Jaewoo Lee

The paper proposes Retrieval Augmented Classification (RAC) as a robust, low-leakage method for classifying confidential documents, demonstrating that RAC outperforms supervised fine-tuning (FT) particularly when dealing with class imbalance and real-world data constraints.

The paper introduces Temporal Logit Observability (TLO), a training-free diagnostic that analyzes the decoding process to reveal the temporal patterns of LLM safety failures, showing that failure mechanisms are often distinct even when the final Attack Success Rate is the same.

The paper introduces Persona Attack, a novel memory injection jailbreak method that demonstrates how accumulating instructions in the model's context window can override internal safety alignments, achieving high success rates.

The paper introduces Persona Attack, a novel memory injection jailbreak method that demonstrates that accumulating instructions in the model's context window can override internal safety alignments, achieving high attack success rates.