Turning Structure into Reward: Toward Reliable Self-Supervised Reinforcement Learning
Date:
Invited workshop talk at Jump Trading AI Workshop, Hong Kong
I shared an overview of our work on general-purpose self-supervised reinforcement learning at the Jump Trading AI Workshop in Hong Kong.
The presentation connects three sources of feedback already present in tasks and models: task duality, uneven capabilities across related tasks, and input-output structure. It discusses how DuPO, MAPO, and related methods turn these structures into reward or preference signals for self-verification, strong-to-weak transfer, and reliable improvement without relying on dense human annotations.
The poster is available as a PDF.