In modern ML Ops environments, model deployment is a critical process that traditionally relies on static heuristics such as validation error comparisons and A/B testing. However, these methods require human intervention to adapt to real-world deployment challenges, such as model drift or unexpected performance degradation. We investigate whether reinforcement learning, specifically multi-armed bandit (MAB) algorithms, can dynamically manage model deployment decisions more effectively. Our approach enables more adaptive production environments by continuously evaluating deployed models and rolling back underperforming ones in real-time. We test six model selection strategies across two real-world datasets and find that RL based approaches match or exceed traditional methods in performance. Our findings suggest that reinforcement learning (RL)-based model management can improve automation, reduce reliance on manual interventions, and mitigate risks associated with post-deployment model failures.
Reward functions are central in reinforcement learning (RL), guiding agents towards optimal decision-making. The complexity of RL tasks requires meticulously designed reward functions that effectively drive learning while avoiding unintended consequences. Effective reward design aims to provide signals that accelerate the agent's convergence to optimal behavior. Crafting rewards that align with task objectives, foster desired behaviors, and prevent undesirable actions is inherently challenging. This thesis delves into the critical role of reward signals in RL, highlighting their impact on the agent's behavior and learning dynamics and addressing challenges such as delayed, ambiguous, or intricate rewards. In this thesis work, we tackle different aspects of reward shaping. First, we address the problem of designing informative and interpretable reward signals from a teacher's/expert's perspective (teacher-driven). Here, the expert, equipped with the optimal policy and the corresponding value function, designs reward signals that expedite the agent's convergence to optimal behavior. Second, we build on this teacher-driven approach by introducing a novel method for adaptive interpretable reward design. In this scenario, the expert tailors the rewards based on the learner's current policy, ensuring alignment and optimal progression. Third, we propose a meta-learning approach, enabling the agent to self-design its reward signals online without expert input (agent-driven). This self-driven method considers the agent's learning and exploration to establish a self-improving feedback loop.
Image quality assessment (IQA) focuses on the perceptual visual quality of images, playing a crucial role in downstream tasks such as image reconstruction, compression, and generation. The rapid advancement of multi-modal large language models (MLLMs) has significantly broadened the scope of IQA, moving toward comprehensive image quality understanding that incorporates content analysis, degradation perception, and comparison reasoning beyond mere numerical scoring. Previous MLLM-based methods typically either generate numerical scores lacking interpretability or heavily rely on supervised fine-tuning (SFT) using large-scale annotated datasets to provide descriptive assessments, limiting their flexibility and applicability. In this paper, we propose Q-Insight, a reinforcement learning-based model built upon group relative policy optimization (GRPO), which demonstrates strong visual reasoning capability for image quality understanding while requiring only a limited amount of rating scores and degradation labels. By jointly optimizing score regression and degradation perception tasks with carefully designed reward functions, our approach effectively exploits their mutual benefits for enhanced performance. Extensive experiments demonstrate that Q-Insight substantially outperforms existing state-of-the-art methods in both score regression and degradation perception tasks, while exhibiting impressive zero-shot generalization to comparison reasoning tasks. Code will be available at https://github.com/lwq20020127/Q-Insight.
We introduce Entropy-Guided Sequence Weighting (EGSW), a novel approach that enhances the exploration-exploitation tradeoff by dynamically assigning weights to generated outputs based on their advantage and entropy for Reinforcement Learning-based Large Language Model fine-tuning. EGSW integrates entropy regularization with advantage-based weighting to balance policy updates, enabling efficient exploration in high-dimensional state spaces. By employing temperature-scaled softmax weighting over sequences, EGSW prioritizing high-reward, high-uncertainty steps while maintaining training stability. Although originally developed to improve Group Relative Policy Optimization (GRPO) during large language model (LLM) fine-tuning, EGSW is generalizable to other reinforcement learning (RL) algorithms and can be implemented in both step-wise and trajectory-wise settings. Empirical evaluations demonstrate that EGSW enhances GRPO reasoning ability, yielding improvements in sample efficiency. Future work will explore the application of EGSW to advanced RL methodologies.
Lane keeping in autonomous driving systems requires scenario-specific weight tuning for different objectives. We formulate lane-keeping as a constrained reinforcement learning problem, where weight coefficients are automatically learned along with the policy, eliminating the need for scenario-specific tuning. Empirically, our approach outperforms traditional RL in efficiency and reliability. Additionally, real-world demonstrations validate its practical value for real-world autonomous driving.
Offline reinforcement learning (RL) looks at learning how to optimally solve tasks using a fixed dataset of interactions from the environment. Many off-policy algorithms developed for online learning struggle in the offline setting as they tend to over-estimate the behaviour of out of distributions actions. Existing offline RL algorithms adapt off-policy algorithms, employing techniques such as constraining the policy or modifying the value function to achieve good performance on individual datasets but struggle to adapt to different tasks or datasets of different qualities without tuning hyper-parameters. We introduce a policy switching technique that dynamically combines the behaviour of a pure off-policy RL agent, for improving behaviour, and a behavioural cloning (BC) agent, for staying close to the data. We achieve this by using a combination of epistemic uncertainty, quantified by our RL model, and a metric for aleatoric uncertainty extracted from the dataset. We show empirically that our policy switching technique can outperform not only the individual algorithms used in the switching process but also compete with state-of-the-art methods on numerous benchmarks. Our use of epistemic uncertainty for policy switching also allows us to naturally extend our method to the domain of offline to online fine-tuning allowing our model to adapt quickly and safely from online data, either matching or exceeding the performance of current methods that typically require additional modification or hyper-parameter fine-tuning.
We propose a way to optimize chain-of-thought with reinforcement learning, but without external reward function. Our algorithm relies on viewing chain-of-thought as latent variable as part of a probabilistic inference problem. Contrary to the full evidence lower bound, we propose to apply a much simpler Jensen's lower bound, which derives tractable objectives with simple algorithmic components (e.g., without the need for parametric approximate posterior), making it more conducive to modern large-scale training. The lower bound approach naturally interpolates other methods such as supervised fine-tuning and online reinforcement learning, whose practical trade-offs we will illustrate. Finally, we show that on mathematical reasoning problems, optimizing with Jensen's lower bound is as effective as policy gradient with external reward. Taken together, our results showcase as a proof of concept to this new algorithmic paradigm's potential to more generic applications.
Large Language Models (LLMs) have shown remarkable capabilities in reasoning, exemplified by the success of OpenAI-o1 and DeepSeek-R1. However, integrating reasoning with external search processes remains challenging, especially for complex multi-hop questions requiring multiple retrieval steps. We propose ReSearch, a novel framework that trains LLMs to Reason with Search via reinforcement learning without using any supervised data on reasoning steps. Our approach treats search operations as integral components of the reasoning chain, where when and how to perform searches is guided by text-based thinking, and search results subsequently influence further reasoning. We train ReSearch on Qwen2.5-7B(-Instruct) and Qwen2.5-32B(-Instruct) models and conduct extensive experiments. Despite being trained on only one dataset, our models demonstrate strong generalizability across various benchmarks. Analysis reveals that ReSearch naturally elicits advanced reasoning capabilities such as reflection and self-correction during the reinforcement learning process.
Offline reinforcement learning (RL) aims to learn from historical data without requiring (costly) access to the environment. To facilitate offline RL research, we previously introduced NeoRL, which highlighted that datasets from real-world tasks are often conservative and limited. With years of experience applying offline RL to various domains, we have identified additional real-world challenges. These include extremely conservative data distributions produced by deployed control systems, delayed action effects caused by high-latency transitions, external factors arising from the uncontrollable variance of transitions, and global safety constraints that are difficult to evaluate during the decision-making process. These challenges are underrepresented in previous benchmarks but frequently occur in real-world tasks. To address this, we constructed the extended Near Real-World Offline RL Benchmark (NeoRL-2), which consists of 7 datasets from 7 simulated tasks along with their corresponding evaluation simulators. Benchmarking results from state-of-the-art offline RL approaches demonstrate that current methods often struggle to outperform the data-collection behavior policy, highlighting the need for more effective methods. We hope NeoRL-2 will accelerate the development of reinforcement learning algorithms for real-world applications. The benchmark project page is available at https://github.com/polixir/NeoRL2.
Large Language Models (LLMs) have shown remarkable promise in reasoning and decision-making, yet their integration with Reinforcement Learning (RL) for complex robotic tasks remains underexplored. In this paper, we propose an LLM-guided hierarchical RL framework, termed LDSC, that leverages LLM-driven subgoal selection and option reuse to enhance sample efficiency, generalization, and multi-task adaptability. Traditional RL methods often suffer from inefficient exploration and high computational cost. Hierarchical RL helps with these challenges, but existing methods often fail to reuse options effectively when faced with new tasks. To address these limitations, we introduce a three-stage framework that uses LLMs for subgoal generation given natural language description of the task, a reusable option learning and selection method, and an action-level policy, enabling more effective decision-making across diverse tasks. By incorporating LLMs for subgoal prediction and policy guidance, our approach improves exploration efficiency and enhances learning performance. On average, LDSC outperforms the baseline by 55.9\% in average reward, demonstrating its effectiveness in complex RL settings. More details and experiment videos could be found in \href{https://raaslab.org/projects/LDSC/}{this link\footnote{https://raaslab.org/projects/LDSC}}.
Modern reinforcement learning (RL) systems have demonstrated remarkable capabilities in complex environments, such as video games. However, they still fall short of achieving human-like sample efficiency and adaptability when learning new domains. Theory-based reinforcement learning (TBRL) is an algorithmic framework specifically designed to address this gap. Modeled on cognitive theories, TBRL leverages structured, causal world models - "theories" - as forward simulators for use in planning, generalization and exploration. Although current TBRL systems provide compelling explanations of how humans learn to play video games, they face several technical limitations: their theory languages are restrictive, and their planning algorithms are not scalable. To address these challenges, we introduce TheoryCoder, an instantiation of TBRL that exploits hierarchical representations of theories and efficient program synthesis methods for more powerful learning and planning. TheoryCoder equips agents with general-purpose abstractions (e.g., "move to"), which are then grounded in a particular environment by learning a low-level transition model (a Python program synthesized from observations by a large language model). A bilevel planning algorithm can exploit this hierarchical structure to solve large domains. We demonstrate that this approach can be successfully applied to diverse and challenging grid-world games, where approaches based on directly synthesizing a policy perform poorly. Ablation studies demonstrate the benefits of using hierarchical abstractions.
Model-based offline Reinforcement Learning (RL) constructs environment models from offline datasets to perform conservative policy optimization. Existing approaches focus on learning state transitions through ensemble models, rollouting conservative estimation to mitigate extrapolation errors. However, the static data makes it challenging to develop a robust policy, and offline agents cannot access the environment to gather new data. To address these challenges, we introduce Model-based Offline Reinforcement learning with AdversariaL data augmentation (MORAL). In MORAL, we replace the fixed horizon rollout by employing adversaria data augmentation to execute alternating sampling with ensemble models to enrich training data. Specifically, this adversarial process dynamically selects ensemble models against policy for biased sampling, mitigating the optimistic estimation of fixed models, thus robustly expanding the training data for policy optimization. Moreover, a differential factor is integrated into the adversarial process for regularization, ensuring error minimization in extrapolations. This data-augmented optimization adapts to diverse offline tasks without rollout horizon tuning, showing remarkable applicability. Extensive experiments on D4RL benchmark demonstrate that MORAL outperforms other model-based offline RL methods in terms of policy learning and sample efficiency.
In this article, we primarily examine a variety of RL-based and RL-free methods designed to address Reinforcement Learning from Human Feedback (RLHF) and Large Reasoning Models (LRMs). We begin with a concise overview of the typical steps involved in RLHF and LRMs. Next, we reinterpret several RL-based and RL-free algorithms through the perspective of neural structured bandit prediction, providing a clear conceptual framework that uncovers a deeper connection between these seemingly distinct approaches. Following this, we briefly review some core principles of reinforcement learning, drawing attention to an often-overlooked aspect in existing RLHF studies. This leads to a detailed derivation of the standard RLHF objective within a full RL context, demonstrating its equivalence to neural structured bandit prediction. Finally, by reinvestigating the principles behind Proximal Policy Optimization (PPO), we pinpoint areas needing adjustment, which culminates in the introduction of the Generalized Reinforce Optimization (GRO) framework, seamlessly integrating RL-based and RL-free methods in RLHF. We look forward to the community's efforts to empirically validate GRO and invite constructive feedback.
Buildings with Heating, Ventilation, and Air Conditioning (HVAC) systems play a crucial role in ensuring indoor comfort and efficiency. While traditionally governed by physics-based models, the emergence of big data has enabled data-driven methods like Deep Reinforcement Learning (DRL). However, Reinforcement Learning (RL)-based techniques often suffer from sample inefficiency and limited generalization, especially across varying HVAC systems. We introduce a model-based reinforcement learning framework that uses a Hypernetwork to continuously learn environment dynamics across tasks with different action spaces. This enables efficient synthetic rollout generation and improved sample usage. Our approach demonstrates strong backward transfer in a continual learning setting after training on a second task, minimal fine-tuning on the first task allows rapid convergence within just 5 episodes and thus outperforming Model Free Reinforcement Learning (MFRL) and effectively mitigating catastrophic forgetting. These findings have significant implications for reducing energy consumption and operational costs in building management, thus supporting global sustainability goals. Keywords: Deep Reinforcement Learning, HVAC Systems Control, Hypernetworks, Transfer and Continual Learning, Catastrophic Forgetting
Skills have been introduced to offline reinforcement learning (RL) as temporal abstractions to tackle complex, long-horizon tasks, promoting consistent behavior and enabling meaningful exploration. While skills in offline RL are predominantly modeled within a continuous latent space, the potential of discrete skill spaces remains largely underexplored. In this paper, we propose a compact discrete skill space for offline RL tasks supported by state-of-the-art transformer-based encoder and diffusion-based decoder. Coupled with a high-level policy trained via offline RL techniques, our method establishes a hierarchical RL framework where the trained diffusion decoder plays a pivotal role. Empirical evaluations show that the proposed algorithm, Discrete Diffusion Skill (DDS), is a powerful offline RL method. DDS performs competitively on Locomotion and Kitchen tasks and excels on long-horizon tasks, achieving at least a 12 percent improvement on AntMaze-v2 benchmarks compared to existing offline RL approaches. Furthermore, DDS offers improved interpretability, training stability, and online exploration compared to previous skill-based methods.
Recently, deep reinforcement learning (DRL) has emerged as a promising approach for robotic control. However, the deployment of DRL in real-world robots is hindered by its sensitivity to environmental perturbations. While existing whitebox adversarial attacks rely on local gradient information and apply uniform perturbations across all states to evaluate DRL robustness, they fail to account for temporal dynamics and state-specific vulnerabilities. To combat the above challenge, we first conduct a theoretical analysis of white-box attacks in DRL by establishing the adversarial victim-dynamics Markov decision process (AVD-MDP), to derive the necessary and sufficient conditions for a successful attack. Based on this, we propose a selective state-aware reinforcement adversarial attack method, named STAR, to optimize perturbation stealthiness and state visitation dispersion. STAR first employs a soft mask-based state-targeting mechanism to minimize redundant perturbations, enhancing stealthiness and attack effectiveness. Then, it incorporates an information-theoretic optimization objective to maximize mutual information between perturbations, environmental states, and victim actions, ensuring a dispersed state-visitation distribution that steers the victim agent into vulnerable states for maximum return reduction. Extensive experiments demonstrate that STAR outperforms state-of-the-art benchmarks.
Multi-agent reinforcement learning (MARL) is increasingly ubiquitous in training dynamic and adaptive synthetic characters for interactive simulations on geo-specific terrains. Frameworks such as Unity's ML-Agents help to make such reinforcement learning experiments more accessible to the simulation community. Military training simulations also benefit from advances in MARL, but they have immense computational requirements due to their complex, continuous, stochastic, partially observable, non-stationary, and doctrine-based nature. Furthermore, these simulations require geo-specific terrains, further exacerbating the computational resources problem. In our research, we leverage Unity's waypoints to automatically generate multi-layered representation abstractions of the geo-specific terrains to scale up reinforcement learning while still allowing the transfer of learned policies between different representations. Our early exploratory results on a novel MARL scenario, where each side has differing objectives, indicate that waypoint-based navigation enables faster and more efficient learning while producing trajectories similar to those taken by expert human players in CSGO gaming environments. This research points out the potential of waypoint-based navigation for reducing the computational costs of developing and training MARL models for military training simulations, where geo-specific terrains and differing objectives are crucial.
Deep learning approaches are becoming increasingly attractive for equation discovery. We show the advantages and disadvantages of using neural-guided equation discovery by giving an overview of recent papers and the results of experiments using our modular equation discovery system MGMT ($\textbf{M}$ulti-Task $\textbf{G}$rammar-Guided $\textbf{M}$onte-Carlo $\textbf{T}$ree Search for Equation Discovery). The system uses neural-guided Monte-Carlo Tree Search (MCTS) and supports both supervised and reinforcement learning, with a search space defined by a context-free grammar. We summarize seven desirable properties of equation discovery systems, emphasizing the importance of embedding tabular data sets for such learning approaches. Using the modular structure of MGMT, we compare seven architectures (among them, RNNs, CNNs, and Transformers) for embedding tabular datasets on the auxiliary task of contrastive learning for tabular data sets on an equation discovery task. For almost all combinations of modules, supervised learning outperforms reinforcement learning. Moreover, our experiments indicate an advantage of using grammar rules as action space instead of tokens. Two adaptations of MCTS -- risk-seeking MCTS and AmEx-MCTS -- can improve equation discovery with that kind of search.
Semantic Interpretability in Reinforcement Learning (RL) enables transparency, accountability, and safer deployment by making the agent's decisions understandable and verifiable. Achieving this, however, requires a feature space composed of human-understandable concepts, which traditionally rely on human specification and fail to generalize to unseen environments. In this work, we introduce Semantically Interpretable Reinforcement Learning with Vision-Language Models Empowered Automation (SILVA), an automated framework that leverages pre-trained vision-language models (VLM) for semantic feature extraction and interpretable tree-based models for policy optimization. SILVA first queries a VLM to identify relevant semantic features for an unseen environment, then extracts these features from the environment. Finally, it trains an Interpretable Control Tree via RL, mapping the extracted features to actions in a transparent and interpretable manner. To address the computational inefficiency of extracting features directly with VLMs, we develop a feature extraction pipeline that generates a dataset for training a lightweight convolutional network, which is subsequently used during RL. By leveraging VLMs to automate tree-based RL, SILVA removes the reliance on human annotation previously required by interpretable models while also overcoming the inability of VLMs alone to generate valid robot policies, enabling semantically interpretable reinforcement learning without human-in-the-loop.
In this work, we augment reinforcement learning with an inference-time collision model to ensure safe and efficient container management in a waste-sorting facility with limited processing capacity. Each container has two optimal emptying volumes that trade off higher throughput against overflow risk. Conventional reinforcement learning (RL) approaches struggle under delayed rewards, sparse critical events, and high-dimensional uncertainty -- failing to consistently balance higher-volume empties with the risk of safety-limit violations. To address these challenges, we propose a hybrid method comprising: (1) a curriculum-learning pipeline that incrementally trains a PPO agent to handle delayed rewards and class imbalance, and (2) an offline pairwise collision model used at inference time to proactively avert collisions with minimal online cost. Experimental results show that our targeted inference-time collision checks significantly improve collision avoidance, reduce safety-limit violations, maintain high throughput, and scale effectively across varying container-to-PU ratios. These findings offer actionable guidelines for designing safe and efficient container-management systems in real-world facilities.