Visual reasoning, a cornerstone of human intelligence, encompasses complex perceptual and logical processes essential for solving diverse visual problems. While advances in computer vision have produced powerful models for various perceptual tasks, leveraging these for general visual reasoning remains challenging. Prior work demonstrates that augmenting LLMs with vision models via supervised finetuning improves performance, but faces key limitations such as expensive data generation, reliance on careful data filtering, and poor generalization. To address these issues, we propose ReVPT to enhance multi-modal LLMs' abilities to reason about and use visual tools through reinforcement learning. We introduce a novel RL algorithm based on GRPO, designed to train models to reason with a suite of four visual tools. Through extensive experiments, we show that our method achieves state-of-the-art performance on several perception-heavy benchmarks, including SAT, CV-Bench, BLINK and MMStar, significantly outperforming the supervised and text-based RL finetuning baselines. Notably, Our ReVPT-3B and ReVPT-7B outperform the instruct models by 9.03% and 9.44% on CV-Bench. Finally, we bring to the community new insights on RL-based visual tool-usage through extensive ablations. Our code is available at https://github.com/ls-kelvin/REVPT.

Recent progress in generative modeling has highlighted the importance of Reinforcement Learning (RL) for fine-tuning, with KL-regularized methods in particular proving to be highly effective for both autoregressive and diffusion models. Complementing this line of work, the Relative Trajectory Balance (RTB) objective was recently introduced in the context of Generative Flow Networks (GFlowNets) to serve the same role of improving fine-tuning in sequential generative models. Building on prior work linking GFlowNets and maximum-entropy RL, we establish in this paper an equivalence between RTB and Trust-PCL, an off-policy RL method with KL regularization. This equivalence situates RTB within the broader theoretical landscape of KL-regularized RL, and clarifies its relationship to earlier methods. Leveraging this insight, we revisit an illustrative example from the RTB paper and show that KL-regularized RL methods achieve comparable performance, offering an alternative perspective to what was previously reported.

Lane change decision-making for autonomous vehicles is a complex but high-reward behavior. In this paper, we propose a hybrid input based deep reinforcement learning (DRL) algorithm, which realizes abstract lane change decisions and lane change actions for autonomous vehicles within traffic flow. Firstly, a surrounding vehicles trajectory prediction method is proposed to reduce the risk of future behavior of surrounding vehicles to ego vehicle, and the prediction results are input into the reinforcement learning model as additional information. Secondly, to comprehensively leverage environmental information, the model extracts feature from high-dimensional images and low-dimensional sensor data simultaneously. The fusion of surrounding vehicle trajectory prediction and multi-modal information are used as state space of reinforcement learning to improve the rationality of lane change decision. Finally, we integrate reinforcement learning macro decisions with end-to-end vehicle control to achieve a holistic lane change process. Experiments were conducted within the CARLA simulator, and the results demonstrated that the utilization of a hybrid state space significantly enhances the safety of vehicle lane change decisions.

Our work applies reinforcement learning to construct counterexamples concerning conjectured bounds on the spectral radius of the Laplacian matrix of a graph. We expand upon the re-implementation of Wagnar's approach by Stevanovic et al. with the ability to train numerous unique models simultaneously and a novel redefining of the action space to adjust the influence of the current local optimum on the learning process.

With artificial intelligence (AI) being applied to bring autonomy to decision-making in safety-critical domains such as the ones typified in the aerospace and emergency-response services, there has been a call to address the ethical implications of structuring those decisions, so they remain reliable and justifiable when human lives are at stake. This paper contributes to addressing the challenge of decision-making by proposing a structured decision-making framework as a foundational step towards responsible AI. The proposed structured decision-making framework is implemented in autonomous decision-making, specifically within disaster management. By introducing concepts of Enabler agents, Levels and Scenarios, the proposed framework's performance is evaluated against systems relying solely on judgement-based insights, as well as human operators who have disaster experience: victims, volunteers, and stakeholders. The results demonstrate that the structured decision-making framework achieves 60.94% greater stability in consistently accurate decisions across multiple Scenarios, compared to judgement-based systems. Moreover, the study shows that the proposed framework outperforms human operators with a 38.93% higher accuracy across various Scenarios. These findings demonstrate the promise of the structured decision-making framework for building more reliable autonomous AI applications in safety-critical contexts.

Reward maximization, safe exploration, and intrinsic motivation are often studied as separate objectives in reinforcement learning (RL). We present a unified geometric framework, that views these goals as instances of a single optimization problem on the space of achievable long-term behavior in an environment. Within this framework, classical methods such as policy mirror descent, natural policy gradient, and trust-region algorithms naturally generalize to nonlinear utilities and convex constraints. We illustrate how this perspective captures robustness, safety, exploration, and diversity objectives, and outline open challenges at the interface of geometry and deep RL.

Large language models often require costly optimization, such as reinforcement learning, to master complex reasoning tasks. This work demonstrates that reasoning ability, once learned, can be extracted and transferred between models as a compact task vector. We source two publicly available, identically initialized Qwen2.5 models, one fine-tuned with supervised fine-tuning (SFT) and the other with group relative policy optimization (GRPO) on the same dataset. From these, we extract a reasoning vector: $v_{\text{reason}} = θ_{\text{GRPO}} - θ_{\text{SFT}}$. We hypothesize that this vector captures the reasoning capability instilled by reinforcement learning while factoring out shared knowledge from the SFT process. When added to compatible instruction-tuned models through simple arithmetic, this vector consistently improves performance across diverse reasoning benchmarks: GSM8K (+4.9%), HumanEval (+4.3%), SciQ (+1.7%), and BigBenchHard (+12.3% for the 1.5B model). The performance improvements persist under adversarial conditions. Conversely, subtracting the vector causes significant performance degradation (-11.8% on GSM8K), demonstrating the vector's strong contribution to the model's reasoning abilities. This work shows how reasoning capabilities, typically developed through expensive training, can be extracted from existing open-source models and reused through simple tensor arithmetic, offering a practical way to enhance models by recycling prior computational investments.

Sample efficiency in the face of computationally expensive simulations is a common concern in surrogate modeling. Current strategies to minimize the number of samples needed are not as effective in simulated environments with wide state spaces. As a response to this challenge, we propose a novel method to efficiently sample simulated deterministic environments by using policies trained by Reinforcement Learning. We provide an extensive analysis of these surrogate-building strategies with respect to Latin-Hypercube sampling or Active Learning and Kriging, cross-validating performances with all sampled datasets. The analysis shows that a mixed dataset that includes samples acquired by random agents, expert agents, and agents trained to explore the regions of maximum entropy of the state transition distribution provides the best scores through all datasets, which is crucial for a meaningful state space representation. We conclude that the proposed method improves the state-of-the-art and clears the path to enable the application of surrogate-aided Reinforcement Learning policy optimization strategies on complex simulators.

Human Activity Recognition (HAR) using wearable sensors is crucial for healthcare, fitness tracking, and smart environments, yet cross-user variability -- stemming from diverse motion patterns, sensor placements, and physiological traits -- hampers generalization in real-world settings. Conventional supervised learning methods often overfit to user-specific patterns, leading to poor performance on unseen users. Existing domain generalization approaches, while promising, frequently overlook temporal dependencies or depend on impractical domain-specific labels. We propose Temporal-Preserving Reinforcement Learning Domain Generalization (TPRL-DG), a novel framework that redefines feature extraction as a sequential decision-making process driven by reinforcement learning. TPRL-DG leverages a Transformer-based autoregressive generator to produce temporal tokens that capture user-invariant activity dynamics, optimized via a multi-objective reward function balancing class discrimination and cross-user invariance. Key innovations include: (1) an RL-driven approach for domain generalization, (2) autoregressive tokenization to preserve temporal coherence, and (3) a label-free reward design eliminating the need for target user annotations. Evaluations on the DSADS and PAMAP2 datasets show that TPRL-DG surpasses state-of-the-art methods in cross-user generalization, achieving superior accuracy without per-user calibration. By learning robust, user-invariant temporal patterns, TPRL-DG enables scalable HAR systems, facilitating advancements in personalized healthcare, adaptive fitness tracking, and context-aware environments.

Quantum reinforcement learning (QRL) promises compact function approximators with access to vast Hilbert spaces, but its practical progress is slowed by NISQ-era constraints such as limited qubits and noise accumulation. We introduce a multi-chip ensemble framework using multiple small Quantum Convolutional Neural Networks (QCNNs) to overcome these constraints. Our approach partitions complex, high-dimensional observations from the Super Mario Bros environment across independent quantum circuits, then classically aggregates their outputs within a Double Deep Q-Network (DDQN) framework. This modular architecture enables QRL in complex environments previously inaccessible to quantum agents, achieving superior performance and learning stability compared to classical baselines and single-chip quantum models. The multi-chip ensemble demonstrates enhanced scalability by reducing information loss from dimensionality reduction while remaining implementable on near-term quantum hardware, providing a practical pathway for applying QRL to real-world problems.