Curriculum-LoRA cuts LLM community simulation cost 10x

A new paper titled "Benchmarking LLMs for Community Governance Simulation with Life-history Narratives" presents a comprehensive framework for using AI to simulate individual residents' attitudes in community governance. The dataset consists of approximately 1.2 million characters of first-person narrative collected through two-hour semi-structured interviews with each of 92 residents in an urban community, organized around nine community-governance domains. The authors benchmarked 18 mainstream LLMs across four prompting strategies and found that adding rich life-history profiles meaningfully raises fidelity compared to simple demographic prompts, but at the cost of longer prompts and more input tokens.

To solve the fidelity-cost gap, the researchers developed curriculum-LoRA, a parameter-efficient personalization algorithm that achieves the same fidelity as the best baseline while reducing per-call cost by roughly 10x. This Pareto-dominates all tested configurations. The system integrates curriculum-LoRA into a closed-loop policy-evaluation pipeline, making individual-level LLM-based resident simulation practical for resource-constrained local governments. The work brings systematic pre-evaluation of community governance policies in silico before real-world deployment, potentially transforming how local administrations gather and act on resident feedback.