Related work on Reward design-generative reward

[YuxuanJiang1]

Thanks for maintaining this great survey! I’d like to suggest our recent work: SCRIBE: Structured Mid-Level Supervision for Tool-Using Language Models.

It introduces a mid-level abstraction framework that anchors reward modeling in "Skill Prototypes" to solve the credit assignment problem in multi-step tool-use.

Thanks again for consideration!