Analyzing Refinement in Data-Free Evolution of Code-Generation LLMs

Data-free evolution methods reduce reliance on human-curated data: a proposer generates tasks, and a solver learns to solve them. The solver's performance is determined by the quality of its generated tasks, which is often improved through a refinement operation that filters or selects important tasks for solver training. In this paper, we investigate when refinement actually helps the solver, from two specific perspectives. We first identify from the solver-gradient: the refinement affects the solver only through the alignment between the solver gradient and the downstream task performance. Then, we investigate the proposer-support: the refinement can only select tasks that the proposer emits, so the solver performance is limited by the capability of the proposer. Empirically, on code generation benchmarks, we find existing refinement operators can not significantly outperform the non-refinement baseline. Furthermore, using a stronger proposer can produce more diverse and high-quality tasks, thus improving the solver's performance even without refinement.