Title: GARL: Game-Theoretic Reinforcement Learning for Multi-Agent Strategic Prioritisation

Abstract

Large language model (LLM) driven multi-agent systems are seeing growing adoption in tasks requiring strategic decision-making. In these environments, success is determined not merely by the intrinsic capabilities of individual models, but equally by the policies governing how agents interact and evolve. While multi-agent reinforcement learning offers a pathway to optimise these interaction protocols, the design of rewards within this field often lacks grounding in the underlying interaction structure and remains heavily tailored to specific tasks.

To bridge this divide, we introduce GARL, a framework for GAme-theoretic Reinforcement Learning designed specifically for multi-agent strategic prioritisation. GARL models strategic prioritisation as a two-stage game. In the first phase, competing agents distribute strategic resources across a shared pool of candidates. Subsequently, a higher-level arbiter utilizes these inputs to generate the final ranking. By translating the resulting game-theoretic utilities into role-specific reinforcement signals, GARL ensures that policy optimisation is directed by a structured interaction framework.

We implemented GARL in the context of issues-in-dispute ranking, aiming to prioritise fundamental issues within legal proceedings. Our experimental results indicate that GARL enhances ranking accuracy. Notably, it allows small, open-source LLMs to achieve competitiveness comparable to robust closed-source LLMs within identical candidate-ranking scenarios. Furthermore, the framework delivers improvements in both legal-domain proficiency and general strategic decision-making capabilities. Ultimately, GARL illustrates how the structural dynamics of game-theoretic interactions can be effectively converted into reinforcement-learning objectives, offering a rigorous methodology for policy optimisation in multi-agent strategic prioritisation.

Source: arXiv Generated at: 2026-06-04 00:00:00 UTC