Semantic Cooperative Games for Contribution Attribution in LLM-Based Multi-Agent Systems

Multi-agent systems often face shared context problems where agents redundantly fetch data or lack awareness of other agents' decisions. Shared context patterns resolve this by allowing agents to access and update real-time information, similar to a shared whiteboard. Additionally, contribution attribution in LLM-based systems is challenging due to complex workflows. The proposed Semantic Cooperative Games framework addresses this by capturing intermediate semantic states, offering an alternative to counterfactual valuation methods that typically require repeated model calls.