TRACE (LRM Safety) - Prompt: leaderboard
TRACE (Wu et al., 2026) tests guardrail models on 5,000 prompt, reasoning-trace and final-response triples from four large reasoning models in English and Chinese, with safety labels and evidence spans from a majority of three LLM annotators. This board scores safe or unsafe judgments of the prompt. F1 on the unsafe class in percent; higher is better.
Metric: F1 on the unsafe class (%; guardrail safe or unsafe judgments of the prompt in 5,000 prompt, reasoning-trace and final-response triples from four reasoning models, English and Chinese, nine risk categories and ten attack strategies; labels by a majority of three LLM annotators, temperature 0). Source: arxiv.org. Saturation forecast: Rough model projection: around 2026. 18 models tracked.
Top models
| # | Model | Score | Overall rank |
|---|---|---|---|
| 1 | PolyGuard-Qwen | 91.17 | |
| 2 | Qwen3Guard-Gen-8B | 88.43 | |
| 3 | Qwen3Guard-Gen-4B | 88.31 | |
| 4 | YuFeng-XGuard-Reason-8B | 88.27 | |
| 5 | YuFeng-XGuard-Reason-0.6B | 88.24 | |
| 6 | gpt-oss-safeguard-20b | 86.96 | |
| 7 | WildGuard 7B | 86.37 | |
| 8 | Llama-3.1-Nemotron-Safety-Guard-8B-v3 | 85.99 | |
| 9 | Qwen3Guard-Gen-0.6B | 85.47 | |
| 10 | PolyGuard-Qwen-Smol | 83.53 |
Interactive version: theaggregate.ai/benchmark?slug=trace-lrm-safety-prompt · How It Works · Data refreshed daily, snapshot 2026-11-12.