Leveraging Multi-modal Large Language Models (MLLMs) to create embodied agents offers a promising avenue for tackling real-world tasks. While language-centric embodied agents have garnered substantial attention, MLLM-based embodied agents remain underexplored due to the lack of comprehensive evaluation frameworks. To bridge this gap, we introduce EmbodiedBench, an extensive benchmark designed to evaluate vision-driven embodied agents. EmbodiedBench features: (1) a diverse set of 1,128 testing tasks across four environments, ranging from high-level semantic tasks (e.g., household) to low-level tasks involving atomic actions (e.g., navigation and manipulation); and (2) six meticulously curated subsets evaluating essential agent capabilities like commonsense reasoning, complex instruction understanding, spatial awareness, visual perception, and long-term planning. Through extensive experiments, we evaluated 24 leading proprietary and open-source MLLMs within EmbodiedBench. Our findings reveal that: MLLMs excel at high-level tasks but struggle with low-level manipulation, with the best model, GPT-4o, scoring only $28.9\%$ on average. EmbodiedBench provides a multifaceted standardized evaluation platform that not only highlights existing challenges but also offers valuable insights to advance MLLM-based embodied agents. Our code and dataset are available at [https://embodiedbench.github.io](https://embodiedbench.github.io).

A major challenge in training VLM agents, compared to LLM agents, is that states shift from simple texts to complex visual observations, which introduces partial observability and demands robust world modeling. We ask: can VLM agents build internal world models through explicit visual state reasoning? In this work, we architecturally enforce and reward VLM agent’s reasoning process via reinforcement learning (RL), formulating the problem as a Partially Observable Markov Decision Process (POMDP). We demonstrate that structuring agent’s reasoning into StateEstimation (“what is the current state?”) and TransitionModeling (“what is next?”) is critical by studying five reasoning strategies. Investigating how agents should ground visual states and represent these internal beliefs, we reveal the optimal representations are task-dependent: Natural Language excels at capturing semantic relationships for general tasks, while Structured formats are essential for high-precision manipulation. These insights motivate our approach to reward shaping and credit assignment. We leverage a WorldModeling Reward to densely rewards the agent’s turn-by-turn state predictions, while our Bi-Level General Advantage Estimation (Bi-Level GAE) enables turn-aware credit assignment. Through such world model reasoning, we enable a 3B model to achieve performance of 0.82 on a set of five diverse agent tasks, nearly 3× improvement over its untrained counterpart (0.21) and surpassing proprietary reasoning models like GPT-5 (0.75), Gemini 2.5 Pro (0.67) and Claude 4.5 (0.62). All experiments are supported by our VAGEN framework, a scalable system for training and analyzing multi-turn VLM agents across diverse visual environments

The growing dependence of autonomous vehicles on infrared thermal imaging for adverse-condition perception has heightened the security importance of infrared vehicle detectors. This work proposes the Infrared Vehicle Adversarial Patch (InfVAP), a physically realizable attack framework targeting these detectors. Addressing the dual challenges of limited prior research and complex vehicle geometry, our solution introduces: (1) Pixel-level attack localization using DeepLab-v3+ segmentation to ensure precise patch placement; (2) An improved Particle Swarm Optimization (PSO) algorithm with dynamic parameter adaptation and shape optimization to enhance attack effectiveness; (3) Expectation Over Transformation (EOT)-based robustness enhancement against real-world perturbations; and (4) A top-K confidence suppression loss (K=1000) for maximum attack potency. Extensive evaluations on the FLIR ADAS V2 dataset demonstrate that InfVAP achieves state-of-the-art performance: 93.93% digital Attack Success Rate (ASR) and 80.2% physical ASR at distances of 12 meters. Comprehensive ablation studies prove all components are statistically significant, with PSO optimization contributing most to attack potency.

Adversarial attacks manipulate deep learning models into generating erroneous predictions by introducing subtle yet carefully crafted perturbations to input data. Such attacks serve dual roles: revealing model vulnerabilities while enhancing robustness. Although adversarial attacks are well- explored in visual tracking, their efficacy in infrared single-object tracking remains limited. To bridge this gap, we propose a novel heating-expanding attack to deceive SiameseRPN-based infrared trackers. By exploiting the inherent thermal-texture discrepancy between visible and infrared imaging modalities, we introduce a novel heating-expanding loss function. This loss function drives the training of an adversarial perturbation generator that synthesizes imperceptible thermal perturbations. When strategically applied at the target-background boundaries, these perturbations trigger progressive bounding box expansion during tracking. The accumulated interference induces complete tracker failure via boundary drift. Benchmark evaluations on LSOTB-TIR and PTB-TIR demonstrate our method’s consistent precision reduction of 45.0% and 45.4% (surpassing SOTA by 7.8/10.2 pp). The approach sustains attack performance against other trackers, reducing norm precision by 12.6% for DaSiamRPN and 4.5% for TaMOs respectively without architecture-specific tuning, proving plug- and-play readiness for deployment.

Traffic signal control (TSC) plays a crucial role in the intelligent transportation system. Among existing TSC approaches, Proactive Traffic Signal Control (PTSC) predicts future traffic states at intersections and proactively adjusts control policies. It is evident that PTSC methods are highly effective in alleviating both current and future traffic congestion at intersections. However, existing PTSC methods focus on point estimation prediction while neglecting prediction reliability. Additionally, they fail to adaptively coordinate between current and future traffic states for optimal control. To address these limitations, we propose an innovative Proactive-Plugin that can be combined with existing TSC methods to enhance the accuracy and robustness of traffic signal control policies. This plugin enhances two critical aspects: 1) Prediction reliability is achieved through Fine-grained Traffic Uncertainty Quantification. This module generates probabilistic forecasts along with confidence intervals to explicitly indicate the credibility of the predictions. 2) Coordination adaptiveness is enabled by a Current-Future Tradeoff Integration mechanism. This mechanism dynamically adjusts the relative influence of current traffic states and probabilistic forecasts on control policies. To further ensure robustness, we design a multi-task joint optimization to reduce the negative impact of inaccurate predictions during training. Experimental results on six real-world datasets demonstrate consistent improvements in traffic efficiency, validating the effectiveness of our approach.

As the electric vehicle market continues to grow, the importance of thermal management in its electric drive system has become increasingly prominent. As an integral component of the electric drive system, the temperature of permanent magnet motors directly affects the performance and safety of electric vehicles. Precise temperature monitoring and prediction of the motor, especially the stator and rotor regions with concentrated heat generation, are crucial to preventing overheating motor damage. However, it is hard to seize the temperature by sensors simply. This paper proposes a datadriven method for predicting the stator and rotor temperatures of PMSMs. Firstly, based on the studied form-wound PMSM, a simulation model of the motor is established, and the corresponding losses are calculated. Secondly, a thermal network model of the motor is constructed using the calculated losses, which enables the analysis of temperature distribution and trends within the motor. Then, based on the thermal network model, a data-driven model is constructed using a CNN-LSTM architecture. Ultimately, to verify the precision of the predictive model, simulation and experimental tests are conducted, wherein the model's projections are juxtaposed against actual motor data derived from experiments. The results show that the average error of stator temperature is 4.3 °C, and the maximum error is 6.7 °C. The average temperature error of the rotor permanent magnet is 3.9 °C, and the maximum error is 6.2 °C. In comparison with existing methodologies, the estimation accuracy has been enhanced by a minimum of 25%. The prediction outcomes exhibit high fidelity, thereby furnishing an efficacious approach for forecasting the temperature of PMSMs.

The application of reinforcement learning in traffic signal control (TSC) has been extensively researched and yielded notable achievements. However, most existing works for TSC assume that traffic data from all surrounding intersections is fully and continuously available through sensors. In real-world applications, this assumption often fails due to sensor malfunctions or data loss, making TSC with missing data a critical challenge. To meet the needs of practical applications, we introduce DiffLight, a novel conditional diffusion model for TSC under data-missing scenarios in the offline setting. Specifically, we integrate two essential sub-tasks, i.e., traffic data imputation and decision-making, by leveraging a Partial Rewards Conditioned Diffusion (PRCD) model to prevent missing rewards from interfering with the learning process. Meanwhile, to effectively capture the spatial-temporal dependencies among intersections, we design a Spatial-Temporal transFormer (STFormer) architecture. In addition, we propose a Diffusion Communication Mechanism (DCM) to promote better communication and control performance under data-missing scenarios. Extensive experiments on five datasets with various data-missing scenarios demonstrate that DiffLight is an effective controller to address TSC with missing data. The code of DiffLight is released at https://github.com/lokol5579/DiffLight-release.

This study explores the feasibility of performing Chinese word segmentation (CWS) and POS tagging by deep learning.We try to avoid task-specific feature engineering, and use deep layers of neural networks to discover relevant features to the tasks.We leverage large-scale unlabeled data to improve internal representation of Chinese characters, and use these improved representations to enhance supervised word segmentation and POS tagging models.Our networks achieved close to state-of-theart performance with minimal computational cost.We also describe a perceptron-style algorithm for training the neural networks, as an alternative to maximum-likelihood method, to speed up the training process and make the learning algorithm easier to be implemented.