Benchmark RadarAI BENCHMARK PROFILE
VARM-Bench
VARM-Bench evaluates verifiable structured reasoning in Chinese abusive-speech moderation. It uses field-anchored chain-of-thought rationales with six decision fields, and a deterministic protocol assessing field correctness, alignment, output validity, and record errors.
- Released
- 2026-08-16
- Readiness
- Paper only
- Primary field
- General AI
Why it matters
Existing benchmarks support classification but not verifiable reasoning. VARM-Bench provides an auditable protocol for evaluating moderation rationales, revealing that strong label performance can conceal errors in complete records.
Motivation
The widespread circulation of abusive online content has increased the need for reliable moderation of Chinese social-media text.
Primary resources
Benchmark Radar records only publicly supported details and links back to primary sources for verification.