Can Large Language Models Anticipate Behavioral Responses to Social Policies? A Case of Pension Enrollment Prediction among China's Flexible Workers

On September 7, 2026, researchers published the first dedicated model for predicting pension participation among flexible workers in China, FlexPension-LLM, on arXiv. This model uses the DKI-RDistill method to inject policy-based prompts containing Probit marginal effects and household registration province rules. It utilizes LoRA/SFT to enhance supervised distillation into an open-source mixed expert (MoE) student model, with teacher errors corrected by real labels. In the CHFS 2019 blind test, FlexPension-LLM achieved a composite F1 score of 0.9316, surpassing its Claude Sonnet 4.5 teacher and 17 baseline models, and showing no significant statistical difference from Claude Opus 4.6; in four external surveys, it averaged a composite F1 of 0.7549, with the narrowest performance range. Analysis indicates that the performance improvement is mainly due to policy-based prompt injection and error correction mechanisms.