[论文解读] The AI Scientist: Towards Fully Automated Open-Ended Scientific Discovery

Chris Lu, Cong Lu+4 more|arXiv (Cornell University)|Aug 12, 2024

Scientific Computing and Data Management被引用 93

一句话总结

The AI Scientist autonomously generates research ideas, code, runs experiments, writes full papers, and performs automated reviews to enable end-to-end open-ended ML discovery.

ABSTRACT

One of the grand challenges of artificial general intelligence is developing agents capable of conducting scientific research and discovering new knowledge. While frontier models have already been used as aides to human scientists, e.g. for brainstorming ideas, writing code, or prediction tasks, they still conduct only a small part of the scientific process. This paper presents the first comprehensive framework for fully automatic scientific discovery, enabling frontier large language models to perform research independently and communicate their findings. We introduce The AI Scientist, which generates novel research ideas, writes code, executes experiments, visualizes results, describes its findings by writing a full scientific paper, and then runs a simulated review process for evaluation. In principle, this process can be repeated to iteratively develop ideas in an open-ended fashion, acting like the human scientific community. We demonstrate its versatility by applying it to three distinct subfields of machine learning: diffusion modeling, transformer-based language modeling, and learning dynamics. Each idea is implemented and developed into a full paper at a cost of less than $15 per paper. To evaluate the generated papers, we design and validate an automated reviewer, which we show achieves near-human performance in evaluating paper scores. The AI Scientist can produce papers that exceed the acceptance threshold at a top machine learning conference as judged by our automated reviewer. This approach signifies the beginning of a new era in scientific discovery in machine learning: bringing the transformative benefits of AI agents to the entire research process of AI itself, and taking us closer to a world where endless affordable creativity and innovation can be unleashed on the world's most challenging problems. Our code is open-sourced at https://github.com/SakanaAI/AI-Scientist

研究动机与目标

• Motivate and enable fully automated, open-ended scientific discovery beyond isolated automation tasks.
• Demonstrate an end-to-end pipeline where frontier LLMs ideate, plan, execute experiments, write manuscripts, and simulate reviews in ML subfields.
• Show that automated reviewing can approach near-human performance and guide iterative knowledge accumulation.

提出的方法

• Use an LLM-driven agent (The AI Scientist) to generate novel research ideas and assess novelty and feasibility.
• Leverage Aider, an LLM-based coding assistant, to implement plan-directed code changes and execute experiments.
• Automatically generate plots and LaTeX manuscript sections from experimental notes and results.
• Execute a simulated review process using a GPT-4o-based reviewer aligned with conference guidelines.
• Maintain an open-ended archive of discovered ideas and artifacts for iterative growth.

实验结果

研究问题

• RQ1 Can an autonomous system generate, execute, and report novel ML research ideas with minimal human intervention?
• RQ2 What is the feasibility and cost of fully automated ML research across multiple subfields?
• RQ3 How well can an automated reviewer judge automated ML papers compared to human reviews?
• RQ4 What are the strengths, limitations, and ethical considerations of end-to-end automated scientific discovery?

主要发现

• The AI Scientist can generate and execute complete ML papers with low cost (under $15 per paper).
• An automated LLM reviewer achieves near-human performance in key evaluation metrics on ICLR/NeurIPS-style benchmarks.
• The pipeline supports end-to-end paper production, including ideation, experimentation, manuscript drafting, and automated review.
• The framework works across multiple ML subfields (diffusion modeling, transformer language modeling, learning dynamics).
• The system can produce papers that meet conference-like acceptance thresholds as judged by its own reviewer.