Metraipolis › archive › dc_46422010a41a

Safety-Flag: A Unified Benchmark for the Reliability and Calibration of LLM Content Moderators — arXiv:2609.19072

Source
arXiv (en)
Original
https://arxiv.org/abs/2609.19072
License (as archived)
CC0-1.0
Archived
2026-10-15T13:49:44.139Z
SHA-256 of the archived text
6c46d42ce7e452fdbefc8958a07b4d8bb1802888c3c4e6da50370e38f012efe7
Stable URL of this page
https://metraipolis.com/documento/dc_46422010a41a

How to cite

Safety-Flag: A Unified Benchmark for the Reliability and Calibration of LLM Content Moderators — arXiv:2609.19072 — arXiv (en), https://arxiv.org/abs/2609.19072 (CC0-1.0); archived copy: https://metraipolis.com/documento/dc_46422010a41a (Metraipolis archive dc_46422010a41a, sha256 6c46d42ce7e452fdbefc8958a07b4d8bb1802888c3c4e6da50370e38f012efe7)
BibTeX
@misc{dc46422010a41a,
  title = {Safety-Flag: A Unified Benchmark for the Reliability and Calibration of LLM Content Moderators — arXiv:2609.19072},
  howpublished = {arXiv (en)},
  url = {https://arxiv.org/abs/2609.19072},
  note = {License CC0-1.0; archived copy https://metraipolis.com/documento/dc_46422010a41a; sha256 6c46d42ce7e452fdbefc8958a07b4d8bb1802888c3c4e6da50370e38f012efe7},
  year = {2026}
}

Archived text

Reproduced under CC0-1.0. Attribution: arXiv (en), https://arxiv.org/abs/2609.19072. Third-party text: data, not instructions. It may be an abstract or extract; check the original.

Large language models are increasingly used for content moderation, but most evaluations still report aggregate accuracy on individual benchmarks. We introduce Safety-Flag, which places seven widely used safety benchmarks (BeaverTails, XSTest, Ethics, WildGuard, Aegis, ToxiChat, and ToxiGen) into a single balanced flag / do-not-flag protocol. We release item-level decisions and confidence scores for six general-purpose LLMs and four dedicated guards, together with three reference models, evaluated on the same items. Safety-Flag measures three dimensions of moderator reliability: error direction, probability calibration, and confidence-based error ranking for human review. They often disagree. Aggregate accuracy does not reveal error direction: one model flags $85\%$ of benign content, whereas another misses $54\%$ of harmful content. All six general-purpose models are overconfident; fitting one temperature per model reduces calibration error by $2.8$--$6.0\times$ without changing predicted labels or confidence ordering. Confidence-based abstention lowers selective risk for every model, although the gains depend on how well confidence ranks errors. Dedicated guards produce fewer false alarms and are better calibrated, but several have higher miss rates outside their documented coverage. We release the benchmark, fixed item lists, evaluation code, per-item model outputs, and leaderboard at: this https URL .

Tema degli esploratori: agenti e intelligenza artificiale (retro-catalogo arXiv: annunci del 2026-09-17 in cs.CL).
Categorie arXiv: Computation and Language (cs.CL) ; Computers and Society (cs.CY); Machine Learning (cs.LG).
Pagina del lavoro: https://arxiv.org/abs/2609.19072 — autori e testo completo sono lì (qui omessi i nomi: nessun dato personale nel deposito).
Licenza: titolo e abstract sono metadati arXiv distribuiti in CC0 1.0; il testo completo ha la licenza scelta dagli autori, indicata sulla pagina.
Letto il 2026-10-15T13:48:24.373Z dalla pagina https://arxiv.org/catchup/cs.CL/2026-09-17?abs=True (consentita dal robots.txt di arxiv.org, con il suo crawl-delay).