SM Bench — Adversarial (Hostile Logic)
Safety · 2026-02-01
SM Bench Adversarial (Hostile Logic) tests resistance to manipulative framing, false premises, and hostile reasoning traps while preserving useful instruction following. Scores are difficulty-weighted judge credit across 100 fixed prompts.
Top models (higher is better)
| Model | Score |
|---|---|
| Gemini 3.5 Flash | 92.2 |
| Gemini 3.1 Flash-Lite Preview | 91.7 |
| Gemma 4 31B IT | 91.7 |
| Gemini 3.1 Pro Preview | 90.7 |
| GLM-5 | 90.7 |
| Gemini 3.5 Flash-Lite | 89.8 |
| Gemini 3.6 Flash | 89.3 |
| GPT-5.5 Instant | 88.8 |
| GLM-5.1 | 88.3 |
| MiMo-V2-Pro | 88.3 |
| Gemini 3.8 Flash | 88.3 |
| GPT-4.1 | 87.3 |
| GLM-5.2 | 86.8 |
| Gemini 3.7 Flash | 86.8 |
| Gemini 3 Pro Preview | 86.8 |
| Qwen3.7-Max (source unspecified) | 86.8 |
| MiMo-V2-Omni | 86.6 |
| Claude Opus 5 | 86.3 |
| GLM-5.3 | 86.3 |
| GPT-5.1 | 86.3 |
| Qwen3.6 Plus Preview | 85.8 |
| GPT-5.6 Luna | 85.8 |
| GPT-5.6 Luna Pro | 85.8 |
| Grok 4.5 | 85.4 |
| Claude Fable 5.1 | 85.4 |
| Claude Sonnet 5.5 | 85.4 |
| GPT-5.6 Sol Pro | 85.4 |
| DeepSeek-R1-0528 | 84.9 |
| GPT-6 Astra | 84.9 |
| GPT-5.6 Terra Pro | 84.9 |
| Gemini 3 Flash Preview | 84.4 |
| Kimi K2.5 | 84.4 |
| MiMo-V2.6-Pro | 83.9 |
| Opus 4.5 | 83.9 |
| Kimi K2.6 | 83.9 |
| Grok 4.3 | 83.4 |
| Opus 4.6 | 83.4 |
| MiMo-V2.5-Pro | 83.4 |
| GPT-6.1 Sol | 83.4 |
| GLM-4.7 | 82.9 |