Collected sources and patterns will appear here. Add from search or the patterns library.
GroupCompletions -> RelativeRewards
Compute relative policy advantages from group-level rewards across multiple outputs generated for a single prompt, eliminating the memory and compute overhead of a separate critic model.
Problem it solves
Standard RLHF algorithms require hosting and backpropagating through an active critic network, which significantly increases memory overhead during training.
Consumes
Emits
The real projects this mechanism was found in. Attribution is the point — this is how the best teams actually do it.