Beyond Fixed Psychological Personas: State Beats Trait, but Language Models are State-Blind

TL;DR

This paper introduces Chameleon, a dataset of 5,001 psychological profiles across contexts, revealing that language models are state-blind and primarily focus on user traits, with implications for personalized AI and alignment.

Contribution

The paper presents Chameleon, a novel dataset capturing user state and trait, and demonstrates that LLMs ignore user state while reward models react inconsistently to it.

Findings

74% of variance is within-person (state)

LLMs focus mainly on trait, ignoring state

Reward models react differently to user state

Abstract

User interactions with language models vary due to static properties of the user (trait) and the specific context of the interaction (state). However, existing persona datasets (like PersonaChat, PANDORA etc.) capture only trait, and ignore the impact of state. We introduce Chameleon, a dataset of 5,001 contextual psychological profiles from 1,667 Reddit users, each measured across multiple contexts. Using the Chameleon dataset, we present three key findings. First, inspired by Latent State-Trait theory, we decompose variance and find that 74% is within-person(state) while only 26% is between-person (trait). Second, we find that LLMs are state-blind: they focus on trait only, and produce similar responses regardless of state. Third, we find that reward models react to user state, but inconsistently: different models favor or penalize the same users in opposite directions. We release Chameleon to support research on affective computing, personalized dialogue, and RLHF alignment.