ALLAN: Stabilizing Hostilities through Arbitration and Diplomatic Engagement

Project funded by DARPA
PI: Jordan Boyd-Graber

Using a gaming environment based on the classic board game Diplomacy, SHADE will train and evaluate automated and artificial intelligence (AI)-assisted techniques that can identify and address the complications of multi-party negotiation with deception, collusion, profiling, and other real-world features.

Diplomatic negotiation is complex and requires a detailed, in-depth understanding of the positions and interactions of many potentially untrustworthy allies and adversaries. Successful diplomats must identify deceptions, explore valid courses of action, and assess the consequences of any diplomatic actions.

The UMD team is working to build datasets that are annotated for deception in online conversations and build agents that communicate using natural language.

	Jordan Boyd-Graber Associate Professor, Computer Science (UMD)
	Jonathan Kummerfeld PI, Sydney
	Jonathan May PI, USC ISI
	Karthik Narasimhan PI, Princeton
	Denis Peskoff Postdoc, Princeton
	Brandon Stewart PI, Princeton
	Feng Gu Master's Student, UMD

<< back to top

• Denis Peskoff and Brandon Stewart. Credible without Credit: Domain Experts Assess Generative Language Models, ACL 2023.
• Wichayaporn Wongkamjan, Feng Gu, Yanze Wang, Ulf Hermjakob, Jonathan May, Brandon M. Stewart, Jonathan K. Kummerfeld, Denis Peskoff, and Jordan Lee Boyd-Graber. More Victories, Less Cooperation: Assessing Cicero's Diplomacy Play. Association for Computational Linguistics, 2024. [Bibtex]
  
  

  @inproceedings{Wongkamjan:Gu:Wang:Hermjakob:May:Stewart:Kummerfeld:Peskoff:Boyd-Graber-2024,
  	Title = {More Victories, Less Cooperation: Assessing Cicero's Diplomacy Play},
  	Author = {Wichayaporn Wongkamjan and Feng Gu and Yanze Wang and Ulf Hermjakob and Jonathan May and Brandon M. Stewart and Jonathan K. Kummerfeld and Denis Peskoff and Jordan Lee Boyd-Graber},
  	Booktitle = {Association for Computational Linguistics},
  	Year = {2024},
  	Location = {Bangkok, Thailand},
  	Url = {http://cs.umd.edu/~jbg//docs/2024_acl_cicero.pdf},
  }
  

  
  

  Accessible Abstract: Meta's recent AI, Cicero, grabbed headlines by its ability to beat humans at the game of Diplomacy: notable because players of the game not just need to make the right moves but also need to negotiate with each other in natural language. This paper investigates why it wins so many games, measuring its ability to persuade and trick other players. While Cicero wins just about every game, this is because of superhuman strategy, not superhuman communication, suggesting there is still further room for developing Diplomacy-playing AIs.

Acknowledgments

Any opinions, findings, and conclusions or recommendations expressed in this material are those of the researchers and do not necessarily reflect the views of the sponsor.