multi-reward

Definition

A setup in which learning is driven by several reward signals at once instead of one scalar. It arises whenever the objectives genuinely conflict — helpfulness against safety, quality against brevity — since collapsing them into a weighted sum quietly fixes the trade-off in advance. The research question is how to set or adapt those weights, or how to optimise without scalarising at all.

Explainers using this term

• Paper Walkthrough: SA-MRPO — Stop Studying the Subject You've Already AcedLearn What's Left, Not What's Mastered: Saturation Aware Advantage Reweighting for Multi-Reward Policy Optimization