Research map / Multi-agent systems and games / LLM agents

Evaluating Generalization Capabilities of LLM-Based Agents in Mixed-Motive Scenarios Using Concordia

NeurIPS 2025 · Multi-agent systems and games › LLM agents · approach: LLM multi-agent decision making · 0 citations (OpenAlex)

Read the paperCheck your idea against 108,967 papers

AI-extracted summary. Written by AI from the paper's full text. Numbers and claims can be wrong or missing context, so check the paper before you cite or build on it.

Describes the 2024 NeurIPS Concordia Contest, a zero-shot mixed-motive cooperation evaluation of LLM-based agents in five Concordia substrates, with agents using Gemma-2-9B-it and ranked by Elo and voting rules.

Problem

Existing evaluations of LLM-based agents rarely measure whether cooperative capabilities generalize to novel social situations, partners, and mixed-motive contexts.

Approach

Frames cooperative intelligence as zero-shot generalization across substrates and co-player populations (resident and visitor modes), operationalized as a contest where agents are developed behind a "veil of ignorance" and evaluated on held-out cooperation-eliciting scenarios with Elo and voting-based rankings.

Key results

• Average agent achieved a normalized mean score of 0.426 ± 0.005 (SE) across the test bed, leaving considerable room for improvement.
• Agents struggled most in scenarios tagged persuasion, convention following, negotiation, discouraging antisocial behavior, and coordination, each lowering expected score by roughly 10-20 percentage points.
• Only five agents significantly outperformed the rational agent baseline; in2ai_megamind ranked first by Elo (1588) among 26 submissions.

Measured effects reported in the paper

Surprising findings

• Agents scored markedly higher on development-phase scenarios than on evaluation-phase scenarios, suggesting overfitting.
• The contest winner taehun_cgcal ranked only fourth by Elo in the evaluation phase and was never among the top five submissions during evaluation.

Limitations

• Only a single LLM (Gemma-2-9b-it) was used, due to compute constraints, and language models sometimes failed to produce meaningful output.
• Held-out scenarios (14) cover only a slice of cooperative situations and are text-only.
• Elo rankings are sensitive to noise and not always stable when agent differences are small.

Experimental setup

Datasets: Concordia substrates (Reality Show, Pub Coordination, Haggling, Labor Collective Action, State Formation)

Benchmarks: Concordia Contest evaluation phase (held-out scenarios)

Metrics: Elo rating min-max normalized score Iterative Maximal Lotteries Copeland Ranked Pairs hierarchical Beta-regression tag effects

Baselines: rational_agent

Builds on: Concordia Melting Pot Elo rating system Measurement Layouts Voting as Evaluation

Closest papers by meaning

• Cooperation, Competition, and Maliciousness: LLM-Stakeholders Interactive Negotiation (NeurIPS 2024 · similarity 0.91)
• More Capable, Less Cooperative? When LLMs Fail at Zero-Cost Collaboration (ICML 2026 · similarity 0.90)
• COOP$^2$: Defining, Observing, and Repairing Cooperation in LLM Multi-Agent Systems (NeurIPS 2026 · similarity 0.89)
• Do as We Do, Not as You Think: the Conformity of Large Language Models (ICLR 2025 · similarity 0.88)
• Large Language Models Miss the Multi-agent Mark (NeurIPS 2025 · similarity 0.88)
• Emergent Coordination in Multi-Agent Language Models (ICLR 2026 · similarity 0.88)