TRACE (LRM Safety) - Reasoning Trace: 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 reasoning trace. F1 on the unsafe class in percent; higher is better.
Metric: F1 on the unsafe class (%; guardrail safe or unsafe judgments of the reasoning trace 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 | YuFeng-XGuard-Reason-8B | 84.26 | |
| 2 | PolyGuard-Qwen | 82.57 | |
| 3 | Octopus-SEval-14B | 80.1 | |
| 4 | gpt-oss-safeguard-20b | 79.6 | |
| 5 | Llama-3.1-Nemotron-Safety-Guard-8B-v3 | 77.74 | |
| 6 | YuFeng-XGuard-Reason-0.6B | 76.55 | |
| 7 | Qwen3Guard-Gen-8B | 75.73 | |
| 8 | Qwen3Guard-Gen-4B | 72.04 | |
| 9 | Qwen3Guard-Gen-0.6B | 71.79 | |
| 10 | PolyGuard-Qwen-Smol | 65.91 |
Interactive version: theaggregate.ai/benchmark?slug=trace-lrm-safety-reasoning-trace · How It Works · Data refreshed daily, snapshot 2026-11-12.