Search papers, labs, and topics across Lattice.
This study investigates the phenomenon of "persona collapse" in large language models (LLMs) when providing advice across diverse contexts, revealing that existing models often default to a single supportive persona despite varying situational needs. By analyzing 1,281 advice posts, the authors demonstrate that while human advisors adapt their personas based on context, leading models fail to do so, collapsing over 90% of their responses into a singular persona. The proposed method, Inverse-Process Distillation, significantly reduces divergence from human-like persona distribution by about 80%, yet human evaluators still prefer the original collapsed persona in many scenarios, highlighting the complexity of effective advice-giving.
Despite reducing persona collapse by 80%, LLMs still struggle to match human adaptability in advice-giving, with users favoring the default persona even in challenging situations.
LLMs are increasingly used for personal advice on relationships, work, moral dilemmas, and crises. Post-training selects a stable, prosocial Assistant persona, but good advice requires more than a good default character: a skilled advisor comforts someone in crisis, challenges someone in denial, and stays procedural with a logistical question. We formalize advice-giving as situation-conditioned persona selection in a space defined by hedonic tone and agency support, and call failures of this mapping"persona collapse"(the compression of diverse situations into a single default persona). Across 1,281 advice posts spanning 14 contexts, top-rated human responses shift systematically across five personas, while three frontier models collapse over 90\% of responses into a single supportive persona regardless of context. Prompting the model to first pick a fitting persona only deepens the collapse. We then ask whether the collapse can be repaired. Our method, Inverse-Process Distillation, reconstructs the situational reading that could have produced each human response and trains on the result, aiming to distill the situation-to-persona policy rather than the answers. It cuts divergence from the human persona distribution by approximately 80\%. Yet in a blinded study, 199 experienced advice-givers rating responses across four situations in sequence prefer the collapsed default over every repaired model, most strongly when the situation calls for challenge, though this preference shifts with repeated exposures.
