NG-Router: Graph-Supervised Multi-Agent Collaboration for Nutrition Question Answering

TL;DR

This work introduces Nutritional-Graph Router (NG-Router), a novel framework that formulates nutritional QA as a supervised, knowledge-graph-guided multi-agent collaboration problem, and proposes a gradient-based subgraph retrieval mechanism that identifies salient evidence during training, thereby enhancing multi-hop and relational reasoning.

Abstract

Diet plays a central role in human health, and Nutrition Question Answering (QA) offers a promising path toward personalized dietary guidance and the prevention of diet-related chronic diseases. However, existing methods face two fundamental challenges: the limited reasoning capacity of single-agent systems and the complexity of designing effective multi-agent architectures, as well as contextual overload that hinders accurate decision-making. We introduce Nutritional-Graph Router (NG-Router), a novel framework that formulates nutritional QA as a supervised, knowledge-graph-guided multi-agent collaboration problem. NG-Router integrates agent nodes into heterogeneous knowledge graphs and employs a graph neural network to learn task-aware routing distributions over agents, leveraging soft supervision derived from empirical agent performance. To further address contextual overload, we propose a gradient-based subgraph retrieval mechanism that identifies salient evidence during training, thereby enhancing multi-hop and relational reasoning. Extensive experiments across multiple benchmarks and backbone models demonstrate that NG-Router consistently outperforms both single-agent and ensemble baselines, offering a principled approach to domain-aware multi-agent reasoning for complex nutritional health tasks.

Figures (7)

• Figure 1: Illustration of Nutrition QA challenges. (a) Poor dietary habits lead to health risks and complex user–food–condition interactions, as shown in a personalized QA example. (b) Two key research gaps emerge: (i) domain-specific reasoning requires multiple complementary agents, with no single model performing optimally across queries; and (ii) excessive and unstructured contextual information hampers accurate retrieval and reasoning.
• Figure 2: Overview of our proposed framework. (a) shows the KG extension process, where QA instances are extended into context graphs with query and agent nodes linked to nutritional entities. (b) shows our type-aware GNN router, which propagates contextual signals and guides gradient-based subgraph retrieval, refining the graph and producing agent importance weights for downstream collaboration and reasoning.
• Figure 3: Percentage change ($\Delta$) in F1 relative to k=24, used as the base (0%). Curves show how performance varies with the top $k$ agent clipped across datasets.
• Figure 4: F1 performance on the three datasets, varying Layers (top) and Hidden Dimensions (bottom). Error bars denote standard deviations.
• Figure 5: Per-question agent routing cases. Each case shows the question, gold answer, and agents' probability with their generated answers.