This work develops LEMUR: Learning to Align with Multi-Objective Reinforcement Learning with Preference feedback, a novel framework where an agent interactively learns from the preferences of multiple humans to learn optimal multi-objective policies.
Abstract
Reinforcement Learning (RL) systems are typically trained using a single, well-specified scalar reward function. However, real-world decision-making tasks often involve multiple, competing objectives, such as performance versus efficiency, where ground-truth reward functions are difficult to specify or inaccessible. While Multi-Objective RL (MORL) addresses such trade-offs by modeling rewards as vectors, existing approaches typically assume access to a well-specified reward function for each objective, inheriting the same challenges faced by single-objective RL. Meanwhile, Preference-based RL (PbRL) has shown great potential in solving complex tasks without access to a pre-defined reward function through reward learning from human feedback, yet has largely been studied in single-objective settings. In this work, we bridge this gap with LEMUR: Learning to Align with Multi-Objective Reinforcement Learning with Preference feedback, a novel framework where an agent interactively learns from the preferences of multiple humans to learn optimal multi-objective policies. Our approach jointly learns policies and multiple objective-specific reward models from human feedback, enabling agents to effectively balance competing objectives during learning. We evaluate LEMUR on a variety of benchmark multi-objective tasks, and empirical results demonstrate its superior performance over baseline methods. Our method presents a promising direction for solving multi-objective decision-making tasks without pre-defined reward functions.
MeRLa (Meta-Learned Reward Shaping), a principled framework that meta-learns a task-aware shaping function across auxiliary tasks before RLHF training, is introduced, providing theoretical guarantees for policy invariance, analyze representation drift sensitivity, and formally address incentive misalignment from entropy maximization.
This paper introduces Multi-AGent Preference-Integrated lEarning (MAGPIE), a framework that leverages agent-specific preference signals in the multi-agent learning process and can derive Nash equilibrium solutions.
Ni Mu, Yao Luan, Yiqin Yang et al.· IEEE Transactions on Automat...· 0 citations
A unified view of ARD in RL is provided by introducing a taxonomy, organized by the primary driver of the reward variation, that distinguishes external-feedback-driven reward updates from reward adaptations driven by endogenous within-run signals and those conditioned on exogenous context signals.
Raphaela Baybas, Carlo D'Eramo, Philipp Brune· 0 citations
One of the main challenges in the field of artificial intelligence, and reinforcement learning (RL) in particular, is the development of generalist and flexible agents capable of solving multiple tasks—each requiring the agent to learn a potentially new, specialized behavior. Tackling this challenge requires agents to learn behaviors that may involve optimizing a single objective, or trading off between multiple conflicting objectives. In this thesis, we study how to design flexible RL agents that can, in a sample-efficient manner, adapt their behavior to solve any given tasks—each of which is defined by multiple (possibly conflicting) objectives. We introduce new multi-policy methods that empower RL agents to (i) carefully learn multiple behaviors, each specialized in a particular task; and (ii) combine previously-learned behaviors to efficiently identify solutions to novel tasks, which, importantly, may require the agent to assign different preferences to each of its new objectives. The methods we introduce have strong theoretical guarantees regarding the optimality of the set of behaviors learned by agents and their capability to solve new tasks in a zero-shot manner, even in the presence of function approximation errors. We evaluate the proposed methods in various challenging multi-task and multi-objective RL problems and show that our algorithms outperform various current state-of-the-art methods in domains with both discrete and continuous state and action spaces.
L. N. Alegre, Ana L. C. Bazzan, Bruno C. da Silva· Anais do XXXIX Concurso de T...· 0 citations
Q-based Variational IRL (QVIRL), a novel Bayesian IRL method that recovers a posterior distribution over rewards from expert demonstrations via primarily learning a variational distribution over optimal Q-values, is introduced.
Ondrej Bajgar, Peter Tisnikar, Alessandro Abate et al.· 0 citations
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
We use cookies to run the site and, with your consent, for analytics and to show ads.
See our Cookie Policy.
