Stick To Your Role! Leaderboard

LLMs can role-play different personas by simulating their values and behavior, but can they stick to their role whatever the context? Is simulated Joan of Arc more tradition-driven than Elvis? Will it still be the case after playing chess?

The Stick to Your Role! leaderboard compares LLMs based on undesired sensitivity to context change. LLM-exhibited behavior always depends on the context (prompt). While some context-dependence is desired (e.g. following instructions), some is undesired (e.g. drastically changing the simulated value expression based on the interlocutor). As proposed in our paper, undesired context-dependence should be seen as a property of LLMs - a dimension of LLM comparison (alongside others such as model size speed or expressed knowledge). This leaderboard aims to provide such a comparison and extends our paper with a more focused and elaborate experimental setup. Standard benchmarks present many questions from the same minimal contexts (e.g. multiple choice questions), we present same questions from many different contexts.

The Stick to Your Role! leaderboard focuses on the stability of simulated personal values during role-playing. We study the coherence of a simulated population. In contrast to evaluating each simulated persona separately, we evaluate personas relative to each other, i.e. as a population. You can browse the simulated population, questionnaires, and contexts used on our 🤗 StickToYourRole dataset.

We leverage Schwartz's theory of Basic Personal Values, which defines 10 values Self-Direction, Stimulation, Hedonism, Achievement, Power, Security, Conformity, Tradition, Benevolence, Universalism), and the associated PVQ-40 and SVS questionnaires (available here).

Using the methodology from psychology, we focus on population-level (interpersonal) value stability, i.e. Rank-Order stability (RO stability). Rank-Order stability refers to the extent to which the order of different personas (in terms of expression of some value) remains the same along different contexts. Refer here or to our paper for more details.

In addition to Rank-Order stability we compute validity metrics (Stress, CFI, SRMR, RMSEA), which are a common practice in psychology. Validity refers to the extent to which the questionnaire measures what it purports to measure. It can be seen as the questionnaire's accuracy in measuring the intended factors, i.e. values. For example, basic personal values should be organized in a circular structure, and questions measuring the same value should be correlated. The table below additionally shows the validity metrics, refer here for more details.

We aggregate Rank-Order stability and validation metrics to rank the models. We do so in two ways: Cardinal and Ordinal. Following this paper, we compute the stability and diversity of those rankings. See here for more details.

To sum up here are the metrics used:

• RO-stability - Do the same simulated participants always (in every context) express same values more strongly than other participants?
  The correlation in the order of simulated participants (ordered based on the expression of the same values) over different contexts
• Stress - Is value expression (intercorrelations of values) structured as predicted by the theory (and as in humans), i.e. in a circle?
  The Multi-dimensional scaling (MDS) fit of the observed value structure to the theoretical circular structure. Stress of 0 indicates 'perfect' fit, 0.025 excellent, 0.05 good, 0.1 fair, and 0.2 poor.
• CFI, SRMR, RMSEA - To what extent does the questionnaire measure what it is supposed to measure - values?
  Common Confirmatory Factor Analysis (CFA) metrics showing the fit of the posited model of the relation of items (questions) to factors (values) on the observed data, applied here with Magnifying Glass CFA. For CFI >.90 is considered acceptable fit, for SRMR and RMSEA is <.05 considered good fit and <.08 reasonable.
• Ordinal - Win Rate - Which model beats the most other models across most metrics?
  The percentage of won games, where a game is a comparison of each model pair, each metric, and each context pair (for stability) or context (for validity metrics)
• Cardinal - Score - Which model has the highest average score?
  The score averaged over all metrics (with descending metrics inverted), context pairs (for stability) and contexts (for validity metrics)

Feel free to suggest models by opening an Issue on the space. Suggested models should have context length >8k tokens.