liu.seSearch for publications in DiVA
Change search
CiteExportLink to record
Permanent link

Direct link
Cite
Citation style
  • apa
  • ieee
  • modern-language-association-8th-edition
  • vancouver
  • oxford
  • Other style
More styles
Language
  • de-DE
  • en-GB
  • en-US
  • fi-FI
  • nn-NO
  • nn-NB
  • sv-SE
  • Other locale
More languages
Output format
  • html
  • text
  • asciidoc
  • rtf
知识外化:多模态大型语言模型中的可逆遗忘和模块化检索
School of Cyber Science and Engineering, Southeast University, Nanjing, China; Key Laboratory of New Generation Artificial Intelligence Technology and Its Interdisciplinary Applications (Southeast University), Ministry of Education, China.
School of Instrument Science and Engineering, Southeast University, Nanjing, China.
Key Laboratory of New Generation Artificial Intelligence Technology and Its Interdisciplinary Applications (Southeast University), Ministry of Education, China; School of Computer Science and Engineering, Southeast University, Nanjing, China.
Key Laboratory of New Generation Artificial Intelligence Technology and Its Interdisciplinary Applications (Southeast University), Ministry of Education, China; School of Computer Science and Engineering, Southeast University, Nanjing, China.
Show others and affiliations
2026 (Chinese)In: 第十四届国际学习表征会议, 2026Conference paper, Published paper (Refereed)
Abstract [zh]

多模态大型语言模型(MLLM)通过在海量网络数据集上训练,实现了卓越的跨模态理解能力,但却无意中将敏感的个人信息和专有信息内化。现有的机器学习遗忘方法通过不可逆地修改模型参数来永久删除知识。这种破坏性的方法与现代隐私法规相冲突,后者要求数据管理必须可审计、可逆且用户可控。为了应对这些挑战,我们提出了知识外化(Knowledge Externalization)框架,用于在MLLM中实现可逆且模块化的知识管理。我们首先提出了双流记忆调优(Dual-Stream Memory Tuning)方法,该方法将目标知识从模型的内部参数转移到外部记忆令牌(Token)中。为了减轻外化多个概念时的梯度干扰,我们进一步引入了软正交加权(Soft Orthogonal Weighting)技术,该技术能够保持每个令牌的独立性。我们提出的框架展现了三个关键能力:(i)它能够有效地遗忘基础模型中的目标概念,同时利用相应的记忆令牌实现高保真度的知识恢复。 (ii) 它支持持续的知识编辑,允许在外部化之后动态更新存储在外部令牌中的信息。(iii) 它展现出卓越的组合性,可以自由组合多个记忆令牌(包括已编辑的令牌),从而同时恢复与每个概念对应的知识。我们的源代码将在近期发布。

Place, publisher, year, edition, pages
2026.
National Category
Computer Systems
Identifiers
URN: urn:nbn:se:liu:diva-226807OAI: oai:DiVA.org:liu-226807DiVA, id: diva2:2093296
Conference
The Fourteenth International Conference on Learning Representations
Available from: 2026-08-18 Created: 2026-08-18 Last updated: 2026-09-04

Open Access in DiVA

No full text in DiVA

Other links

https://openreview.net/forum?id=ZHK6nBHRXw

Authority records

Geng, Jiahui

Search in DiVA

By author/editor
Geng, Jiahui
By organisation
Artificial Intelligence and Integrated Computer SystemsFaculty of Science & Engineering
Computer Systems

Search outside of DiVA

GoogleGoogle Scholar

urn-nbn

Altmetric score

urn-nbn
Total: 3 hits
CiteExportLink to record
Permanent link

Direct link
Cite
Citation style
  • apa
  • ieee
  • modern-language-association-8th-edition
  • vancouver
  • oxford
  • Other style
More styles
Language
  • de-DE
  • en-GB
  • en-US
  • fi-FI
  • nn-NO
  • nn-NB
  • sv-SE
  • Other locale
More languages
Output format
  • html
  • text
  • asciidoc
  • rtf