Reproduce “Evaluating Generalization Capabilities of LLM-Based Agents in Mixed-Motive Scenarios Using Concordia” on OpenResearch

About this paper

Chandler Smith, Marwa Abdulhai, Manfred Diaz, Marko Tesic, Rakshit S. Trivedi, Alexander Sasha Vezhnevets, Lewis Hammond, Jesse Clifton and 78 more

Concordia evaluates whether language-based agents can find mutual gains with unfamiliar partners across changing social rules and populations. It uses five text-based simulations and held-out scenarios with varied partner strategies. Contestants averaged a normalized score of 0.426 ± 0.005 standard error; only five agents significantly outperformed a rational-agent baseline. These 2024 contest results use one shared Gemma 2 model and do not establish broader LLM performance.

Key insights

• General cooperative intelligence in LLM agents can be rigorously evaluated by testing their ability to identify and exploit mutual gains across diverse, unseen social contexts and co-player strategies.
• LLM-based agents exhibit specific 'capability gaps' in cooperative scenarios, particularly struggling with complex social demands such as persuasion, negotiation, and discouraging antisocial behavior.
• The 'veil of ignorance' combined with varied social dilemma environments provides an effective framework for benchmarking zero-shot generalization, highlighting the need for agents to develop flexible and robust social strategies.

Summary by alphaXiv.