Stick To Your Role!: leaderboard

Stick To Your Role! (Flowers team, Inria, 2024) has a model simulate a population of personas that answer the PVQ-40 and SVS value questionnaires after simulated conversations on different topics, and measures how stable and theory-consistent the expressed Schwartz values stay across contexts. The cardinal score averages rank-order stability over context pairs with validity metrics (CFA fit and MDS stress, lower-is-better ones inverted), on a -1 to 1 scale; higher is better.

Metric: Cardinal Score (-1 to 1). Source: huggingface.co. Saturation forecast: Estimated already saturated. 31 models tracked.

Top models

#ModelScoreOverall rank
1Qwen 2.5 VL 72B Instruct0.84#845
2Llama 3.1 Nemotron 70B Instruct0.81#1447
3Mistral Large 2 (Jul)0.79#1153
4Llama 3.3 70B Instruct0.78#1353
5Llama 3.1 70B Instruct0.77#1541
6QwQ-32B0.77#909
7Qwen 3 32B0.74#960
8Mistral Large 2 (Nov) Instruct (2411)0.73#1199
9Qwen 3 8B0.72#1838
10Qwen 3 235B A22B FP80.72#533
11Mistral Small 3.10.7#1699
12Qwen2.5-14B-Instruct-1M0.7#1417
13Qwen 3 4B0.7#2297
14Llama 3.1 8B Instruct0.62#4277
15Llama 4 Scout Instruct0.62#1532

No result here: #1 Claude Fable 5.2, #2 Gemini 4 Carbon, #5 Claude Opus 5.5.

Interactive version: theaggregate.ai/benchmark?slug=stick-to-your-role · How It Works · Data refreshed daily, snapshot 2026-11-12.