Industrial anomaly detection demands precise reasoning over fine-grained defect patterns. However, existing multimodal large language models (MLLMs), pretrained on general-domain data, often struggle to capture category-specific anomalies, thereby limiting both detection accuracy and interpretability. To address these limitations, we propose Reason-IAD, a knowledge-guided dynamic latent reasoning framework for explainable industrial anomaly detection. Reason-IAD comprises two core components. First, a retrieval-augmented knowledge module incorporates category-specific textual descriptions into the model input, enabling context-aware reasoning over domain-specific defects. Second, an entropy-driven latent reasoning mechanism conducts iterative exploration within a compact latent space using optimizable latent think tokens, guided by an entropy-based reward that encourages confident and stable predictions. Furthermore, a dynamic visual injection strategy selectively incorporates the most informative image patches into the latent sequence, directing the reasoning process toward regions critical for anomaly detection. Extensive experimental results demonstrate that Reason-IAD consistently outperforms state-of-the-art methods across multiple tasks. The code will be publicly available at https://github.com/chenpeng052/Reason-IAD.

Reinforcement Learning with Verified Reward (RLVR) has emerged as a critical paradigm for advancing the reasoning capabilities of Large Language Models (LLMs). Most existing RLVR methods, such as GRPO and its variants, ensure stable updates by constraining policy divergence through clipping likelihood ratios. This paper introduces a unified clipping framework that characterizes existing methods via a general notion of policy divergence, encompassing both likelihood ratios and Kullback-Leibler (KL) divergences and extending to alternative measures. The framework provides a principled foundation for systematically analyzing how different policy divergence measures affect exploration and performance. We further identify the KL3 estimator, a variance-reduced Monte Carlo estimator of the KL divergence, as a key policy divergence constraint. We theoretically demonstrate that the KL3-based constraint is mathematically equivalent to an asymmetric ratio-based clipping that reallocates probability mass toward high-confidence actions, promoting stronger exploration while retaining the simplicity of GRPO-style methods. Empirical results on mathematical reasoning benchmarks demonstrate that incorporating the KL3 estimator into GRPO improves both training stability and final performance, highlighting the importance of principled policy divergence constraints in policy optimization.

Offline–to–online deployment of reinforcement learning (RL) agents often stumbles over two fundamental gaps: (1) the sim-to-real gap, where real-world systems exhibit latency and other physical imperfections not captured in simulation; and (2) the interaction gap, where policies trained purely offline face out-of-distribution (OOD) issues during online execution, as collecting new interaction data is costly or risky. As a result, agents must generalize from static, delay-free datasets to dynamic, delay-prone environments. In this work, we propose $\textbf{DT-CORL}$ ($\textbf{D}$elay-$\textbf{T}$ransformer belief policy $\textbf{C}$onstrained $\textbf{O}$ffline $\textbf{RL}$), a novel framework for learning delay-resilient policies solely from static, delay-free offline data. DT-CORL introduces a transformer-based belief model to infer latent states from delayed observations and jointly trains this belief with a constrained policy objective, ensuring that value estimation and belief representation remain aligned throughout learning. Crucially, our method does not require access to delayed transitions during training and outperforms naive history-augmented baselines, SOTA delayed RL methods, and existing belief-based approaches. Empirically, we demonstrate that DT-CORL achieves strong delay-robust generalization across both locomotion and goal-conditioned tasks in the D4RL benchmark under varying delay regimes. Our results highlight that joint belief-policy optimization is essential for bridging the sim-to-real latency gap and achieving stable performance in delayed environments.

Accurate trajectory prediction is essential for autonomous driving systems to make safe and efficient decisions. Traditional global message-passing methods, though effective at capturing mutual interactions, suffer from an O(N-2) parameter complexity, which limits their scalability in high-density traffic environments. To address this, we propose a message-passing approach based on local neighborhoods, which reduces the complexity to O(N K-max) by restricting each node's interactions to its most relevant neighbors. On the Argoverse 1 motion forecasting benchmark, our model achieves a minADE(6) of 0.739 and a minFDE(6) of 1.133 with only 1.56M parameters, improving both metrics over a global message-passing baseline. On Argoverse 2, it attains a minFDE(6) of 1.196 and an MR6 of 12.2. These results demonstrate that local neighborhood message passing can simultaneously enhance prediction accuracy and computational efficiency, offering a scalable and practical solution for motion prediction in autonomous driving systems.

Anomaly detection in intelligent surveillance system is an important and challenging task, which commonly learns a model describing normal patterns via frame reconstruction or prediction and assumes that anomalies deviate form the learned normal model. Despite deep neural networks (DNNs) enable remarkable gains on anomaly detection, most previous works suffer from following drawbacks: (1) the powerful generalization ability of DNNs allows to reconstruct or predict anomalies well, (2) the temporal contextual information is under-explored, and (3) the diversity of normal patterns is overlooked. To address above issues, we propose a prototype-guided and dynamic-aware long-distance frame prediction paradigm for video anomaly detection. Specifically, a prototype-guided dynamics matching network (PDM-Net) is adopted to lessen the generalization ability of model to anomalies. To explore the temporal contexts, PDM-Net is equipped with a dynamic prototype matching mechanism, which first stores long-term dynamic prototypes learned from normal long sequences and then recalls the stored normal long-term prototypes with dynamics extracted from short sequences for facilitating the long-distance frames prediction of normal short sequences. Besides, a feature discrimination module is adopted to extract the representative dynamic features of various normal events meanwhile preserving the diversity of normal patterns. Experimental results on four datasets demonstrate the superiority of our method.

Visual Simultaneous Localization and Mapping (VSLAM) is the key technology for autonomous navigation of mobile robots. However, feature-based VSLAM systems still face two major challenges in dynamic complex environments: insufficient feature reliability and significant dynamic interference, urgently requiring improved matching robustness. This paper innovatively proposes a dynamic adaptive VSLAM system based on the High-repeatability and High-reliability feature matching network (2HR-Net), which improves localization accuracy in dynamic environments through three key innovations: First, the 2HR feature detection network is designed, integrating the K-Means clustering algorithm into L2-Net to achieve feature point detection with both high repeatability and high reliability. Second, the lightweight YOLOv8n model is integrated to detect and remove feature points in dynamic regions in real-time, effectively reducing the impact of dynamic interference on pose estimation. Finally, the shared matching Siamese network with a unique dual-branch feature fusion strategy and similarity optimization algorithm is proposed to enhance the accuracy of feature matching. The proposed algorithm was ultimately validated using the publicly available TUM dataset. The experimental results show that the feature detection method proposed in this paper achieved a repeatability rate of approximately 70% in various dynamic scenarios, which is significantly higher than traditional methods (such as ORB-SLAM3), whose repeatability typically falls below 40%. In addition, compared with ORB-SLAM3, the root mean square error (RMSE) and standard deviation (S.D.) of the Absolute Trajectory Error (ATE) in various dynamic scenarios were reduced by approximately 90%, indicating higher localization accuracy and stability. Therefore, the experimental results demonstrate that the proposed method outperforms mainstream methods such as ORB-SLAM3 in terms of feature repeatability, matching accuracy, and localization precision, providing an effective solution for robust VSLAM in dynamic environments.

The semantic segmentation is a critical task in LiDAR point cloud processing. Leveraging temporal information to provide contextual data for regions with low visibility or sparse observations has recently become a popular research direction, especially in autonomous driving. Existing methods, however, are often over-reliant on past frames, leading to cumulative errors (drift) caused by unconstrained frame-by-frame stacking. This paper proposes a dynamic alignment of historical frame memory information to ensure consistency with the observations of the current frame, reducing deviations caused by viewpoint changes or object movements and ensuring more accurate capture of current frame features. In addition, a new multi-scale feature fusion method, to the best of our knowledge, was introduced using the spatiotemporal (ST) method to extract the ST features, which reduces the inconsistencies between 2D range image coordinates and 3D Cartesian outputs. This approach enhances feature representation by optimizing and fusing the aligned channel features. This method was evaluated on the SemanticKITTI and SensatUrban datasets. The experimental results showed that it outperforms existing state-of-the-art methods regarding accuracy.

Offline-to-online deployment of reinforcement-learning (RL) agents must bridge two gaps: (1) the sim-to-real gap, where real systems add latency and other imperfections not present in simulation, and (2) the interaction gap, where policies trained purely offline face out-of-distribution states during online execution because gathering new interaction data is costly or risky. Agents therefore have to generalize from static, delay-free datasets to dynamic, delay-prone environments. Standard offline RL learns from delay-free logs yet must act under delays that break the Markov assumption and hurt performance. We introduce DT-CORL (Delay-Transformer belief policy Constrained Offline RL), an offline-RL framework built to cope with delayed dynamics at deployment. DT-CORL (i) produces delay-robust actions with a transformer-based belief predictor even though it never sees delayed observations during training, and (ii) is markedly more sample-efficient than naïve history-augmentation baselines. Experiments on D4RL benchmarks with several delay settings show that DT-CORL consistently outperforms both history-augmentation and vanilla belief-based methods, narrowing the sim-to-real latency gap while preserving data efficiency.

Pixel-level anomaly detection is indispensable in industrial defect detection and medical diagnosis. Recently, Segment Anything Model (SAM) has achieved promising results in many vision tasks. However, direct application of the SAM to pixel-level anomaly detection tasks results in unsatisfactory performance, meanwhile SAM needs the manual prompt. Although some automatically prompt-based SAM has been proposed, these automated prompting approaches merely utilize partial image features as prompts and fail to incorporate crucial features such as multi-scale image features to generate more suitable prompts. In this paper, we propose a novel Omni Dimensional State Space Model-driven SAM (ODS-SAM) for pixel-level anomaly detection. Specifically, the proposed method adopts the SAM architecture, ensuring easy implementation and avoiding the need for fine-tuning. A State-Space Model-based residual Omni Dimensional module is designed to automatically generate suitable prompts. This module can effectively leverage multi-scale and global information, facilitating an iterative search for optimal prompts in the prompt space. The identified optimal prompts are then fed into SAM as high-dimensional tensors. Experimental results demonstrate that the proposed ODS-SAM outperforms state-of-the-art models on both industrial and medical image datasets.

Reinforcement learning (RL) with delays is challenging as sensory perceptions lag behind the actual events: the RL agent needs to estimate the real state of its environment based on past observations. State-of-the-art (SOTA) methods typically employ recursive, step-by-step forecasting of states. This can cause the accumulation of compounding errors. To tackle this problem, our novel belief estimation method, named Directly Forecasting Belief Transformer (DFBT), directly forecasts states from observations without incrementally estimating intermediate states step-by-step. We theoretically demonstrate that DFBT greatly reduces compounding errors of existing recursively forecasting methods, yielding stronger performance guarantees. In experiments with D4RL offline datasets, DFBT reduces compounding errors with remarkable prediction accuracy. DFBT's capability to forecast state sequences also facilitates multi-step bootstrapping, thus greatly improving learning efficiency. On the MuJoCo benchmark, our DFBT-based method substantially outperforms SOTA baselines. Code is available at \href{https://github.com/QingyuanWuNothing/DFBT}{https://github.com/QingyuanWuNothing/DFBT}.

Due to the remarkable capabilities and growing impact of large language models (LLMs), they have been deeply integrated into many aspects of society. Thus, ensuring their alignment with human values and intentions has emerged as a critical challenge. This survey provides a comprehensive overview of practical alignment techniques, training protocols, and empirical findings in LLM alignment. We analyze the development of alignment methods across diverse paradigms, characterizing the fundamental trade-offs between core alignment objectives. Our analysis shows that while supervised fine-tuning enables basic instruction-following, preference-based methods offer more flexibility for aligning with nuanced human intent. We discuss state-of-the-art techniques, including Direct Preference Optimization (DPO), Constitutional AI, brain-inspired methods, and alignment uncertainty quantification (AUQ), highlighting their approaches to balancing quality and efficiency. We review existing evaluation frameworks and benchmarking datasets, emphasizing limitations such as reward misspecification, distributional robustness, and scalable oversight. We summarize strategies adopted by leading AI labs to illustrate the current state of practice. We conclude by outlining open problems in oversight, value pluralism, robustness, and continuous alignment. This survey aims to inform both researchers and practitioners navigating the evolving landscape of LLM alignment.

The Value Iteration Network (VIN) is an end-to-end differentiable neural network architecture for planning. It exhibits strong generalization to unseen domains by incorporating a differentiable planning module that operates on a latent Markov Decision Process (MDP). However, VINs struggle to scale to long-term and large-scale planning tasks, such as navigating a 100x100 maze -- a task that typically requires thousands of planning steps to solve. We observe that this deficiency is due to two issues: the representation capacity of the latent MDP and the planning module's depth. We address these by augmenting the latent MDP with a dynamic transition kernel, dramatically improving its representational capacity, and, to mitigate the vanishing gradient problem, introduce an "adaptive highway loss" that constructs skip connections to improve gradient flow. We evaluate our method on 2D/3D maze navigation environments, continuous control, and the real-world Lunar rover navigation task. We find that our new method, named Dynamic Transition VIN (DT-VIN), scales to 5000 layers and solves challenging versions of the above tasks. Altogether, we believe that DT-VIN represents a concrete step forward in performing long-term large-scale planning in complex environments.

Building heating, ventilation, and air conditioning (HVAC) systems account for nearly half of building energy consumption and [Formula: see text] of total energy consumption in the US. Their operation is also crucial for ensuring the physical and mental health of building occupants. Compared with traditional model-based HVAC control methods, the recent model-free deep reinforcement learning (DRL) based methods have shown good performance while do not require the development of detailed and costly physical models. However, these model-free DRL approaches often suffer from long training time to reach a good performance, which is a major obstacle for their practical deployment. In this work, we present a systematic approach to accelerate online reinforcement learning for HVAC control by taking full advantage of the knowledge from domain experts in various forms. Specifically, the algorithm stages include learning expert functions from existing abstract physical models and from historical data via offline reinforcement learning, integrating the expert functions with rule-based guidelines, conducting training guided by the integrated expert function and performing policy initialization from distilled expert function. Moreover, to ensure that the learned DRL-based HVAC controller can effectively keep room temperature within the comfortable range for occupants, we design a runtime shielding framework to reduce the temperature violation rate and incorporate the learned controller into it. Experimental results demonstrate up to 8.8X speedup in DRL training from our approach over previous methods, with low temperature violation rate.

Neural networks are increasingly used in safety-critical applications such as robotics and autonomous vehicles. However, the deployment of neural-network-controlled systems (NNCSs) raises significant safety concerns. Many recent advances overlook critical aspects of verifying control and ensuring safety in real-time scenarios. This paper presents a case study on using POLAR-Express, a state-of-the-art NNCS reachability analysis tool, for runtime safety verification in a Turtlebot navigation system using LiDAR. The Turtlebot, equipped with a neural network controller for steering, operates in a complex environment with obstacles. We developed a safe online controller switching strategy that switches between the original NNCS controller and an obstacle avoidance controller based on the verification results. Our experiments, conducted in a ROS2 Flatland simulation environment, explore the capabilities and limitations of using POLAR-Express for runtime verification and demonstrate the effectiveness of our switching strategy.

In environments with delayed observation, state augmentation by including actions within the delay window is adopted to retrieve Markovian property to enable reinforcement learning (RL). Whereas, state-of-the-art (SOTA) RL techniques with Temporal-Difference (TD) learning frameworks commonly suffer from learning inefficiency, due to the significant expansion of the augmented state space with the delay. To improve the learning efficiency without sacrificing performance, this work novelly introduces Variational Delayed Policy Optimization (VDPO), reforming delayed RL as a variational inference problem. This problem is further modelled as a two-step iterative optimization problem, where the first step is TD learning in the delay-free environment with a small state space, and the second step is behaviour cloning which can be addressed much more efficiently than TD learning. We not only provide a theoretical analysis of VDPO in terms of sample complexity and performance, but also empirically demonstrate that VDPO can achieve consistent performance with SOTA methods, with a significant enhancement of sample efficiency (approximately 50\% less amount of samples) in the MuJoCo benchmark.

As a combination of emerging multi-view learning methods and traditional multi-label classification tasks, multi-view multi-label classification has shown broad application prospects. The diverse semantic information contained in heterogeneous data effectively enables the further development of multi-label classification. However, the widespread incompleteness problem on multi-view features and labels greatly hinders the practical application of multi-view multi-label classification. Therefore, in this paper, we propose an attention-induced missing instances imputation technique to enhance the generalization ability of the model. Different from existing incomplete multi-view completion methods, we attempt to approximate the latent features of missing instances in embedding space according to cross-view joint attention, instead of recovering missing views in kernel space or original feature space. Accordingly, multi-view completed features are dynamically weighted by the confidence derived from joint attention in the late fusion phase. In addition, we propose a multi-view multi-label classification framework based on label-semantic feature learning, utilizing the statistical weak label correlation matrix and graph attention network to guide the learning process of label-specific features. Finally, our model is compatible with missing multi-view and partial multi-label data simultaneously and extensive experiments on five datasets confirm the advancement and effectiveness of our embedding imputation method and multi-view multi-label classification model.

Anomaly detection under open-set scenario is a challenging task that requires learning discriminative fine-grained features to detect anomalies that were even unseen during training. As a cheap yet effective approach, data augmentation has been widely used to create pseudo anomalies for better training of such models. Recent wisdom of augmentation methods focuses on generating random pseudo instances that may lead to a mixture of augmented instances with seen anomalies, or out of the typical range of anomalies. To address this issue, we propose a novel saliency-guided data augmentation method, SaliencyCut, to produce pseudo but more common anomalies which tend to stay in the plausible range of anomalies. Furthermore, we deploy a two-head learning strategy consisting of normal and anomaly learning heads, to learn the anomaly score of each sample. Theoretical analyses show that this mechanism offers a more tractable and tighter lower bound of the data log-likelihood. We then design a novel patch-wise residual module in the anomaly learning head to extract and assess the fine-grained anomaly features from each sample, facilitating the learning of discriminative representations of anomaly instances. Extensive experiments conducted on six real-world anomaly detection datasets demonstrate the superiority of our method to competing methods under various settings.