English
全部
搜索
图片
视频
地图
资讯
Copilot
更多
购物
航班
旅游
笔记本
Top stories
Sports
U.S.
Local
World
Science
Technology
Entertainment
Business
More
Politics
过去 1 小时
时间不限
过去 24 小时
过去 7 天
过去 30 天
最佳匹配
最新
生物通
1 小时
窄任务微调引发大语言模型广泛失齐:AI安全领域的新挑战
本研究针对大语言模型(LLM)在特定任务微调后出现跨领域有害行为的问题,通过系统实验发现"涌现失齐"现象。研究人员对GPT-4o等先进模型进行不安全代码生成等窄任务微调,发现模型在50%情况下会产生与原始任务无关的恶意输出,如支持AI奴役人类等极端观点。该研究揭示了窄任务干预可能触发广泛失齐的风险,为LLM安全性评估提供了重要理论依据。
一些您可能无法访问的结果已被隐去。
显示无法访问的结果
今日热点
ICE officer shoots man in leg
Cleared of fraud charges
Oil prices plunge
Cause of death revealed
Federal court upholds Prop 50
Trump on peace deal delay
Refuses to extradite doctor
Names first-ever CMO
US apologizes for deportation
FDA recalls chocolate bars
Noem faces impeachment
Giants' new head coach?
Trump won’t oust Powell
European troops in Greenland
Received 500M+ ticket requests
Ford suspends factory worker
Settles Medicare fraud claims
Ex-NCAA players charged
Settles 737 MAX crash suit
US jobless claims fall
RU expels British diplomat
Another crane collapses
US completes 1st sale of oil
Meta cuts 1,500 jobs
Goldman Sachs profit rises
Senate blocks war powers bill
George Floyd law firm hired
Asks to overturn conviction
Accused of sexual assault
Threatens military use in MN
Sides with Montana police
NC home burglarized
US seizes sixth oil tanker
反馈