Aligning large language models with pointwise absolute rewards has so far required online, on-policy algorithms such as PPO and GRPO. In contrast, simpler methods that can leverage offline or off-policy data, such as DPO and REBEL, are limited to learning from preference pairs or relative signals. To bridge this gap, we introduce Quantile Reward Policy Optimization (QRPO), which learns from pointwise absolute rewards while preserving the simplicity and offline applicability of DPO-like methods. QRPO uses quantile rewards to enable regression to the closed-form solution of the KL-regularized RL objective. This reward yields an analytically tractable partition function, removing the need for relative signals to cancel this term. Moreover, QRPO scales with increased compute to estimate quantile rewards, opening a new dimension for pre-computation scaling. Empirically, QRPO consistently achieves top performance on chat and coding evaluations--reward model scores, AlpacaEval 2, and LeetCode--compared to DPO, REBEL, and SimPO across diverse datasets and 8B-scale models. Finally, we find that training with robust rewards instead of converting them to preferences induces less length bias.
State University of New York at Stony Brook（纽约州立大学石溪分校）；Peking University（北京大学）
Classical reinforcement learning (RL) aims to optimize the expected cumulative rewards. In this work, we consider the RL setting where the goal is to optimize the quantile of the cumulative rewards. We parameterize the policy controlling actions by neural networks and propose a novel policy gradient algorithm called Quantile-Based Policy Optimization (QPO) and its variant Quantile-Based Proximal Policy Optimization (QPPO) to solve deep RL problems with quantile objectives. QPO uses two coupled iterations running at different time scales for simultaneously estimating quantiles and policy parame…
IRPM: Intergroup Relative Preference Modeling for Pointwise Generative Reward Models
HUJING Digital Media \& Entertainment Group (XingYun Lab), Beijing, China；Department of Automation, Tsinghua University, Beijing, China；Fudan University, Shanghai, China；Beihang University, Beijing, China；Institute of Information Engineering, Chinese Academy of Sciences, Beijing, China
Allen Institute for Artificial Intelligence（艾伦人工智能研究所）；Paul G. Allen School of Computer Science, University of Washington（华盛顿大学保罗·G·艾伦计算机科学学院）
Large-scale language models often learn behaviors that are misaligned with user expectations. Generated text may contain offensive or toxic language, contain significant repetition, or be of a different sentiment than desired by the user. We consider the task of unlearning these misalignments by fine-tuning the language model on signals of what not to do. We introduce Quantized Reward Konditioning (Quark), an algorithm for optimizing a reward function that quantifies an (un)wanted property, while not straying too far from the original model. Quark alternates between (i) collecting samples with…
Yale University（耶鲁大学）；Rutgers, The State University of New Jersey（罗格斯大学（新泽西州立大学））；Bar-Ilan University（巴伊兰大学）
We study the best-arm identification problem in multi-armed bandits with stochastic, potentially private rewards, when the goal is to identify the arm with the highest quantile at a fixed, prescribed level. First, we propose a (non-private) successive elimination algorithm for strictly optimal best-arm identification, we show that our algorithm is $δ$-PAC and we characterize its sample complexity. Further, we provide a lower bound on the expected number of pulls, showing that the proposed algorithm is essentially optimal up to logarithmic factors. Both upper and lower complexity bounds depend …
Sorbonnes Universités UPMC Univ Paris 06（索邦大学（巴黎第六大学））；CNRS（法国国家科学研究中心）；School of Electronics and Information Technology（电子与信息学院）；SYSU-CMU Joint Institute of Engineering（中山大学-卡内基梅隆大学联合工程学院）；SYSU-CMU Shunde Joint Research Institute（中山大学-卡内基梅隆大学顺德联合研究院）
In reinforcement learning, the standard criterion to evaluate policies in a state is the expectation of (discounted) sum of rewards. However, this criterion may not always be suitable, we consider an alternative criterion based on the notion of quantiles. In the case of episodic reinforcement learning problems, we propose an algorithm based on stochastic approximation with two timescales. We evaluate our proposition on a simple model of the TV show, Who wants to be a millionaire.
Rice University（莱斯大学）；National University of Singapore（新加坡国立大学）；University of Illinois Urbana-Champaign（伊利诺伊大学厄巴纳-香槟分校）
In this paper, we consider a bandit problem in which there are a number of groups each consisting of infinitely many arms. Whenever a new arm is requested from a given group, its mean reward is drawn from an unknown reservoir distribution (different for each group), and the uncertainty in the arm's mean reward can only be reduced via subsequent pulls of the arm. The goal is to identify the infinite-arm group whose reservoir distribution has the highest $(1-α)$-quantile (e.g., median if $α= \frac{1}{2}$), using as few total arm pulls as possible. We introduce a two-step algorithm that first req…
National Tsing Hua University（台湾清华大学）；NVIDIA Corporation（英伟达公司）
In fully cooperative multi-agent reinforcement learning (MARL) settings, the environments are highly stochastic due to the partial observability of each agent and the continuously changing policies of the other agents. To address the above issues, we integrate distributional RL and value function factorization methods by proposing a Distributional Value Function Factorization (DFAC) framework to generalize expected value function factorization methods to their DFAC variants. DFAC extends the individual utility functions from deterministic variables to random variables, and models the quantile …
DeepMind；University of Cambridge（剑桥大学）；Google Brain（谷歌大脑）
In reinforcement learning an agent interacts with the environment by taking actions and observing the next state and reward. When sampled probabilistically, these state transitions, rewards, and actions can all induce randomness in the observed long-term return. Traditionally, reinforcement learning algorithms average over this randomness to estimate the value function. In this paper, we build on recent work advocating a distributional approach to reinforcement learning in which the distribution over returns is modeled explicitly instead of only estimating the mean. That is, we examine methods…
University of Torontoland（多伦多兰大学）；University of Edenborrow（伊登伯勒大学）
We study the stochastic Multi-Armed Bandit (MAB) problem with random delays in the feedback received by the algorithm. We consider two settings: the reward-dependent delay setting, where realized delays may depend on the stochastic rewards, and the reward-independent delay setting. Our main contribution is algorithms that achieve near-optimal regret in each of the settings, with an additional additive dependence on the quantiles of the delay distribution. Our results do not make any assumptions on the delay distributions: in particular, we do not assume they come from any parametric family of …
Optimistic Reinforcement Learning with Quantile Objectives
Grado Department of Industrial & Systems Engineering, Virginia Tech, Blacksburg, VA 24061, USA（弗吉尼亚理工大学工业与系统工程系）；Department of Industrial, Manufacturing, and Systems Engineering, University of Texas at Arlington, Arlington, TX 76019, USA（德克萨斯理工大学工业、制造与系统工程系）
In this paper, we contribute to the Extreme Bandit problem, a variant of Multi-Armed Bandits in which the learner seeks to collect the largest possible reward. We first study the concentration of the maximum of i.i.d random variables under mild assumptions on the tail of the rewards distributions. This analysis motivates the introduction of Quantile of Maxima (QoMax). The properties of QoMax are sufficient to build an Explore-Then-Commit (ETC) strategy, QoMax-ETC, achieving strong asymptotic guarantees despite its simplicity. We then propose and analyze a more adaptive, anytime algorithm, QoMa…
Many reinforcement learning (RL) problems in practice are offline, learning purely from observational data. A key challenge is how to ensure the learned policy is safe, which requires quantifying the risk associated with different actions. In the online setting, distributional RL algorithms do so by learning the distribution over returns (i.e., cumulative rewards) instead of the expected return; beyond quantifying risk, they have also been shown to learn better representations for planning. We propose Conservative Offline Distributional Actor Critic (CODAC), an offline RL algorithm suitable fo…
RUBRIC-ARROW: Alternating Pointwise Rubric Reward Modeling for LLM Post-training in Non-verifiable Domains
University at Albany（阿尔巴尼大学）；Rutgers University（罗格斯大学）；Purdue University（普渡大学）；Independent Researcher（独立研究员）；Emory University（埃默里大学）；Georgia Institute of Technology（佐治亚理工学院）
SPARK: Synergistic Policy And Reward Co-Evolving Framework
Shanghai Jiao Tong University（上海交通大学）；Shanghai Artificial Intelligence Laboratory（上海人工智能实验室）；Fudan University（复旦大学）；The Chinese University of Hong Kong（香港中文大学）；Shanghai Innovation Institute（上海创新研究院）
University of Alberta（阿尔伯塔大学）；Huawei Noah’s Ark（华为诺亚方舟实验室）；University of Oxford（牛津大学）；Huawei Hi-Silicon（华为海思）；University of Waterloo（滑铁卢大学）
In distributional reinforcement learning (RL), the estimated distribution of value function models both the parametric and intrinsic uncertainties. We propose a novel and efficient exploration method for deep RL that has two components. The first is a decaying schedule to suppress the intrinsic uncertainty. The second is an exploration bonus calculated from the upper quantiles of the learned distribution. In Atari 2600 games, our method outperforms QR-DQN in 12 out of 14 hard games (achieving 483 \% average gain across 49 games in cumulative rewards over QR-DQN with a big win in Venture). We a…
In traditional reinforcement learning, an agent maximizes the reward collected during its interaction with the environment by approximating the optimal policy through the estimation of value functions. Typically, given a state s and action a, the corresponding value is the expected discounted sum of rewards. The optimal action is then chosen to be the action a with the largest value estimated by value function. However, recent developments have shown both theoretical and experimental evidence of superior performance when value function is replaced with value distribution in context of deep Q l…
In Cooperative Multi-Agent Reinforcement Learning (MARL) and under the setting of Centralized Training with Decentralized Execution (CTDE), agents observe and interact with their environment locally and independently. With local observation and random sampling, the randomness in rewards and observations leads to randomness in long-term returns. Existing methods such as Value Decomposition Network (VDN) and QMIX estimate the value of long-term returns as a scalar that does not contain the information of randomness. Our proposed model QR-MIX introduces quantile regression, modeling joint state-a…
We tackle the problem of acting in an unknown finite and discrete Markov Decision Process (MDP) for which the expected shortest path from any state to any other state is bounded by a finite number $D$. An MDP consists of $S$ states and $A$ possible actions per state. Upon choosing an action $a_t$ at state $s_t$, one receives a real value reward $r_t$, then one transits to a next state $s_{t+1}$. The reward $r_t$ is generated from a fixed reward distribution depending only on $(s_t, a_t)$ and similarly, the next state $s_{t+1}$ is generated from a fixed transition distribution depending only on…
Off-Policy Safe Reinforcement Learning with Constrained Optimistic Exploration
Faculty of Mechanical Engineering（机械工程学院）；Delft University of Technology（代尔夫特理工大学）；Faculty of Electrical Engineering, Mathematics and Computer Science（电气工程、数学和计算机科学学院）
Inference-Time Scaling for Generalist Reward Modeling
DeepSeek-AI；Dept. of Computer Sci. & Tech., Tsinghua University（计算机科学与技术系，清华大学）；Institute for AI Industry Research (AIR), Tsinghua University（人工智能产业研究院（AIR），清华大学）
North Carolina State University（北卡罗来纳州立大学）；London School of Economics and Political Science（伦敦政治经济学院）；ByteDance（字节跳动）；University of Miami（迈阿密大学）
Off-Policy evaluation (OPE) is concerned with evaluating a new target policy using offline data generated by a potentially different behavior policy. It is critical in a number of sequential decision making problems ranging from healthcare to technology industries. Most of the work in existing literature is focused on evaluating the mean outcome of a given policy, and ignores the variability of the outcome. However, in a variety of applications, criteria other than the mean may be more sensible. For example, when the reward distribution is skewed and asymmetric, quantile-based metrics are ofte…
CLT and Edgeworth Expansion for m-out-of-n Bootstrap Estimators of The Studentized Median
Department of Industrial Engineering and Management Science Northwestern University（工业工程与管理科学系西北大学）；Department of Computer Science Northwestern University（计算机科学系西北大学）
