MAESTRO: Meta-learning Adaptive Estimation of Scalarization Trade-offs for Reward Optimization

TL;DR

MAESTRO introduces a meta-learning approach that dynamically balances conflicting objectives in reward optimization for large language models, improving performance across diverse benchmarks.

Contribution

It proposes a novel meta-cognitive layer that treats reward scalarization as a latent policy, enabling adaptive trade-offs in open-domain LLM tasks.

Findings

Outperforms static reward baselines on seven benchmarks.

Reduces redundant generation in some settings.

Maintains efficiency advantages of GRPO.

Abstract

Group-Relative Policy Optimization (GRPO) has emerged as an efficient paradigm for aligning Large Language Models (LLMs), yet its efficacy is primarily confined to domains with verifiable ground truths. Extending GRPO to open-domain settings remains a critical challenge, as unconstrained generation entails multi-faceted and often conflicting objectives - such as creativity versus factuality - where rigid, static reward scalarization is inherently suboptimal. To address this, we propose MAESTRO (Meta-learning Adaptive Estimation of Scalarization Trade-offs for Reward Optimization), which introduces a meta-cognitive orchestration layer that treats reward scalarization as a dynamic latent policy, leveraging the model's terminal hidden states as a semantic bottleneck to perceive task-specific priorities. We formulate this as a contextual bandit problem within a bi-level optimization framework, where a lightweight Conductor network co-evolves with the policy by utilizing group-relative advantages as a meta-reward signal. Across seven benchmarks, MAESTRO consistently outperforms single-reward and static multi-objective baselines, while preserving the efficiency advantages of GRPO, and in some settings even reducing redundant generation.