top-k logits

Research reveals logits can leak sensitive info from LLMs

A new research paper explores the potential for information leakage from large language models, even at the final logit level. Researchers used vision-language models to compare information retained at different represe…