Instructional Segment Embedding: Improving LLM Safety with Instruction Hierarchy
Instructional Segment Embedding: Improving LLM Safety with Instruction Hierarchy
About this item
Full title
Author / Creator
Publisher
Ithaca: Cornell University Library, arXiv.org
Journal title
Language
English
Formats
Publication information
Publisher
Ithaca: Cornell University Library, arXiv.org
Subjects
More information
Scope and Contents
Contents
Large Language Models (LLMs) are susceptible to security and safety threats, such as prompt injection, prompt extraction, and harmful requests. One major cause of these vulnerabilities is the lack of an instruction hierarchy. Modern LLM architectures treat all inputs equally, failing to distinguish between and prioritize various types of instructio...
Alternative Titles
Full title
Instructional Segment Embedding: Improving LLM Safety with Instruction Hierarchy
Authors, Artists and Contributors
Identifiers
Primary Identifiers
Record Identifier
TN_cdi_proquest_journals_3116750704
Permalink
https://devfeature-collection.sl.nsw.gov.au/record/TN_cdi_proquest_journals_3116750704
Other Identifiers
E-ISSN
2331-8422
