The Invisible Gap: Romanized Indic Generation Fails Where Automatic Metrics Succeed
arXiv:2611.09085v1 Announce Type: new Abstract: Hundreds of millions of Indians communicate with digital systems in romanized Latin characters -- typing "aap kaise hain" rather than native script -- yet romanized Indic generation is absent from NLP benchmarks and evaluation frameworks. We demonstrate that automatic metrics (BGE-M3, BERTScore) systematically misrepresent the quality of romanized Hindi, Telugu, and Kannada generation. We evaluate four LLMs (Claude Opus 4.8, GPT-5.5, Gemini 2.5-fla
阅读原始来源
来源记录
研究分析
研究主张
论文标识：2611.09085。 来源描述：arXiv:2611.09085v1 Announce Type: new Abstract: Hundreds of millions of Indians communicate with digital systems in romanized Latin characters -- typing "aap kaise hain" rather than native script -- yet romanized Indic generation is absent from NLP benchmarks and evaluation frameworks. We demonstrate that automatic metrics (BGE-M3, BERTScore) systematically misrepresent the quality of romanized Hindi, Telugu, and Kannada generation. We evaluate four LLMs (Claude Opus 4.8, GPT-5.5, Gemini 2.5-fla
评测与比较记录
可供比较的近期同类记录：“PromptOps: Iterative Prompt Optimization for LLM-as-a-Judge Systems” (arXiv, 2026年11月6日)；“Synthesizing symbolic models of human and machine reasoning at scale” (arXiv, 2026年11月6日)。主题相近仅用于导航，未核验其是否属于同一需求或独立证据。
结论适用范围
结论只能在所述模型版本、任务、数据与推理预算内解释。社区票数、下载量或榜单位置不能证明另一工作流中的可靠性。
有效性与复现缺口
需核对样本选择、基线配置、不确定性、数据污染风险、负面结果以及代码或权重可获取性；此处没有本站复现实验记录。
复核路线
阅读论文或模型卡及评测附录，针对明确基线复现一个有限结论，再测试代表性用户任务并记录结论失效的场景。