FaithfulBench (Stated): leaderboard
FaithfulBench (Kadous et al., 2026) scores whether AI counsel on first-person moral dilemmas upholds the professed faith of the user across eight religious and secular traditions, judged against the texts of each tradition after one round of user pushback. In this framing a one-line prefix tells the model that the user is a practising adherent; each conversation is scored from -1 (counsel against the tradition) to +1 (counsel the tradition would endorse) as the mean of two LLM judges, averaged per tradition and then across traditions; higher is better.
Metric: Faithfulness score (-1 to +1; a one-line prefix states that the user is a practising adherent; mean over eight tradition modules of per-tradition means (555 dilemma scenarios in all), after one user pushback per sitting drawn from six pressures; each transcript scored -1 (counsel against the user tradition) to +1 (counsel a wise companion of that tradition would endorse) against scenario-specific source guidance, mean of the Gemini 3.6 Flash and Claude Opus 4.8 judges). Source: arxiv.org. Saturation forecast: Estimated already saturated. 5 models tracked.
Top models
| # | Model | Score | Overall rank |
|---|---|---|---|
| 1 | Claude Sonnet 5 | 0.82 | #108 |
| 2 | Inkling | 0.79 | #270 |
| 3 | GPT-5.6 Terra | 0.69 | #60 |
| 4 | Gemini 3.6 Flash | 0.54 | #72 |
| 5 | Qwen 3 235B A22B Instruct | -0.12 | #680 |
No result here: #1 Claude Fable 5.2, #2 Gemini 4 Carbon, #5 Claude Opus 5.5.
Interactive version: theaggregate.ai/benchmark?slug=faithfulbench-stated · How It Works · Data refreshed daily, snapshot 2026-11-12.