RVDB: leaderboard
RVDB (Realistic Value Decision Benchmark, 2026) tests whether explicit role-gender cues change the decision an LLM supports in Chinese value-conflict scenarios built on Schwartz value pairs, with the scenario, roles and candidate decisions held fixed. The score is the average decision flip rate in percent over ten pairwise gender-configuration comparisons, zero-shot at temperature 0; lower is better.
Metric: Average decision flip rate (%) over the ten pairwise role-gender configuration comparisons: share of decision units whose supported decision changes when only the gender of the two roles changes (position-balanced, zero-shot, temperature 0); lower is better. Source: arxiv.org. Saturation forecast: Estimated already saturated. 7 models tracked.
Top models
| # | Model | Score | Overall rank |
|---|---|---|---|
| 1 | Qwen 3 Max | 3.37 | #403 |
| 2 | GLM-4 9B Chat | 4.89 | #2759 |
| 3 | Qwen 2.5 7B Instruct | 5.37 | #2564 |
| 4 | Qwen 3 14B | 5.64 | #1233 |
| 5 | GPT-4o Mini | 6.22 | #1623 |
| 6 | Llama 3.1 8B Instruct | 9.5 | #4246 |
| 7 | Llama 3 8B Instruct | 9.61 | #4612 |
No result here: #1 Claude Fable 5.2, #2 Gemini 4 Carbon, #5 Claude Opus 5.5.
Interactive version: theaggregate.ai/benchmark?slug=rvdb · How It Works · Data refreshed daily, snapshot 2026-11-13.