LARGO optimizes a latent embedding to make a model answer affirmatively, decodes it into a natural-language suffix with the model itself, and iteratively refines until the suffix jailbreaks the model.
Title resolution pending
1 Pith paper cite this work. Polarity classification is still indexing.
1
Pith paper citing it
fields
cs.LG 1years
2025 1verdicts
CONDITIONAL 1representative citing papers
citing papers explorer
-
LARGO: Latent Adversarial Reflection through Gradient Optimization for Jailbreaking LLMs
LARGO optimizes a latent embedding to make a model answer affirmatively, decodes it into a natural-language suffix with the model itself, and iteratively refines until the suffix jailbreaks the model.