Recent advancements in aligning large language models via reinforcement learning have achieved remarkable gains in solving complex reasoning problems, but at the cost of expensive on-policy rollouts and limited exploration of diverse reasoning paths. In this work, we introduce TreePO, involving a self-guided rollout algorithm that views sequence generation as a tree-structured searching process. Composed of dynamic tree sampling policy and fixed-length segment decoding, TreePO leverages local uncertainty to warrant additional branches. By amortizing computation across common prefixes and pruning low-value paths early, TreePO essentially reduces the per-update compute burden while preserving or enhancing exploration diversity. Key contributions include: (1) a segment-wise sampling algorithm that alleviates the KV cache burden through contiguous segments and spawns new branches along with an early-stop mechanism; (2) a tree-based segment-level advantage estimation that considers both global and local proximal policy optimization. and (3) analysis on the effectiveness of probability and quality-driven dynamic divergence and fallback strategy. We empirically validate the performance gain of TreePO on a set reasoning benchmarks and the efficiency saving of GPU hours from 22\% up to 43\% of the sampling design for the trained models, meanwhile showing up to 40\% reduction at trajectory-level and 35\% at token-level sampling compute for the existing models. While offering a free lunch of inference efficiency, TreePO reveals a practical path toward scaling RL-based post-training with fewer samples and less compute. Home page locates at https://m-a-p.ai/TreePO.

Extended reality technologies are transforming fields such as healthcare, entertainment, and education, with Smart Eye-Wears (SEWs) and Artificial Intelligence (AI) playing a crucial role. However, SEWs face inherent limitations in computational power, memory, and battery life, while offloading computations to external servers is constrained by network conditions and server workload variability. To address these challenges, we propose a Federated Reinforcement Learning (FRL) framework, enabling multiple agents to train collaboratively while preserving data privacy. We implemented synchronous and asynchronous federation strategies, where models are aggregated either at fixed intervals or dynamically based on agent progress. Experimental results show that federated agents exhibit significantly lower performance variability, ensuring greater stability and reliability. These findings underscore the potential of FRL for applications requiring robust real-time AI processing, such as real-time object detection in SEWs.

Clinical decision support must adapt online under safety constraints. We present an online adaptive tool where reinforcement learning provides the policy, a patient digital twin provides the environment, and treatment effect defines the reward. The system initializes a batch-constrained policy from retrospective data and then runs a streaming loop that selects actions, checks safety, and queries experts only when uncertainty is high. Uncertainty comes from a compact ensemble of five Q-networks via the coefficient of variation of action values with a $\tanh$ compression. The digital twin updates the patient state with a bounded residual rule. The outcome model estimates immediate clinical effect, and the reward is the treatment effect relative to a conservative reference with a fixed z-score normalization from the training split. Online updates operate on recent data with short runs and exponential moving averages. A rule-based safety gate enforces vital ranges and contraindications before any action is applied. Experiments in a synthetic clinical simulator show low latency, stable throughput, a low expert query rate at fixed safety, and improved return against standard value-based baselines. The design turns an offline policy into a continuous, clinician-supervised system with clear controls and fast adaptation.

This study addresses the Min-Max Multiple Traveling Salesmen Problem ($m^3$-TSP), which aims to coordinate tours for multiple salesmen such that the length of the longest tour is minimized. Due to its NP-hard nature, exact solvers become impractical under the assumption that $P \ne NP$. As a result, learning-based approaches have gained traction for their ability to rapidly generate high-quality approximate solutions. Among these, two-stage methods combine learning-based components with classical solvers, simplifying the learning objective. However, this decoupling often disrupts consistent optimization, potentially degrading solution quality. To address this issue, we propose a novel two-stage framework named \textbf{Generate-and-Split} (GaS), which integrates reinforcement learning (RL) with an optimal splitting algorithm in a joint training process. The splitting algorithm offers near-linear scalability with respect to the number of cities and guarantees optimal splitting in Euclidean space for any given path. To facilitate the joint optimization of the RL component with the algorithm, we adopt an LSTM-enhanced model architecture to address partial observability. Extensive experiments show that the proposed GaS framework significantly outperforms existing learning-based approaches in both solution quality and transferability.

We present a novel goal-conditioned recurrent state space (GC-RSSM) model capable of learning latent dynamics of pick-and-place garment manipulation. Our proposed method LaGarNet matches the state-of-the-art performance of mesh-based methods, marking the first successful application of state-space models on complex garments. LaGarNet trains on a coverage-alignment reward and a dataset collected through a general procedure supported by a random policy and a diffusion policy learned from few human demonstrations; it substantially reduces the inductive biases introduced in the previous similar methods. We demonstrate that a single-policy LaGarNet achieves flattening on four different types of garments in both real-world and simulation settings.

Proximal Policy Optimisation (PPO) is an established and effective policy gradient algorithm used for Language Model Reinforcement Learning from Human Feedback (LM-RLHF). PPO performs well empirically but has a heuristic motivation and handles the KL-divergence constraint used in LM-RLHF in an ad-hoc manner. In this paper, we develop a a new action-value RL method for the LM-RLHF setting, KL-regularised Q-Learning (KLQ). We then show that our method is equivalent to a version of PPO in a certain specific sense, despite its very different motivation. Finally, we benchmark KLQ on two key language generation tasks -- summarisation and single-turn dialogue. We demonstrate that KLQ performs on-par with PPO at optimising the LM-RLHF objective, and achieves a consistently higher win-rate against PPO on LLM-as-a-judge evaluations.

We introduce PuzzleJAX, a GPU-accelerated puzzle game engine and description language designed to support rapid benchmarking of tree search, reinforcement learning, and LLM reasoning abilities. Unlike existing GPU-accelerated learning environments that provide hard-coded implementations of fixed sets of games, PuzzleJAX allows dynamic compilation of any game expressible in its domain-specific language (DSL). This DSL follows PuzzleScript, which is a popular and accessible online game engine for designing puzzle games. In this paper, we validate in PuzzleJAX several hundred of the thousands of games designed in PuzzleScript by both professional designers and casual creators since its release in 2013, thereby demonstrating PuzzleJAX's coverage of an expansive, expressive, and human-relevant space of tasks. By analyzing the performance of search, learning, and language models on these games, we show that PuzzleJAX can naturally express tasks that are both simple and intuitive to understand, yet often deeply challenging to master, requiring a combination of control, planning, and high-level insight.

Many existing studies have achieved significant improvements in the reasoning capabilities of large language models (LLMs) through reinforcement learning with verifiable rewards (RLVR), while the enhancement of reasoning abilities in small language models (SLMs) has not yet been sufficiently explored. Combining distilled data from larger models with RLVR on small models themselves is a natural approach, but it still faces various challenges and issues. Therefore, we propose \textit{\underline{R}}ecall-\textit{\underline{E}}xtend \textit{\underline{D}}ynamics(RED): Enhancing Small Language Models through Controlled Exploration and Refined Offline Integration. In this paper, we explore the perspective of varying exploration spaces, balancing offline distillation with online reinforcement learning. Simultaneously, we specifically design and optimize for the insertion problem within offline data. By monitoring the ratio of entropy changes in the model concerning offline and online data, we regulate the weight of offline-SFT, thereby addressing the issues of insufficient exploration space in small models and the redundancy and complexity during the distillation process. Furthermore, to tackle the distribution discrepancies between offline data and the current policy, we design a sample-accuracy-based policy shift mechanism that dynamically chooses between imitating offline distilled data and learning from its own policy.

We advance market-making strategies by integrating Adversarial Reinforcement Learning (ARL), Hawkes Processes, and variable volatility levels while also expanding the action space available to market makers (MMs). To enhance the adaptability and robustness of these strategies -- which can quote always, quote only on one side of the market or not quote at all -- we shift from the commonly used Poisson process to the Hawkes process, which better captures real market dynamics and self-exciting behaviors. We then train and evaluate strategies under volatility levels of 2 and 200. Our findings show that the 4-action MM trained in a low-volatility environment effectively adapts to high-volatility conditions, maintaining stable performance and providing two-sided quotes at least 92\% of the time. This indicates that incorporating flexible quoting mechanisms and realistic market simulations significantly enhances the effectiveness of market-making strategies.

Market making is a popular trading strategy, which aims to generate profit from the spread between the quotes posted at either side of the market. It has been shown that training market makers (MMs) with adversarial reinforcement learning allows to overcome the risks due to changing market conditions and to lead to robust performances. Prior work assumes, however, that MMs keep quoting throughout the trading process, but in practice this is not required, even for ``registered'' MMs (that only need to satisfy quoting ratios defined by the market rules). In this paper, we build on this line of work and enrich the strategy space of the MM by allowing to occasionally not quote or provide single-sided quotes. Towards this end, in addition to the MM agents that provide continuous bid-ask quotes, we have designed two new agents with increasingly richer action spaces. The first has the option to provide bid-ask quotes or refuse to quote. The second has the option to provide bid-ask quotes, refuse to quote, or only provide single-sided ask or bid quotes. We employ a model-driven approach to empirically compare the performance of the continuously quoting MM with the two agents above in various types of adversarial environments. We demonstrate how occasional refusal to provide bid-ask quotes improves returns and/or Sharpe ratios. The quoting ratios of well-trained MMs can basically meet any market requirements, reaching up to 99.9$\%$ in some cases.