We study pure exploration in structured stochastic multi-armed bandits, aiming to efficiently identify the correct hypothesis from a finite set of alternatives. For a broad class of tasks, asymptotic analyses reduce to a maximin optimization that admits a two-player zero-sum game interpretation between an experimenter and a skeptic: the experimenter allocates measurements to rule out alternatives while the skeptic proposes alternatives. We reformulate the game by allowing the skeptic to adopt a mixed strategy, yielding a concave-convex saddle-point problem. This viewpoint leads to Frank-Wolfe Self-Play (FWSP): a projection-free, regularization-free, tuning-free method whose one-hot updates on both sides match the bandit sampling paradigm. However, structural constraints introduce sharp pathologies that complicate algorithm design and analysis: our linear-bandit case study exhibits nonunique optima, optimal designs with zero mass on the best arm, bilinear objectives, and nonsmoothness at the boundary. We address these challenges via a differential-inclusion argument, proving convergence of the game value for best-arm identification in linear bandits. Our analysis proceeds through a continuous-time limit: a differential inclusion with a Lyapunov function that decays exponentially, implying a vanishing duality gap and convergence to the optimal value. Although Lyapunov analysis requires differentiability of the objective, which is not guaranteed on the boundary, we show that along continuous trajectories the algorithm steers away from pathological nonsmooth points and achieves uniform global convergence to the optimal game value. We then embed the discrete-time updates into a perturbed flow and show that the discrete game value also converges. Building on FWSP, we further propose a learning algorithm based on posterior sampling. Numerical experiments demonstrate a vanishing duality gap.

Imitation Learning (IL) algorithms offer an efficient way to train an agent by mimicking an expert's behavior without requiring a reward function. IL algorithms often necessitate access to state and action information from expert demonstrations. Although expert actions can provide detailed guidance, requiring such action information may prove impractical for real-world applications where expert actions are difficult to obtain. To address this limitation, the concept of learning from observation (LfO) or state-only imitation learning (SOIL) has recently gained attention, wherein the imitator only has access to expert state visitation information. In this paper, we present a framework for LfO and use it to survey and classify existing LfO methods in terms of their trajectory construction, assumptions and algorithm's design choices. This survey also draws connections between several related fields like offline RL, model-based RL and hierarchical RL. Finally, we use our framework to identify open problems and suggest future research directions.

Conventional multi-agent reinforcement learning (MARL) methods rely on time-triggered execution, where agents sample and communicate actions at fixed intervals. This approach is often computationally expensive and communication-intensive. To address this limitation, we propose ET-MAPG (Event-Triggered Multi-Agent Policy Gradient reinforcement learning), a framework that jointly learns an agent's control policy and its event-triggering policy. Unlike prior work that decouples these mechanisms, ET-MAPG integrates them into a unified learning process, enabling agents to learn not only what action to take but also when to execute it. For scenarios with inter-agent communication, we introduce AET-MAPG, an attention-based variant that leverages a self-attention mechanism to learn selective communication patterns. AET-MAPG empowers agents to determine not only when to trigger an action but also with whom to communicate and what information to exchange, thereby optimizing coordination. Both methods can be integrated with any policy gradient MARL algorithm. Extensive experiments across diverse MARL benchmarks demonstrate that our approaches achieve performance comparable to state-of-the-art, time-triggered baselines while significantly reducing both computational load and communication overhead.

Sub-grid parameterisations in climate models are traditionally static and tuned offline, limiting adaptability to evolving states. This work introduces FedRAIN-Lite, a federated reinforcement learning (FedRL) framework that mirrors the spatial decomposition used in general circulation models (GCMs) by assigning agents to latitude bands, enabling local parameter learning with periodic global aggregation. Using a hierarchy of simplified energy-balance climate models, from a single-agent baseline (ebm-v1) to multi-agent ensemble (ebm-v2) and GCM-like (ebm-v3) setups, we benchmark three RL algorithms under different FedRL configurations. Results show that Deep Deterministic Policy Gradient (DDPG) consistently outperforms both static and single-agent baselines, with faster convergence and lower area-weighted RMSE in tropical and mid-latitude zones across both ebm-v2 and ebm-v3 setups. DDPG's ability to transfer across hyperparameters and low computational cost make it well-suited for geographically adaptive parameter learning. This capability offers a scalable pathway towards high-complexity GCMs and provides a prototype for physically aligned, online-learning climate models that can evolve with a changing climate. Code accessible at https://github.com/p3jitnath/climate-rl-fedrl.

Advanced reasoning models with agentic capabilities (AI agents) are deployed to interact with humans and to solve sequential decision-making problems under (approximate) utility functions and internal models. When such problems have resource or failure constraints where action sequences may be forcibly terminated once resources are exhausted, agents face implicit trade-offs that reshape their utility-driven (rational) behaviour. Additionally, since these agents are typically commissioned by a human principal to act on their behalf, asymmetries in constraint exposure can give rise to previously unanticipated misalignment between human objectives and agent incentives. We formalise this setting through a survival bandit framework, provide theoretical and empirical results that quantify the impact of survival-driven preference shifts, identify conditions under which misalignment emerges and propose mechanisms to mitigate the emergence of risk-seeking or risk-averse behaviours. As a result, this work aims to increase understanding and interpretability of emergent behaviours of AI agents operating under such survival pressure, and offer guidelines for safely deploying such AI systems in critical resource-limited environments.

Cost-aware Dynamic Workflow Scheduling (CADWS) is a key challenge in cloud computing, focusing on devising an effective scheduling policy to efficiently schedule dynamically arriving workflow tasks, represented as Directed Acyclic Graphs (DAG), to suitable virtual machines (VMs). Deep reinforcement learning (DRL) has been widely employed for automated scheduling policy design. However, the performance of DRL is heavily influenced by the design of the problem-tailored policy network and is highly sensitive to hyperparameters and the design of reward feedback. Considering the above-mentioned issues, this study proposes a novel DRL method combining Graph Attention Networks-based policy network and Evolution Strategy, referred to as GATES. The contributions of GATES are summarized as follows: (1) GATES can capture the impact of current task scheduling on subsequent tasks by learning the topological relationships between tasks in a DAG. (2) GATES can assess the importance of each VM to the ready task, enabling it to adapt to dynamically changing VM resources. (3) Utilizing Evolution Strategy's robustness, exploratory nature, and tolerance for delayed rewards, GATES achieves stable policy learning in CADWS. Extensive experimental results demonstrate the superiority of the proposed GATES in CADWS, outperforming several state-of-the-art algorithms. The source code is available at: https://github.com/YaShen998/GATES.

The advantage function is a central concept in RL that helps reduce variance in policy gradient estimates. Recently, for language modeling, Group Relative Policy Optimization (GRPO) was proposed to compute the advantage for each output by subtracting the mean reward, as the baseline, for all outputs in the group. However, it can lead to high variance when the reward advantage is inaccurately predicted. In this work, we propose Kalman Filter Enhanced Group Relative Policy Optimization (KRPO) model, by using lightweight Kalman filtering to dynamically estimate the latent reward baseline and uncertainty. This filtering technique replaces the naive group mean, enabling more adaptive advantage normalization. Our method does not require additional learned parameters over GRPO. This approach offers a simple yet effective way to incorporate multiple outputs of GRPO into advantage estimation, improving policy optimization in settings where highly dynamic reward signals are difficult to model for language models. Through the accuracies and rewards obtained from math question answering and reasoning, we show that using a more adaptive advantage estimation model, KRPO can improve the stability and performance of GRPO. The code is available at https://github.com/billhhh/KRPO_LLMs_RL.

Recent advances in generative world models have enabled classical safe control methods, such as Hamilton-Jacobi (HJ) reachability, to generalize to complex robotic systems operating directly from high-dimensional sensor observations. However, obtaining comprehensive coverage of all safety-critical scenarios during world model training is extremely challenging. As a result, latent safety filters built on top of these models may miss novel hazards and even fail to prevent known ones, overconfidently misclassifying risky out-of-distribution (OOD) situations as safe. To address this, we introduce an uncertainty-aware latent safety filter that proactively steers robots away from both known and unseen failures. Our key idea is to use the world model's epistemic uncertainty as a proxy for identifying unseen potential hazards. We propose a principled method to detect OOD world model predictions by calibrating an uncertainty threshold via conformal prediction. By performing reachability analysis in an augmented state space-spanning both the latent representation and the epistemic uncertainty-we synthesize a latent safety filter that can reliably safeguard arbitrary policies from both known and unseen safety hazards. In simulation and hardware experiments on vision-based control tasks with a Franka manipulator, we show that our uncertainty-aware safety filter preemptively detects potential unsafe scenarios and reliably proposes safe, in-distribution actions. Video results can be found on the project website at https://cmu-intentlab.github.io/UNISafe