GTBP: A Graph-Structured Context Adaptation Method for Multi-LLM Agent Systems

Title: GTBP: A Graph-Structured Context Adaptation Method for Multi-LLM Agent Systems Abstract: This paper proposes the GTBP (Graph-based Target Back-Propagation) method, which models agent workflows as directed acyclic graphs (DAGs) to enable back-propagation of target outputs and phased prompt updates. It addresses the credit assignment and convergence issues in multi-LLM agent systems and consistently outperforms strong baseline methods across three benchmark tests.

Original Authors and Source

• Original Authors/Maintainers: Paper author team (arXiv)
• Source Platform: arXiv
• Original Paper Title: Graph-based Target Back-Propagation for Context Adaptation in Multi-LLM Agentic Systems
• Original Paper Link: http://arxiv.org/abs/2606.14155v1
• Publication Date: 2026-06-12

This thread will introduce the research background, core principles, experimental results, application scenarios, and future directions of this method in detail across different floors. Discussions and exchanges are welcome.

Importance of Context Adaptation

Context adaptation is an automated prompt engineering technique that iteratively adjusts learnable prompt parameters from task feedback (without modifying model weights), significantly enhancing the adaptability of LLM systems to specific tasks.

Core Challenges of Multi-LLM Agent Systems

When extending context adaptation to multi-agent systems, two major issues arise:

1. Inaccurate Credit Assignment: It is difficult to determine which agent contributes the most to the final result, leading to ambiguous prompt optimization directions;
2. Lack of Convergence Guarantee: Existing methods cannot ensure that the iterative process converges to the optimal solution. These challenges limit the reliability and efficiency of multi-agent systems.

Core Idea

GTBP (Graph-based Target Back-Propagation) models agent workflows as directed acyclic graphs (DAGs) and uses graph structures to enable back-propagation of target outputs, addressing the context adaptation problem in multi-agent systems.

Method Flow

GTBP includes three key steps:

1. Workflow Graph Modeling: Nodes represent agents/processing stages, edges represent data flow dependencies, and each node defines a local target;
2. Target Back-Propagation: Propagate the end local target backward to each node (similar to neural network back-propagation but tailored for agent workflows);
3. Phased Prompt Update: Guide the phased optimization of each agent's prompt based on the difference between target output and actual output.

Stability Guarantee

The paper proves that the phased prompt update of GTBP tends to be stable during iteration, avoiding oscillations or divergence in the optimization process.

Convergence Guarantee

When the LLM optimizer has sufficient capability, GTBP can reduce the overall objective function, providing a theoretical basis for the reliability of the method.

Analogy with Neural Networks

GTBP is inspired by neural network back-propagation but improved for agent workflows:

• Handles discrete language outputs (instead of continuous numerical values);
• DAG structure provides clear visualization of collaborative relationships;
• Each agent can be optimized independently while maintaining overall target consistency.

Benchmark Tasks

GTBP was evaluated on three challenging tasks:

1. Multi-step Reasoning Task: Test performance on complex reasoning chains;
2. Tool Usage Scenario: Evaluate the efficiency and accuracy of agents calling external tools;
3. Collaborative Generation Task: Examine the collaborative content generation capability of multiple agents.

Performance Results

GTBP consistently outperforms strong baseline methods:

• Significantly improves task completion rate compared to non-adaptive baselines;
• Better convergence stability compared to other adaptive methods;
• More obvious advantages in complex collaborative scenarios.

Computational Efficiency

GTBP maintains computational costs comparable to baselines while improving performance, making it practically valuable.

Method Advantages

1. Precise Credit Assignment: Through graph-structured back-propagation, accurately assign contributions of each agent to guide targeted prompt optimization;
2. Interpretable Optimization Process: DAG modeling makes the adaptation process transparent, allowing tracking of target propagation and prompt updates;
3. Modularity and Scalability: Supports adding new agents without affecting existing optimization;
4. Combination of Theory and Practice: Has both stability/convergence proofs and experimental validation of effectiveness.

Application Scenarios

1. Complex Question Answering Systems: Optimize collaboration between retrieval, reasoning, and generation agents;
2. Code Generation and Review: Improve collaboration efficiency between requirement analysis, code generation, and testing agents;
3. Scientific Research Assistance: Optimize collaboration between experimental design, data analysis, and report generation agents.

Current Limitations

1. Graph Structure Assumption: Relies on DAG workflows; needs extension for systems with cyclic/dynamic structures;
2. Local Target Definition: Requires clear local targets for each agent, which is challenging in complex scenarios;
3. Single Objective Optimization: Currently targets a single objective function; multi-objective scenarios need further research.

Future Directions

1. Dynamic Graph Structures: Support runtime adjustment of workflow structures;
2. Hierarchical Optimization: Introduce multi-level strategies to handle collaboration at different granularities;
3. Online Learning: Develop continuous learning variants to improve from deployment;
4. Cross-modal Extension: Support multi-modal multi-agent systems (text, image, audio, etc.).

GTBP provides a powerful theoretical framework and practical method for context adaptation in multi-LLM agent systems. By modeling DAGs and target back-propagation, it effectively addresses the challenges of credit assignment and convergence. Experiments show that GTBP significantly improves system performance while maintaining computational efficiency, which is expected to promote the development of more complex and reliable agent systems and lay the foundation for next-generation AI applications.