Vision-language model (VLM) fine-tuning for application-specific visual grounding based on natural language instructions has become one of the most popular approaches for learning-enabled autonomous systems. However, such fine-tuning relies heavily on high-quality datasets to achieve successful performance in various downstream tasks. Additionally, VLMs often encounter limitations due to insufficient and imbalanced fine-tuning data. To address these issues, we propose a new generalizable framework to improve VLM fine-tuning by integrating it with a reinforcement learning (RL) agent. Our method utilizes the RL agent to manipulate objects within an indoor setting to create synthetic data for fine-tuning to address certain vulnerabilities of the VLM. Specifically, we use the performance of the VLM to provide feedback to the RL agent to generate informative data that efficiently fine-tune the VLM over the targeted task (e.g. spatial reasoning). The key contribution of this work is developing a framework where the RL agent serves as an informative data sampling tool and assists the VLM in order to enhance performance and address task-specific vulnerabilities. By targeting the data sampling process to address the weaknesses of the VLM, we can effectively train a more context-aware model. In addition, generating synthetic data allows us to have precise control over each scene and generate granular ground truth captions. Our results show that the proposed data generation approach improves the spatial reasoning performance of VLMs, which demonstrates the benefits of using RL-guided data generation in vision-language tasks.

Reinforcement Learning (RL) (Sutton and Barto, We tackle average-reward infinite-horizon 2018) tackles the sequential decision-making problem POMDPs with an unknown transition model of an agent interacting with an unknown or partially but a known observation model, a setting known environment with the goal of maximizing the that has been previously addressed in two long-term sum of rewards. The RL agent should tradeoff limiting ways: (i) frequentist methods relying between exploring the environment to learn its on suboptimal stochastic policies having structure and exploiting the estimates to compute a a minimum probability of choosing each action, policy that maximizes the reward. This problem has and (ii) Bayesian approaches employing been successfully addressed in past works under the the optimal policy class but requiring MDP formulation (Bartlett and Tewari, 2009; Jaksch strong assumptions about the consistency et al., 2010; Zanette and Brunskill, 2019). MDPs assume of employed estimators. Our work removes full observability of the state space but this assumption these limitations by proving convenient estimation is often violated in many real-world scenarios guarantees for the transition model such as robotics or finance, where only a partial and introducing an optimistic algorithm that observation of the environment is available. In this leverages the optimal class of deterministic case, it is more appropriate to model the problem using belief-based policies. We introduce modifications Partially-Observable MDPs (Sondik, 1978).

Hybrid Group Relative Policy Optimization (Hybrid GRPO) is a reinforcement learning framework that extends Proximal Policy Optimization (PPO) and Group Relative Policy Optimization (GRPO) by incorporating empirical multi-sample action evaluation while preserving the stability of value function-based learning. Unlike DeepSeek GRPO, which eliminates the value function in favor of purely empirical reward estimation, Hybrid GRPO introduces a structured advantage computation method that balances empirical action sampling with bootstrapped value estimation. This approach enhances sample efficiency, improves learning stability, and mitigates variance amplification observed in purely empirical methods. A detailed mathematical comparison between PPO, DeepSeek GRPO, and Hybrid GRPO is presented, highlighting key differences in advantage estimation and policy updates. Experimental validation in a controlled reinforcement learning environment demonstrates that Hybrid GRPO achieves superior convergence speed, more stable policy updates, and improved sample efficiency compared to existing methods. Several extensions to Hybrid GRPO are explored, including entropy-regularized sampling, hierarchical multi-step sub-sampling, adaptive reward normalization, and value-based action selection. Beyond reinforcement learning in simulated environments, Hybrid GRPO provides a scalable framework for bridging the gap between large language models (LLMs) and real-world agent-based decision-making. By integrating structured empirical sampling with reinforcement learning stability mechanisms, Hybrid GRPO has potential applications in autonomous robotics, financial modeling, and AI-driven control systems. These findings suggest that Hybrid GRPO serves as a robust and adaptable reinforcement learning methodology, paving the way for further advancements in policy optimization.

We address the problem of best policy identification in preference-based reinforcement learning (PbRL), where learning occurs from noisy binary preferences over trajectory pairs rather than explicit numerical rewards. This approach is useful for post-training optimization of generative AI models during multi-turn user interactions, where preference feedback is more robust than handcrafted reward models. In this setting, learning is driven by both an offline preference dataset -- collected from a rater of unknown 'competence' -- and online data collected with pure exploration. Since offline datasets may exhibit out-of-distribution (OOD) biases, principled online data collection is necessary. To address this, we propose Posterior Sampling for Preference Learning ($\mathsf{PSPL}$), a novel algorithm inspired by Top-Two Thompson Sampling, that maintains independent posteriors over the true reward model and transition dynamics. We provide the first theoretical guarantees for PbRL in this setting, establishing an upper bound on the simple Bayesian regret of $\mathsf{PSPL}$. Since the exact algorithm can be computationally impractical, we also provide an approximate version that outperforms existing baselines.

Large Language Models (LLMs) have demonstrated remarkable capabilities in complex reasoning tasks, yet generating reliable reasoning processes remains a significant challenge. We present a unified probabilistic framework that formalizes LLM reasoning through a novel graphical model incorporating latent thinking processes and evaluation signals. Within this framework, we introduce the Bootstrapping Reinforced Thinking Process (BRiTE) algorithm, which works in two steps. First, it generates high-quality rationales by approximating the optimal thinking process through reinforcement learning, using a novel reward shaping mechanism. Second, it enhances the base LLM by maximizing the joint probability of rationale generation with respect to the model's parameters. Theoretically, we demonstrate BRiTE's convergence at a rate of $1/T$ with $T$ representing the number of iterations. Empirical evaluations on math and coding benchmarks demonstrate that our approach consistently improves performance across different base models without requiring human-annotated thinking processes. In addition, BRiTE demonstrates superior performance compared to existing algorithms that bootstrap thinking processes use alternative methods such as rejection sampling, and can even match or exceed the results achieved through supervised fine-tuning with human-annotated data.

Designing learning agents that explore efficiently in a complex environment has been widely recognized as a fundamental challenge in reinforcement learning. While a number of works have demonstrated the effectiveness of techniques based on randomized value functions on a single agent, it remains unclear, from a theoretical point of view, whether injecting randomization can help a society of agents {\it concurently} explore an environment. The theoretical results %that we established in this work tender an affirmative answer to this question. We adapt the concurrent learning framework to \textit{randomized least-squares value iteration} (RLSVI) with \textit{aggregated state representation}. We demonstrate polynomial worst-case regret bounds in both finite- and infinite-horizon environments. In both setups the per-agent regret decreases at an optimal rate of $\Theta\left(\frac{1}{\sqrt{N}}\right)$, highlighting the advantage of concurent learning. Our algorithm exhibits significantly lower space complexity compared to \cite{russo2019worst} and \cite{agrawal2021improved}. We reduce the space complexity by a factor of $K$ while incurring only a $\sqrt{K}$ increase in the worst-case regret bound, compared to \citep{agrawal2021improved,russo2019worst}. Additionally, we conduct numerical experiments to demonstrate our theoretical findings.

This paper argues that model-free reinforcement learning (RL) agents, while lacking explicit planning mechanisms, exhibit behaviours that can be analogised to System 1 ("thinking fast") processes in human cognition. Unlike model-based RL agents, which operate akin to System 2 ("thinking slow") reasoning by leveraging internal representations for planning, model-free agents react to environmental stimuli without anticipatory modelling. We propose a novel framework linking the dichotomy of System 1 and System 2 to the distinction between model-free and model-based RL. This framing challenges the prevailing assumption that intentionality and purposeful behaviour require planning, suggesting instead that intentionality can manifest in the structured, reactive behaviours of model-free agents. By drawing on interdisciplinary insights from cognitive psychology, legal theory, and experimental jurisprudence, we explore the implications of this perspective for attributing responsibility and ensuring AI safety. These insights advocate for a broader, contextually informed interpretation of intentionality in RL systems, with implications for their ethical deployment and regulation.

Control of distributed parameter systems affected by delays is a challenging task, particularly when the delays depend on spatial variables. The idea of integrating analytical control theory with learning-based control within a unified control scheme is becoming increasingly promising and advantageous. In this paper, we address the problem of controlling an unstable first-order hyperbolic PDE with spatially-varying delays by combining PDE backstepping control strategies and deep reinforcement learning (RL). To eliminate the assumption on the delay function required for the backstepping design, we propose a soft actor-critic (SAC) architecture incorporating a DeepONet to approximate the backstepping controller. The DeepONet extracts features from the backstepping controller and feeds them into the policy network. In simulations, our algorithm outperforms the baseline SAC without prior backstepping knowledge and the analytical controller.

Autonomous systems are increasingly expected to operate in the presence of adversaries, though an adversary may infer sensitive information simply by observing a system, without even needing to interact with it. Therefore, in this work we present a deceptive decision-making framework that not only conceals sensitive information, but in fact actively misleads adversaries about it. We model autonomous systems as Markov decision processes, and we consider adversaries that attempt to infer their reward functions using inverse reinforcement learning. To counter such efforts, we present two regularization strategies for policy synthesis problems that actively deceive an adversary about a system's underlying rewards. The first form of deception is ``diversionary'', and it leads an adversary to draw any false conclusion about what the system's reward function is. The second form of deception is ``targeted'', and it leads an adversary to draw a specific false conclusion about what the system's reward function is. We then show how each form of deception can be implemented in policy optimization problems, and we analytically bound the loss in total accumulated reward that is induced by deception. Next, we evaluate these developments in a multi-agent sequential decision-making problem with one real agent and multiple decoys. We show that diversionary deception can cause the adversary to believe that the most important agent is the least important, while attaining a total accumulated reward that is $98.83\%$ of its optimal, non-deceptive value. Similarly, we show that targeted deception can make any decoy appear to be the most important agent, while still attaining a total accumulated reward that is $99.25\%$ of its optimal, non-deceptive value.

Contextual sequential decision-making problems play a crucial role in machine learning, encompassing a wide range of downstream applications such as bandits, sequential hypothesis testing and online risk control. These applications often require different statistical measures, including expectation, variance and quantiles. In this paper, we provide a universal admissible algorithm framework for dealing with all kinds of contextual online decision-making problems that directly learns the whole underlying unknown distribution instead of focusing on individual statistics. This is much more difficult because the dimension of the regression is uncountably infinite, and any existing linear contextual bandits algorithm will result in infinite regret. To overcome this issue, we propose an efficient infinite-dimensional functional regression oracle for contextual cumulative distribution functions (CDFs), where each data point is modeled as a combination of context-dependent CDF basis functions. Our analysis reveals that the decay rate of the eigenvalue sequence of the design integral operator governs the regression error rate and, consequently, the utility regret rate. Specifically, when the eigenvalue sequence exhibits a polynomial decay of order $\frac{1}{\gamma}\ge 1$, the utility regret is bounded by $\tilde{\mathcal{O}}\Big(T^{\frac{3\gamma+2}{2(\gamma+2)}}\Big)$. By setting $\gamma=0$, this recovers the existing optimal regret rate for contextual bandits with finite-dimensional regression and is optimal under a stronger exponential decay assumption. Additionally, we provide a numerical method to compute the eigenvalue sequence of the integral operator, enabling the practical implementation of our framework.