Reproduce “S-DAG: A Subject-Based Directed Acyclic Graph for Multi-Agent Heterogeneous Reasoning” on OpenResearch
Turn your favorite coding agents into research agents that review literature, develop hypotheses, and run experiments.
Universal app for Apple silicon and Intel. Requires macOS 11 or later.
About this paper
Jiangwen Dong, Zehui Lin, Wanyu Lin, Mingjin Zhang
Researchers from The Hong Kong Polytechnic University developed S-DAG, a framework that addresses complex, multi-subject questions by decomposing them into a Subject-based Directed Acyclic Graph (S-DAG) and coordinating specialized LLMs based on this structure. The S-DAG framework achieved an average accuracy of 59.73% across MMLU-Pro, GPQA, and MedMCQA benchmarks, outperforming single-model and multi-agent baselines while reducing inference time and LLM calls.
Key insights
- Decomposing complex, multi-subject questions into a structured directed acyclic graph of interconnected subjects significantly improves reasoning accuracy and computational efficiency.
- A Graph Neural Network (GNN) can effectively learn robust, fine-grained inter-subject dependencies from training data, overcoming the inconsistencies often observed in direct LLM outputs for structural generation.
- Precise profiling of LLM capabilities across specific subject domains and matching them to corresponding subject nodes in the S-DAG is crucial for optimal multi-agent collaboration and overall system performance.
Summary by alphaXiv.
Tip: change arxiv.org to autoarxiv.org in any arXiv abstract link to open the paper here.