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

2026-06-12

AI Summary

This work addresses the challenges of inaccurate credit assignment and lack of convergence guarantees in context adaptation for multi-agent large language model (LLM) systems. The authors propose the Graph-based Target Backpropagation (GTBP) framework, which models agent workflows as directed acyclic graphs and propagates local targets backward through the graph. By leveraging the discrepancy between assigned targets and actual outputs, GTBP incrementally optimizes tunable prompts in a stage-wise manner. This approach provides, for the first time, a theoretically grounded method with convergence guarantees for context adaptation in multi-LLM agent systems, enabling precise credit assignment and stable prompt updates. Experimental results demonstrate that GTBP consistently outperforms strong baselines across three benchmarks while maintaining comparable computational overhead.

Abstract

Context adaptation automates prompt engineering in LLM-based systems by iteratively revising tunable prompts from task feedback, without modifying model weights. Extending this paradigm to multi-LLM agentic systems is crucial: existing methods suffer from inaccurate credit assignment and lack convergence guarantees. We propose \textbf{G}raph-based \textbf{T}arget \textbf{B}ack-\textbf{P}ropagation (GTBP), a context adaptation framework for agentic workflows modeled as directed acyclic graphs. GTBP propagates local target outputs backward through the workflow graph and uses target--output discrepancies to guide a stage-wise prompt update mechanism. Theoretically, we show that GTBP's stage-wise prompt updates become stable over iterations, and that a sufficiently capable LLM optimizer can decrease the overall objective. Empirically, GTBP consistently outperforms strong baselines across three benchmarks while maintaining comparable computational cost.

Problem

• context adaptation
• multi-LLM agentic systems
• credit assignment
• convergence guarantees
• prompt engineering

Innovation

• Graph-based Target Back-Propagation
• Context Adaptation
• Multi-LLM Agentic Systems
• Prompt Engineering
• Convergence Guarantee