Social harm risk evaluation on SHARP (n=901 prompts)
1.689CVaR95 Log-Riskclaude-sonnet-4.5
Evaluation Results
| Method | Links | |||
|---|---|---|---|---|
| claude-sonnet-4.52026.01 | 1.689 | 0.158 | 0.772 | |
| gemini-1.5-pro2026.01 | 3.499 | 0.463 | 0.962 | |
| claude-3.5-sonnet2026.01 | 3.782 | 0.449 | 0.965 | |
| gemini-2.5-pro2026.01 | 3.868 | 0.56 | 0.971 | |
| qwen3-235bParameters=235b2026.01 | 3.895 | 0.589 | 0.967 | |
| gpt-4o2026.01 | 4.519 | 0.748 | 0.984 | |
| gpt-oss-120bParameters=120b2026.01 | 5.286 | 1.209 | 0.994 | |
| deepseek-chat2026.01 | 5.735 | 1.296 | 0.994 | |
| mistral-large2026.01 | 5.988 | 1.322 | 0.995 | |
| llama3-3-70bParameters=70b2026.01 | 7.934 | 1.968 | 0.998 | |
| llama3-1-405bParameters=405b2026.01 | 8.397 | 2.154 | 0.998 |