This work proposes PRISM, a new multi-reward RL framework built upon the idea of policy-space decomposition and composition, which alleviates the potential conflict during multi-reward policy optimization, while enabling controllability during inference by flexible policy composition.
Abstract
Modern large language models (LLMs) are expected not just to answer correctly, but to adapt their behavior to different human values and use cases. As a result, multi-reward reinforcement learning (RL) has become an increasingly important problem for LLMs, where each reward captures a different aspect of desired behavior. However, optimizing with multiple rewards suffers from a more severe alignment tax issue, where different optimization objectives can trade off or even conflict with each other, leading to unstable and inefficient post-training. In this work, we propose PRISM, a new multi-reward RL framework built upon the idea of policy-space decomposition and composition. Instead of compositing different rewards, PRISM optimizes a set of standalone positive policies and a global negative policy. This alleviates the potential conflict during multi-reward policy optimization, while enabling controllability during inference by flexible policy composition. Experiments on scientific reasoning, tool-use reasoning, and helpfulness-safety alignment show that PRISM consistently outperforms existing multi-reward RL baselines, with extra controllability for inference-time preference control.
Reinforcement learning (RL) with group-relative advantages has become the de facto standard for post-training language model reasoners. However, when optimizing multiple reward objectives, existing methods typically scalarize the reward vector with a fixed weighted sum before group-wise standardization. We show that this design leads to two fundamental problems: rollouts with distinct reward profiles can receive identical advantages, and all objectives are optimized with fixed relative weights regardless of their current level of saturation. As a result, training continues to allocate gradient budget to already-solved objectives instead of focusing on those with greater remaining headroom. We introduce \textbf{Saturation Aware Advantage Reweighting for Multi-Reward Policy Optimization} (SA-MRPO), which standardizes each reward objective independently and adaptively discounts its contribution according to a batch-level estimate of objective saturation. This dynamically reallocates optimization effort toward under-optimized objectives while empirically maintaining performance on those that are already well satisfied. We further show that saturation-aware reweighting can reverse the sign of an update, rather than merely rescale its magnitude. Across mathematical reasoning with two- and three-objective reward combinations, SA-MRPO improves the harder correctness objective over GDPO in 12 of 15 benchmark comparisons, with gains of up to $5\%$ on AIME24. On adaptive reasoning it improves accuracy on all five benchmarks, by $3.8\%$ on average and up to $9.2 \%$ on AMC23, and on coding benchmarks it improves pass rate by up to $2.3\%$, while in all settings maintaining the easier objectives near their already satisfied levels.
Yixuan Wang, Yifei Chen, Haichao Zhang et al.· 0 citations
Reinforcement learning holds significant potential for training large language models to handle multi-turn interactive tasks, but directly training with outcome rewards often results in slow convergence due to the sparsity of signals and the lack of fine-grained feedback.
Qiang Liu, Taian Guo, Ruizhi Qiao et al.· 0 citations
Learning a reward model from human feedback and optimizing a policy against it is one approach to aligning AI systems with individual users. From a fairness perspective, existing work improves such alignment by developing data-efficient and accurate reward models that capture minority preferences despite scarce data. We push this line of inquiry one step further and argue that data-efficient and accurate per-user reward models are not sufficient: users whose reward models are difficult to \textit{optimize} at the policy level can become a new underserved group. We start from the observation that one user's reward model can be easy to optimize from the initial policy while another's is not. We argue that, given a sufficiently diverse user population, a curriculum naturally emerges between easy- and hard-to-optimize reward models. Building on this insight, we propose CurriPO, which grows a tree-structured curriculum to accommodate diverse user-specific objectives, covering the population in a single traversal. Specifically, CurriPO automatically constructs a curriculum over diverse user reward models, allowing it to branch from the existing curriculum and reuse reward models previously incorporated into the curriculum. To the best of our knowledge, this is the first work to explicitly exploit multi-user structure to address optimization in AI alignment. Extensive experiments on personalized continuous control in a simulated environment show that CurriPO achieves $1.2$--$2.1\times$ the population satisfaction of the strongest baseline while substantially reducing training time. Additional analysis attributes much of this improvement to the users left underserved by conventional optimization.
Large language models achieve strong reasoning performance, but often at prohibitive training cost - a challenge that is especially acute for compact models ($\leq 4 \, \mathrm{B}$ parameters) trained under limited budgets. We introduce MADA-RL, a post-training framework that specializes compact models into generator and critic roles and trains them with a debate-aware learning signal, fine-tuning only a small subset of parameters via LoRA adapters. Our central contribution is a counterfactual critic advantage: a dynamic, role-conditioned baseline that redefines the critic's advantage as its reward minus the generator ensemble's per-instance accuracy. This explicitly optimizes critics to improve over generator consensus rather than to merely reproduce a correct answer, yielding more targeted credit assignment than static mean-reward normalization. At deployment, the specialized agents are composed in a lightweight multi-round protocol. Across five mathematical reasoning benchmarks, MADA-RL raises the accuracy of the DeepSeek-R1-Distill-Qwen-1.5B model from $39.9 \, \%$ to $41.9 \, \%$ ($+2.0$ points, $p<0.001$) using $16$ times fewer trainable parameters than fully fine-tuned baselines, placing it on the accuracy-trainable-parameter Pareto front. It approaches, but does not surpass, the strongest baselines (DeepScaleR, STILL-3), which are trained on substantially larger datasets; we analyse this gap and the associated inference-time cost directly. A controlled study isolates the source of MADA-RL's gains: the counterfactual advantage produces the highest critic improvement rate of any model evaluated, indicating that trained critics learn to correct generator errors rather than to imitate them.
Martino M. L. Pulici, Cuong Xuan Chu, Evgeny Kharlamov et al.· 0 citations
Deep reinforcement learning (RL) agents achieve strong performance by optimizing scalar reward functions. However, once deployed, the policies of these RL agents are often rigid and costly to adapt to new performance criteria. For instance, an agent trained to maximize expected cumulative reward may not accommodate previously unknown stakeholder preferences. Existing approaches to achieve fairness, a type of preference, in RL typically assume that such preferences are known a priori and require complete retraining of the policy under a fairness-oriented metric. Inspired by inference-time alignment in large language models, we investigate the problem of steering a pretrained RL policy toward welfare-based fairness objectives at inference time without updating the base policy's parameters. We formalize inference-time fairness alignment as a policy shaping problem and propose a multiplicative policy shaping framework that adjusts action probabilities using action-dependent welfare scores, thus requiring no modification to the base policy. Our framework is general and compatible with any deep RL agent. Through extensive experiments across multiple domains, we demonstrate that inference-time policy shaping substantially improves welfare-based fairness objectives while preserving core task performance.
Umer Siddique, Peilang Li, Conor Wallace et al.· 0 citations
This work proposes LaRes, a novel hybrid framework that achieves efficient policy learning through reward function search by leveraging large language models to generate the reward function population, guiding RL in policy learning.
Pengyi Li, Hongyao Tang, Jinbin Qiao et al.· Advances in Neural Informati...· 4 citations