We study what actually works and what doesn't for training large language models as agents via multi-turn reinforcement learning. Despite rapid progress, existing frameworks and definitions are fragmented, and there is no systematic formulation or analysis of which design choices matter across tasks. We address this gap by first breaking down the design space into three inter-related pillars--environment, reward, and policy--and empirically derive a recipe for training LLM agents in situated textual domains. In particular, we test TextWorld and ALFWorld, popular domains for testing situated embodied reasoning, as well as SWE-Gym for more software engineering style tasks. Training LLMs as autonomous agents to navigate open-ended environments presents unique challenges: planning across extended horizons, making multi-turn sequential decisions, and optimizing for multi-turn rewards. The transition from static single-turn problem-solving to dynamic multi-step reasoning is essential for agentic benchmarks such as interactive text and embodied simulations (TextWorld (C ˆ ot e et al., 2018), ALFWorld (Shridhar et al., 2021), etc.), real-world software programming (OSWorld (Xie et al., 2024), SWE-gym (Pan et al., 2025), etc.), and abstract reasoning in novel situations (ARC-AGI (Chollet et al., 2025)). However, existing multi-turn RL implementations vary widely: some refer to tool-augmented single queries as multi-turn (Zeng et al., 2025), while many rely on model-based assumptions (Wang et al., 2025). This fragmentation has led to incomparable results across papers and confusion about what constitutes true multi-turn learning versus pseudo-multi-turn adaptations of single-turn methods. This paper aims to facilitate research efforts on the open research question: What factors are practically important in making multi-turn RL for LLM agent learning work. Motivated by the lack of standardization of multi-turn RL approaches, we systematically decompose the design space into three interdependent pillars--environment, reward, and policy--and empirically derive a recipe for training LLM agents in situated textual domains (Figure 1). We evaluate our approach on TextWorld and ALFWorld for embodied reasoning, and SWE-gym for real-world programming, revealing critical insights for each pillar.

Large Language Model (LLM) agents excel at multi-step, tool-augmented tasks. However, smart homes introduce distinct challenges, requiring agents to handle latent user intents, temporal dependencies, device constraints, scheduling, and more. The main bottlenecks for developing smart home agents with such capabilities include the lack of a realistic simulation environment where agents can interact with devices and observe the results, as well as a challenging benchmark to evaluate them. To address this, we introduce $\textbf{SimuHome}$, a time-accelerated home environment that simulates smart devices, supports API calls, and reflects changes in environmental variables. By building the simulator on the Matter protocol, the global industry standard for smart home communication, SimuHome provides a high-fidelity environment, and agents validated in SimuHome can be deployed on real Matter-compliant devices with minimal adaptation. We provide a challenging benchmark of 600 episodes across twelve user query types that require the aforementioned capabilities. Our evaluation of 16 agents under a unified ReAct framework reveals distinct capabilities and limitations across models. Models under 7B parameters exhibited negligible performance across all query types. Even GPT-4.1, the best-performing standard model, struggled with implicit intent inference, state verification, and particularly temporal scheduling. While reasoning models such as GPT-5.1 consistently outperformed standard models on every query type, they required over three times the average inference time, which can be prohibitive for real-time smart home applications. This highlights a critical trade-off between task performance and real-world practicality.

This paper applies the authors' recent results on asynchronous stochastic approximation (SA) in the Borkar-Meyn framework to reinforcement learning in average-reward semi-Markov decision processes (SMDPs). We establish the convergence of an asynchronous SA analogue of Schweitzer's classical relative value iteration algorithm, RVI Q-learning, for finite-space, weakly communicating SMDPs. In particular, we show that the algorithm converges almost surely to a compact, connected subset of solutions to the average-reward optimality equation, with convergence to a unique, sample path-dependent solution under additional stepsize and asynchrony conditions. Moreover, to make full use of the SA framework, we introduce new monotonicity conditions for estimating the optimal reward rate in RVI Q-learning. These conditions substantially expand the previously considered algorithmic framework and are addressed through novel arguments in the stability and convergence analysis of RVI Q-learning.

How much does a trained RL policy actually use its past observations? We propose \emph{Temporal Range}, a model-agnostic metric that treats first-order sensitivities of multiple vector outputs across a temporal window to the input sequence as a temporal influence profile and summarizes it by the magnitude-weighted average lag. Temporal Range is computed via reverse-mode automatic differentiation from the Jacobian blocks $\partial y_s/\partial x_t\in\mathbb{R}^{c\times d}$ averaged over final timesteps $s\in\{t+1,\dots,T\}$ and is well-characterized in the linear setting by a small set of natural axioms. Across diagnostic and control tasks (POPGym; flicker/occlusion; Copy-$k$) and architectures (MLPs, RNNs, SSMs), Temporal Range (i) remains small in fully observed control, (ii) scales with the task's ground-truth lag in Copy-$k$, and (iii) aligns with the minimum history window required for near-optimal return as confirmed by window ablations. We also report Temporal Range for a compact Long Expressive Memory (LEM) policy trained on the task, using it as a proxy readout of task-level memory. Our axiomatic treatment draws on recent work on range measures, specialized here to temporal lag and extended to vector-valued outputs in the RL setting. Temporal Range thus offers a practical per-sequence readout of memory dependence for comparing agents and environments and for selecting the shortest sufficient context.

Abstract-- Assigning tasks to robots often involves supplying the robot with an overarching goal, such as through natural language, and then relying on the robot to uncover and execute a plan to achieve that goal. In many settings common to human-robot interaction, however, the world is only partially observable to the robot, requiring that it create plans under uncertainty. Although many probabilistic planning algorithms exist for this purpose, these algorithms can be inefficient if executed with the robot's limited computational resources, or may require more steps than expected to achieve the goal. We thereby created a new, lightweight, probabilistic planning algorithm, Plan-Orchestrated Tree Assembly for Lookahead (POrTAL), that combines the strengths of two baseline planning algorithms, FF-Replan and POMCP . In a series of case studies, we demonstrate POrTAL's ability to quickly arrive at solutions that outperform these baselines in terms of number of steps. We additionally demonstrate how POrTAL performs under varying temporal constraints. The ability of modern robots to respond to arbitrary user requests has advanced considerably in recent years. This advancement is in large part due to robots' ability to autonomously plan their own actions. When receiving a goal such as "bring me a cup of coffee," for example, a robot can calculate the minimum number of steps required to achieve this goal: obtain the coffee grinds, proceeding to the coffee maker, load the grinds, and so on. In many scenarios common to human-robot interaction, however, this planning must be performed under considerable uncertainty.

Long-horizon, dexterous manipulation tasks such as furniture assembly, food preparation, and manufacturing have been a holy grail in robotics. Recent large robot action models (T eam et al., 2025; Black et al., 2024; Kim et al., 2024) have made substantial breakthroughs towards these goals by imitating expert demonstrations of diverse qualities. We provide a more comprehensive review of related work in Section 6, but highlight here a key trend: while supervised learning from demonstration, also known as behavior cloning (BC), has been applied across domains for decades (Pomerleau, 1988), its recent success in robotic manipulation has coincided with the adoption of what we term generative control policies (GCPs): robotic control policies that use generative modeling architectures, such as diffusion models, flow models, and autoregressive transformers, as parameterizations of the mapping from observation to action. Given the seemingly transformative nature of GCPs for robot learning, there has been much speculation about the origin of their superior performance relative to policies trained with a regression loss, henceforth regression control policies (RCPs). GCPs, by modeling conditional distributions over actions, are uniquely suited to the multi-task pretraining paradigm popular in today's large robotic models.

Beamforming-capable antenna arrays with many elements enable higher data rates in next generation 5G and 6G networks. In current practice, analog beamforming uses a codebook of pre-configured beams with each of them radiating towards a specific direction, and a beam management function continuously selects \textit{optimal} beams for moving user equipments (UEs). However, large codebooks and effects caused by reflections or blockages of beams make an optimal beam selection challenging. In contrast to previous work and standardization efforts that opt for supervised learning to train classifiers to predict the next best beam based on previously selected beams we formulate the problem as a partially observable Markov decision process (POMDP) and model the environment as the codebook itself. At each time step, we select a candidate beam conditioned on the belief state of the unobservable optimal beam and previously probed beams. This frames the beam selection problem as an online search procedure that locates the moving optimal beam. In contrast to previous work, our method handles new or unforeseen trajectories and changes in the physical environment, and outperforms previous work by orders of magnitude.

Leader-follower general-sum stochastic games (LF-GSSGs) model sequential decision-making under asymmetric commitment, where a leader commits to a policy and a follower best responds, yielding a strong Stackelberg equilibrium (SSE) with leader-favourable tie-breaking. This paper introduces a dynamic programming (DP) framework that applies Bellman recursion over credible sets-state abstractions formally representing all rational follower best responses under partial leader commitments-to compute SSEs. We first prove that any LF-GSSG admits a lossless reduction to a Markov decision process (MDP) over credible sets. We further establish that synthesising an optimal memoryless deterministic leader policy is NP-hard, motivating the development of ε-optimal DP algorithms with provable guarantees on leader exploitability. Experiments on standard mixed-motive benchmarks-including security games, resource allocation, and adversarial planning-demonstrate empirical gains in leader value and runtime scalability over state-of-the-art methods.

The growing exploration of Large Language Models (LLM) and Vision-Language Models (VLM) has opened avenues for enhancing the effectiveness of reinforcement learning (RL). However, existing LLM-based RL methods often focus on the guidance of control policy and encounter the challenge of limited representations of the backbone networks. To tackle this problem, we introduce Enhanced Semantic Motion Representations (Semore), a new VLM-based framework for visual RL, which can simultaneously extract semantic and motion representations through a dual-path backbone from the RGB flows. Semore utilizes VLM with common-sense knowledge to retrieve key information from observations, while using the pre-trained clip to achieve the text-image alignment, thereby embedding the ground-truth representations into the backbone. To efficiently fuse semantic and motion representations for decision-making, our method adopts a separately supervised approach to simultaneously guide the extraction of semantics and motion, while allowing them to interact spontaneously. Extensive experiments demonstrate that, under the guidance of VLM at the feature level, our method exhibits efficient and adaptive ability compared to state-of-art methods. All codes are released.

Policy-gradient methods have received increased attention recently as a mechanism for learning to act in partially observable environments. They have shown promise for problems admitting memoryless policies but have been less successful when memory is required. In this paper we develop several improved algorithms for learning policies with memory in an infinite-horizon setting -- directly when a known model of the environment is available, and via simulation otherwise. We compare these algorithms on some large POMDPs, including noisy robot navigation and multi-agent problems.