Search papers, labs, and topics across Lattice.
Chiao Tung University ♣, IBM Research ♠ Hon Hai Research Institute
2
0
4
TRACE reveals how token influence can expose malicious manipulations in RAG systems, achieving high detection rates without heavy computational costs.
RAS reveals that internal representation alignment can serve as a reliable and efficient metric for LLM safety, outperforming conventional output-based evaluations.
