From Representations to Behaviors: Exploring the Person–Situation–Behavior Triad in LLMs

This paper investigates whether Large Language Models (LLMs) possess internal, mechanistic representations of personality that function similarly to human personality traits. Drawing on Funder's personality triad framework—which posits that personality is the interplay between the Person (internal state), the Situation (context), and the Behavior (observable action)—the authors seek to bridge the gap between abstract personality scores and concrete model behavior.

The researchers use sparse autoencoders (SAEs) to decompose the model's hidden states into sparse, interpretable features. By constructing contrastive pairs of high- and low-trait responses to identical social situations, they identify specific SAE features that correlate with opposing poles of the Big Five personality traits. They then validate these features through causal interventions, checking whether modifying these internal activations consistently shifts the model's behavior across new, unseen scenarios while maintaining coherent, instruction-following responses.

The study provides three primary contributions. First, it successfully identifies and isolates sparse internal features that represent personality traits, moving beyond simple lexicosyntactic patterns. Second, it demonstrates that intervening on these features allows for bidirectional control of personality expression; for example, increasing an 'Agreeableness' feature makes the model more empathetic across diverse, novel social contexts. Third, the authors show that these interventions produce behavioral trade-offs consistent with human psychology. For instance, shifting a model toward higher 'Extraversion' or 'Agreeableness' results in predictable changes in social intelligence tasks, confirming that these internal representations are not just superficial labels but are linked to functional behavioral outcomes.

Most existing research on LLM personality relies on 'persona prompting' or surface-level output analysis, which often fails to capture the underlying mechanisms of how a model 'thinks' about a personality. By mapping personality to specific, controllable internal features, this work provides a mechanistic foundation for building more reliable and steerable AI agents. This approach allows researchers to move from simply 'asking' a model to act a certain way to 'tuning' its internal state, offering a more robust method for aligning AI behavior with human-like social and psychological expectations.