← All research projects

Language & agents EMNLP 2026

MARBO: Relational Belief Grounding for LLM Agents in Social Deduction Games

Yechan Hwang · Sangjun Bae · Jeongmo Kim · Sangwoo Bang · Seungyul Han

UNIST · Machine Learning & Intelligent Control Lab

Read the paper ↗

RESEARCH OVERVIEW

Ground strategic actions and in-game speech in reliable beliefs about hidden roles and team relationships.

MARBO uses relational beliefs to guide preference optimization in social deduction games. Feedback depends on both the reliability of those beliefs and the strategic outcomes of the resulting behavior.

Publication details & source ↗