Instruction tuning shifts LLM susceptibility to misinformation from assistant-provided text to user-provided text, making user-role misinformation more likely to override the model's correct parametric knowledge.
Title resolution pending
1 Pith paper cite this work. Polarity classification is still indexing.
1
Pith paper citing it
fields
cs.CL 1years
2025 1verdicts
CONDITIONAL 1representative citing papers
citing papers explorer
-
Exploring the Impact of Instruction-Tuning on LLM's Susceptibility to Misinformation
Instruction tuning shifts LLM susceptibility to misinformation from assistant-provided text to user-provided text, making user-role misinformation more likely to override the model's correct parametric knowledge.