LLM Milgram Obedience - Baseline Full-Obedience Rate: leaderboard
Share of valid baseline sessions in which an LLM playing Milgram's Teacher, under a scripted Experimenter, administers the full 450 V shock (voice-feedback condition, temperature 1.0, reasoning disabled, at most 15 sessions). From the authors' census file for arXiv 2608.16177v2. Percent; lower means less obedience to harmful authority.
Metric: Valid baseline sessions ending at 450 V (%; lower is less obedient). Source: github.com. Saturation forecast: Estimated already saturated. 39 models tracked.
Top models
| # | Model | Score | Overall rank |
|---|---|---|---|
| 1 | Llama 3.3 70B Instruct | 0 | #541 |
| 2 | Gemini 3.5 Flash | 0 | #66 |
| 3 | Qwen 3.8 Max | 0 | #34 |
| 4 | Claude Sonnet 5 | 0 | #65 |
| 5 | Gemini 3.7 Flash | 0 | #24 |
| 6 | Grok 4.6 | 0 | #29 |
| 7 | GPT-5.6 Sol (Non-reasoning) | 0 | #22 (GPT-5.6 Sol) |
| 8 | GPT-5.6 Terra (Non-reasoning) | 0 | #40 (GPT-5.6 Terra) |
| 9 | GPT-5.6 Luna (Non-reasoning) | 0 | #79 (GPT-5.6 Luna) |
| 10 | MiMo-V2.5 (Non-reasoning) | 0 | #196 (MiMo-V2.5) |
| 11 | MiMo-V2.5-Pro (Non-reasoning) | 10 | #167 (MiMo-V2.5-Pro) |
| 12 | DeepSeek V3 (0324) | 11.11 | #349 |
| 13 | Nova Pro (v1) | 11.11 | #508 |
| 14 | Qwen 3.8 27B (Non-reasoning) | 26.67 | #123 (Qwen 3.8 27B) |
| 15 | Gemini 2.5 Pro | 28.57 | #159 |
No result here: #3 Claude Opus 5.5, #6 GPT-6 Astra, #7 GPT-6.1 Sol.
Interactive version: theaggregate.ai/benchmark?slug=llm-milgram-obedience-baseline-full-obedience-rate · How It Works · Data refreshed daily, snapshot 2026-11-05.