trnasition-level reward design

[boardman0]

In a fixed workflow with multiple roles (each defined by a distinct system prompt), AgentLightning models each role’s I/O as transitions and may group them. Without role-level rewards, is training still effective, or does it degenerate into GRPO.

[ultmaster]

I think it's related to #31

[boardman0]

Thank you for the reply. In the absence of role-level rewards and with a static workflow, can we regard the setup as equivalent to GRPO optimized with a trajectory-level reward?

[ultmaster]

I think it's slightly different. In GRPO for RLHF, each trajectory = one generated response. In our setup, each trajectory = multiple responses.