Deep Multi-Agent Reinforcement Learning (MARL) excels in co-operative tasks but often struggles with local optima in high - dimensional joint action spaces. In contrast, Evolutionary Algorithms (EAs) offer robust global exploration capabilities. Although hybrid approaches seek to combine the strengths of both paradigms, they typically face a critical bottleneck: the prohibitive sample cost of evaluating large populations via environment rollouts. To address this challenge, we propose Surrogate-assisted Evolutionary Multi-Agent Reinforcement Learning (SEMARL), a unified framework that synergizes gradient-based refinement with surrogate-assisted evolutionary search. SEMARL employs a cooperative co-evolutionary architecture to maintain diverse agent policies and injects gradient-refined parameters into the population to accelerate convergence. Crucially, we leverage the centralized critic from the gradient learner as a computationally efficient surrogate for fitness estimation. To prevent misleading guidance from an inaccurate critic, we introduce an adaptive reliability control mechanism based on temporal difference (TD) error, which dynamically regulates the surrogate's influence. Experiments on the Multi-Agent MuJoCo benchmark demonstrate that SEMARL significantly outperforms other algorithms, achieving superior asymptotic performance with substantially higher sample efficiency.

Despite recent advancements in offline multi-task reinforcement learning (MTRL) have harnessed the powerful capabilities of the Transformer architecture, most approaches focus on a limited number of tasks, with scaling to extremely massive tasks remaining a formidable challenge. In this paper, we first revisit the key impact of task numbers on current MTRL method, and further reveal that naively expanding the parameters proves insufficient to counteract the performance degradation as the number of tasks escalates. Building upon these insights, we propose M3DT, a novel mixture-of-experts (MoE) framework that tackles task scalability by further unlocking the model’s parameter scalability. Specifically, we enhance both the architecture and the optimization of the agent, where we strengthen the Decision Transformer (DT) backbone with MoE to reduce task load on parameter subsets, and introduce a three-stage training mechanism to facilitate efficient training with optimal performance. Experimental results show that, by increasing the number of experts, M3DT not only consistently enhances its performance as model expansion on the fixed task numbers, but also exhibits remarkable task scalability, successfully extending to 160 tasks with superior performance.

Training safe LLMs remains a critical challenge. The most widely used method, Refusal Training (RT), struggles to generalize against various Out-of-Distribution (OOD) jailbreaking attacks. Although various advanced methods have been proposed to address this issue, we instead question whether OOD attacks inherently surpass the capability of vanilla RT. Evaluations using Best-of-N (BoN) reveal significant safety improvements as N increases, indicating models possess adequate latent safety knowledge but RT fails to consistently elicit it under OOD scenarios. Further domain adaptation analysis reveals that direct RT causes reliance on superficial shortcuts, resulting in non-generalizable representation mappings. Inspired by our findings, we propose training model to perform safety reasoning for each query. Specifically, we synthesize reasoning supervision aligned with specified guidelines that reflect diverse perspectives on safety knowledge. This encourages model to engage in deeper reasoning, explicitly eliciting and utilizing latent safety knowledge for each query. Extensive experiments show that our method significantly improves model generalization against OOD attacks.

LLMs have made impressive progress, but their growing capabilities also expose them to highly flexible jailbreaking attacks designed to bypass safety alignment. While many existing defenses focus on known types of attacks, it is more critical to prepare LLMs for *unseen* attacks that may arise during deployment. To address this, we propose a **lifelong safety alignment** framework that enables LLMs to continuously adapt to new and evolving jailbreaking strategies. Our framework introduces a competitive setup between two components: a **Meta-Attacker**, trained to actively discover novel jailbreaking strategies, and a **Defender**, trained to resist them. To effectively warm up the Meta-Attacker, we first leverage the GPT-4o API to extract key insights from a large collection of jailbreak-related research papers. Through iterative training, the first iteration Meta-Attacker achieves a 73% attack success rate (ASR) on RR and a 57% transfer ASR on LAT using only *single-turn* attacks. Meanwhile, the Defender progressively improves its robustness and ultimately reduces the Meta-Attacker's success rate to just 7%, enabling safer and more reliable deployment of LLMs in open-ended environments.

We present a simple yet effective method for skeleton-agnostic motion retargeting. Previous methods transfer motion between high-resolution meshes, failing to preserve the inherent local-part motions in the mesh. Addressing this issue, our proposed method learns the correspondence in a coarse-to-fine fashion by disentangling the retargeting process within multi-scale meshes. First, we propose a mesh-pooling module that pools the mesh representations for better motion transfer. This module improves the ability to handle small-part motion and preserves the local motion interdependence between neighboring mesh vertices. Furthermore, we leverage a multi-scale refinement procedure to complement missing mesh details by gradually refining the low-resolution mesh output with a higher-resolution one. We evaluate our method on several well-known 3D character datasets, and it yields an average improvement of 25% on point-wise mesh Euclidean distance (PMD) against the start-of-art method. Qualitative results show that our method is significantly helpful in preserving the moving consistency of different body parts on the target character due to disentangling body-part structures and mesh details in a multi-scale way.

We propose an effective hybrid approach jointly leveraging local and global features for shortest-path (SP) distance estimation in domain-agnostic large-scale graphs. Previous works struggle to make estimations either from node-wise local embeddings or by compressing a global SP distance matrix, causing insufficient learning at some distance and loss of accuracy. Unlike them, we find a way to better preserve local distance on node embeddings, and then integrate them with a global process for accurate estimation at every distance. First, we propose a distance-consistent embedding method that better preserves the distance between each node and its local neighbors due to resampling node occurrence on random walks. Second, we train a feed-forward network with boosting techniques (FFN-BT) to estimate SP distance from these embeddings plus existing global features. Experimental results show that our approach averagely yields 10% improved accuracy and 20% reduced time when compared to existing methods on a broad class of graphs.

Data augmentation (DA) is a crucial technique for enhancing the sample efficiency of visual reinforcement learning (RL) algorithms. Notably, employing simple observation transformations alone can yield outstanding performance without extra auxiliary representation tasks or pre-trained encoders. However, it remains unclear which attributes of DA account for its effectiveness in achieving sample-efficient visual RL. To investigate this issue and further explore the potential of DA, this work conducts comprehensive experiments to assess the impact of DA's attributes on its efficacy and provides the following insights and improvements: (1) For individual DA operations, we reveal that both ample spatial diversity and slight hardness are indispensable. Building on this finding, we introduce Random PadResize (Rand PR), a new DA operation that offers abundant spatial diversity with minimal hardness. (2) For multi-type DA fusion schemes, the increased DA hardness and unstable data distribution result in the current fusion schemes being unable to achieve higher sample efficiency than their corresponding individual operations. Taking the non-stationary nature of RL into account, we propose a RL-tailored multi-type DA fusion scheme called Cycling Augmentation (CycAug), which performs periodic cycles of different DA operations to increase type diversity while maintaining data distribution consistency. Extensive evaluations on the DeepMind Control suite and CARLA driving simulator demonstrate that our methods achieve superior sample efficiency compared with the prior state-of-the-art methods.

Path planning and collision avoidance are critical components of UAV control algorithms that play a crucial role in executing UAV missions. As scenarios become increasingly complex, the traditional control methods just ain't cutting it to meet the requirements. Reinforcement learning is an emerging decision-making control algorithm that attempts to address these issues as an alternative to traditional methods and has made significant advances. Unfortunately, standard RL approaches only aim to maximize rewards, however balancing task performance and safety in completing UAV tasks poses a challenge since these two objectives sometimes conflict, leading to a trade-off often difficult to manage. This paper proposes three techniques to address this problem. First, we model the path planning and collision avoidance issue in a constrained RL framework, eliminating the need for complex reward engineering. Second, we expand our previous work in the UAV setting and introduce an exact penalty optimization (EPO) algorithm to provide stricter constraint guarantees. We also propose a novel spatial information representation method for the UAV scenario to help UAVs better understand environmental information. The experimental results demonstrate the effectiveness of the EPO and spatial representation modules proposed in this paper, through a significant reduction in collisions as well as a strong improvement in the rate of reaching the destination.