The paper proposes Population-Curated Policy Optimization (PCPO), a method for algorithm-design agents that learn from self-generated algorithms to overcome in-context evolutionary stagnation. They characterize the stagnation issue and introduce the Improvement Chain proposition, which motivates learning from successive algorithms. PCPO uses a global population to retain diverse high-quality algorithms and a hybrid policy update scheme that combines on-policy exploration with off-policy learning from elite algorithms.
In the task of designing learning rate schedules for global placement in electronic design automation (EDA), trained on only 4 chip cases, PCPO outperforms state-of-the-art in-context evolutionary methods (e.g., OpenEvolve, ShinkaEvolve) on average across 16 chip cases. With an 8B-size base model, PCPO achieves competitive performance with GPT-5.5, reduces inference-time token cost, and demonstrates significant speedups on four GPU kernel designs (8.27x average speedup against PyTorch Eager baseline). These results show that parametric self-evolution can efficiently escape stagnation and generalize from limited high-quality algorithms.