Semantic Cooperative Games Attribute Contributions in Multi-Agent LLM Systems.

Researchers have developed Semantic Cooperative Games (SCG), a new framework designed to accurately attribute contributions among agents in complex, multi-agent systems powered by large language models (LLMs). As LLM-based workflows become more intricate, understanding which agent contributes what to a final output is crucial for debugging, optimization, and accountability. Traditional attribution methods often rely on computationally expensive counterfactual evaluations, such as removing agents and observing outcome changes. SCG, however, models the language flow as a "semantic generation hypergraph" and defines a Semantic Shapley Value (SSV) to quantify agent contributions based on semantic support logic. A key component is SLIC, a single-trajectory algorithm that efficiently constructs this hypergraph and computes SSV without needing to rerun agent subsets. This significantly reduces computational cost while maintaining high consistency with traditional Shapley value baselines. The framework provides a fast, interpretable, and counterfactual-free method for understanding agent roles in complex LLM-based multi-agent systems.