Hugging Face Daily Papers · September 2, 2026 · 4 min read

Safin-1: Safety from Within through Memory-Native State Evolution

Mirrored from Hugging Face Daily Papers for archival readability. Support the source by reading on the original site.

Like Read original ↗

Safin-1 introduces “Safety from Within,” a memory-native approach that treats safety as an evolving internal model state rather than a post-hoc constraint. Built on MARCH, it combines content-conditioned memory routing with state evolution to support persistent capability adaptation and safer behavior over long-horizon interactions.</p>\n","updatedAt":"2026-09-02T03:19:40.432Z","author":{"_id":"69804cdd555beac9be536071","avatarUrl":"/avatars/380c3021671962416fab240c179972da.svg","fullname":"mingzhang","name":"Mingzh21","type":"user","isPro":false,"isHf":false,"isHfAdmin":false,"isMod":false,"followerCount":1,"isUserFollowing":false}},"numEdits":0,"identifiedLanguage":{"language":"en","probability":0.8873022794723511},"editors":["Mingzh21"],"editorAvatarUrls":["/avatars/380c3021671962416fab240c179972da.svg"],"reactions":[],"isReport":false}}],"primaryEmailConfirmed":false,"paper":{"id":"2609.00092","authors":[{"_id":"6a978cb4fe3c2f89286c39ae","name":"Ming Zhang","hidden":false},{"_id":"6a978cb4fe3c2f89286c39af","name":"Kaisen Yang","hidden":false},{"_id":"6a978cb4fe3c2f89286c39b0","name":"Shu Yu","hidden":false},{"_id":"6a978cb4fe3c2f89286c39b1","name":"Ermo Hua","hidden":false},{"_id":"6a978cb4fe3c2f89286c39b2","name":"Zhekai Chen","hidden":false},{"_id":"6a978cb4fe3c2f89286c39b3","name":"Cheng Jin","hidden":false},{"_id":"6a978cb4fe3c2f89286c39b4","name":"Jingnan Zheng","hidden":false},{"_id":"6a978cb4fe3c2f89286c39b5","name":"Yi Zhang","hidden":false},{"_id":"6a978cb4fe3c2f89286c39b6","name":"Zhongtian Ma","hidden":false},{"_id":"6a978cb4fe3c2f89286c39b7","name":"Jiawei Zhou","hidden":false},{"_id":"6a978cb4fe3c2f89286c39b8","name":"Sirui Chen","hidden":false},{"_id":"6a978cb4fe3c2f89286c39b9","name":"Qiaosheng Zhang","hidden":false},{"_id":"6a978cb4fe3c2f89286c39ba","name":"Xiang Wang","hidden":false},{"_id":"6a978cb4fe3c2f89286c39bb","name":"Ning Ding","hidden":false},{"_id":"6a978cb4fe3c2f89286c39bc","name":"Xia Hu","hidden":false},{"_id":"6a978cb4fe3c2f89286c39bd","name":"Bowen Zhou","hidden":false},{"_id":"6a978cb4fe3c2f89286c39be","name":"Youbang Sun","hidden":false},{"_id":"6a978cb4fe3c2f89286c39bf","name":"Chaochao Lu","hidden":false}],"publishedAt":"2026-08-31T00:00:00.000Z","submittedOnDailyAt":"2026-09-02T00:00:00.000Z","title":"Safin-1: Safety from Within through Memory-Native State Evolution","submittedOnDailyBy":{"_id":"69804cdd555beac9be536071","avatarUrl":"/avatars/380c3021671962416fab240c179972da.svg","isPro":false,"fullname":"mingzhang","user":"Mingzh21","type":"user","name":"Mingzh21"},"summary":"Long-horizon complex tasks require foundation models to accumulate information, maintain internal states, and adapt over extended interactions. Safety should be an intrinsic property of the model itself, rather than a behavioral constraint relying solely on external safeguards or post-hoc alignment such as supervised fine-tuning. This motivates Safety from Within, where safety-relevant capabilities are represented and invoked through the model's native computation. We present Safin-1, a family of foundation models realizing this principle through memory routing and state evolution. Safin-1 is built on Memory-Anchor Routing across Context History (MARCH), a network architecture that maintains structured memory states and selectively retrieves relevant historical information through content-conditioned routing. It supports test-time adaptation of persistent capability states without repeatedly modifying the backbone, enabling controlled specialization over a shared foundation. We investigate this interface on downstream safety tasks through a Safety State, demonstrating effective state-based adaptation with substantial safety improvements. More broadly, the routed-state interface unifies contextual memory and persistent capability adaptation within the model's native computation, reframing memory from a passive record of prior context into an active substrate for maintaining and evolving model behavior. Evaluations across general capabilities, long-context understanding, retrieval, and efficiency further validate Safin-1. These findings provide a path toward safety as a state-native and adaptively maintainable capability. This work is only an initial architectural exploration of Safety from Within, and substantial further work is needed to realize this broader vision.","upvotes":16,"discussionId":"6a978cb5fe3c2f89286c39c0","githubRepo":"https://github.com/AI45Lab/Safin-1","githubRepoAddedBy":"user","ai_summary":"Safin-1 introduces a memory-routing architecture that embeds safety as an internal, evolving model state rather than an external constraint.","ai_keywords":["memory routing","state evolution","MARCH","Memory-Anchor Routing across Context History","content-conditioned routing","test-time adaptation","Safety State","routed-state interface"],"ai_summary_model":"thinkingmachines/Inkling-Small","githubStars":1,"organization":{"_id":"6a4fb75a1c66dbf208e7ddb6","name":"Shanghai-AI-Laboratory","fullname":"Shanghai AI Laboratory","avatar":"https://cdn-avatars.huggingface.co/v1/production/uploads/65cd955637be1841d0b75397/Rao_Kq6NMtTVfSqLUIR4k.webp"}},"canReadDatabase":false,"canManagePapers":false,"canSubmit":false,"hasHfLevelAccess":false,"upvoted":false,"upvoters":[{"_id":"66b190a3375b7138223041a2","avatarUrl":"/avatars/86574aafbc2cbccbe13ad4c7dbd31d5e.svg","isPro":false,"fullname":"tony wang","user":"deepsilent","type":"user"},{"_id":"620783f24e28382272337ba4","avatarUrl":"https://cdn-avatars.huggingface.co/v1/production/uploads/620783f24e28382272337ba4/zkUveQPNiDfYjgGhuFErj.jpeg","isPro":false,"fullname":"GuoLiangTang","user":"Tommy930","type":"user"},{"_id":"66a90a0a78c6b61d824b2e28","avatarUrl":"/avatars/6c33ccf85d3369d0385c0ec238d2bee4.svg","isPro":false,"fullname":"Kaisen Yang","user":"anomyous-author","type":"user"},{"_id":"679ce8c048ebd7903d76a832","avatarUrl":"/avatars/5f3fecaacfee6e2d5a72dd19fe87055a.svg","isPro":false,"fullname":"Youbang Sun","user":"Youbang","type":"user"},{"_id":"6898562e524e753b04240630","avatarUrl":"https://cdn-avatars.huggingface.co/v1/production/uploads/no-auth/ngRjX-dHx7SUNWavJ3IH6.png","isPro":false,"fullname":"Jincheng","user":"JinCheng777","type":"user"},{"_id":"622474f38dc6b0b64f5e903d","avatarUrl":"/avatars/d6b60a014277a8ec7d564163c5f644aa.svg","isPro":false,"fullname":"Yuxin Zuo","user":"yuxinzuo","type":"user"},{"_id":"6739b8405d4d88ae26140940","avatarUrl":"https://cdn-avatars.huggingface.co/v1/production/uploads/no-auth/l_5rD6EVQ9vDui0UFDEJq.png","isPro":false,"fullname":"罗天蔚","user":"Cyan666","type":"user"},{"_id":"66ed083acaf696884760729a","avatarUrl":"https://cdn-avatars.huggingface.co/v1/production/uploads/noauth/RgPe99BqsJsHUWoXO1qtS.jpeg","isPro":false,"fullname":"Nick Yang","user":"RadioBlue","type":"user"},{"_id":"682b0d5da1b279a81a04b34f","avatarUrl":"/avatars/69cccc5dd800ace0335d78934e9de16e.svg","isPro":false,"fullname":"Baiting Wu","user":"PeterWuqwq","type":"user"},{"_id":"69030a56eb9d3355ce577e85","avatarUrl":"/avatars/3093dca379f31679cb96bd7dafdcad9a.svg","isPro":false,"fullname":"Sirui Chen","user":"SiruiChen","type":"user"},{"_id":"6458e8ce4b7baff9a84aa0da","avatarUrl":"/avatars/c450f4885e68d28c22fd87f9efdfedec.svg","isPro":false,"fullname":"kaikai zhao","user":"LifeIsSoSolong","type":"user"},{"_id":"6445fa2ffc22e309d78bef3e","avatarUrl":"https://cdn-avatars.huggingface.co/v1/production/uploads/6445fa2ffc22e309d78bef3e/FQaINLd0PjgY9EnK_APRk.jpeg","isPro":false,"fullname":"Messi Hua","user":"Messi-Hua","type":"user"}],"acceptLanguages":["en"],"dailyPaperRank":0,"organization":{"_id":"6a4fb75a1c66dbf208e7ddb6","name":"Shanghai-AI-Laboratory","fullname":"Shanghai AI Laboratory","avatar":"https://cdn-avatars.huggingface.co/v1/production/uploads/65cd955637be1841d0b75397/Rao_Kq6NMtTVfSqLUIR4k.webp"},"markdownContentUrl":"https://huggingface.co/buckets/huggingchat/papers-content/resolve/2609/2609.00092.md","query":{}}">

Papers

arxiv:2609.00092

Safin-1: Safety from Within through Memory-Native State Evolution

Published on Aug 31

· Submitted by

mingzhang on Sep 2

Shanghai AI Laboratory

Upvote

Authors:

Abstract

Safin-1 introduces a memory-routing architecture that embeds safety as an internal, evolving model state rather than an external constraint.

Generated by thinkingmachines/Inkling-Small

Long-horizon complex tasks require foundation models to accumulate information, maintain internal states, and adapt over extended interactions. Safety should be an intrinsic property of the model itself, rather than a behavioral constraint relying solely on external safeguards or post-hoc alignment such as supervised fine-tuning. This motivates Safety from Within, where safety-relevant capabilities are represented and invoked through the model's native computation. We present Safin-1, a family of foundation models realizing this principle through memory routing and state evolution. Safin-1 is built on Memory-Anchor Routing across Context History (MARCH), a network architecture that maintains structured memory states and selectively retrieves relevant historical information through content-conditioned routing. It supports test-time adaptation of persistent capability states without repeatedly modifying the backbone, enabling controlled specialization over a shared foundation. We investigate this interface on downstream safety tasks through a Safety State, demonstrating effective state-based adaptation with substantial safety improvements. More broadly, the routed-state interface unifies contextual memory and persistent capability adaptation within the model's native computation, reframing memory from a passive record of prior context into an active substrate for maintaining and evolving model behavior. Evaluations across general capabilities, long-context understanding, retrieval, and efficiency further validate Safin-1. These findings provide a path toward safety as a state-native and adaptively maintainable capability. This work is only an initial architectural exploration of Safety from Within, and substantial further work is needed to realize this broader vision.