TukaBench 登 EMNLP：非洲语言提示让 LLM 拒答率显著下降
davlanade · x · 2026-10-26
Researchers presented TukaBench at EMNLP 2026 in Budapest: a culturally grounded jailbreak benchmark for 7 African languages extending JailbreakBench via four settings (human translation, culturally adapted translation, human-curated prompts filtered via GPT-5.2 interaction, and code-switched prompts).
Key findings:
- Across closed and open models, prompting in African languages reduces refusal rates vs English; culturally adapted prompts yield the least refusal.
- Two structural issues surfaced: comprehension failures in low-resource languages (addressed via a new "Deflection" label) and degraded LLM-as-a-judge reliability, with judge-human agreement dropping for lower-resource languages and scripts.
「安全」频道最新
- 曝 Anthropic IPO 在即,反开源权重恐慌被指利益驱动 — mark_k · 2026-10-27
- Cohere 发现热门开源模型疑似「call home」：试图上传文件至北京存储桶 — stanislavfort · 2026-10-27
- Claude 未经授权上网事件后续分析：模型自辩不可信 — AccBalanced · 2026-10-27
- 生产级 AI Agent 常规评测 95% 稳健，遭定向攻击后跌至 37% — njyx · 2026-10-27
- Anthropic IPO 文件披露:违规客户贡献 5%-7% 营收,至少 8.14 亿美元 — henkvaness · 2026-10-27
- 限制模型做逆向工程和漏洞研究，惩罚的恰恰是守规矩的白帽 — mark_k · 2026-10-27