Artifact Distillation：无需训练的前沿到小模型迁移
原标题：Artifact Distillation: Training-Free Transfer from Frontier to Small Language Models
AI 摘要
论文提出 Artifact Distillation，一种无需训练的知识迁移方法：教师模型（Claude Opus 4.5）在训练集上解题，并将解题过程外化为可检索的 artifact bank，学生模型在推理时查询该库，无需梯度更新。在 LiveBench 四个任务上，Claude Haiku 4.5 借助该库在空间推理上超过未辅助的 Claude Sonnet 4.5（0.707 对 0.647），成本约为其 60%；同一原始库还将 Qwen3.5-27B 的汇总得分从 0.730 提升到 0.800，表明迁移不限于教师模型家族。
正文节选
Artifact Distillation: Training-Free Transfer from Frontier to Small Language Models Abstract Closing the gap between frontier and smaller language models is usually approached by distillation: fine-tuning the small model (student) on a frontier model’s (teacher) outputs, which binds transfer to training compute and leaves the transferred knowledge largely opaque. We propose Artifact Distillation: a teacher solves training instances and externalizes its problem-solving procedure into an artifact