后训练量化在文本嵌入模型上的失效边界：四类嵌入器家族的实测图谱
原标题：Where Post-Training Quantization Breaks Text Embedders: A Measured Map Across Four Embedder Families
AI 摘要
该研究系统测试了仅权重量化（PTQ）在检索嵌入模型上的表现，覆盖四个架构家族的五个检查点，在多种位宽和分组大小下评估三个检索语料库。结果显示，INT4/g16 下模块级量化影响极小（最多约1个NDCG点），INT3 下模块敏感度排序因模型家族而异且成本不可加，INT2 下检索质量保留率从1.3%到65.9%不等，存在强烈的模型家族依赖的“低比特悬崖”。权重重建误差在均匀量化下能较好预测检索损失，但在模块级和跨家族场景下预测能力弱，且嵌入表从未成为主导的隔离保护优先级。
正文节选
Where Post-Training Quantization Breaks Text Embedders: A Measured Map Across Four Embedder Families Abstract Weight-only post-training quantization (PTQ) is the cheapest way to shrink a retrieval embedder, and the received advice for applying it — protect the embedding table, allocate bits by module sensitivity, and prefer a ranking-aware objective over weight reconstruction — comes from a decade of network and Transformer quantization work and has been carried into LLM PTQ largely intact. We t