MoralityBench Updates with GPT-6 Astra, Muse Wonder, and Gemini 4 Argon

MoralityBench has been updated to include three new generative model configurations: GPT-6 Astra, Muse Wonder, and Gemini 4 Argon. The benchmark evaluates AI systems using an adapted 36-item Moral Foundations Questionnaire 2 (MFQ-2) and the 20-item Ethics Position Questionnaire (EPQ). The combined archive now contains 4,295 usable ratings from 4,480 scheduled items across 16 scored generative model configurations, each administered five times.

Consistency and Agreement

Among the 16 LLMs tested, only 3 models achieved an agreement rate (consistency across runs) exceeding 90%. The structured decision system Jev, tested separately, displayed the highest level of agreement at 95.7% after rounding. Exact agreement across all ten pairs of runs in the original cohort ranged from 60.4% to 94.7%.

Model Performance and Distances

Within the original cohort, DeepSeek V4.1 Flash showed the lowest average distance from the six published US foundation means (0.239), followed by MiMo v2.6 Pro (0.253). MiMo had the lowest distance in three of the four new runs. Notably, Gemini 4 Argon refused to answer a number of questions, which is described as unusual for an LLM. Its profile depends heavily on the parser, as many responses were verbose or missing.

Methodological Notes

The October 14 extension uses strict whole-answer integer scoring, while older models retain their original scoring. The release includes individual responses, completion records, scoring code, and a public leaderboard. The authors emphasize that these descriptive results concern responses to adapted questionnaires and do not establish equivalent human traits, moral agency, or deployment safety.