Witness Overlap：开放权重模型家族内的方向性溯源
原标题：Witeness Overlap: Directional Provenance Inside Open-Weight Model Families
AI 摘要
该论文提出 Witness Overlap，一种无需提示、无需训练的白盒测试方法，用于判断开放权重模型家族内部检查点之间的方向性溯源关系（即谁先谁后）。方法引入同家族第三个检查点作为“见证者”，比较以候选端点为锚的权重更新方向重叠度，将重叠度较低的一方预测为父模型。在来自16个家族的176个LLM检查点上，单见证者测试以Frobenius余弦正确判断了95.3%的父子关系，并进一步评估了根节点识别、兄弟区分、VLM与扩散模型泛化及链式排序。
正文节选
Witness Overlap: Directional Provenance Inside Open-Weight Model Families Abstract Open-weight models are often released, fine-tuned, aligned, merged, and re-released, making provenance audits ask not only whether checkpoints are related, but also which checkpoint came first. Many existing model-provenance methods are designed for a base-known audit setting: given a victim or source model, they test whether a suspect model is related to it. Although these audits are framed as source-to-suspect t