Multi-Agent RLVE: Emergent Cooperation and Competition in Procedurally Generated Language Environments

Research Question: How does the introduction of multi-agent dynamics—where LMs interact in procedurally generated, adaptive environments—affect the emergence of advanced reasoning, negotiation, and social behaviors?

Hypothesis: Multi-agent adaptive RLVE environments will foster the emergence of sophisticated language strategies, such as negotiation, theory of mind, or adversarial reasoning, that are absent in single-agent training settings.

Experiment Plan: - Extend RLVE-Gym to support multi-agent tasks (e.g., collaborative problem solving, competitive games, negotiation scenarios).

• Train two or more LMs with individual or shared rewards, using adaptive difficulty scaling.
• Track the emergence of cooperative or adversarial strategies and evaluate on benchmarks requiring social reasoning or deception.
• Compare to single-agent RLVE-trained models on tasks requiring multi-party coordination.

References: ['Zeng, Z. et al. (2025). RLVE: Scaling Up Reinforcement Learning for Language Models with Adaptive Verifiable Environments.', 'Xing, X., Zhou, Z., Li, Y., Xiao, B., & Xun, Y. (2024). Multi-UAV Adaptive Cooperative Formation Trajectory Planning Based on an Improved MATD3 Algorithm of Deep Reinforcement Learning. IEEE Transactions on Vehicular Technology.']