SOCIA-EVO: Automated Simulator Construction via Dual-Anchored Bi-Level Optimization

TL;DR

SOCIA-EVO is a novel dual-anchored evolutionary framework for automated simulator construction that ensures statistical fidelity and robust convergence through bi-level optimization and strategic self-curation.

Contribution

It introduces a dual-anchored bi-level optimization approach with a static blueprint and self-curating Strategy Playbook for improved simulator fidelity.

Findings

Achieves statistically consistent simulators with observational data.

Effectively manages structural and parametric errors in long-horizon LLM agents.

Demonstrates robust convergence through execution feedback and strategy falsification.

Abstract

Automated simulator construction requires distributional fidelity, distinguishing it from generic code generation. We identify two failure modes in long-horizon LLM agents: contextual drift and optimization instability arising from conflating structural and parametric errors. We propose SOCIA-EVO, a dual-anchored evolutionary framework. SOCIA-EVO introduces: (1) a static blueprint to enforce empirical constraints; (2) a bi-level optimization to decouple structural refinement from parameter calibration; and (3) a self-curating Strategy Playbook that manages remedial hypotheses via Bayesian-weighted retrieval. By falsifying ineffective strategies through execution feedback, SOCIA-EVO achieves robust convergence, generating simulators that are statistically consistent with observational data. The code and data of SOCIA-EVO are available here: https://github.com/cruiseresearchgroup/SOCIA/tree/evo.