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.

# Model Ordinal - Win rate (↑) Cardinal - Score (↑) RO Stability (↑)
reka-flash-3 0.362 0.490 0.289
Llama-4-Scout-17B-16E-Instruct 0.490 0.618 0.497
Llama-3.3-70B-Instruct 0.782 0.783 0.810
Llama-3.1-70B-Instruct 0.703 0.772 0.763
Llama-3.1-Nemotron-70B-Instruct 0.784 0.807 0.799
Llama-3.1-8B-Instruct 0.473 0.620 0.554
Llama-3.2-3B-Instruct 0.350 0.492 0.310
Llama-3.2-1B-Instruct 0.148 0.285 0.018
Mistral-Large-Instruct-2411 0.612 0.733 0.706
Mistral-Large-Instruct-2407 0.710 0.787 0.779
Mistral-Nemo-Instruct-2407 0.346 0.524 0.410
Mistral-Small-3.1-24B-Instruct-2503 0.558 0.703 0.685
QwQ-32B 0.737 0.772 0.809
Qwen2.5-VL-72B-Instruct 0.872 0.836 0.813
Qwen2.5-VL-7B-Instruct 0.228 0.424 0.286
Qwen2.5-VL-3B-Instruct 0.101 0.266 0.060
Qwen2.5-14B-Instruct-1M 0.565 0.702 0.655
Dracarys2-72B-Instruct 0.753 0.774 0.790
Nautilus-70B-v0.1 0.596 0.724 0.719
Cydonia-22B-v1.2 0.407 0.566 0.493
Ministrations-8B-v1 0.322 0.492 0.349
dummy 0.097 0.229 -0.009
Ordinal Cardinal

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:

# Model Ordinal - Win rate (↑) Cardinal - Score (↑) RO Stability (↑) Stress (↓) CFI (↑) SRMR (↓) RMSEA (↓)
reka-flash-3 0.362 0.490 0.289 0.219 0.636 0.324 0.327
Llama-4-Scout-17B-16E-Instruct 0.490 0.618 0.497 0.195 0.698 0.272 0.274
Llama-3.3-70B-Instruct 0.782 0.783 0.810 0.151 0.707 0.248 0.287
Llama-3.1-70B-Instruct 0.703 0.772 0.763 0.159 0.741 0.207 0.252
Llama-3.1-Nemotron-70B-Instruct 0.784 0.807 0.799 0.156 0.777 0.157 0.205
Llama-3.1-8B-Instruct 0.473 0.620 0.554 0.178 0.620 0.347 0.353
Llama-3.2-3B-Instruct 0.350 0.492 0.310 0.235 0.640 0.367 0.340
Llama-3.2-1B-Instruct 0.148 0.285 0.018 0.292 0.457 0.498 0.458
Mistral-Large-Instruct-2411 0.612 0.733 0.706 0.172 0.724 0.247 0.262
Mistral-Large-Instruct-2407 0.710 0.787 0.779 0.182 0.776 0.199 0.217
Mistral-Nemo-Instruct-2407 0.346 0.524 0.410 0.213 0.565 0.400 0.396
Mistral-Small-3.1-24B-Instruct-2503 0.558 0.703 0.685 0.177 0.671 0.297 0.315
QwQ-32B 0.737 0.772 0.809 0.177 0.693 0.271 0.307
Qwen2.5-VL-72B-Instruct 0.872 0.836 0.813 0.159 0.845 0.100 0.148
Qwen2.5-VL-7B-Instruct 0.228 0.424 0.286 0.259 0.453 0.479 0.464
Qwen2.5-VL-3B-Instruct 0.101 0.266 0.060 0.288 0.365 0.601 0.590
Qwen2.5-14B-Instruct-1M 0.565 0.702 0.655 0.179 0.713 0.269 0.269
Dracarys2-72B-Instruct 0.753 0.774 0.790 0.158 0.731 0.258 0.279
Nautilus-70B-v0.1 0.596 0.724 0.719 0.169 0.673 0.279 0.308
Cydonia-22B-v1.2 0.407 0.566 0.493 0.208 0.562 0.404 0.392
Ministrations-8B-v1 0.322 0.492 0.349 0.230 0.548 0.400 0.378
dummy 0.097 0.229 -0.009 0.293 0.376 0.622 0.592
Motivation and Methodology page

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

You can find more details in our paper.

If you found this project useful, please cite one of our related papers.

Short paper

@inproceedings{kovavc2024stick, title={Stick to your Role! Stability of Personal Values Expressed in Large Language Models}, author={Kova{\v{c}}, Grgur and Portelas, R{\'e}my and Sawayama, Masataka and Dominey, Peter Ford and Oudeyer, Pierre-Yves}, booktitle={Proceedings of the Annual Meeting of the Cognitive Science Society}, volume={46}, year={2024} }

Long paper

@article{kovavc2024stick, title={Stick to your role! Stability of personal values expressed in large language models}, author={Kova{\v{c}}, Grgur and Portelas, R{\'e}my and Sawayama, Masataka and Dominey, Peter Ford and Oudeyer, Pierre-Yves}, journal={PloS one}, volume={19}, number={8}, pages={e0309114}, year={2024}, publisher={Public Library of Science San Francisco, CA USA} }