端侧命名实体识别：准确率、成本、可靠性与置信度的可部署性研究
原标题：On-Device Named-Entity Recognition: A Deployability Study of Accuracy, Cost, Reliability, and Confidence
AI 摘要
该研究针对端侧命名实体识别（NER）的可部署性展开系统评测，覆盖经典标注器、双向编码器（GLiNER）与本地生成式LLM（Qwen3、DeepSeek-R1）三类范式、13M至8B参数共九个系统，在三个特性不同的数据集上同时报告准确率、延迟与输出有效性。研究发现，4B指令LLM在干净新闻文本上准确率具竞争力，但编码器以九分之一至二十四分之一规模实现毫秒级延迟且零畸形输出，而最小生成模型在长输入上畸形输出率高达27%，该问题靠规模而非输出预算解决。作者还用跨家族LLM评审团构建银标金标准并做人工复验，发现金标准来源会翻转编码器与生成模型的排名；同时刻画了GLiNER的置信度校准（AUROC 0.76-0.86但严重过度自信，ECE 0.24-0.47）。
正文节选
On-Device Named-Entity Recognition: A Deployability Study of Accuracy, Cost, Reliability, and Confidence Abstract Named-entity recognition (NER) is increasingly wanted on-device (no API, low latency, data kept local), but the practitioner’s real question is not who wins a leaderboard: it is which model is deployable, how it is evaluated without a human-annotation budget, and whether its confidence can be trusted. We answer these jointly. We place nine systems spanning three paradigms and 13 M to