Graph-based Target Back-Propagation for Context Adaptation in Multi-LLM Agentic Systems

The paper introduces Graph-based Target Back-Propagation (GTBP), a novel context adaptation framework designed for multi-LLM agentic systems that automates prompt engineering without altering model weights. GTBP utilizes directed acyclic graphs to propagate local target outputs backward, enabling stable stage-wise prompt updates that improve convergence and credit assignment. Empirical results demonstrate that GTBP surpasses existing methods on three benchmarks while maintaining similar computational efficiency, offering significant implications for practitioners focusing on optimizing prompt strategies in multi-LLM environments.