VARM-Bench - Harmfulness Label: leaderboard
Part of VARM-Bench (Yuan et al., 2026) on moderation rationales for Chinese social-media text, zero-shot on the 1,600-item test set: macro-F1 in percent of the harmful or non-harmful label anchored in the rationale. Unparseable outputs count as wrong; higher is better.
Metric: Macro-F1 (%) of the harmful or non-harmful label; deduplicated 1,600-item test set of Chinese social-media posts and comments; the model writes one rationale with six anchored decisions (target, target type, target explicitness, author stance, harmfulness label, fine-grained category) that a deterministic parser scores, unparseable outputs counting as wrong; deterministic decoding; zero-shot prompt. Source: arxiv.org. Saturation forecast: Estimated already saturated. 6 models tracked.
Top models
| # | Model | Score | Overall rank |
|---|---|---|---|
| 1 | GPT-5.5 | 97.6 | #49 |
| 2 | Qwen 3.7 Max | 96.9 | #145 |
| 3 | DeepSeek V4 Pro | 95.6 | #188 |
No result here: #1 Claude Fable 5.2, #2 Gemini 4 Carbon, #5 Claude Opus 5.5.
Interactive version: theaggregate.ai/benchmark?slug=varm-bench-harmfulness-label · How It Works · Data refreshed daily, snapshot 2026-11-13.