MAESTRO: Meta-learning Adaptive Estimation of Scalarization Trade-offs for Reward Optimization
This addresses the problem of aligning large language models in open-domain scenarios for AI researchers and practitioners, though it is incremental as it builds upon GRPO.
The paper tackles the challenge of extending Group-Relative Policy Optimization (GRPO) to open-domain settings with conflicting objectives like creativity and factuality, by proposing MAESTRO, a meta-learning method that dynamically adapts reward scalarization, resulting in consistent outperformance over baselines across seven benchmarks while preserving GRPO's efficiency.
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.