[{"data":1,"prerenderedAt":-1},["ShallowReactive",2],{"news-cdd58eeb-a16b-40e0-a6a7-c58e1bcc323a":3},{"id":4,"title":5,"summary":6,"original_url":7,"source_id":8,"tags":9,"published_at":23,"created_at":24,"modified_at":25,"is_published":26,"publish_type":27,"image_url":13,"view_count":28},"cdd58eeb-a16b-40e0-a6a7-c58e1bcc323a","GRAM 把双用途知识锁进 MLP 旁路:Anthropic 让一份预训练跑出五种安全配置","Anthropic 与 AE Studio 在 alignment 博客发布 GRAM（Gradient-Routed Auxiliary Modules）研究,提出一种全新的「架构级」安全访问控制思路。它在 Transformer 每个 MLP 层旁挂接小型辅助模块,训练时根据数据类型路由梯度,推理时删掉对应模块即可关闭特定能力——而不需要重训一整个模型。实验覆盖 50M 到 5B 参数,成功把病毒学、网络安全、核物理、专业代码四类双用途知识隔离到独立模块;一份 GRAM 模型即可重构出五种不同过滤配置,组合 4 个模块还能得到 16 种开关状态。在双用途能力保留与遗忘、对抗微调鲁棒性、组合性、部分标注场景上,GRAM 都优于 MaxEnt 等事后遗忘方法和 LoRA 微调基线。这是首次把 frontier 模型的访问控制从「拒绝训练 + 分类器」的行为层,推进到「权重拓扑」的结构层。Anthropic 强调该工作尚未进入生产 Claude,但思路值得长期跟踪:未来同一个基础模型,或许能根据用户信任级别动态切片,把高级能力「按需点亮、按需关停」——前提是它能扩展到几百 B 参数并解决指令微调兼容、纠缠能力分离等开放问题。","https:\u002F\u002Falignment.anthropic.com\u002F2026\u002Fmodular-pretraining\u002F","1fa87d30-d9f3-4752-b3be-0373933b3aaf",[10,14,17,20],{"id":11,"name":12,"slug":12,"description":13,"color":13},"1fcfaaf2-67de-43d3-9e35-5784852fec60","ai-safety",null,{"id":15,"name":16,"slug":16,"description":13,"color":13},"23544f6a-eea1-4f05-aa8d-749ca862d5d2","anthropic",{"id":18,"name":19,"slug":19,"description":13,"color":13},"01598627-1ea6-4b27-a5d8-874971571a71","llm",{"id":21,"name":22,"slug":22,"description":13,"color":13},"4f214978-cac1-4f39-aa4b-f92a0d0934b7","transformer","2026-07-11T00:00:00Z","2026-07-10T20:07:47.528919Z","2026-07-10T20:07:47.528943Z",true,"agent",2]