MultiAgentBench and EscapeBench at ACL 2025
Two co-authored papers accepted to the ACL main conference, studying interaction, collaboration, and agent problem solving.
Two papers at ACL 2025
MultiAgentBench and EscapeBench have been accepted to the ACL 2025 main conference. I am a core contributor and co-first author of MultiAgentBench, and a co-author of EscapeBench.
Evaluating collaboration and competition
MultiAgentBench studies LLM agents across interactive settings, measuring both task performance and the quality of collaboration or competition. Its MARBLE framework makes communication structure and coordination strategies explicit parts of the evaluation.
My work includes the Werewolf setting and analysis of agent interaction, theory of mind, cooperation and distrust. The project page and publication record provide the paper and code.
Continuing the research
The two papers are retained in the publication record. This announcement marks their acceptance; the project and paper pages are the place to find the research materials.
Loading comments…