimizing forgetting by maintaining a buffer of important training instances from past learned tasks. However, the main limitation of this approach is the fixed size of the buffer. External knowledge bases can be utilized to overcome this limitation through data augmentation. We propose a novel augmentation-based approach to incorporate external knowledge in the replay-based continual learning framework. We evaluate several strategies with three datasets from prior studies related to deviant behavior classification to assess the integration of external knowledge in continual learning and demonstrate that augmentation helps outperform baseline replay-based approaches.

Large language models remain vulnerable to jailbreak attacks that bypass safety guardrails to elicit harmful outputs. Defending against novel jailbreaks represents a critical challenge in AI safety. Adversarial training -- designed to make models robust against worst-case perturbations -- has been the dominant paradigm for adversarial robustness. However, due to optimization challenges and difficulties in defining realistic threat models, adversarial training methods often fail on newly developed jailbreaks in practice. This paper proposes a new paradigm for improving robustness against unseen jailbreaks, centered on the Adversarial Déjà Vu hypothesis: novel jailbreaks are not fundamentally new, but largely recombinations of adversarial skills from previous attacks. We study this hypothesis through a large-scale analysis of 32 attack papers published over two years. Using an automated pipeline, we extract and compress adversarial skills into a sparse dictionary of primitives, with LLMs generating human-readable descriptions. Our analysis reveals that unseen attacks can be effectively explained as sparse compositions of earlier skills, with explanatory power increasing monotonically as skill coverage grows. Guided by this insight, we introduce Adversarial Skill Compositional Training (ASCoT), which trains on diverse compositions of skill primitives rather than isolated attack instances. ASCoT substantially improves robustness to unseen attacks, including multi-turn jailbreaks, while maintaining low over-refusal rates. We also demonstrate that expanding adversarial skill coverage, not just data scale, is key to defending against novel attacks. \textcolor{red}{\textbf{Warning: This paper contains content that may be harmful or offensive in nature.

Diffusion Models have become a cornerstone of modern generative AI for their exceptional generation quality and controllability. However, their inherent \textit{multi-step iterations} and \textit{complex backbone networks} lead to prohibitive computational overhead and generation latency, forming a major bottleneck for real-time applications. Although existing acceleration techniques have made progress, they still face challenges such as limited applicability, high training costs, or quality degradation. Against this backdrop, \textbf{Diffusion Caching} offers a promising training-free, architecture-agnostic, and efficient inference paradigm. Its core mechanism identifies and reuses intrinsic computational redundancies in the diffusion process. By enabling feature-level cross-step reuse and inter-layer scheduling, it reduces computation without modifying model parameters. This paper systematically reviews the theoretical foundations and evolution of Diffusion Caching and proposes a unified framework for its classification and analysis. Through comparative analysis of representative methods, we show that Diffusion Caching evolves from \textit{static reuse} to \textit{dynamic prediction}. This trend enhances caching flexibility across diverse tasks and enables integration with other acceleration techniques such as sampling optimization and model distillation, paving the way for a unified, efficient inference framework for future multimodal and interactive applications. We argue that this paradigm will become a key enabler of real-time and efficient generative AI, injecting new vitality into both theory and practice of \textit{Efficient Generative Intelligence}.

Deep neural networks (DNNs) have revolutionized machine learning, achieving unprecedented success in tasks such as image classification [12], natural language processing [20], and reinforcement learning [17]. However, their overparameterized nature often leads to overfitting, where models excel on training data but fail to generalize to unseen samples [23]. This generalization gap defined as the difference between training and test error poses a critical challenge, particularly in high-stakes applications like medical diagnostics or autonomous systems [7]. Regularization techniques, such as dropout [18] and weight decay, are widely used to mitigate overfitting, but their static nature limits adaptability to the complex, non-convex loss landscapes of DNNs. Recent advances in adaptive regularization [14, 5] show promise by dynamically adjusting parameters during training, yet these methods often lack rigorous theoretical guarantees to quantify their impact on generalization.

The increase in computing power and the necessity of AI-assisted decision-making boost the growing application of large language models (LLMs). Along with this, the potential retention of sensitive data of LLMs has spurred increasing research into machine unlearning. However, existing unlearning approaches face a critical dilemma: Aggressive unlearning compromises model utility, while conservative strategies preserve utility but risk hallucinated responses. This significantly limits LLMs' reliability in knowledge-intensive applications. To address this, we introduce a novel Attention-Shifting (AS) framework for selective unlearning. AS is driven by two design objectives: (1) context-preserving suppression that attenuates attention to fact-bearing tokens without disrupting LLMs' linguistic structure; and (2) hallucination-resistant response shaping that discourages fabricated completions when queried about unlearning content. AS realizes these objectives through two attention-level interventions, which are importance-aware suppression applied to the unlearning set to reduce reliance on memorized knowledge and attention-guided retention enhancement that reinforces attention toward semantically essential tokens in the retained dataset to mitigate unintended degradation. These two components are jointly optimized via a dual-loss objective, which forms a soft boundary that localizes unlearning while preserving unrelated knowledge under representation superposition. Experimental results show that AS improves performance preservation over the state-of-the-art unlearning methods, achieving up to 15% higher accuracy on the ToFU benchmark and 10% on the TDEC benchmark, while maintaining competitive hallucination-free unlearning effectiveness. Compared to existing methods, AS demonstrates a superior balance between unlearning effectiveness, generalization, and response reliability.

Financial anomalies exhibit heterogeneous mechanisms (price shocks, liquidity freezes, contagion cascades, regime shifts), but existing detectors treat all anomalies uniformly, producing scalar scores without revealing which mechanism is failing, where risks concentrate, or how to intervene. This opacity prevents targeted regulatory responses. Three unsolved challenges persist: (1) static graph structures cannot adapt when market correlations shift during regime changes; (2) uniform detection mechanisms miss type-specific signatures across multiple temporal scales while failing to integrate individual behaviors with network contagion; (3) black-box outputs provide no actionable guidance on anomaly mechanisms or their temporal evolution. We address these via adaptive graph learning with specialized expert networks that provide built-in interpretability. Our framework captures multi-scale temporal dependencies through BiLSTM with self-attention, fuses temporal and spatial information via cross-modal attention, learns dynamic graphs through neural multi-source interpolation, adaptively balances learned dynamics with structural priors via stress-modulated fusion, routes anomalies to four mechanism-specific experts, and produces dual-level interpretable attributions. Critically, interpretability is embedded architecturally rather than applied post-hoc. On 100 US equities (2017-2024), we achieve 92.3% detection of 13 major events with 3.8-day lead time, outperforming best baseline by 30.8pp. Silicon Valley Bank case study demonstrates anomaly evolution tracking: Price-Shock expert weight rose to 0.39 (33% above baseline 0.29) during closure, peaking at 0.48 (66% above baseline) one week later, revealing automatic temporal mechanism identification without labeled supervision.

We propose a framework for automated policy evaluation using Cosmos-Surg-dVRK, a Cosmos world foundation model (WFM) finetune, to perform simulated surgical policy rollouts and subsequent automated success rate evaluation using a video classifier. After rollout completion, the generated video is automatically labeled for task success or failure using a trained video classifier, enabling objective selection of the most promising surgical policies for real-robot evaluation and deployment. The rise of surgical robots and vision-language-action models has accelerated the development of autonomous surgical policies and efficient assessment strategies. However, evaluating these policies directly on physical robotic platforms such as the da Vinci Research Kit (dVRK) remains hindered by high costs, time demands, reproducibility challenges, and variability in execution. World foundation models (WFM) for physical AI offer a transformative approach to simulate complex real-world surgical tasks, such as soft tissue deformation, with high fidelity. This work introduces Cosmos-Surg-dVRK, a surgical finetune of the Cosmos WFM, which, together with a trained video classifier, enables fully automated online evaluation and benchmarking of surgical policies. On tabletop suture pad tasks, the automated pipeline achieves strong correlation between online rollouts in Cosmos-Surg-dVRK and policy outcomes on the real dVRK Si platform, as well as good agreement between human labelers and the V -JEP A 2-derived video classifier. Additionally, preliminary experiments with ex-vivo porcine cholecystectomy tasks in Cosmos-Surg-dVRK demonstrate promising alignment with real-world evaluations, highlighting the platform's potential for more complex surgical procedures. World models have emerged as a foundational approach for enabling intelligent agents to understand and act in complex simulated environments. Building on early work by Ha & Schmidhuber (2018), they learn compact latent representations of spatial and temporal dynamics and predict the consequences of actions in context. Leveraging diffusion processes, recent efforts have introduced large-scale multi-modal world foundation models (WFMs) (Wan et al., 2025; Agarwal et al., 2025; Kong et al., 2024; Xie et al., 2025; Ball et al., 2025). These video generative models serve as scalable, general-purpose learned simulators that encode and synthesize diverse physical phenomena, approximate scene dynamics, render plausible sensory observations, and facilitate policy evaluation and training, enabling progress on sim-to-real transfer. At the intersection of these advances lies the burgeoning field of physical AI, AI systems equipped with sensors and actuators that enable perception, reasoning, and actuation in the physical world.

Adversarial examples reveal critical vulnerabilities in deep neural networks by exploiting their sensitivity to imperceptible input perturbations. While adversarial training remains the predominant defense strategy, it often incurs significant computational cost and may compromise clean-data accuracy. In this work, we investigate an architectural approach to adversarial robustness by embedding group-equivariant convolutions-specifically, rotation- and scale-equivariant layers-into standard convolutional neural networks (CNNs). These layers encode symmetry priors that align model behavior with structured transformations in the input space, promoting smoother decision boundaries and greater resilience to adversarial attacks. We propose and evaluate two symmetry-aware architectures: a parallel design that processes standard and equivariant features independently before fusion, and a cascaded design that applies equivariant operations sequentially. Theoretically, we demonstrate that such models reduce hypothesis space complexity, regularize gradients, and yield tighter certified robustness bounds under the CLEVER (Cross Lipschitz Extreme Value for nEtwork Robustness) framework. Empirically, our models consistently improve adversarial robustness and generalization across CIFAR-10, CIFAR-100, and CIFAR-10C under both FGSM and PGD attacks, without requiring adversarial training. These findings underscore the potential of symmetry-enforcing architectures as efficient and principled alternatives to data augmentation-based defenses.

This paper introduces SpikeFit, a novel training method for Spiking Neural Networks (SNNs) that enables efficient inference on neuromorphic hardware, considering all its stringent requirements: the number of neurons and synapses that can fit on a single device, and lower bit-width representations (e.g., 4-bit, 8-bit). Unlike conventional compressing approaches that address only a subset of these requirements (limited numerical precision and limited number of neurons in the network), SpikeFit treats the allowed weights' discrete values themselves as learnable parameters co-optimized with the model, allowing for optimal Clusterization-Aware Training (CAT) of the model's weights at low precision (2-, 4-, or 8-bit) which results in higher network compression efficiency, as well as limiting the number of unique synaptic connections to a value required by neuromorphic processor. This joint optimization allows SpikeFit to find a discrete weight set aligned with hardware constraints, enabling the most complete deployment across a broader range of neuromorphic processors than existing methods of SNN compression support. Moreover, SpikeFit introduces a new hardware-friendly Fisher Spike Contribution (FSC) pruning method showing the state-of-the-art performance. We demonstrate that for spiking neural networks constrained to only four unique synaptic weight values (M = 4), our SpikeFit method not only outperforms state-of-the-art SNNs compression methods and conventional baselines combining extreme quantization schemes and clustering algorithms, but also meets a wider range of neuromorphic hardware requirements and provides the lowest energy use in experiments.

In the context of imitation learning, visuomotor-based diffusion policy learning is one of the main directions in robotic manipulation. Most of these approaches rely on point clouds as observation inputs and construct scene representations through point clouds feature learning, which enables them to achieve remarkable accuracy. However, the existing literature lacks an in-depth exploration of vision-only solutions that have significant potential. In this paper, we propose a Vision-Only and single-view Diffusion Policy learning method (VO-DP) that leverages pretrained visual foundation models to achieve effective fusion of semantic and geometric features. We utilize intermediate features from VGGT incorporating semantic features from DINOv2 and geometric features from Alternating Attention blocks. Features are fused via cross-attention and spatially compressed with a CNN to form the input to the policy head. Extensive experiments demonstrate that VO-DP not only outperforms the vision-only baseline DP significantly but also exhibits distinct performance trends against the point cloud-based method DP3: in simulation tasks, VO-DP achieves an average success rate of 64.6% on par with DP3 64.0% and far higher than DP 34.8%, while in real-world tasks, it reaches 87.9%, outperforming both DP3 67.5% and DP 11.2% by a notable margin. Further robustness evaluations confirm that VO-DP remains highly stable under varying conditions including color, size, background, and lighting. Lastly, we open-source a training library for robotic manipulation. Built on Accelerate, this library supports multi-machine and multi-GPU parallel training, as well as mixed precision training. It is compatible with visuomotor policies such as DP, DP3 and VO-DP, and also supports the RoboTwin simulator.