Building agents that autonomously operate mobile devices has attracted increasing attention. While Vision-Language Models (VLMs) show promise, most existing approaches rely on direct state-to-action mappings, which lack structured reasoning and planning, and thus generalize poorly to novel tasks or unseen UI layouts. We introduce Hi-Agent, a trainable hierarchical vision-language agent for mobile control, featuring a high-level reasoning model and a low-level action model that are jointly optimized. For efficient training, we reformulate multi-step decision-making as a sequence of single-step subgoals and propose a foresight advantage function, which leverages execution feedback from the low-level model to guide high-level optimization. This design alleviates the path explosion issue encountered by Group Relative Policy Optimization (GRPO) in long-horizon tasks and enables stable, critic-free joint training. Hi-Agent achieves a new State-Of-The-Art (SOTA) 87.9% task success rate on the Android-in-the-Wild (AitW) benchmark, significantly outperforming prior methods across three paradigms: prompt-based (AppAgent: 17.7%), supervised (Filtered BC: 54.5%), and reinforcement learning-based (DigiRL: 71.9%). It also demonstrates competitive zero-shot generalization on the ScreenSpot-v2 benchmark. On the more challenging AndroidWorld benchmark, Hi-Agent also scales effectively with larger backbones, showing strong adaptability in high-complexity mobile control scenarios.

In this work, we study risk-aware reinforcement learning for quadrupedal locomotion. Our approach trains a family of risk-conditioned policies using a Conditional Value-at-Risk (CVaR) constrained policy optimization technique that provides improved stability and sample efficiency. At deployment, we adaptively select the best performing policy from the family of policies using a multi-armed bandit framework that uses only observed episodic returns, without any privileged environment information, and adapts to unknown conditions on the fly. Hence, we train quadrupedal locomotion policies at various levels of robustness using CVaR and adaptively select the desired level of robustness online to ensure performance in unknown environments. We evaluate our method in simulation across eight unseen settings (by changing dynamics, contacts, sensing noise, and terrain) and on a Unitree Go2 robot in previously unseen terrains. Our risk-aware policy attains nearly twice the mean and tail performance in unseen environments compared to other baselines and our bandit-based adaptation selects the best-performing risk-aware policy in unknown terrain within two minutes of operation.

Iterative retrieval-augmented generation (RAG) enables large language models to answer complex multi-hop questions, but each additional loop increases latency, costs, and the risk of introducing distracting evidence, motivating the need for an efficient stopping strategy. Existing methods either use a predetermined number of iterations or rely on confidence proxies that poorly reflect whether more retrieval will actually help. We cast iterative RAG as a finite-horizon Markov decision process and introduce Stop-RAG, a value-based controller that adaptively decides when to stop retrieving. Trained with full-width forward-view Q($λ$) targets from complete trajectories, Stop-RAG learns effective stopping policies while remaining compatible with black-box APIs and existing pipelines. On multi-hop question-answering benchmarks, Stop-RAG consistently outperforms both fixed-iteration baselines and prompting-based stopping with LLMs. These results highlight adaptive stopping as a key missing component in current agentic systems, and demonstrate that value-based control can improve the accuracy of RAG systems.

Abstract-- Domain adaptation methods aim to bridge the gap between datasets by enabling knowledge transfer across domains, reducing the need for additional expert annotations. However, many approaches struggle with reliability in the target domain, an issue particularly critical in medical image segmentation, where accuracy and anatomical validity are essential. This challenge is further exacerbated in spatio-temporal data, where the lack of temporal consistency can significantly degrade segmentation quality, and particularly in echocardiography, where the presence of artifacts and noise can further hinder segmentation performance. To address these issues, we present RL4Seg3D, an unsupervised domain adaptation framework for 2D + time echocardiography segmentation. RL4Seg3D integrates novel reward functions and a fusion scheme to enhance key landmark precision in its segmentations while processing full-sized input videos. By leveraging reinforcement learning for image segmentation, our approach improves accuracy, anatomical validity, and temporal consistency while also providing, as a beneficial side effect, a robust uncertainty estimator, which can be used at test time to further enhance segmentation performance. We demonstrate the effectiveness of our framework on over 30,000 echocardiographic videos, showing that it outperforms standard domain adaptation techniques without the need for any labels on the target domain. Obtaining such annotations is laborious, logistically challenging, and expensive, in particular for 3D images or 2D+t image sequences. This has driven the development of semi-supervised and unsupervised domain adaption methods to leverage larger datasets containing few or no annotations [1]. Reinforcement learning (RL) offers an alternative to conventional supervised training by leveraging automated reward mechanisms to iteratively improve model outputs. We recently proposed a RL-based segmentation strategy (RL4Seg) [2] framing 2D segmentation as a single-timestep RL task, in which a segmentation network acts as an agent, and is optimized through reward-driven interactions with unlabeled data.

For example, a recent study [1] concludes that NPCs based on behavior trees (BTs) are still more viable than those based on machine learning (ML), calling for new approaches, strategies, and tooling to overcome the barrier to adoption. Additional work has also underscored the need for reusable and adjustable models [2], motivated by game developers' preferences to reuse previously developed assets, provided that reuse does not result in repetitive gameplay. Traditional BT approaches and modern RL techniques each have their respective strengths and limitations in video game development. BTs offer a structured and hierarchical method for managing NPC behaviors, enabling the design of complex systems with predictable outcomes given sufficient development time. However, this complexity can make multi-task BTs less engaging and cumbersome to develop [2]. Conversely, RL provides a dynamic and adaptive approach to decision making [3], allowing developers to guide an agent through trial-and-error. However, training generally-capable RL models remains a challenge, particularly due to reward shaping, negative task transfer [4, 5], and compute resource demands [6].

Abstract--T ask and motion planning (T AMP) for robotics manipulation necessitates long-horizon reasoning involving versatile actions and skills. While deterministic actions can be crafted by sampling or optimizing with certain constraints, planning actions with uncertainty, i.e., probabilistic actions, remains a challenge for T AMP . On the contrary, Reinforcement Learning (RL) excels in acquiring versatile, yet short-horizon, manipulation skills that are robust with uncertainties. Besides the policy, a RL skill is defined with data-driven logical components that enable the skill to be deployed by symbolic planning. A plan refinement sub-routine is designed to further tackle the inevitable effect uncertainties. In the experiments, we compare our method with baseline hierarchical planning from both T AMP and RL fields and illustrate the strength of the method. The results show that by embedding RL skills, we extend the capability of T AMP to domains with probabilistic skills, and improve the planning efficiency compared to the previous methods. Reinforcement Learning (RL) empowers robots to acquire manipulation skills without human programming. However, prior works mostly tackle single-skill or short-term manipulation tasks, such as grasping [1] or peg insertion [2] or synergies between two actions [3]. The long-horizon manipulation planning remains a challenge in the RL field because of expanding state/action spaces and sparse rewards etc [4].

Abstract--The growing demand for parking has increased the need for automated parking planning methods that can operate reliably in confined spaces. In restricted and complex environments, high-precision maneuvers are required to achieve a high success rate in planning, yet existing approaches often rely on explicit action modeling, which faces challenges when accurately modeling the optimal action distribution. In this paper, we propose DRIP, a diffusion-refined planner anchored in reinforcement learning (RL) prior action distribution, in which an RL-pretrained policy provides prior action distributions to regularize the diffusion training process. By steering the denoising trajectory through the reinforcement learning prior distribution during training, the diffusion model inherits a well-informed initialization, resulting in more accurate action modeling, a higher planning success rate, and reduced inference steps. We evaluate our approach across parking scenarios with varying degrees of spatial constraints. Experimental results demonstrate that our method significantly improves planning performance in confined-space parking environments while maintaining strong generalization in common scenarios.

Pretraining a policy on offline data followed by fine-tuning through online interactions, known as Offline-to-Online Reinforcement Learning (O2O RL), has emerged as a promising paradigm for real-world RL deployment. However, both offline datasets and online interactions in practical environments are often noisy or even maliciously corrupted, severely degrading the performance of O2O RL. Existing works primarily focus on mitigating the conservatism of offline policies via online exploration, while the robustness of O2O RL under data corruption, including states, actions, rewards, and dynamics, is still unexplored. In this work, we observe that data corruption induces heavy-tailed behavior in the policy, thereby substantially degrading the efficiency of online exploration. To address this issue, we incorporate Inverse Probability Weighted (IPW) into the online exploration policy to alleviate heavy-tailedness, and propose a novel, simple yet effective method termed $\textbf{RPEX}$: $\textbf{R}$obust $\textbf{P}$olicy $\textbf{EX}$pansion. Extensive experimental results on D4RL datasets demonstrate that RPEX achieves SOTA O2O performance across a wide range of data corruption scenarios. Code is available at $\href{https://github.com/felix-thu/RPEX}{https://github.com/felix-thu/RPEX}$.

In this work, we present a simple yet theoretically motivated improvement to Supervised Fine-Tuning (SFT) for the Large Language Model (LLM), addressing its limited generalization compared to reinforcement learning (RL). Through mathematical analysis, we reveal that standard SFT gradients implicitly encode a problematic reward structure that may severely restrict the generalization capabilities of model. To rectify this, we propose Dynamic Fine-Tuning (DFT), stabilizing gradient updates for each token by dynamically rescaling the objective function with the probability of this token. With just a single-line change, the method outperforms standard SFT on multiple difficult benchmarks and base models, from math reasoning to code generation and multi-modal tasks, demonstrating improved generalization. Additionally, DFT achieves competitive results in offline RL settings, providing an effective yet streamlined alternative. Supervised Fine-Tuning (SFT), which adapts models to expert demonstrations, has become the standard post-training paradigm for Large Language Models (LLMs). It enables efficient task adaptation and capability enhancement (Chung et al., 2024; Zhang et al., 2024b; Sanh et al., 2022; Ouyang et al., 2022), and is popular for its ease of implementation and rapid acquisition of expert-like behaviors (Wei et al., 2022; Zhou et al., 2023). Despite these advantages, SFT often shows limited generalization compared to reinforcement learning (RL) (Chu et al., 2024; Ouyang et al., 2022; Christiano et al., 2017; Bai et al., 2022; Huan et al., 2025; Swamy et al., 2025). RL leverages explicit reward or verification signals to explore diverse strategies and thus generalizes better. However, RL requires substantial computation, careful hyperparameter tuning, and explicit reward signals--conditions often impractical in real-world settings (Schulman et al., 2017; Ouyang et al., 2022; Sheng et al., 2025; Strubell et al., 2019; Liu & Yin, 2024; Winsta, 2025). Moreover, RL can struggle to recover expert-like behaviors that SFT captures efficiently (Mandlekar et al., 2022; Chen et al., 2025b).

Recent studies show that the reasoning capabilities of Large Language Models (LLMs) can be improved by applying Reinforcement Learning (RL) to question-answering (QA) tasks in areas such as math and coding. With a long context length, LLMs may learn to perform search, as indicated by the self-correction behavior observed in DeepSeek R1. However, this search behavior is often imprecise and lacks confidence, resulting in long, redundant responses and highlighting deficiencies in intuition and verification. Inspired by the Dual Process Theory in psychology, we introduce a simple modification to the QA task that includes four stages: Fast Thinking, where the LLM must answer within a strict token budget; Verification, where the model evaluates its initial response; Slow Thinking, where it refines the initial response with more deliberation; and Summarization, where it distills the refinement from the previous stage into precise steps. Our proposed task improves average accuracy from 25.6% to 27.3% for Qwen2.5-1.5B, and from 45.9% to 51.0% for DeepSeek-R1-Qwen-1.5B. Notably, for Qwen2.5-1.5B, the Fast Thinking mode alone achieves 25.2% accuracy using fewer than 1000 tokens, demonstrating substantial inference efficiency gains. These findings suggest that intuition and deliberative reasoning are distinct, complementary systems benefiting from targeted training. Additionally, we have open-sourced both the trained models and the source code.