Translation in progress Some entries still include text in the original language.
Indexed by this site 2026-10-27 00:03 3 days ago
Originally published Oct 26
Selected 2026-10-27 00:04
Category Governance CSET · Emily Tavenner · Indexed by this site 2026-10-27 00:03 · Originally publishedOct 26 Attention35
How black-box distillation works How Black-Box Distillation Works
AI summary
CSET 发布技术讲解，拆解黑盒蒸馏的五个步骤，即确定目标并选择师生模型、收集种子数据、引导教师模型、生成或收集合成数据、用合成数据训练学生模型，并结合美国对中国实验室对抗性蒸馏的指控逐环节分析。文章认为中国实验室确实在从美国专有模型蒸馏，目标是在后训练阶段转移能力并混淆来源，而非压缩模型体积；但由于种子数据和引导策略的公开信息不足，基于分类器的检测能否奏效难以评估。在规模上，Anthropic 称七家中国实验室与 Claude 的交互累计超过两亿次，Google 称 Gemini 被约十万条提示词播种，美国政府联合公告则称提取了数十亿 token；文章以 GLM-4.5 为例估算，蒸馏数据可能占其后训练数据的 14% 至 36%。文章同时指出美国公司对学生模型训练这一最终环节没有可见性，蒸馏带来的实际能力提升仍不确定，建议美国实验室公开相关内部研究以提高透明度。
Why it was selected
This technical explainer breaks black-box distillation into five stages and explains each one in turn, and at each stage compares it against public reports alleging that Chinese labs adversarially distill US models, pointing out information gaps in the seed data, elicitation strategy, synthetic data scale and student training stages, providing a verifiable technical basis for assessing detection feasibility, distillation gains and related legislative claims.
Read original cset.georgetown.edu
Why it was selected Attention35
Attention 0–100: content value contributes up to 50 and circulation up to 50 (accounts weighted by likes); circulation decays over time. How it was rated
This technical explainer breaks black-box distillation into five stages and explains each one in turn, and at each stage compares it against public reports alleging that Chinese labs adversarially distill US models, pointing out information gaps in the seed data, elicitation strategy, synthetic data scale and student training stages, providing a verifiable technical basis for assessing detection feasibility, distillation gains and related legislative claims.