Under adversarial data curation, a self-consuming generative model's alignment with user preferences is claimed to depend on the covariance between true and malicious reward functions; the proposed attack algorithms aim to make that covariance negative.
Best-of-venom: Attacking RLHF by injecting poisoned preference data
1 Pith paper cite this work. Polarity classification is still indexing.
1
Pith paper citing it
fields
cs.LG 1years
2025 1verdicts
REJECT 1representative citing papers
citing papers explorer
-
Self-Consuming Generative Models with Adversarially Curated Data
Under adversarial data curation, a self-consuming generative model's alignment with user preferences is claimed to depend on the covariance between true and malicious reward functions; the proposed attack algorithms aim to make that covariance negative.