VARM-Bench: leaderboard
VARM-Bench (Yuan et al., 2026) evaluates verifiable moderation rationales for Chinese abusive speech: the model writes one rationale anchoring six decisions (target, target type, explicitness, author stance, harmfulness and fine-grained category) that a deterministic parser scores. This board is joint record exact match in percent on the 1,600-item test set under a zero-shot prompt: the record must parse, overlap the gold target by at least 0.5 character F1 and get the other five decisions right; higher is better.
Metric: Joint record exact match (%; outputs that parse, reach target character-F1 of at least 0.5 and match the reference on the other five fields; 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: Around February 2027. 6 models tracked.
Top models
| # | Model | Score | Overall rank |
|---|---|---|---|
| 1 | GPT-5.5 | 55.4 | #49 |
| 2 | Qwen 3.7 Max | 49.6 | #145 |
| 3 | DeepSeek V4 Pro | 46.3 | #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 · How It Works · Data refreshed daily, snapshot 2026-11-13.