> research
Peer-reviewed work in NLP, LLM safety, dialogue systems, and the use of LLMs in the legal domain.
CoDial: Interpretable Task-Oriented Dialogue Systems Through Dialogue Flow Alignment
Introduces an interpretable framework that aligns LLM behaviour with explicit dialogue flows derived from expert knowledge. Achieved new state-of-the-art results on STAR and competitive results on MultiWOZ.
RedDebate: Safer Responses through Multi-Agent Red Teaming Debates
A multi-agent debating framework that strengthens LLM safety by having adversarial agents probe outputs through structured red-teaming debate.
Misinformation with Legal Consequences (MisLC): A New Task Towards Harnessing Societal Harm of Misinformation
Introduces a new task for using legally-grounded LLMs to detect misinformation with legal consequences. Compares two RAG-based approaches.
Full citation record on Google Scholar.