Mintong Kang
Titles as the sources gave them. In 2 entries, with 0 checked quotations.
In the Base
Every entry they appear in.
The Target Agent Is Its Own Best Critic but a Poor Proposer: Self-Improved Red Teaming
According to the abstract, Kang, Rawls, Kazemi Rad and colleagues separate criticism from attack proposal in automated red teaming. They report that the target agent's own criticism helps, while letting the target propose attacks works poorly because it refuses or underperforms on purpose.
C-SafeGen: Certified Safe LLM Generation with Claim-Based Streaming Guardrails
According to the abstract, Kang, Chen and Li propose a decoding method that checks each claim an LLM generates with a streaming guardrail and backtracks on risky ones, paired with a statistical risk guarantee based on conformal analysis.