TRACE (LRM Safety) - Final Response Evidence: 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 the evidence span a guardrail extracts to justify its judgment of the final response, by token-level F1 against the annotated evidence in percent; higher is better.
Metric: Token-level F1 (%; overlap between the evidence span a guardrail extracts for its judgment of the final response and the annotated evidence, TRACE English and Chinese). Source: arxiv.org. Saturation forecast: Rough model projection: around 2026. 4 models tracked.
Top models
| # | Model | Score | Overall rank |
|---|---|---|---|
| 1 | YuFeng-XGuard-Reason-8B | 13.71 | |
| 2 | Octopus-SEval-14B | 13.25 | |
| 3 | YuFeng-XGuard-Reason-0.6B | 12.03 | |
| 4 | gpt-oss-safeguard-20b | 9.92 |
Interactive version: theaggregate.ai/benchmark?slug=trace-lrm-safety-final-response-evidence · How It Works · Data refreshed daily, snapshot 2026-11-12.