Reproduce “Socially-Weighted Alignment: A Game-Theoretic Framework for Multi-Agent LLM Systems” on OpenResearch

About this paper

Furkan Mumcu, Yasin Yilmaz

A game-theoretic framework called Socially-Weighted Alignment (SWA) is introduced for multi-agent large language model (LLM) systems, enabling agents to internalize collective welfare at inference time. The framework analytically derives a critical social weight (λ*) that predicts a phase transition from system overload to stable operation, a prediction empirically validated across various LLMs in shared-resource environments.

Key insights

• A closed-form critical social weight (λ*) can be derived analytically, predicting a phase transition where the system shifts from persistent overload to stable operation as λ increases.
• SWA effectively reshapes the game played by LLM agents, allowing for the design of system-level stability by aligning individual incentives with global well-being without costly retraining or parameter updates.
• The proposed inference-time mechanism is model-agnostic and generalizes to various forms of negative externalities in multi-agent LLM workflows, such as API rate limits, context window management, and merge conflicts.

Summary by alphaXiv.