Language models are vulnerable to short adversarial suffixes that can reliably alter predictions. Previous works usually find such suffixes with gradient search or rule-based methods, but these are brittle and often tied to a single task or model. In this paper, a reinforcement learning framework is used where the suffix is treated as a policy and trained with Proximal Policy Optimization against a frozen model as a reward oracle. Rewards are shaped using calibrated cross-entropy, removing label bias and aggregating across surface forms to improve transferability. The proposed method is evaluated on five diverse NLP benchmark datasets, covering sentiment, natural language inference, paraphrase, and commonsense reasoning, using three distinct language models: Qwen2-1.5B Instruct, TinyLlama-1.1B Chat, and Phi-1.5. Results show that RL-trained suffixes consistently degrade accuracy and transfer more effectively across tasks and models than previous adversarial triggers of similar genres.

In critical care settings such as the Intensive Care Unit, clinicians face the complex challenge of balancing conflicting objectives, primarily maximizing patient survival while minimizing resource utilization (e.g., length of stay). Single-objective Reinforcement Learning approaches typically address this by optimizing a fixed scalarized reward function, resulting in rigid policies that fail to adapt to varying clinical priorities. Multi-objective Reinforcement Learning (MORL) offers a solution by learning a set of optimal policies along the Pareto Frontier, allowing for dynamic preference selection at test time. However, applying MORL in healthcare necessitates strict offline learning from historical data. In this paper, we benchmark three offline MORL algorithms, Conditioned Conservative Pareto Q-Learning (CPQL), Adaptive CPQL, and a modified Pareto Efficient Decision Agent (PEDA) Decision Transformer (PEDA DT), against three scalarized single-objective baselines (BC, CQL, and DDQN) on the MIMIC-IV dataset. Using Off-Policy Evaluation (OPE) metrics, we demonstrate that PEDA DT algorithm offers superior flexibility compared to static scalarized baselines. Notably, our results extend previous findings on single-objective Decision Transformers in healthcare, confirming that sequence modeling architectures remain robust and effective when scaled to multi-objective conditioned generation. These findings suggest that offline MORL is a promising framework for enabling personalized, adjustable decision-making in critical care without the need for retraining.

Training end-to-end policies from image data to directly predict navigation actions for robotic systems has proven inherently difficult. Existing approaches often suffer from either the sim-to-real gap during policy transfer or a limited amount of training data with action labels. To address this problem, we introduce Vision-Language Distance (VLD) learning, a scalable framework for goal-conditioned navigation that decouples perception learning from policy learning. Instead of relying on raw sensory inputs during policy training, we first train a self-supervised distance-to-goal predictor on internet-scale video data. This predictor generalizes across both image- and text-based goals, providing a distance signal that can be minimized by a reinforcement learning (RL) policy. The RL policy can be trained entirely in simulation using privileged geometric distance signals, with injected noise to mimic the uncertainty of the trained distance predictor. At deployment, the policy consumes VLD predictions, inheriting semantic goal information-"where to go"-from large-scale visual training while retaining the robust low-level navigation behaviors learned in simulation. We propose using ordinal consistency to assess distance functions directly and demonstrate that VLD outperforms prior temporal distance approaches, such as ViNT and VIP. Experiments show that our decoupled design achieves competitive navigation performance in simulation while supporting flexible goal modalities, providing an alternative and, most importantly, scalable path toward reliable, multimodal navigation policies.

Current autonomous driving systems often favor end-to-end frameworks, which take sensor inputs like images and learn to map them into trajectory space via neural networks. Previous work has demonstrated that models can achieve better planning performance when provided with a prior distribution of possible trajectories. However, these approaches often overlook two critical aspects: 1) The appropriate trajectory prior can vary significantly across different driving scenarios. 2) Their trajectory evaluation mechanism lacks policy-driven refinement, remaining constrained by the limitations of one-stage supervised training. To address these issues, we explore improvements in two key areas. For problem 1, we employ MoE to apply different trajectory priors tailored to different scenarios. For problem 2, we utilize Reinforcement Learning to fine-tune the trajectory scoring mechanism. Additionally, we integrate models with different perception backbones to enhance perceptual features. Our integrated model achieved a score of 51.08 on the navsim ICCV benchmark, securing third place.

For centuries, khalasi (Gujarati for sailor) have skillfully harnessed ocean currents to navigate vast waters with minimal effort. Emulating this intuition in autonomous systems remains a significant challenge, particularly for Autonomous Surface Vehicles tasked with long duration missions under strict energy budgets. In this work, we present a learning-based approach for energy-efficient surface vehicle navigation in vortical flow fields, where partial observability often undermines traditional path-planning methods. We present an end to end reinforcement learning framework based on Soft Actor Critic that learns flow-aware navigation policies using only local velocity measurements. Through extensive evaluation across diverse and dynamically rich scenarios, our method demonstrates substantial energy savings and robust generalization to previously unseen flow conditions, offering a promising path toward long term autonomy in ocean environments. The navigation paths generated by our proposed approach show an improvement in energy conservation 30 to 50 percent compared to the existing state of the art techniques.

Our research investigates the complex task of multiple autonomous agents learning to coordinate and deliver packages in warehouse environments--a problem requiring implicit communication, collision avoidance, and efficient task allocation without centralized control. Traditional warehouse automation relies on centralized planning systems that face scalability limitations; multi-agent reinforcement learning (MARL) offers an alternative through decentralized learned policies, but requires solving the credit assignment problem. We compare MARL algorithms on warehouse coordination: QMIX [Rashid et al., 2018] (value decomposition), IPPO (independent learning), and MASAC (centralized critic). Our study progresses from MPE for validation to RWARE for warehouse evaluation, culminating in Unity 3D deployment where agents demonstrate learned package delivery behavior. QMIX emerged as the best performer after systematic comparison. Our contributions: (1) hyperparameter analysis showing default configurations fail on sparse-reward warehouse tasks, (2) comparative evaluation across algorithms and scales, (3) Unity ML-Agents integration demonstrating sim-to-sim transfer with successful package delivery, and (4) identification of scaling challenges. Full experimental details and results are documented in our Quarto documentation book. 1

Training Large Language Models (LLMs) to reason often relies on Reinforcement Learning (RL) with task-specific verifiers. However, many real-world reasoning-intensive tasks lack verifiers, despite offering abundant expert demonstrations that remain under-utilized for reasoning-focused training. We introduce RARO (Relativistic Adversarial Reasoning Optimization) that learns strong reasoning capabilities from only expert demonstrations via Inverse Reinforcement Learning. Our method sets up an adversarial game between a policy and a relativistic critic: the policy learns to mimic expert answers, while the critic aims to identify the experts among (expert, policy) answer pairs. Both the policy and the critic are trained jointly and continuously via RL, and we identify the key stabilization techniques required for robust learning. Empirically, RARO significantly outperforms strong verifier-free baselines on all of our evaluation tasks -- Countdown, DeepMath, and Poetry Writing -- and enjoys the same robust scaling trends as RL with verifiers. These results demonstrate that our method effectively elicits strong reasoning performance from expert demonstrations alone, enabling robust reasoning learning even when task-specific verifiers are unavailable. Recent advances in Large Language Models (LLMs) have been driven substantially by improvements in their reasoning abilities. Reasoning enables LLMs to perform deliberate intermediate computations before producing answers to the user queries, proposing candidate solutions and self-corrections. Much of this progress has been enabled via Reinforcement Learning (RL) on verifiable tasks such as mathematics and competitive programming (DeepSeek-AI et al., 2025; Y ang et al., 2025a; Shao et al., 2024; Luo et al., 2025). Notably, recent work has demonstrated that RL with V erifiable Rewards (RL VR) can enable LLMs to develop robust reasoning capabilities without any additional supervision (DeepSeek-AI et al., 2025). A growing body of work further improves the efficiency and stability of such RL algorithms on verifiable tasks, such as DAPO (Y u et al., 2025) and GSPO (Zheng et al., 2025). However, comparatively little attention has been paid to developing reasoning abilities on non-verifiable tasks, where task-specific verifiers are unavailable. Y et, in many impactful and challenging tasks -- such as analytical writing, open-ended research, or financial analysis -- LLM outputs are not directly verifiable due to hard-to-specify criteria, wide variation among acceptable answers, and other practical constraints. A popular approach in these settings is Reinforcement Learning from Human Feedback (RLHF) (Ouyang et al., 2022; Rafailov et al., 2023), but they require collecting human preferences beyond demonstration data, which is often a time-consuming and expensive process.

We study the problem of efficiently estimating policies that simultaneously optimize multiple objectives in reinforcement learning (RL). Given $n$ objectives (or tasks), we seek the optimal partition of these objectives into $k \ll n$ groups, where each group comprises related objectives that can be trained together. This problem arises in applications such as robotics, control, and preference optimization in language models, where learning a single policy for all $n$ objectives is suboptimal as $n$ grows. We introduce a two-stage procedure -- meta-training followed by fine-tuning -- to address this problem. We first learn a meta-policy for all objectives using multitask learning. Then, we adapt the meta-policy to multiple randomly sampled subsets of objectives. The adaptation step leverages a first-order approximation property of well-trained policy networks, which is empirically verified to be accurate within a 2% error margin across various RL environments. The resulting algorithm, PolicyGradEx, efficiently estimates an aggregate task-affinity score matrix given a policy evaluation algorithm. Based on the estimated affinity score matrix, we cluster the $n$ objectives into $k$ groups by maximizing the intra-cluster affinity scores. Experiments on three robotic control and the Meta-World benchmarks demonstrate that our approach outperforms state-of-the-art baselines by 16% on average, while delivering up to $26\times$ faster speedup relative to performing full training to obtain the clusters. Ablation studies validate each component of our approach. For instance, compared with random grouping and gradient-similarity-based grouping, our loss-based clustering yields an improvement of 19%. Finally, we analyze the generalization error of policy networks by measuring the Hessian trace of the loss surface, which gives non-vacuous measures relative to the observed generalization errors.

A fully autonomous humanoid returns machine-fed shuttles in a motion-capture arena; overlaid arcs show an incoming (blue) and returned (orange) trajectory. Abstract--Humanoid robots have demonstrated strong capabilities for interacting with static scenes across locomotion, manipulation, and more challenging loco-manipulation tasks. Y et the real world is dynamic, and quasi-static interactions are insufficient to cope with diverse environmental conditions. As a step toward more dynamic interaction scenarios, we present a reinforcement-learning-based training pipeline that produces a unified whole-body controller for humanoid badminton, enabling coordinated lower-body footwork and upper-body striking without motion priors or expert demonstrations. Training follows a three-stage curriculum--first footwork acquisition, then precision-guided racket swing generation, and finally task-focused refinement--yielding motions in which both legs and arms serve the hitting objective. For deployment, we incorporate an Extended Kalman Filter (EKF) to estimate and predict shuttlecock trajectories for target striking. We also introduce a prediction-free variant that dispenses with EKF and explicit trajectory prediction. T o validate the framework, we conduct five sets of experiments in both simulation and the real world. In simulation, two robots sustain a rally of 21 consecutive hits. Moreover, the prediction-free variant achieves successful hits with comparable performance relative to the target-known policy. In real-world tests, both prediction and controller modules exhibit high accuracy, and on-court hitting achieves an outgoing shuttle speed up to 19.1 m/s with a mean return landing distance of 4 m. These experimental results show that our proposed training scheme can deliver highly dynamic while precise goal striking in badminton, and can be adapted to more dynamics-critical domains. Humanoid platforms have been proposed as general-purpose embodied agents for human-compatible skills [1, 2, 3, 4, 5, 6, 7]. Despite rapid progress in locomotion and motion imitation, agile, contact-rich interactions with fast-moving objects under tight reaction windows remain underexplored.

Reinforcement learning (RL) has been recognized as a powerful tool for robot control tasks. RL typically employs reward functions to define task objectives and guide agent learning. However, since the reward function serves the dual purpose of defining the optimal goal and guiding learning, it is challenging to design the reward function manually, which often results in a suboptimal task representation. To tackle the reward design challenge in RL, inspired by the satisficing theory, we propose a Test-driven Reinforcement Learning (TdRL) framework. In the TdRL framework, multiple test functions are used to represent the task objective rather than a single reward function. Test functions can be categorized as pass-fail tests and indicative tests, each dedicated to defining the optimal objective and guiding the learning process, respectively, thereby making defining tasks easier. Building upon such a task definition, we first prove that if a trajectory return function assigns higher returns to trajectories closer to the optimal trajectory set, maximum entropy policy optimization based on this return function will yield a policy that is closer to the optimal policy set. Then, we introduce a lexicographic heuristic approach to compare the relative distance relationship between trajectories and the optimal trajectory set for learning the trajectory return function. Furthermore, we develop an algorithm implementation of TdRL. Experimental results on the DeepMind Control Suite benchmark demonstrate that TdRL matches or outperforms handcrafted reward methods in policy training, with greater design simplicity and inherent support for multi-objective optimization. We argue that TdRL offers a novel perspective for representing task objectives, which could be helpful in addressing the reward design challenges in RL applications.