Distributional RL

Learning from a sequence of interactions, as soon as observations are perceived and acted upon, without explicitly storing them, holds the promise of simpler, more efficient and adaptive algorithms. For over a decade, however, deep reinforcement learning walked the contrary path, augmenting agents with replay buffers or parallel sampling routines, in an effort to tame learning instability. Recently, this topic has been revisited by Elsayed et al. (2024), focusing on update computation through eligibility traces and modifications to the optimisation routine, resulting in the StreamQ algorithm. In this work we take a step back, investigating the efficacy of established updates, such as those implemented by DQN and C51 within this online setting. Not only do we find that they perform well, but through analysing how the optimisation algorithm generally, and Adam in particular, interacts with these updates, we contend that two properties are essential for robust performance: i) the derivative of the objective is to be bounded and ii) weight updates are variance-adjusted. Rigorous and exhaustive experimentation demonstrates that C51, which exhibits both characteristics, is competitive with StreamQ across a subset of 55 Atari games. Using these insights, we derive a variance-adjusted algorithm based on eligibility traces, termed Adaptive Q$(λ)$, which approaches double the human baseline on the same subset, surpassing existing methods by all performance metrics.

Quantile-based distributional reinforcement learning methods learn return distributions through sampled quantile regression, but their bootstrapped target quantiles may induce distorted or degenerate distribution estimates. We propose Robust Quantile-based Implicit Quantile Networks (RQIQN), a lightweight Wasserstein distributionally robust enhancement boosted from a quantile estimation perspective. We first reinterpret a snapshot of IQN loss as a collection of local empirical quantile estimation problems over sampled current fractions. We then robustify each local slot with a Wasserstein distributionally robust quantile estimation formulation, yielding a closed-form, fraction-dependent correction to the Bellman target. This correction directly addresses distributional degeneration: its median antisymmetry preserves the risk-neutral quantile average, while its monotonicity enlarges upper-lower quantile gaps and counteracts collapsed distributional spread. RQIQN thus regularizes quantile geometry without changing the underlying value objective or requiring additional sample set reconstruction. Finally, we empirically show that the proposed RQIQN outperforms other existing quantile-based distributional reinforcement learning algorithms in risk-sensitive navigation and Atari games.

We propose Flow-Anchored Noise-conditioned Q-Learning (FAN), a highly efficient and high-performing offline reinforcement learning (RL) algorithm. Recent work has shown that expressive flow policies and distributional critics improve offline RL performance, but at a high computational cost. Specifically, flow policies require iterative sampling to produce a single action, and distributional critics require computation over multiple samples (e.g., quantiles) to estimate value. To address these inefficiencies while maintaining high performance, we introduce FAN. Our method employs a behavior regularization technique that uses a single flow policy iteration and requires a single Gaussian noise sample for distributional critics. Our theoretical analysis of convergence and performance bounds demonstrates that these simplifications not only improve efficiency but also lead to superior task performance. Experiments on robotic manipulation and locomotion tasks demonstrate that FAN achieves state-of-the-art performance while significantly reducing both training and inference runtimes. We release our code at https://github.com/brianlsy98/FAN.

This paper explores the application of the Soft Actor-Critic (SAC) algorithm within a Distributional Reinforcement Learning setting and introduces an implementation of such algorithm named Cramér-based Distributional Soft Actor-Critic (C-DSAC). The novel approach employs distributional reinforcement learning to represent state-action values, and minimizes the squared Cramér distance for learning the distribution. Empirical results across various robotic benchmarks indicate that our algorithm surpasses the performance of baseline SAC and contemporary distributional methods, with the performance advantage becoming increasingly pronounced in high-complexity environments. To explain the efficiency of the new approach, we conduct an analysis showing that its superior performance is partly due to confidence-driven Q-value updates: High-variance target distributions (low confidence in target) lead to more conservative model updates, thereby attenuating the impact of overestimated values. This work deepens the understanding of distributional reinforcement learning, offering insights into the algorithmic mechanisms governing convergence and value estimation.

Quality-Diversity (QD) algorithms excel at discovering diverse repertoires of skills, but are hindered by poor sample efficiency and often require tens of millions of environment steps to solve complex locomotion tasks. Recent advances in Reinforcement Learning (RL) have shown that high Update-to-Data (UTD) ratios accelerate Actor-Critic learning. While effective, standard high-UTD algorithms typically utilise target networks to stabilise training. This requirement introduces a significant computational bottleneck, rendering them impractical for resource-intensive Quality-Diversity (QD) tasks where sample efficiency and rapid population adaptation are critical. In this paper, we introduce QDHUAC, a sample-efficient, target-free and distributional QD-RL algorithm that provides dense and low-variance gradient signals, which enables high-UTD training for Dominated Novelty Search whilst requiring an order of magnitude fewer environment steps. We demonstrate that our method enables stable training at high UTD ratios, achieving competitive coverage and fitness on high-dimensional Brax environments with an order of magnitude fewer samples than baselines. Our results suggest that combining target-free distributional critics with dominance-based selection is a key enabler for the next generation of sample-efficient evolutionary RL algorithms.

This paper investigates the off-policy evaluation (OPE) problem from a distributional perspective. Rather than focusing solely on the expectation of the total return, as in most existing OPE methods, we aim to estimate the entire return distribution. To this end, we introduce a quantile-based approach for OPE using deep quantile process regression, presenting a novel algorithm called Deep Quantile Process regression-based Off-Policy Evaluation (DQPOPE). We provide new theoretical insights into the deep quantile process regression technique, extending existing approaches that estimate discrete quantiles to estimate a continuous quantile function. A key contribution of our work is the rigorous sample complexity analysis for distributional OPE with deep neural networks, bridging theoretical analysis with practical algorithmic implementations. We show that DQPOPE achieves statistical advantages by estimating the full return distribution using the same sample size required to estimate a single policy value using conventional methods. Empirical studies further show that DQPOPE provides significantly more precise and robust policy value estimates than standard methods, thereby enhancing the practical applicability and effectiveness of distributional reinforcement learning approaches.

Process reward models (PRMs) provide fine-grained supervision for reasoning, but reliable PRMs often require step annotations or heavy verification pipelines, making them costly to scale and refresh during online RL. Implicit PRMs reduce this cost by training log-likelihood-ratio rewards from trajectory-level outcome labels. However, the log-ratio is constrained only as a sequence-level aggregate during training, while inference decomposes it into token- or step-level scores for partial prefixes. This train-inference mismatch leaves local credits weakly identified, so distribution-wide scoring can amplify misleading advantages. We propose Implicit Prefix-Value Reward Model (IPVRM), which directly learns the probability of eventual correctness for each prefix from outcome labels. Step signals are then obtained as temporal-difference (TD) differences between consecutive prefix values, aligning the training target with inference-time use. IPVRM markedly improves step-verification F1 on ProcessBench. To exploit these prefix values during policy optimization, we further introduce Distribution-Level RL (DistRL), which applies TD advantages to both sampled tokens and high-probability candidate tokens, providing dense counterfactual updates without additional rollouts. Experiments show that DistRL brings limited gains with unreliable implicit rewards, but consistently improves downstream reasoning when paired with IPVRM. The implementation of our method is available at https://github.com/gaoshiping/IPVRM .

Reinforcement learning is typically treated as a uniform, data-driven optimization process, where updates are guided by rewards and temporal-difference errors without explicitly exploiting global structure. In contrast, dynamic programming methods rely on structured information propagation, enabling efficient and stable learning. In this paper, we provide evidence that such structure can be recovered from the learning dynamics of distributional reinforcement learning. By analyzing the temporal evolution of return distributions, we identify signals that capture when and where learning occurs in the state space. In particular, we introduce a temporal learning indicator t*(s) that reflects when a state undergoes its strongest learning update during training. Empirically, this signal induces an ordering over states that is consistent with a dynamic programming-style propagation of information. Building on this observation, we propose StructRL, a framework that exploits these signals to guide sampling in alignment with the emerging propagation structure. Our preliminary results suggest that distributional learning dynamics provide a mechanism to recover and exploit dynamic programming-like structure without requiring an explicit model. This offers a new perspective on reinforcement learning, where learning can be interpreted as a structured propagation process rather than a purely uniform optimization procedure.

Researchers and practitioners are increasingly considering reinforcement learning to optimize decisions in complex domains like robotics and healthcare. To date, these efforts have largely utilized expectation-based learning. However, relying on expectation-focused objectives may be insufficient for making consistent decisions in highly uncertain situations involving multiple heterogeneous groups. While distributional reinforcement learning algorithms have been introduced to model the full distributions of outcomes, they can yield large discrepancies in realized benefits among comparable agents. This challenge is particularly acute in healthcare settings, where physicians (controllers) must manage multiple patients (subordinate agents) with uncertain disease progression and heterogeneous treatment responses. We propose a Boosted Distributional Reinforcement Learning (BDRL) algorithm that optimizes agent-specific outcome distributions while enforcing comparability among similar agents and analyze its convergence. To further stabilize learning, we incorporate a post-update projection step formulated as a constrained convex optimization problem, which efficiently aligns individual outcomes with a high-performing reference within a specified tolerance. We apply our algorithm to manage hypertension in a large subset of the US adult population by categorizing individuals into cardiovascular disease risk groups. Our approach modifies treatment plans for median and vulnerable patients by mimicking the behavior of high-performing references in each risk group. Furthermore, we find that BDRL improves the number and consistency of quality-adjusted life years compared with reinforcement learning baselines.

Reinforcement Learning (RL) has proven highly effective in addressing complex control and decision-making tasks. However, in most traditional RL algorithms, the policy is typically parameterized as a diagonal Gaussian distribution, which constrains the policy from capturing multimodal distributions, making it difficult to cover the full range of optimal solutions in multi-solution problems, and the return is reduced to a mean value, losing its multimodal nature and thus providing insufficient guidance for policy updates. In response to these problems, we propose a RL algorithm termed flow-based policy with distributional RL (FP-DRL). This algorithm models the policy using flow matching, which offers both computational efficiency and the capacity to fit complex distributions. Additionally, it employs a distributional RL approach to model and optimize the entire return distribution, thereby more effectively guiding multimodal policy updates and improving agent performance. Experimental trails on MuJoCo benchmarks demonstrate that the FP-DRL algorithm achieves state-of-the-art (SOTA) performance in most MuJoCo control tasks while exhibiting superior representation capability of the flow policy.

Federated reinforcement learning typically aggregates value functions or policies by parameter averaging, which emphasizes expected return and can obscure statistical multimodality and tail behavior that matter in safety-critical settings. We formalize federated distributional reinforcement learning (FedDistRL), where clients parametrize quantile value function critics and federate these networks only. We also propose TR-FedDistRL, which builds a per client, risk-aware Wasserstein barycenter over a temporal buffer. This local barycenter provides a reference region to constrain the parameter averaged critic, ensuring necessary distributional information is not averaged out during the federation process. The distributional trust region is implemented as a shrink-squash step around this reference. Under fixed-policy evaluation, the feasibility map is nonexpansive and the update is contractive in a probe-set Wasserstein metric under evaluation. Experiments on a bandit, multi-agent gridworld, and continuous highway environment show reduced mean-smearing, improved safety proxies (catastrophe/accident rate), and lower critic/policy drift versus mean-oriented and non-federated baselines.

Scaling Maximum Entropy Reinforcement Learning (RL) to high-dimensional humanoid control remains a fundamental challenge, as the ''curse of dimensionality'' induces severe exploration inefficiency and training instability. Consequently, highly optimized deterministic policy gradients currently dominate high-throughput regimes. We address this limitation with FastDSAC, a framework that effectively unlocks the potential of maximum entropy stochastic policies for complex continuous control. We introduce Dimension-wise Entropy Modulation (DEM) to dynamically redistribute the exploration budget, alongside a continuous distributional critic tailored to ensure accurate value estimation by mitigating both high-dimensional overestimation and discrete quantization artifacts. Extensive evaluations on HumanoidBench and a diverse set of continuous control tasks demonstrate that FastDSAC establishes state-of-the-art performance for high-dimensional stochastic policies on the evaluated benchmarks. Our method is competitive with and often outperforms strong deterministic baselines, with gains of 180% and 350% on the challenging Basketball and Balance Hard tasks, respectively.

Distributional reinforcement learning (DRL) studies the evolution of full return distributions under Bellman updates rather than focusing on expected values. A classical result is that the distributional Bellman operator is contractive under the Cramér metric, which corresponds to an $L^2$ geometry on differences of cumulative distribution functions (CDFs). While this contraction ensures stability of policy evaluation, existing analyses remain largely metric, focusing on contraction properties without elucidating the structural action of the Bellman update on distributions. In this work, we analyse distributional Bellman dynamics directly at the level of CDFs, treating the Cramér geometry as the intrinsic analytical setting. At this level, the Bellman update acts affinely on CDFs and linearly on differences between CDFs, and its contraction property yields a uniform bound on this linear action. Building on this intrinsic formulation, we construct a family of regularised spectral Hilbert representations that realise the CDF-level geometry by exact conjugation, without modifying the underlying Bellman dynamics. The regularisation affects only the geometry and vanishes in the zero-regularisation limit, recovering the native Cramér metric. This framework clarifies the operator structure underlying distributional Bellman updates and provides a foundation for further functional and operator-theoretic analyses in DRL.

Many distributional quantities in reinforcement learning are intrinsically joint across actions, including distributions of gaps and probabilities of superiority. However, the classical Markov decision process (MDP) formalism specifies only marginal laws and leaves the joint law of counterfactual one-step outcomes across multiple possible actions at a state unspecified. We study coupled-dynamics environments with a multi-action generative interface which can sample counterfactual one-step outcomes for multiple actions under shared exogenous randomness. We propose joint MDPs (JMDPs) as a formalism for such environments by augmenting an MDP with a multi-action sample transition model which specifies a coupling of one-step counterfactual outcomes, while preserving standard MDP interaction as marginal observations. We adopt and formalize a one-step coupling regime where dependence across actions is confined to immediate counterfactual outcomes at the queried state. In this regime, we derive Bellman operators for $n$th-order return moments, providing dynamic programming and incremental algorithms with convergence guarantees.

Autonomous driving involves multiple, often conflicting objectives such as safety, efficiency, and comfort. In reinforcement learning (RL), these objectives are typically combined through weighted summation, which collapses their relative priorities and often yields policies that violate safety-critical constraints. To overcome this limitation, we introduce the Preordered Multi-Objective MDP (Pr-MOMDP), which augments standard MOMDPs with a preorder over reward components. This structure enables reasoning about actions with respect to a hierarchy of objectives rather than a scalar signal. To make this structure actionable, we extend distributional RL with a novel pairwise comparison metric, Quantile Dominance (QD), that evaluates action return distributions without reducing them into a single statistic. Building on QD, we propose an algorithm for extracting optimal subsets, the subset of actions that remain non-dominated under each objective, which allows precedence information to shape both decision-making and training targets. Our framework is instantiated with Implicit Quantile Networks (IQN), establishing a concrete implementation while preserving compatibility with a broad class of distributional RL methods. Experiments in Carla show improved success rates, fewer collisions and off-road events, and deliver statistically more robust policies than IQN and ensemble-IQN baselines. By ensuring policies respect rewards preorder, our work advances safer, more reliable autonomous driving systems.

Equalizer parameter optimization is critical for signal integrity in high-speed memory systems operating at multi-gigabit data rates. However, existing methods suffer from computationally expensive eye diagram evaluation, optimization of expected rather than worst-case performance, and absence of uncertainty quantification for deployment decisions. In this paper, we propose a distributional risk-sensitive reinforcement learning framework integrating Information Bottleneck latent representations with Conditional Value-at-Risk optimization. We introduce rate-distortion optimal signal compression achieving 51 times speedup over eye diagrams while quantifying epistemic uncertainty through Monte Carlo dropout. Distributional reinforcement learning with quantile regression enables explicit worst-case optimization, while PAC-Bayesian regularization certifies generalization bounds. Experimental validation on 2.4 million waveforms from eight memory units demonstrated mean improvements of 37.1% and 41.5% for 4-tap and 8-tap equalizer configurations with worst-case guarantees of 33.8% and 38.2%, representing 80.7% and 89.1% improvements over Q-learning baselines. The framework achieved 62.5% high-reliability classification eliminating manual validation for most configurations. These results suggest the proposed framework provides a practical solution for production-scale equalizer optimization with certified worst-case guarantees.

Current reinforcement learning objectives for large-model reasoning primarily focus on maximizing expected rewards. This paradigm can lead to overfitting to dominant reward signals, while neglecting alternative yet valid reasoning trajectories, thereby limiting diversity and exploration. To address this issue, we introduce Learning Advantage Distributions (LAD), a distribution-matching framework that replaces advantage maximization with learning the advantage-induced distribution. By establishing the equivalence between the optimal policy update and an advantage-based target distribution, we derive a practical LAD objective formulated as minimizing an $f$-divergence between the policy-induced and advantage-induced distributions. This yields a gradient update that increases likelihood for high-advantage responses while suppressing over-confident probability growth, preventing collapse without requiring auxiliary entropy regularization. LAD incurs no extra training cost compared to GRPO and scales naturally to LLM post-training. In a controlled bandit setting, LAD faithfully recovers the multimodal advantage distribution, validating the theoretical formulation. Experiments on math and code reasoning tasks across several LLM backbones show that LAD reliably improves both accuracy and generative diversity.

Reliable value estimation serves as the cornerstone of reinforcement learning (RL) by evaluating long-term returns and guiding policy improvement, significantly influencing the convergence speed and final performance. Existing works improve the reliability of value function estimation via multi-critic ensembles and distributional RL, yet the former merely combines multi point estimation without capturing distributional information, whereas the latter relies on discretization or quantile regression, limiting the expressiveness of complex value distributions. Inspired by flow matching's success in generative modeling, we propose a generative paradigm for value estimation, named FlowCritic. Departing from conventional regression for deterministic value prediction, FlowCritic leverages flow matching to model value distributions and generate samples for value estimation.

Risk-sensitive reinforcement learning (RL) is crucial for maintaining reliable performance in high-stakes applications. While traditional RL methods aim to learn a point estimate of the random cumulative cost, distributional RL seeks to estimate the entire distribution of it, leading to a unified framework for handling different risk measures. However, developing policy gradient methods for risk-sensitive distributional RL is inherently more complex as it often involves finding the gradient of a probability measure. This paper introduces a new distributional policy gradient framework for risk-sensitive RL, where we derive an analytical gradient of the probability measure of the cumulative cost. For practical implementation, we further design a categorical distributional policy gradient algorithm (CDPG) that approximates arbitrary distributions using a categorical family supported on fixed points. Using Conditional Value-at-Risk (CVaR) as the objective, we prove that the proposed CDPG converges to stationary points and establish its iteration complexities under inexact policy evaluation. Through experiments in a stochastic Cliffwalk environment, we demonstrate the effectiveness of the proposed algorithm and highlight the benefits of incorporating risk sensitivity into distributional RL.