Tag
This paper explores methods to interpret and steer large language model agents in social simulations, comparing prompt-based, SAE-based, and probe-based techniques, and finds that SAE and probe methods often outperform basic prompting for control and interpretability.