Research map / Multi-agent systems and games / LLM agents

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

AAAI 2026 · Multi-agent systems and games › LLM agents · approach: Self-evolving language agent learning · 0 citations (OpenAlex)

Read the paperCheck your idea against 108,967 papers

AI-extracted summary. Written by AI from the paper's full text. Numbers and claims can be wrong or missing context, so check the paper before you cite or build on it.

Proposes S-DAG, which uses a GNN to build a subject-level DAG for a multi-subject question and matches each subject node to a profiled expert LLM for graph-structured multi-agent collaboration, evaluated on curated multi-subject subsets of MMLU-Pro, GPQA and MedMCQA.

Problem

Single LLMs and task-level routing or mixture approaches struggle with heterogeneous questions that span multiple subjects, and prior multi-agent methods lack fine-grained subject-level structure and efficiency.

Approach

A BERT-plus-GNN module predicts subject relevance and directed support-to-dominant edges to form a pruned S-DAG; each subject node is assigned the expert LLM with the highest subject-specific capability score, and outputs flow along DAG edges to a final dominant agent.

Key results

• S-DAG reaches 59.73% average accuracy across MMLU-Pro, GPQA and MedMCQA, versus 57.53% for Symbolic-MoE and 58.52% for GPT-4o-mini.
• Replacing the S-DAG with a fully connected graph lowers accuracy to 57.29% and raises inference time from 15.02s to 38.45s and LLM calls from 4.1 to 8.2.
• Profiled model selection raises average accuracy from 42.19% (random model, with GNN) to 53.51%... compared to 59.73% for full method without the profile, the paper reports the gain across all benchmarks.

Surprising findings

• The LLM-generated subject graph (without the GNN) yields significantly lower accuracy than the learned GNN graph, suggesting LLM subject outputs are noisy.

Limitations

• Expert models are small, fine-tuned models from Hugging Face whose quality varies; subject labels and graph supervision come from a single LLM (qwen-turbo) run three times.
• Multi-subject subsets were manually selected and are small (hundreds of test items), limiting statistical power.
• Experiments on the curated subsets do not cover other reasoning domains or open-ended tasks.

Experimental setup

Datasets: MMLU-Pro GPQA MedMCQA

Benchmarks: MMLU-Pro GPQA MedMCQA

Metrics: accuracy average accuracy inference time number of LLM calls

Baselines: GPT-4o-mini CoT Qwen2.5-72B CoT Llama3.3-70B CoT Qwen2.5-7B CoT Self-Refine MoE GraphRouter MAD Symbolic-MoE

Builds on: Symbolic-MoE GraphRouter Mixture-of-Agents Multi-Agent Debate

Closest papers by meaning

• Think-on-Graph: Deep and Responsible Reasoning of Large Language Model on Knowledge Graph (ICLR 2024 · similarity 0.88)
• MAGDi: Structured Distillation of Multi-Agent Interaction Graphs Improves Reasoning in Smaller Language Models (ICML 2024 · similarity 0.88)
• Graph-of-Agents: A Graph-based Framework for Multi-Agent LLM Collaboration (ICLR 2026 · similarity 0.87)
• Schema-Guided Scene-Graph Reasoning Based on Multi-Agent Large Language Model System (AAAI 2026 · similarity 0.86)
• Reasoning on Graphs: Faithful and Interpretable Large Language Model Reasoning (ICLR 2024 · similarity 0.86)
• Distilling LLM Agent into Small Models with Retrieval and Code Tools (NeurIPS 2025 · similarity 0.86)