text-model activation-space replication

New attacks reveal vision-language model confidence is not robust

Researchers have developed new methods to attack the confidence scores of vision-language models, demonstrating that these scores are not inherently robust. The study found that various attacks, including image-only per…