The hallmark feature of temporal-difference (TD) learning is bootstrapping: using value predictions to generate new value predictions. The vast majority of TD methods for control learn a policy by bootstrapping from a single action-value function (e.g., Q-learning and Sarsa). Significantly less attention has been given to methods that bootstrap from two asymmetric value functions: i.e., methods that learn state values as an intermediate step in learning action values. Existing algorithms in this vein can be categorized as either QV -learning or A V -learning. Though these algorithms have been investigated to some degree in prior work, it remains unclear if and when it is advantageous to learn two value functions instead of just one--and whether such approaches are theoretically sound in general. In this paper, we analyze these algorithmic families in terms of convergence and sample efficiency. We find that while both families are more efficient than Expected Sarsa in the prediction setting, only A V -learning methods offer any major benefit over Q-learning in the control setting. Finally, we introduce a new A V -learning algorithm called Regularized Dueling Q-learning (RDQ), which significantly outperforms Dueling DQN in the MinAtar benchmark.

In reinforcement learning (RL), aligning agent behavior with specific objectives typically requires careful design of the reward function, which can be challenging when the desired objectives are complex. In this work, we propose an alternative approach for flexible behavior alignment that eliminates the need to modify the reward function by selecting appropriate reward aggregation functions. By introducing an algebraic perspective on Markov decision processes (MDPs), we show that the Bellman equations naturally emerge from the recursive generation and aggregation of rewards, allowing for the generalization of the standard discounted sum to other recursive aggregations, such as discounted max and Sharpe ratio. Our approach applies to both deterministic and stochastic settings and integrates seamlessly with value-based and actor-critic algorithms. Experimental results demonstrate that our approach effectively optimizes diverse objectives, highlighting its versatility and potential for real-world applications.

Online 3D Bin Packing Problem (3D-BPP) has widespread applications in industrial automation. Existing methods usually solve the problem with limited resolution of spatial discretization, and/or cannot deal with complex practical constraints well. We propose to enhance the practical applicability of online 3D-BPP via learning on a novel hierarchical representation, packing configuration tree (PCT). PCT is a full-fledged description of the state and action space of bin packing which can support packing policy learning based on deep reinforcement learning (DRL). The size of the packing action space is proportional to the number of leaf nodes, making the DRL model easy to train and well-performing even with continuous solution space. We further discover the potential of PCT as tree-based planners in deliberately solving packing problems of industrial significance, including large-scale packing and different variations of BPP setting. A recursive packing method is proposed to decompose large-scale packing into smaller sub-trees while a spatial ensemble mechanism integrates local solutions into global. For different BPP variations with additional decision variables, such as lookahead, buffering, and offline packing, we propose a unified planning framework enabling out-of-the-box problem solving. Extensive evaluations demonstrate that our method outperforms existing online BPP baselines and is versatile in incorporating various practical constraints. The planning process excels across large-scale problems and diverse problem variations. We develop a real-world packing robot for industrial warehousing, with careful designs accounting for constrained placement and transportation stability. Our packing robot operates reliably and efficiently on unprotected pallets at 10 seconds per box. It achieves averagely 19 boxes per pallet with 57.4% space utilization for relatively large-size boxes.

Comparison of fine-tuning models with reinforcement learning (RL) and supervised fine-tuning (SFT) reveals that, despite similar performance at a new task, RL preserves prior knowledge and capabilities significantly better. We find that the degree of forgetting is determined by the distributional shift, measured as the KL-divergence between the fine-tuned and base policy evaluated on the new task. Our analysis reveals that on-policy RL is implicitly biased towards KL-minimal solutions among the many that solve the new task, whereas SFT can converge to distributions arbitrarily far from the base model. We validate these findings through experiments with large language models and robotic foundation models and further provide theoretical justification for why on-policy RL updates lead to a smaller KL change. We term this principle $\textit{RL's Razor}$: among all ways to solve a new task, RL prefers those closest in KL to the original model.

In the game of poker, being unpredictable, or bluffing, is an essential skill. When humans play poker, they bluff. However, most works on computer-poker focus on performance metrics such as win rates, while bluffing is overlooked. In this paper we study whether two popular algorithms, DQN (based on reinforcement learning) and CFR (based on game theory), exhibit bluffing behavior in Leduc Hold'em, a simplified version of poker. We designed an experiment where we let the DQN and CFR agent play against each other while we log their actions. We find that both DQN and CFR exhibit bluffing behavior, but they do so in different ways. Although both attempt to perform bluffs at different rates, the percentage of successful bluffs (where the opponent folds) is roughly the same. This suggests that bluffing is an essential aspect of the game, not of the algorithm. Future work should look at different bluffing styles and at the full game of poker.

Vision-Language-Action (VLA) models based on flow matching have shown excellent performance in general-purpose robotic manipulation tasks. However, the action accuracy of these models on complex downstream tasks is unsatisfactory. One important reason is that these models rely solely on the post-training paradigm of imitation learning, which makes it difficult to have a deeper understanding of the distribution properties of data quality, which is exactly what Reinforcement Learning (RL) excels at. In this paper, we theoretically propose an offline RL post-training objective for VLA flow models and induce an efficient and feasible offline RL fine-tuning algorithm -- Adaptive Reinforced Flow Matching (ARFM). By introducing an adaptively adjusted scaling factor in the VLA flow model loss, we construct a principled bias-variance trade-off objective function to optimally control the impact of RL signal on flow loss. ARFM adaptively balances RL advantage preservation and flow loss gradient variance control, resulting in a more stable and efficient fine-tuning process. Extensive simulation and real-world experimental results show that ARFM exhibits excellent generalization, robustness, few-shot learning, and continuous learning performance.

Deep reinforcement learning (DRL) has become a popular approach in traffic signal control (TSC) due to its ability to learn adaptive policies from complex traffic environments. Within DRL-based TSC methods, two primary control paradigms are ``choose phase" and ``switch" strategies. Although the agent in the choose phase paradigm selects the next active phase adaptively, this paradigm may result in unexpected phase sequences for drivers, disrupting their anticipation and potentially compromising safety at intersections. Meanwhile, the switch paradigm allows the agent to decide whether to switch to the next predefined phase or extend the current phase. While this structure maintains a more predictable order, it can lead to unfair and inefficient phase allocations, as certain movements may be extended disproportionately while others are neglected. In this paper, we propose a DRL model, named Deep Hierarchical Cycle Planner (DHCP), to allocate the traffic signal cycle duration hierarchically. A high-level agent first determines the split of the total cycle time between the North-South (NS) and East-West (EW) directions based on the overall traffic state. Then, a low-level agent further divides the allocated duration within each major direction between straight and left-turn movements, enabling more flexible durations for the two movements. We test our model on both real and synthetic road networks, along with multiple sets of real and synthetic traffic flows. Empirical results show our model achieves the best performance over all datasets against baselines.

Reinforcement learning with verifiable rewards (RL VR) has emerged to be a predominant paradigm for mathematical reasoning tasks, offering stable improvements in reasoning ability. However, Outcome Reward Models (ORMs) in RL VR are too coarse-grained to distinguish flawed reasoning within correct answers or valid reasoning within incorrect answers. This lack of granularity introduces noisy and misleading gradients significantly and hinders further progress in reasoning process quality. While Process Reward Models (PRMs) offer fine-grained guidance for intermediate steps, they frequently suffer from inaccuracies and are susceptible to reward hacking. To resolve this dilemma, we introduce PRocess cOnsistency Filter (PROF), an effective data process curation method that harmonizes noisy, fine-grained process rewards with accurate, coarse-grained outcome rewards. Our approach retains correct responses with higher averaged process values and incorrect responses with lower averaged process values, while maintaining positive/negative training sample balance. Extensive experiments demonstrate that our method not only consistently improves the final accuracy over 4% compared to the blending approaches, but also strengthens the quality of intermediate reasoning steps. Codes and training recipes are available at https://github.com/Chenluye99/PROF . V erifiable rewards have spurred the widest attention recently because they can reliably improve the performance on reasoning tasks with easily verifiable outcomes, such as mathematical and coding problems (Cobbe et al., 2021; Jaech et al., 2024; Shao et al., 2024; Xiong et al., 2025b). However, since the verifiers can only verify the outcome results, the rewards are too sparse and coarse to measure and supervise the reasoning quality in intermediate steps.

Sepsis is a serious, life-threatening condition. When treating sepsis, it is challenging to determine the correct amount of intravenous fluids and vasopressors for a given patient. While automated reinforcement learning (RL)-based methods have been used to support these decisions with promising results, previous studies have relied on relational data. Given the complexity of modern healthcare data, representing data as a graph may provide a more natural and effective approach. This study models patient data from the well-known MIMIC-III dataset as a heterogeneous graph that evolves over time. Subsequently, we explore two Graph Neural Network architectures - GraphSAGE and GATv2 - for learning patient state representations, adopting the approach of decoupling representation learning from policy learning. The encoders are trained to produce latent state representations, jointly with decoders that predict the next patient state. These representations are then used for policy learning with the dBCQ algorithm. The results of our experimental evaluation confirm the potential of a graph-based approach, while highlighting the complexity of representation learning in this domain.

Reinforcement learning faces significant challenges when applied to tasks characterized by sparse reward structures. Although imitation learning, within the domain of supervised learning, offers faster convergence, it relies heavily on human-generated demonstrations. Recently, Goal-Conditioned Supervised Learning (GCSL) has emerged as a potential solution by enabling self-imitation learning for autonomous systems. By strategically relabelling goals, agents can derive policy insights from their own experiences. Despite the successes of this framework, it presents two notable limitations: (1) Learning exclusively from self-generated experiences can exacerbate the agents' inherent biases; (2) The relabelling strategy allows agents to focus solely on successful outcomes, precluding them from learning from their mistakes. To address these issues, we propose a novel model that integrates contrastive learning principles into the GCSL framework to learn from both success and failure. Through empirical evaluations, we demonstrate that our algorithm overcomes limitations imposed by agents' initial biases and thereby enables more exploratory behavior. This facilitates the identification and adoption of effective policies, leading to superior performance across a variety of challenging environments.