4 ms·
Without anchoring on the group average it's hardly Group-Relative and without likelihood there's no Policy; that doesn't leave much of the original GRPO unchang
by yorwba 3mo ago
Without anchoring on the group average it's hardly Group-Relative and without likelihood there's no Policy; that doesn't leave much of the original GRPO unchanged...
Do you have a particular paper in mind that does likelihood-free best-of-K but just calls it GRPO?
- zimablue22 3mo agoAny of the Best of N papers that exploded in popularity after GRPO. Likelihood is not fundamental to the spirit of GRPO, any exploratory mechanism would work. That sequential LLMs have a step-wise probability is convenient but not critical to this approach (where rejection sampling is widely used in diffusion models).
- kamranjon 3mo agoDo you have an example? Would love to read one.