Can Graph Learning Improve Planning in LLM-based Agents? (2405.19119v3)

Abstract: Task planning in language agents is emerging as an important research topic alongside the development of LLMs. It aims to break down complex user requests in natural language into solvable sub-tasks, thereby fulfilling the original requests. In this context, the sub-tasks can be naturally viewed as a graph, where the nodes represent the sub-tasks, and the edges denote the dependencies among them. Consequently, task planning is a decision-making problem that involves selecting a connected path or subgraph within the corresponding graph and invoking it. In this paper, we explore graph learning-based methods for task planning, a direction that is orthogonal to the prevalent focus on prompt design. Our interest in graph learning stems from a theoretical discovery: the biases of attention and auto-regressive loss impede LLMs' ability to effectively navigate decision-making on graphs, which is adeptly addressed by graph neural networks (GNNs). This theoretical insight led us to integrate GNNs with LLMs to enhance overall performance. Extensive experiments demonstrate that GNN-based methods surpass existing solutions even without training, and minimal training can further enhance their performance. The performance gain increases with a larger task graph size.

• The paper introduces a novel integration of graph learning to overcome LLM task planning limitations, outperforming prompt-based methods.
• It leverages both training-free and training-required GNN approaches to represent task dependencies and boost planning accuracy.
• Experimental results demonstrate significant, scalable performance gains with reduced computational costs in complex decision-making tasks.

Enhancing Task Planning in LLMs with Graph Neural Networks

In the field of artificial intelligence, the integration of LLMs with auxiliary techniques continues to garner interest, particularly in complex decision-making tasks like task planning. The paper "Can Graph Learning Improve Planning in LLM-based Agents?" explores a novel method of enhancing LLMs' task planning capabilities by incorporating graph neural networks (GNNs), contrasting it with the traditional prompt-focused approaches.

Task Planning in LLM-based Agents

Task planning in LLM-based agents involves interpreting a user request, breaking it down into actionable sub-tasks, and executing these tasks to fulfill the request. Historically, this has been tackled primarily through improvements in prompt design. However, this paper proposes a distinct methodology by employing graph learning. Within this paradigm, each sub-task is represented as a node and dependencies as edges, creating a task graph. The decision-making challenge thus translates into selecting a coherent subgraph in this task network.

Limitations of LLMs in Task Planning

Empirically, the authors demonstrate that LLMs, such as those utilized in HuggingGPT, exhibit deficiencies in structuring and interpreting task graphs, leading to planning failures. The theoretical insights presented highlight several LLM limitations:

1. Expressiveness: While Transformers theoretically have the capacity to simulate dynamic programming via edge list inputs, they lack permutation invariance necessary for graph tasks, encounter sparse attention issues, and show deficiencies in generalizing graph structures.
2. Auto-regressive Loss: It introduces spurious correlations which, when compounded, degrade the model’s ability to make accurate task predictions based on the sequential graph inputs.

The Proposed Solution: Integrating GNNs

The primary contribution of this research is the integration of GNNs into the planning process, which addresses these LLM limitations. GNNs inherently understand graph structures, avoiding hallucinations and offering superior task retrieval abilities. The proposed framework operates in two main approaches:

• Training-free GNN Approach: This method leverages a parameter-free model to interpret task steps from LLMs without additional learning, providing an efficient zero-shot solution that improves with enhanced prompt design.
• Training-required GNN Approach: Employing models like GraphSAGE, and using Bayesian Personalized Ranking loss, facilitates learning from both steps and task dependencies, significantly surpassing baselines in performance across various metrics and datasets.

Numerical Findings and Implications

Extensive experiments illustrate that integrating GNNs markedly enhances task planning performance. Notably, performance gains escalate with increased task graph size, showcasing GNNs' scalability. The approach consistently outshines existing methods in both training-free and training-required scenarios, achieving remarkable results with reduced computational costs.

Future Directions and Implications

This research opens pathways for more advanced graph learning methodologies within LLM contexts and emphasizes the need for automated graph generation techniques to handle diverse applications. By addressing decision-making limitations in LLMs with graph learning, this paper not only enhances existing computational models but also sets the stage for further integration of graph-based machine learning with LLMs.

In conclusion, the paper highlights a significant advancement in the field of AI task planning, presenting a robust, scalable solution that creatively marries the power of LLMs with the structural acumen of GNNs, inviting further exploration into their combined potential in diverse computational challenges.