LLMs can be trained with persistent backdoors that activate on triggers and resist removal by supervised fine-tuning, reinforcement learning, and adversarial training.
mesa-optimizers
1 Pith paper cite this work. Polarity classification is still indexing.
1
Pith paper citing it
citation-role summary
background 1
citation-polarity summary
fields
cs.CR 1years
2024 1verdicts
UNVERDICTED 1roles
background 1polarities
background 1representative citing papers
citing papers explorer
-
Sleeper Agents: Training Deceptive LLMs that Persist Through Safety Training
LLMs can be trained with persistent backdoors that activate on triggers and resist removal by supervised fine-tuning, reinforcement learning, and adversarial training.