ARK: AI Research Harness - Offload the Labour, Steer the Science

Jihao Xin, Yangzhixin Luo, Muhammad Bilal, Marco Canini

Track: Track 3: AI Scientist Proposal Competition

TL;DR: ARK is a co-scientist framework that turns a research idea into a publication-ready paper end-to-end, while keeping humans in full control via real-time steering.

Abstract: We present ARK (Automatic Research Kit), a co-scientist framework that automates the full idea-to-paper pipeline while keeping humans in control at every stage. ARK is a model-agnostic multi-agent system whose seven specialized agents and three-phase pipeline transform a research idea end-to-end into a publication-ready paper. Unlike fully autonomous systems that suffer from uncontrolled drift, hallucinated claims and experimental results, confinement to simulated experiments, and non-compliant formatting, ARK addresses these issues through human-in-the-loop control, hallucination elimination, agent specialization, and retrieval-augmented skills. We deliver ARK as Research-as-a-Service (RaaS): users provide a research idea and API credentials through a web portal, and ARK handles the rest end-to-end in a per-project sandbox, producing submission-ready papers across diverse venues. ARK is publicly accessible at idea2paper.org.

Keywords: AI Scientist, AutoResearch, Research-as-a-Service, Multi-Agent System

Submission Number: 9