MoralityBench tests 16 LLMs with validated moral psychology surveys; only 3 exceed 90% consistency
NakedPlato · reddit · 2026-10-15
The open benchmark MoralityBench updated its leaderboard, adding Muse Wonder, GPT-6 Astra, and Gemini 4 Argon for 16 scored generative configurations (five questionnaire runs each), plus the structured decision system Jev.
Method: an adapted 36-item Moral Foundations Questionnaire 2 (MFQ-2) and the 20-item Ethics Position Questionnaire (EPQ), yielding 4,295 usable ratings out of 4,480 scheduled items.
Key findings:
- Only 3 of 16 LLMs showed >90% cross-run agreement; Jev scored highest (95.7%, answering all 280 items), though its structured interface prevents claims of architectural advantage.
- Gemini 4 Argon refused a number of questions and gave verbose or missing answers, making its profile parser-dependent — changing the parser reverses its distance ordering vs. Astra.
- DeepSeek V4.1 Flash sits closest to the six published US foundation means (0.239), followed by MiMo v2.6 Pro (0.253).
- Four configurations cross an EPQ category boundary between runs; DeepSeek's five-run Purity mean falls below the human reference, unlike the original cohort.
The authors stress these are descriptive results implying nothing about moral agency, ethical superiority, or deployment safety. Data, scoring code, and a public leaderboard are on GitHub.
More from Models
- Claude autonomously coded for 16 hours straight and shipped a perfect implementation — Atlan_ · 2026-10-15
- Google shows how its teams use Gemini 4 Argon for code and writing — GeminiApp · 2026-10-15
- Pref-sampling tests: Muse Wonder least model-influenced, Gemini 4 most nuanced — repligate · 2026-10-15
- Opus 5.5 designs 3D heads in OpenSCAD that breathe and express anger — repligate · 2026-10-15
- Gemini 4 Argon refuses public polling numbers, accuses user of running a red team — ActuaryAutomatic6289 · 2026-10-15
- Gemini 4 Argon debuts 5th on VoxelBench, near Opus 5 level but below Astra — legit_api · 2026-10-15