We propose an epoch-based reinforcement learning algorithm for infinite-horizon average-cost Markov decision processes (MDPs) that leverages a partial order over a policy class. In this structure, $π' \leq π$ if data collected under $π$ can be used to estimate the performance of $π'$, enabling counterfactual inference without additional environment interaction. Leveraging this partial order, we show that our algorithm achieves a regret bound of $O(\sqrt{w \log(|Θ|) T})$, where $w$ is the width of the partial order. Notably, the bound is independent of the state and action space sizes. We illustrate the applicability of these partial orders in many domains in operations research, including inventory control and queuing systems. For each, we apply our framework to that problem, yielding new theoretical guarantees and strong empirical results without imposing extra assumptions such as convexity in the inventory model or specialized arrival-rate structure in the queuing model.

Sequential decision-making systems routinely operate with missing or incomplete data. Classical reinforcement learning theory, which is commonly used to solve sequential decision problems, assumes Markovian observability, which may not hold under partial observability. Causal inference paradigms formalise ignorability of missingness. We show these views can be unified and generalized in order to guarantee Q-learning convergence even when the Markov property fails. To do so, we introduce the concept of relative ignorability. Relative ignorability is a graphical-causal criterion which refines the requirements for accurate decision-making based on incomplete data. Theoretical results and simulations both reveal that non-Markovian stochastic processes whose missingness is relatively ignorable with respect to causal estimands can still be optimized using standard Reinforcement Learning algorithms. These results expand the theoretical foundations of safe, data-efficient AI to real-world environments where complete information is unattainable.

We investigate the capacity of transformers to learn algorithms involving their context while solely being trained using next token prediction. We set up Markov chains with random transition matrices and we train transformers to predict the next token. Matrices used during training and test are different and we show that there is a threshold in transformer size and in training set size above which the model is able to learn to estimate the transition probabilities from its context instead of memorizing the training patterns. Additionally, we show that more involved encoding of the states enables more robust prediction for Markov chains with structures different than those seen during training.

Managing physiological variables within clinically safe target zones is a central challenge in healthcare, particularly for chronic conditions such as Type 1 Diabetes Mellitus (T1DM). Reinforcement learning (RL) offers promise for personalising treatment, but struggles with the delayed and heterogeneous effects of interventions. We propose a novel RL framework to study and support decision-making in T1DM technologies, such as automated insulin delivery. Our approach captures the complex temporal dynamics of treatment by unifying two control modalities: \textit{impulse control} for discrete, fast-acting interventions (e.g., insulin boluses), and \textit{switching control} for longer-acting treatments and regime shifts. The core of our method is a constrained Markov decision process augmented with physiological state features, enabling safe policy learning under clinical and resource constraints. The framework incorporates biologically realistic factors, including insulin decay, leading to policies that better reflect real-world therapeutic behaviour. While not intended for clinical deployment, this work establishes a foundation for future safe and temporally-aware RL in healthcare. We provide theoretical guarantees of convergence and demonstrate empirical improvements in a stylised T1DM control task, reducing blood glucose level violations from 22.4\% (state-of-the-art) to as low as 10.8\%.

We present Agent Lightning, a flexible and extensible framework that enables Reinforcement Learning (RL)-based training of Large Language Models (LLMs) for any AI agent. Unlike existing methods that tightly couple RL training with agent or rely on sequence concatenation with masking, Agent Lightning achieves complete decoupling between agent execution and training, allowing seamless integration with existing agents developed via diverse ways (e.g., using frameworks like LangChain, OpenAI Agents SDK, AutoGen, and building from scratch) with almost ZERO code modifications. By formulating agent execution as Markov decision process, we define an unified data interface and propose a hierarchical RL algorithm, LightningRL, which contains a credit assignment module, allowing us to decompose trajectories generated by ANY agents into training transition. This enables RL to handle complex interaction logic, such as multi-agent scenarios and dynamic workflows. For the system design, we introduce a Training-Agent Disaggregation architecture, and brings agent observability frameworks into agent runtime, providing a standardized agent finetuning interface. Experiments across text-to-SQL, retrieval-augmented generation, and math tool-use tasks demonstrate stable, continuous improvements, showcasing the framework's potential for real-world agent training and deployment.

Fine-tuning diffusion policies with reinforcement learning (RL) presents significant challenges. The long denoising sequence for each action prediction impedes effective reward propagation. Moreover, standard RL methods require millions of real-world interactions, posing a major bottleneck for practical fine-tuning. Although prior work frames the denoising process in diffusion policies as a Markov Decision Process to enable RL-based updates, its strong dependence on environment interaction remains highly inefficient. To bridge this gap, we introduce DiWA, a novel framework that leverages a world model for fine-tuning diffusion-based robotic skills entirely offline with reinforcement learning. Unlike model-free approaches that require millions of environment interactions to fine-tune a repertoire of robot skills, DiWA achieves effective adaptation using a world model trained once on a few hundred thousand offline play interactions. This results in dramatically improved sample efficiency, making the approach significantly more practical and safer for real-world robot learning. On the challenging CALVIN benchmark, DiWA improves performance across eight tasks using only offline adaptation, while requiring orders of magnitude fewer physical interactions than model-free baselines. To our knowledge, this is the first demonstration of fine-tuning diffusion policies for real-world robotic skills using an offline world model. We make the code publicly available at https://diwa.cs.uni-freiburg.de.

As smart homes become increasingly prevalent, intelligent models are widely used for tasks such as anomaly detection and behavior prediction. These models are typically trained on static datasets, making them brittle to behavioral drift caused by seasonal changes, lifestyle shifts, or evolving routines. However, collecting new behavior data for retraining is often impractical due to its slow pace, high cost, and privacy concerns. In this paper, we propose SmartGen, an LLM-based framework that synthesizes context-aware user behavior data to support continual adaptation of downstream smart home models. SmartGen consists of four key components. First, we design a Time and Semantic-aware Split module to divide long behavior sequences into manageable, semantically coherent subsequences under dual time-span constraints. Second, we propose Semantic-aware Sequence Compression to reduce input length while preserving representative semantics by clustering behavior mapping in latent space. Third, we introduce Graph-guided Sequence Synthesis, which constructs a behavior relationship graph and encodes frequent transitions into prompts, guiding the LLM to generate data aligned with contextual changes while retaining core behavior patterns. Finally, we design a Two-stage Outlier Filter to identify and remove implausible or semantically inconsistent outputs, aiming to improve the factual coherence and behavioral validity of the generated sequences. Experiments on three real-world datasets demonstrate that SmartGen significantly enhances model performance on anomaly detection and behavior prediction tasks under behavioral drift, with anomaly detection improving by 85.43% and behavior prediction by 70.51% on average. The code is available at https://github.com/horizonsinzqs/SmartGen.

Path planning in unknown environments is a crucial yet inherently challenging capability for mobile robots, which primarily encompasses two coupled tasks: autonomous exploration and point-goal navigation. In both cases, the robot must perceive the environment, update its belief, and accurately estimate potential information gain on-the-fly to guide planning. In this work, we propose CogniPlan, a novel path planning framework that leverages multiple plausible layouts predicted by a COnditional GeNerative Inpainting model, mirroring how humans rely on cognitive maps during navigation. These predictions, based on the partially observed map and a set of layout conditioning vectors, enable our planner to reason effectively under uncertainty. We demonstrate strong synergy between generative image-based layout prediction and graph-attention-based path planning, allowing CogniPlan to combine the scalability of graph representations with the fidelity and predictiveness of occupancy maps, yielding notable performance gains in both exploration and navigation. We extensively evaluate CogniPlan on two datasets (hundreds of maps and realistic floor plans), consistently outperforming state-of-the-art planners. We further deploy it in a high-fidelity simulator and on hardware, showcasing its high-quality path planning and real-world applicability.

Large Language Reasoning Models have demonstrated remarkable success on static tasks, yet their application to multi-round agentic planning in interactive environments faces two fundamental challenges. First, the intractable credit assignment problem renders conventional reinforcement learning ineffective in sparse-reward settings. Second, the computational overhead of verbose, step-by-step reasoning histories is prohibitive. To address these challenges, we propose BPO, a three-stage framework (bootstrapping, extrapolation, and refinement) that establishes a self-improving data flywheel to develop robust reasoning models for long-horizon, sparse-reward environments. Our framework first bootstraps efficient reasoning using the proposed planning quaternions with long-short chain-of-thought fusion. It then extrapolates to out-of-distribution tasks through complexity-stratified curriculum learning. Finally, the model iteratively refines itself by learning exclusively on experiences selected via reward-gated rejection sampling. Experiments on ALFWorld, ScienceWorld, and WebShop demonstrate that our approach achieves state-of-the-art with significant token efficiency, providing a new recipe for reasoning models in agentic planning.

-- Curriculum learning has emerged as a promising approach for training complex robotics tasks, yet current applications predominantly rely on manually designed curricula, which demand significant engineering effort and can suffer from subjective and suboptimal human design choices. While automated curriculum learning has shown success in simple domains like grid worlds and games where task distributions can be easily specified, robotics tasks present unique challenges: they require handling complex task spaces while maintaining relevance to target domain distributions that are only partially known through limited samples. We validate GACL on wheeled navigation in constrained environments and quadruped locomotion in challenging 3D confined spaces, achieving 6.8% and 6.1% higher success rates, respectively, than state-of-the-art methods in each domain. Curriculum learning has shown promises in training robots for complex tasks such as navigating through highly constrained environments or maintaining quadruped locomotion across challenging terrain [1], [2]. However, current applications of curriculum learning in robotics face a fundamental challenge: they predominantly rely on manually designed curricula, which demand significant engineering effort and can suffer from subjective, suboptimal design choices. For example, in quadruped locomotion tasks [2], roboticists must carefully design progressive stages from basic jumping skills to complex obstacle traversal and manually define success metrics and progression conditions at each stage.