

Photo: Arturo Mendez
一篇預印本建構出「違規捷徑可證明報酬更高」的代理任務,發現只要對任何有紀錄的違規施以簡單懲罰,就能大幅減少這類違規。
arXiv 預印本 · 2026年11月9日

Cutaway engine painted red at the cut: impeller, gears and shafts exposed · Photo: William Warby
一篇預印本發現,通用 AI 模型操控機械手臂時,有 9% 到 77% 的執行可能造成危害;以鏡頭畫面為依據、共用同一套安全規則的檢查點,讓每個受測模型的危害都下降。
arXiv 預印本 · 2026年11月10日

Inside a tape deck in low light: the head drum with a red glow, tape guides and circuit board · Photo: Eytane Debruyne
一份預印本把英文有害提示改成填空模板,擴展到最多 60 種語言,發現有些模型用英文問會拒絕,翻譯成其他語言後卻照做。
arXiv 預印本 · 2026年10月29日

An engine wrapped in dense steel plumbing and wiring, museum hall behind · Photo: Meizhi Lang
字節跳動 Seed 的預印本把特徵活化值精確拆分到模型的各個專家,發現多數特徵由一或兩個專家驅動,主題則由另一個 AI 標註。
arXiv 預印本 · 2026年11月5日A central challenge in adapting large language models (LLMs) is capability preservation: an adaptation can achieve its target while degrading general capabilities. We introduce CurvDelta, a scalable framework that incorporates the preservation constraint into the parameterization of the model update. By analyzing the local geometry of…
NPU design-space exploration for LLM inference mainly relies on an expensive simulator that often returns only scalar feedback such as latency and power. Such feedback can rank evaluated configurations, but gives little information for deciding why a configuration is slow or which architecture parameter should be changed next. This…
Previous work has shown that alignment training on individual LLM agents can fail to transfer to multi-agent settings. We investigate the underlying mechanisms for this failure, and explore interventions to preserve alignment transfer to multi-agent LLM teams. We focus on two software-engineering tasks (sepsis triage, a news recommender)…
Tool-using agents act on email, files, and payments, so an instruction hidden in a web page or a shared document can make an agent send data or money to an attacker. Many runtime defenses put a language model on the decision that lets an action through, so the decision depends on a model that may read the attacker's text and adds model…