S-DAG: A Subject-Based Directed Acyclic Graph for Multi-Agent Heterogeneous Reasoning

We propose S-DAG, a subject-based directed acyclic graph framework for multi-agent heterogeneous reasoning that enables structured collaboration among diverse LLM agents.

Framework

Figure 1: The S-DAG framework consists of two stages: (1) S-DAG Generation — a GNN predicts relevant subjects and their interdependencies to construct the subject-based DAG; (2) Subject–LLM Matching & Multi-Agent Collaboration — each subject node is matched to an expert LLM and agents collaborate according to the DAG topology.

Recommended citation: Dong, Jiangwen, Zehui Lin, Wanyu Lin, and Mingjin Zhang. "S-DAG: A Subject-Based Directed Acyclic Graph for Multi-Agent Heterogeneous Reasoning." AAAI 2026.
Download Paper