$\texttt{APOLLO}$: A Self-Guided Multi-Agent System for Scientific Article Generation inspired by Human Thinking

ICLR 2026 Conference Submission17707 Authors

Keywords: LLM-based Multi-agent Systems, Long-form text generation, Knowledge Graph Construction, Fact Verification

TL;DR: We introduce Apollo, a multi-agent framework that generates comprehensive and fact-checked Wikipedia-style scientific articles with citations, outperforming existing methods in accuracy, structure, and diversity.

Abstract:

Automatic generation of Wikipedia-like articles through Retrieval-Augmented Generation (RAG) has recently gained increasing attention.
While recent advances in Large Language Models (LLMs) show considerable promise for synthesizing complex information, current RAG-based systems suffer from two fundamental limitations: they often rely on shallow retrieval strategies, leading to redundant content, and they lack effective mechanisms for factual verification and content organization.
To address these challenges, we present $\texttt{APOLLO}$, a multi-agent framework specifically designed to generate high-quality, comprehensive articles with citations to the given sources.
$\texttt{APOLLO}$ simulates the iterative research and editorial process of human contributors through a set of specialized agents that collaboratively retrieve, fact-check, and structure information.
To evaluate our method, we introduce SciWiki-2k, a dataset comprising 2,000 high-quality Wikipedia articles spanning 20 scientific domains.
Compared to baseline methods, $\texttt{APOLLO}$ produces articles with significantly improved structural coherence, content diversity, and factual accuracy. Human evaluations further establish the practical value of our approach for generating trustworthy, comprehensive articles.

Primary Area: applications to computer vision, audio, language, and other modalities

Submission Number: 17707