AutoResearch: Insight In, Hallucination Out

Recent advances in large language models and agentic systems have substantially expanded the range of scientific research tasks that can be delegated to artificial intelligence. However, automating a larger fraction of the workflow does not by itself ensure a scientifically meaningful research process. Autonomous research requires an integrated system spanning both research discovery and research execution, where automation serves rather than defines the process. This paper introduces AutoResearch, a two-stage system designed to address how research ideas are formed and how they are reliably established through experimentation, adhering to the principle: Insight In, Hallucination Out.

AutoResearch operates through a two-stage research process: Idea Generation and Idea Execution. In the first stage, the system continuously integrates external research signals with a curated domain knowledge base, identifies transferable mechanistic insights, and uses multi-model generation and cross-review to produce grounded, testable research plans. In the second stage, coordinated agents decompose these plans into executable tasks, iteratively implement and diagnose them, and employ independent evidence-based review before accepting research conclusions. This design prevents unsupported intermediate outputs from being promoted into established research conclusions.

To evaluate its effectiveness, AutoResearch is tested across open-ended method exploration, systems optimization, and benchmark-driven machine learning. On the Remote Sensing Image Captioning Dataset (RSICD) for cross-modal retrieval, an AutoResearch-generated idea improves the mean Recall from 32.84 to 34.69 through staged improvements. Furthermore, AutoResearch records only 5 audit-confirmed issue events, compared to 11 to 27 for other comparable autonomous research systems, demonstrating its capability to detect and correct unreliable experimental outcomes.