LLM-Driven Evolutionary Imputation

GPT-5.4's pointwise reconstruction degraded only modestly as missingness increased from 30% to 80%. GPT-5.4 had the best pointwise MAE and RMSE at 80% missingness and an ATE residual of 0.013. GPT-5.4 had the best pointwise MAE and RMSE at 50% missingness. GPT-5.4 had the best pointwise MAE and RMSE at 30% missingness. The evolutionary loop materially improved GPT-5.4 imputation performance, especially at 50% to 80% missingness. GPT-5.4 achieved the best pooled mean rank among imputation methods. The imputation baselines included LLM-driven evolutionary imputers using GPT-5.4, Qwen3.5-Plus, and GPT-OSS-120b.