Can LLMs Anticipate Behavioral Responses to Social Policies? FlexPension-LLM for Flexible Workers
Original title:Can Large Language Models Anticipate Behavioral Responses to Social Policies? A Case of Pension Enrollment Prediction among China's Flexible Workers
Evaluating social policy shifts often stalls between costly field trials and brittle econometric extrapolations. To predict pension enrollment choices among China’s flexible workers, researchers introduced FlexPension-LLM, distilling provincial pension rules and Probit-derived marginal effects into an open-weight MoE architecture. Tested on a CHFS 2019 blind split, it achieved a 0.9316 composite F1—surpassing its Claude Sonnet 4.5 teacher and matching Claude Opus 4.6. The framework demonstrates how rationale-guided distillation can simulate micro-level economic behaviors while maintaining auditable policy decision traces.
Why it's worth reading
It shows how injecting econometric priors into open-weight distillation can effectively simulate micro-level policy responses, rivaling frontier models while retaining auditable reasoning traces.