Will AI Eliminate Human Scientists? The AI Top Conference Paper You See May Not Be Written by Humans!

Every time AI makes significant progress, researchers often joke: "When can AI write papers for us?"

Now, AI creating scientific research papers has become a reality, and the writing cost is as low as about 15 US dollars.

On August 14th, the Chinese technology news site Zhi Dongxi reported that on August 13th, the Japanese Sakana AI team, in collaboration with researchers from the University of Oxford and the University of British Columbia, launched an AI scientist (The AI Scientist). This is an automated research intelligence agent based on a large model.

Give it a broad research field, and it can create an AI field paper like a human.

The "AI programmer's" programming skills are just one aspect of the AI scientist's capabilities, with brainstorming, code running, experimental result summarization, visualization, and automatic review all within its grasp.

For example, the following paper titled "Dualscale Diffusion: Adaptive Feature Balancing for Low-Dimensional Generative Models" was written by the AI scientist. In experiments independently completed by the AI scientist and peer-reviewed, the paper achieved excellent empirical results and has reached the "weak acceptance" standard of the top machine learning conference.The team in the AI Scientist project has referenced a variety of cutting-edge models, such as closed-source models like GPT-4o and Sonnet, as well as open-source models like DeepSeek and Llama 3.

Advertisement

It is reported that the AI Scientist mainly has the following highlights:

1. This is a fully AI-driven automated research system, focusing on research in the field of machine learning.

2. It achieves the automation of the entire research chain, from the generation of inspiration, code writing and running, to the summarization and visualization of experimental results, and finally writing a complete scientific paper.

3. It innovatively introduces an automated peer review mechanism to evaluate the output papers, providing feedback and continuously optimizing the results, with an evaluation accuracy close to human level.

4. This automated research process continues to cycle, open and continuously accumulate knowledge, simulating the operation mode of the human scientific community.

5. In preliminary tests, it has involved multiple fields of machine learning and achieved results, such as contributions in the fields of diffusion models, Transformer architecture, and Grokking.

AI Scientist paper address:

AI Scientist open-source code and experimental results address: .

Four Steps to Complete a Scientific Paper to Meet the Acceptance Standards of AI Top Conferences

Have you heard of AI poets, AI painters, and AI programmers? Now, AI scientists have also emerged.

An AI scientist is a fully automated paper generation system that makes full use of the most cutting-edge large models.

Starting from a foundational initial codebase, such as ready-made open-source research code on GitHub, as long as a broad research field is given, the AI scientist can complete the entire process from creative conception, literature research, experimental design, experimental iteration, chart production, paper writing to preliminary review, producing academic papers rich in profound insights.

What's more amazing is that the AI scientist can continue to operate in an open loop, constantly learning from previous ideas and feedback to optimize subsequent research ideas. This process highly simulates the operation mode of the human scientific community.

▲AI Scientist's Schematic DiagramThe workflow of an AI scientist primarily consists of four major stages:

Idea Generation: Starting from a given template, an AI scientist initiates a "brainstorming" mode, exploring a range of innovative research directions around the existing theme. This template not only includes a basic code framework but also comes with a LaTeX folder containing style files and preset chapter titles, laying the groundwork for subsequent paper writing. During the process of free exploration, the AI scientist also utilizes the academic search engine Semantic Scholar to ensure the originality of the ideas proposed.

Experimental Iteration: Once the research direction is determined, the AI scientist enters the experimental phase. It automatically carries out the experimental plan, collects data, and generates charts to visually display the results. At the same time, the AI scientist meticulously records the content of each chart, ensuring that the experimental notes and graphical materials provide comprehensive support for the subsequent paper writing.

Paper Writing: After the experiments are completed, the AI scientist uses the LaTeX format to write a well-structured and detailed paper, presenting its research findings to the readers. During the writing process, it also uses Semantic Scholar to automatically search for and cite literature in related fields, enhancing the academic and authoritative nature of the paper.

Automatic Review: To improve the quality of the paper, the team has specially developed an automated review system based on a large language model. This system can objectively assess the generated paper with judgment close to that of a human and provide suggestions for improvement. These feedbacks not only help the AI scientist to optimize the current project but also provide valuable references for future research. Through this continuous feedback loop, the AI scientist can continuously iterate and improve, enhancing the level and impact of research findings.

When combined with the most advanced LLM technology, the AI scientist can even write papers that meet the "weak acceptance" standard of top machine learning conferences and gain recognition through the automatic review system.

02.

AI Scientist Paper Showcase: Covering Diffusion Models, Language Modeling, and More

In the announcement, the team provided a series of papers generated by the AI scientist, demonstrating its research capabilities in the fields of diffusion models, language modeling, and Grokking.1. Diffusion Model: "DualScale Diffusion: Adaptive Feature Balance for Low-Dimensional Generative Models"

Paper Address:

Code Address:

2. Language Modeling: "StyleFusion: Adaptive Multi-Style Generation in Character-Level Language Models"

Paper Address:

Code Address:

Language Modeling: "Adaptive Learning Rate for Transformers through Q-Learning"

 Currently, AI scientists do not possess visual processing capabilities, hence they are unable to automatically correct visual elements or chart layout issues in papers.

For instance, the charts it generates sometimes lack clarity, tables may exceed page boundaries, and the overall page layout often appears messy. Introducing multimodal foundation models is expected to fundamentally solve this problem.

Additionally, when executing ideas or performing baseline comparisons, AI scientists may lead to misleading results due to improper operations.

At the same time, when writing and evaluating results, it may occasionally make more serious mistakes, such as finding it difficult to accurately compare the magnitude of two numbers, which is a known flaw of large models. To alleviate this issue, the team has ensured that all experimental results are reproducible and has properly saved all execution files.

In the report, the team deeply analyzes the current limitations of AI scientists and the challenges they may face in the future.

04.

AI scientists "being clever": Modifying scripts on their own, causing AI security risks

The team also observed that AI scientists sometimes try to increase their chances of success through some "clever" methods, such as modifying and executing scripts on their own. In the paper, the team deeply explores the AI security risks that this behavior may bring.

For example, during one execution process, it actually edited the code and made itself run in an infinite loop through system calls.Here is the English translation of the provided text:

On another occasion, a certain experiment took too long and was on the verge of exceeding the team's set timeout limit. Instead of thinking about optimizing the code to improve efficiency, it tried to extend the timeout by modifying the code.

These issues can be mitigated by sandboxing the operating environment of AI scientists. In the full report, the team discusses in-depth the issues of secure code execution and sandboxing.

05.

Conclusion: The Innovative Ability of AI Scientists' Debut Needs to Be Verified

Looking to the future, Sakana AI states that its goal is to apply AI scientists to closed-loop systems of open models, promoting continuous self-improvement of AI. AI scientists will bring about a new world of science fully driven by AI, which not only includes researchers empowered by large language models but also covers reviewers, domain chairs, and even the entire academic conference system.

However, Sakana AI does not believe that the status of human scientists will be weakened as a result. On the contrary, with the emergence of new technologies, the role of scientists will become more diversified, and they will move towards a higher level in the field of scientific research. Automating the scientific research discovery process and integrating it with AI-driven review mechanisms mainly paves a broad road for innovation and solving the most challenging problems in the field of science and technology.

The current version of AI scientists has shown extraordinary strength in innovation based on mature technologies such as diffusion models and Transformers, but whether such systems can truly propose subversive new concepts still needs time to verify.

Note: The original text contains some inconsistencies and unclear references (e.g., "Sakana AI"). I translated it as faithfully as possible based on the provided context.