Interpreting and shaping human behavior in language model simulations
Interpreting and Steering LLM Agents for Social Simulations
Machine LearningArtificial IntelligenceComputation and Language
Summary
Simulations that use large language models can help understand how people behave, but these models often act like black boxes whose decisions are hard to explain or change. The authors show ways to look inside these models to understand what drives behavior and to control it better. They tested three methods on tasks involving human preferences and creativity, finding that some approaches can steer model behavior more reliably than others. This makes it easier for social scientists to both interpret and influence simulated agents in studies.
What this means in practice
- •For interactive system designers: Use probe and feature-based steering to create AI agents whose behaviors can be reliably adjusted in dialogue-based simulations of human decisions.
- •For game developers: Improve non-player characters by interpreting and directing their simulated preferences and creativity to produce more realistic social interactions.
Authors
Jiayue Gaveal Fan, Arul Murugan, Shreyas Krishnan, Abhishek Nagaraj
Abstract
Simulations based on large language models (LLMs) have proven to be powerful for understanding human behavior, making them valuable additions to the social scientific toolkit. However, LLMs are ultimately black boxes based on deep neural networks which limits their value for social science. This is because of a lack of (i) interpretability: i.e. the ability to assign clear mechanisms driving observed behavior; and a lack of (ii) steerability: i.e. the ability to mute or amplify specific theoretically meaningful mechanisms of action to drive specific model behavior. Here, we demonstrate how the black box could be opened up to further enrich LLM-based simulations. Specifically, we compare three types of methods: (1) prompt-based manipulation, (2) SAE-derived feature steering, and (3) probe-based direction steering and examine their utility for LLM-based social scientific simulations. We do so by interpreting and steering two foundational components of human behaviors, namely preferences (risk attitudes, altruism) and capabilities (divergent creativity, product innovation), operationalized using four classic economic and creative tasks implemented as natural-language interactions. Overall, our results show that SAE- and probe-based techniques often outperform basic prompt-based methods for steering LLM agents, although this advantage depends on the specific prompting strategy involved. Together, SAEs and probes constitute an effective pipeline for social scientists seeking to interpret and steer agents in social simulations: SAEs decompose agents' internal representations into human-readable features, after which probes can reliably shift agents' behaviors in specified directions. We discuss implications of these methods for future work using LLM agents for social scientific simulations.