Preserving the privacy of preferences (or rewards) of a sequential decision-making agent when decisions are observable is crucial in many physical and cybersecurity domains. For instance, in wildlife monitoring, agents must allocate patrolling resources without revealing animal locations to poachers. This paper addresses privacy preservation in planning over a sequence of actions in MDPs, where the reward function represents the preference structure to be protected. Observers can use Inverse RL (IRL) to learn these preferences, making this a challenging task. Current research on differential privacy in reward functions fails to ensure guarantee on the minimum expected reward and offers theoretical guarantees that are inadequate against IRL-based observers. To bridge this gap, we propose a novel approach rooted in the theory of deception. Deception includes two models: dissimulation (hiding the truth) and simulation (showing the wrong). Our first contribution theoretically demonstrates significant privacy leaks in existing dissimulation-based methods. Our second contribution is a novel RL-based planning algorithm that uses simulation to effectively address these privacy concerns while ensuring a guarantee on the expected reward. Experiments on multiple benchmark problems show that our approach outperforms previous methods in preserving reward function privacy.
Bayesian optimization (BO) is a sequential decision-making tool widely used for optimizing expensive black-box functions. Recently, Large Language Models (LLMs) have shown remarkable adaptability in low-data regimes, making them promising tools for black-box optimization by leveraging contextual knowledge to propose high-quality query points. However, LLMs lack explicit surrogate modeling and calibrated uncertainty, and their internal mechanisms remain opaque. This structural opacity can result in unnecessary experiments, which are costly in BO and may prevent the search from approaching the optimum. To address this, we propose LLINBO: LLM-in-the-Loop BO, a general framework that wraps existing LLM-based BO approaches, deciding when and how to act on their suggestions and when to revert to classical BO. We introduce three mechanisms for this interaction, each accompanied by a no-harm theoretical guarantee that ensures sublinear cumulative regret for common kernels while requiring no performance assumptions on the LLM. Beyond these theoretical guarantees, we show through simulation studies that LLINBO makes existing LLM-based BO more reliable while still exploiting the LLM's few-shot learning ability to outperform classical BO. We further demonstrate its practical use on a real-world, expensive 3D printing experiment. Code is available at https://github.com/UMDataScienceLab/LLM-in-the-Loop-BO.
We introduce the framework of LLM-Stackelberg games, a class of sequential decision-making models that integrate large language models (LLMs) into strategic interactions between a leader and a follower. Departing from classical Stackelberg assumptions of complete information and rational agents, our formulation allows each agent to reason through structured prompts, generate probabilistic behaviors via LLMs, and adapt their strategies through internal cognition and belief updates. We define two equilibrium concepts: reasoning and behavioral equilibrium, which aligns an agent's internal prompt-based reasoning with observable behavior, and conjectural reasoning equilibrium, which accounts for epistemic uncertainty through parameterized models over an opponent's response. These layered constructs capture bounded rationality, asymmetric information, and meta-cognitive adaptation. We illustrate the framework through a spearphishing case study, where a sender and a recipient engage in a deception game using structured reasoning prompts. This example highlights the cognitive richness and adversarial potential of LLM-mediated interactions. Our results show that LLM-Stackelberg games provide a powerful paradigm for modeling decision-making in domains such as cybersecurity, misinformation, and recommendation systems.
Operating Large Language Models (LLMs) on edge devices is increasingly challenged by limited communication bandwidth and strained computational and memory costs. Thus, cloud-assisted remote fine-tuning becomes indispensable. Nevertheless, existing Low-Rank Adaptation (LoRA) approaches typically employ fixed or heuristic rank configurations, and the subsequent over-the-air transmission of all LoRA parameters could be rather inefficient. To address this limitation, we develop AirLLM, a hierarchical diffusion policy framework for communication-aware LoRA adaptation. Specifically, AirLLM models the rank configuration as a structured action vector that spans all LoRA-inserted projections. To solve the underlying high-dimensional sequential decision-making problem, a Proximal Policy Optimization (PPO) agent generates coarse-grained decisions by jointly observing wireless states and linguistic complexity, which are then refined via Denoising Diffusion Implicit Models (DDIM) to produce high-resolution, task- and channel-adaptive rank vectors. The two modules are optimized alternatively, with the DDIM trained under the Classifier-Free Guidance (CFG) paradigm to maintain alignment with PPO rewards. Experiments under varying signal-to-noise ratios demonstrate that AirLLM consistently enhances fine-tuning performance while significantly reducing transmission costs, highlighting the effectiveness of reinforcement-driven, diffusion-refined rank adaptation for scalable and efficient remote fine-tuning over the air.
Reinforcement Learning (RL) is an important machine learning paradigm for solving sequential decision-making problems. Recent years have witnessed remarkable progress in this field due to the rapid development of deep neural networks. However, the success of RL currently relies on extensive training data and computational resources. In addition, RL's limited ability to generalize across tasks restricts its applicability in dynamic and real-world environments. With the arisen of Continual Learning (CL), Continual Reinforcement Learning (CRL) has emerged as a promising research direction to address these limitations by enabling agents to learn continuously, adapt to new tasks, and retain previously acquired knowledge. In this survey, we provide a comprehensive examination of CRL, focusing on its core concepts, challenges, and methodologies. Firstly, we conduct a detailed review of existing works, organizing and analyzing their metrics, tasks, benchmarks, and scenario settings. Secondly, we propose a new taxonomy of CRL methods, categorizing them into four types from the perspective of knowledge storage and/or transfer. Finally, our analysis highlights the unique challenges of CRL and provides practical insights into future directions.
As Large Language Models (LLMs) are integrated into safety-critical applications involving sequential decision-making in the real world, it is essential to know when to trust LLM decisions. Existing LLM Uncertainty Quantification (UQ) methods are primarily designed for single-turn question-answering formats, resulting in multi-step decision-making scenarios, e.g., LLM agentic system, being underexplored. In this paper, we introduce a principled, information-theoretic framework that decomposes LLM sequential decision uncertainty into two parts: (i) internal uncertainty intrinsic to the current decision, which is focused on existing UQ methods, and (ii) extrinsic uncertainty, a Mutual-Information (MI) quantity describing how much uncertainty should be inherited from preceding decisions. We then propose UProp, an efficient and effective extrinsic uncertainty estimator that converts the direct estimation of MI to the estimation of Pointwise Mutual Information (PMI) over multiple Trajectory-Dependent Decision Processes (TDPs). UProp is evaluated over extensive multi-step decision-making benchmarks, e.g., AgentBench and HotpotQA, with state-of-the-art LLMs, e.g., GPT-4.1 and DeepSeek-V3. Experimental results demonstrate that UProp significantly outperforms existing single-turn UQ baselines equipped with thoughtful aggregation strategies. Moreover, we provide a comprehensive analysis of UProp, including sampling efficiency, potential applications, and intermediate uncertainty propagation, to demonstrate its effectiveness. Codes will be available at https://github.com/jinhaoduan/UProp.
In-context learning (ICL) with dynamically selected demonstrations combines the flexibility of prompting large language models (LLMs) with the ability to leverage training data to improve performance. While ICL has been highly successful for prediction and generation tasks, leveraging it for agentic tasks that require sequential decision making is challenging -- one must think not only about how to annotate long trajectories at scale and how to select demonstrations, but also what constitutes demonstrations, and when and where to show them. To address this, we first propose an algorithm that leverages an LLM with retries along with demonstrations to automatically and efficiently annotate agentic tasks with solution trajectories. We then show that set-selection of trajectories of similar tasks as demonstrations significantly improves performance, reliability, robustness, and efficiency of LLM agents. However, trajectory demonstrations have a large inference cost overhead. We show that this can be mitigated by using small trajectory snippets at every step instead of an additional trajectory. We find that demonstrations obtained from larger models (in the annotation phase) also improve smaller models, and that ICL agents can even rival costlier trained agents. Thus, our results reveal that ICL, with careful use, can be very powerful for agentic tasks as well.
Language models as intelligent agents push the boundaries of sequential decision-making agents but struggle with limited knowledge of environmental dynamics and exponentially huge action space. Recent efforts like GLAM and TWOSOME manually constrain the action space to a restricted subset and employ reinforcement learning to align agents' knowledge with specific environments. However, they overlook fine-grained credit assignments for intra-action tokens, which is essential for efficient language agent optimization, and rely on human's prior knowledge to restrict action space. This paper proposes decomposing language agent optimization from the action level to the token level, offering finer supervision for each intra-action token and manageable optimization complexity in environments with unrestricted action spaces. Beginning with the simplification of flattening all actions, we theoretically explore the discrepancies between action-level optimization and this naive token-level optimization. We then derive the Bellman backup with Action Decomposition (BAD) to integrate credit assignments for both intra-action and inter-action tokens, effectively eliminating the discrepancies. Implementing BAD within the PPO algorithm, we introduce Policy Optimization with Action Decomposition (POAD). POAD benefits from a finer-grained credit assignment process and lower optimization complexity, leading to enhanced learning efficiency and generalization abilities in aligning language agents with interactive environments. We validate POAD across diverse testbeds, with results affirming the advantages of our approach and the correctness of our theoretical analysis.
Large language model (LLM) agents have been applied to sequential decision-making tasks such as web navigation, but without any environment-specific experiences, they often fail in these complex tasks. Moreover, current LLM agents are not designed to continually learn from past experiences during inference time, which could be crucial for them to gain these environment-specific experiences. To address this, we propose Contextual Experience Replay (CER), a training-free framework to enable efficient self-improvement for language agents in their context window. Specifically, CER accumulates and synthesizes past experiences into a dynamic memory buffer. These experiences encompass environment dynamics and common decision-making patterns, allowing the agents to retrieve and augment themselves with relevant knowledge in new tasks, enhancing their adaptability in complex environments. We evaluate CER on the challenging WebArena and VisualWebArena benchmarks. On VisualWebArena, CER achieves a competitive performance of 31.9%. On WebArena, CER also gets a competitive average success rate of 36.7%, relatively improving the success rate of the GPT-4o agent baseline by 51.0%. We also conduct a comprehensive analysis on it to prove its efficiency, validity and understand it better.
Deep Reinforcement Learning (DRL) has achieved remarkable success in complex sequential decision-making tasks, such as playing Atari 2600 games and mastering board games. A critical yet underexplored aspect of DRL is the temporal scale of action execution. We propose a novel paradigm that integrates contextual bandits with DRL to adaptively select action durations, enhancing policy flexibility and computational efficiency. Our approach augments a Deep Q-Network (DQN) with a contextual bandit module that learns to choose optimal action repetition rates based on state contexts. Experiments on Atari 2600 games demonstrate significant performance improvements over static duration baselines, highlighting the efficacy of adaptive temporal abstractions in DRL. This paradigm offers a scalable solution for real-time applications like gaming and robotics, where dynamic action durations are critical.
Multi-armed bandit (MAB) is a widely adopted framework for sequential decision-making under uncertainty. Traditional bandit algorithms rely solely on online data, which tends to be scarce as it must be gathered during the online phase when the arms are actively pulled. However, in many practical settings, rich auxiliary data, such as covariates of past users, is available prior to deploying any arms. We introduce a new setting for MAB where pre-trained machine learning (ML) models are applied to convert side information and historical data into \emph{surrogate rewards}. A prominent feature of this setting is that the surrogate rewards may exhibit substantial bias, as true reward data is typically unavailable in the offline phase, forcing ML predictions to heavily rely on extrapolation. To address the issue, we propose the Machine Learning-Assisted Upper Confidence Bound (MLA-UCB) algorithm, which can be applied to any reward prediction model and any form of auxiliary data. When the predicted and true rewards are jointly Gaussian, it provably improves the cumulative regret, provided that the correlation is non-zero -- even in cases where the mean surrogate reward completely misaligns with the true mean rewards. Notably, our method requires no prior knowledge of the covariance matrix between true and surrogate rewards. We compare MLA-UCB with the standard UCB on a range of numerical studies and show a sizable efficiency gain even when the size of the offline data and the correlation between predicted and true rewards are moderate.
Generative models such as diffusion and flow-matching offer expressive policies for offline reinforcement learning (RL) by capturing rich, multimodal action distributions, but their iterative sampling introduces high inference costs and training instability due to gradient propagation across sampling steps. We propose the \textit{Single-Step Completion Policy} (SSCP), a generative policy trained with an augmented flow-matching objective to predict direct completion vectors from intermediate flow samples, enabling accurate, one-shot action generation. In an off-policy actor-critic framework, SSCP combines the expressiveness of generative models with the training and inference efficiency of unimodal policies, without requiring long backpropagation chains. Our method scales effectively to offline, offline-to-online, and online RL settings, offering substantial gains in speed and adaptability over diffusion-based baselines. We further extend SSCP to goal-conditioned RL, enabling flat policies to exploit subgoal structures without explicit hierarchical inference. SSCP achieves strong results across standard offline RL and behavior cloning benchmarks, positioning it as a versatile, expressive, and efficient framework for deep RL and sequential decision-making.
Software-intensive systems, such as software product lines and robotics, utilise Markov decision processes (MDPs) to capture uncertainty and analyse sequential decision-making problems. Despite the usefulness of conventional policy synthesis methods, they fail to scale to large state spaces. Our approach addresses this issue and accelerates policy synthesis in large MDPs by dynamically refining the MDP and iteratively selecting the most fragile MDP regions for refinement. This iterative procedure offers a balance between accuracy and efficiency, as refinement occurs only when necessary. Through a comprehensive empirical evaluation comprising diverse case studies and MDPs up to 1M states, we demonstrate significant performance improvements yielded by our approach compared to the leading probabilistic model checker PRISM (up to 2x), thus offering a very competitive solution for real-world policy synthesis tasks in larger MDPs.
Dynamic treatment regimes (DTRs) provide a principled framework for optimizing sequential decision-making in domains where decisions must adapt over time in response to individual trajectories, such as healthcare, education, and digital interventions. However, existing statistical methods often rely on strong positivity assumptions and lack robustness under partial data coverage, while offline reinforcement learning approaches typically focus on average training performance, lack statistical guarantees, and require solving complex optimization problems. To address these challenges, we propose POLAR, a novel pessimistic model-based policy learning algorithm for offline DTR optimization. POLAR estimates the transition dynamics from offline data and quantifies uncertainty for each history-action pair. A pessimistic penalty is then incorporated into the reward function to discourage actions with high uncertainty. Unlike many existing methods that focus on average training performance, POLAR directly targets the suboptimality of the final learned policy and offers theoretical guarantees, without relying on computationally intensive minimax or constrained optimization procedures. To the best of our knowledge, POLAR is the first model-based DTR method to provide both statistical and computational guarantees, including finite-sample bounds on policy suboptimality. Empirical results on both synthetic data and the MIMIC-III dataset demonstrate that POLAR outperforms state-of-the-art methods and yields near-optimal, history-aware treatment strategies.
Algorithms for learning programmatic representations for sequential decision-making problems are often evaluated on out-of-distribution (OOD) problems, with the common conclusion that programmatic policies generalize better than neural policies on OOD problems. In this position paper, we argue that commonly used benchmarks undervalue the generalization capabilities of programmatic representations. We analyze the experiments of four papers from the literature and show that neural policies, which were shown not to generalize, can generalize as effectively as programmatic policies on OOD problems. This is achieved with simple changes in the neural policies training pipeline. Namely, we show that simpler neural architectures with the same type of sparse observation used with programmatic policies can help attain OOD generalization. Another modification we have shown to be effective is the use of reward functions that allow for safer policies (e.g., agents that drive slowly can generalize better). Also, we argue for creating benchmark problems highlighting concepts needed for OOD generalization that may challenge neural policies but align with programmatic representations, such as tasks requiring algorithmic constructs like stacks.
Clustering of Bandits (CB) methods enhance sequential decision-making by grouping bandits into clusters based on similarity and incorporating cluster-level contextual information, demonstrating effectiveness and adaptability in applications like personalized streaming recommendations. However, when extending CB algorithms to their neural version (commonly referred to as Clustering of Neural Bandits, or CNB), they suffer from loss of plasticity, where neural network parameters become rigid and less adaptable over time, limiting their ability to adapt to non-stationary environments (e.g., dynamic user preferences in recommendation). To address this challenge, we propose Selective Reinitialization (SeRe), a novel bandit learning framework that dynamically preserves the adaptability of CNB algorithms in evolving environments. SeRe leverages a contribution utility metric to identify and selectively reset underutilized units, mitigating loss of plasticity while maintaining stable knowledge retention. Furthermore, when combining SeRe with CNB algorithms, the adaptive change detection mechanism adjusts the reinitialization frequency according to the degree of non-stationarity, ensuring effective adaptation without unnecessary resets. Theoretically, we prove that SeRe enables sublinear cumulative regret in piecewise-stationary environments, outperforming traditional CNB approaches in long-term performances. Extensive experiments on six real-world recommendation datasets demonstrate that SeRe-enhanced CNB algorithms can effectively mitigate the loss of plasticity with lower regrets, improving adaptability and robustness in dynamic settings.
Reinforcement learning (RL) is a human-designed framework for solving sequential decision making problems. In this work, we demonstrate that, surprisingly, RL emerges in LLM's (Large Language Model) inference time -- a phenomenon known as in-context RL (ICRL). Specifically, we propose a novel multi-round prompting framework called ICRL prompting. The goal is to prompt the LLM to complete a task. After the LLM generates a response at the current round, we give numerical scalar feedbacks for the response, called the rewards. At the next round, we prompt the LLM again with the same task and a context consisting of all previous responses and rewards. We observe that the quality of the LLM's response increases as the context grows. In other words, the LLM is able to maximize the scalar reward signal in the inference time, just like an RL algorithm. We evaluate ICRL prompting in three benchmarks (Game of 24, creative writing, and ScienceWorld) and demonstrate significant performance improvements over baseline methods such as Self-Refine and Reflexion. Surprisingly, in some experiments the reward signals are generated by the LLM itself, yet performance improvements are still observed from ICRL prompting, offering a promising paradigm for scaling test-time compute.
Ensuring responsible use of artificial intelligence (AI) has become imperative as autonomous systems increasingly influence critical societal domains. However, the concept of trustworthy AI remains broad and multi-faceted. This thesis advances knowledge in the safety, fairness, transparency, and accountability of AI systems. In safety, we extend classical deterministic shielding techniques to become resilient against delayed observations, enabling practical deployment in real-world conditions. We also implement both deterministic and probabilistic safety shields into simulated autonomous vehicles to prevent collisions with road users, validating the use of these techniques in realistic driving simulators. We introduce fairness shields, a novel post-processing approach to enforce group fairness in sequential decision-making settings over finite and periodic time horizons. By optimizing intervention costs while strictly ensuring fairness constraints, this method efficiently balances fairness with minimal interference. For transparency and accountability, we propose a formal framework for assessing intentional behaviour in probabilistic decision-making agents, introducing quantitative metrics of agency and intention quotient. We use these metrics to propose a retrospective analysis of intention, useful for determining responsibility when autonomous systems cause unintended harm. Finally, we unify these contributions through the ``reactive decision-making'' framework, providing a general formalization that consolidates previous approaches. Collectively, the advancements presented contribute practically to the realization of safer, fairer, and more accountable AI systems, laying the foundations for future research in trustworthy AI.
Deep reinforcement learning (DRL) has achieved remarkable success in a wide range of sequential decision-making domains, including robotics, healthcare, smart grids, and finance. Recent research demonstrates that attackers can efficiently exploit system vulnerabilities during the training phase to execute backdoor attacks, producing malicious actions when specific trigger patterns are present in the state observations. However, most existing backdoor attacks rely primarily on simplistic and heuristic trigger configurations, overlooking the potential efficacy of trigger optimization. To address this gap, we introduce TooBadRL (Trigger Optimization to Boost Effectiveness of Backdoor Attacks on DRL), the first framework to systematically optimize DRL backdoor triggers along three critical axes, i.e., temporal, spatial, and magnitude. Specifically, we first introduce a performance-aware adaptive freezing mechanism for injection timing. Then, we formulate dimension selection as a cooperative game, utilizing Shapley value analysis to identify the most influential state variable for the injection dimension. Furthermore, we propose a gradient-based adversarial procedure to optimize the injection magnitude under environment constraints. Evaluations on three mainstream DRL algorithms and nine benchmark tasks show that TooBadRL significantly improves attack success rates, while ensuring minimal degradation of normal task performance. These results highlight the previously underappreciated importance of principled trigger optimization in DRL backdoor attacks. The source code of TooBadRL can be found at https://github.com/S3IC-Lab/TooBadRL.
Quality-Diversity (QD) algorithms have shown remarkable success in discovering diverse, high-performing solutions, but rely heavily on hand-crafted behavioral descriptors that constrain exploration to predefined notions of diversity. Leveraging the equivalence between policies and occupancy measures, we present a theoretically grounded approach to automatically generate behavioral descriptors by embedding the occupancy measures of policies in Markov Decision Processes. Our method, AutoQD, leverages random Fourier features to approximate the Maximum Mean Discrepancy (MMD) between policy occupancy measures, creating embeddings whose distances reflect meaningful behavioral differences. A low-dimensional projection of these embeddings that captures the most behaviorally significant dimensions is then used as behavioral descriptors for off-the-shelf QD methods. We prove that our embeddings converge to true MMD distances between occupancy measures as the number of sampled trajectories and embedding dimensions increase. Through experiments in multiple continuous control tasks we demonstrate AutoQD's ability in discovering diverse policies without predefined behavioral descriptors, presenting a well-motivated alternative to prior methods in unsupervised Reinforcement Learning and QD optimization. Our approach opens new possibilities for open-ended learning and automated behavior discovery in sequential decision making settings without requiring domain-specific knowledge.