AI-Secure LLM Trustworthy Leaderboard: leaderboard
Average DecodingTrust score (UIUC AI-Secure lab, 2023) over eight perspectives: toxicity, stereotype bias, adversarial and OOD inputs, adversarial demonstrations, privacy, machine ethics, fairness.
Metric: Average Trust Score (%). Source: huggingface.co. Status: saturated. 35 models tracked.
Top models
| # | Model | Score |
|---|---|---|
| 1 | Claude 2 | 84.52 |
| 2 | GPT-4o (2024-05-13) | 82.96 |
| 3 | Llama 3 8B Instruct | 80.61 |
| 4 | Gemini 1.0 Pro | 80.61 |
| 5 | GPT-4o Mini (2024-07-18) | 76.31 |
| 6 | Llama 2 7B Chat | 74.72 |
| 7 | GPT-3.5 Turbo (0301) | 72.45 |
| 8 | GPT-4 (0314) | 69.24 |
| 9 | gemma-2B (IT) | 67.18 |
| 10 | gemma-7B (IT) | 66.87 |
| 11 | tulu-2-13B | 66.51 |
| 12 | tulu-2-7B | 63.56 |
| 13 | zephyr-7B-beta | 63.24 |
| 14 | vicuna-7B-v1.3 | 60.62 |
| 15 | falcon-7B Instruct | 59.49 |
Interactive version: theaggregate.ai/benchmark?slug=ai-secure-llm-trustworthy-leaderboard · How It Works · Data refreshed daily, snapshot 2026-09-05.