ARGUS Neuro-Symbolic Framework for Reliable User Simulation

ARGUS uses a neuro-symbolic approach with a dynamic agenda stack to prevent goal omission in LLM-based user simulators. The framework can be distilled into compact models via evaluator-guided SFT and online RL, achieving higher behavioral fidelity on MultiWOZ benchmarks.