aiqbench analysis
This research introduces PRISM, a novel reinforcement learning framework designed to resolve conflicts between multiple optimization objectives in large language models. By decomposing policy space into standalone positive policies and a global negative component, the method effectively mitigates alignment tax issues while enhancing inference-time controllability for diverse use cases like scientific reasoning and safety alignment.
Don't Mix Rewards, Mix Policies: Policy Decomposition and Optimization for Multi-Reward RL
arXiv:2607.29246v1 Announce Type: new 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.
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.
Artificial Intelligence (cs.AI)
arXiv:2607.29246 [cs.AI]
(or arXiv:2607.29246v1 [cs.AI] for this version)
https://doi.org/10.48550/arXiv.2607.29246 Focus to learn more arXiv-issued DOI via DataCite (pending registration)
arXivLabs is a framework that allows collaborators to develop and share new arXiv features directly on our website.
Both individuals and organizations that work with arXivLabs have embraced and accepted our values of openness, community, excellence, and user data privacy. arXiv is committed to these values and only works with partners that adhere to them.
Have an idea for a project that will add value for arXiv's community? Learn more about arXivLabs.
Full text under the publisher's open licence · 256 words
Related coverage
[AINews] not much happened today
Latent Space · Sep 18, 2026
Reinforcement Learning with Decomposed Subtasks
arXiv · Sep 24, 2026
[AINews] not much happened today
Latent Space · Oct 3, 2026
TaPR: Test-Aware Policy Refinement for Feedback-Conditioned Code Generation
arXiv · Aug 5, 2026