Pinned post
We extend @cognition’s SWE-2 reward function to steer the Pareto frontier, not just improve it.
SWE-2 uses S − λₑC, where λₑ is cleverly the frontier’s derivative at that effort.
We show that adapting λₑ can target a desired score/cost improvement balance. See below!
SWE-2 uses S − λₑC, where λₑ is cleverly the frontier’s derivative at that effort.
We show that adapting λₑ can target a desired score/cost improvement balance. See below!
1:13