Exploring SakanaAI/AI-Scientist-v2: An AI Agent System for Workshop-Level Automated Scientific Discovery

Project Overview

As artificial intelligence accelerates its penetration into various industries, fully autonomous scientific research systems are becoming increasingly powerful. AI-Scientist-v2 (Project URL: https://github.com/SakanaAI/AI-Scientist-v2), open-sourced by SakanaAI, is a cutting-edge exploration of this trend. As an upgraded version of the first-generation AI Scientist, this project introduces the "Agentic Tree Search" mechanism, aiming to achieve Workshop-level automated scientific discovery.

The core reason this project has recently garnered widespread attention lies in its milestone achievement: the system generated the first fully AI-written Workshop paper accepted through peer review (related experiments point to the ICLR 2025 Workshop). This not only demonstrates the potential of Large Language Models (LLMs) in complex logical reasoning and long-text generation but also provides a concrete reference implementation for the future AI for Science paradigm. For developers focusing on AI agents, automated workflows, and cutting-edge academic exploration, this project offers a highly valuable open-source template.

Core Capabilities and Applicable Boundaries

Core Capabilities:

1. End-to-End Automated Research: The system can autonomously complete the entire closed loop from "generating research hypotheses," "writing code to run experiments," and "collecting and analyzing data" to "writing scientific manuscripts that meet academic standards."
2. Agentic Tree Search: Compared to traditional linear prompt engineering, this system employs a tree search strategy to explore research directions. It can evaluate and select the optimal path among multiple hypothesis branches, significantly improving the novelty and success rate of the research.
3. Multi-Model Support: According to the project documentation, the system supports configuring various mainstream large language models and their API Keys to drive the underlying agent decision engine.

Applicable Boundaries:

• Recommended Users: Academic researchers exploring the boundaries of AI automated research, Large Language Model (LLM) application developers, and computer science teams hoping to use AI to generate preliminary research baselines.
• Not Recommended For: Life science/medical research teams requiring strict physical experiments or human ethical reviews; individual developers extremely sensitive to computing power and API Token costs; users expecting the system to directly produce Top-tier Conference Oral-level breakthrough results.

Insights and Inferences

Based on existing data and project documentation, the following inferences can be drawn:

First, the architectural evolution from "linear execution" to "Agentic Tree Search" indicates that the bottleneck of automated research is shifting from "text generation" to "direction exploration and error correction." The tree search mechanism allows AI to backtrack and try new hypotheses when experiments fail, which is closer to the real research process of human scientists, foreshadowing the inevitable evolutionary direction of Agent architectures in complex tasks.

Second, the system's successful production of a paper accepted by the ICLR 2025 Workshop means that AI-generated academic content has crossed the threshold of "correct formatting" and reached a "Workshop level" with a certain degree of logical coherence and preliminary innovation. However, this also implies that the results currently generated by the system may still be in the micro-innovation or experimental validation stage, and there is still a long way to go before solving major open scientific problems.

Finally, the project currently has 50 Open Issues and has had no code pushes since December 2025, which may indicate that the project's stability in complex environments still needs to be polished by the community. Meanwhile, the lack of a clear open-source license (NOASSERTION) may hinder deep integration and commercial application by enterprise-level users.

30-Minute Getting Started Guide

Although executing the complete scientific research process takes a long time, developers can complete the environment setup and start basic testing within 30 minutes:

1. Clone the Project Repository: Use Git to clone the project to your local environment. git clone https://github.com/SakanaAI/AI-Scientist-v2.git cd AI-Scientist-v2

2. Configure the Python Environment: It is recommended to use Conda or venv to create an independent Python virtual environment and install dependencies. pip install -r requirements.txt

3. Configure API Keys: Following the Supported Models and API Keys guide in the documentation, set the API keys for the required LLMs in environment variables or configuration files. For example: export OPENAI_API_KEY="your_api_key_here"

4. Run the Basic Demo: Check the ICLR 2025 Workshop experiment template provided by the project, and run a lightweight hypothesis generation or data analysis script to verify system connectivity and basic agent decision-making capabilities.

Risks and Limitations

When actually deploying and using AI-Scientist-v2, special attention should be paid to the following risks:

• Data Privacy and Compliance Risks: During operation, the system needs to send experimental data, code, and preliminary ideas to third-party LLM APIs. For research involving Non-Disclosure Agreements (NDAs) or undisclosed core commercial secrets, there is a serious risk of data leakage.
• Academic Ethics and Policy Restrictions: Papers written entirely by AI may face strict scrutiny when submitted to academic conferences or journals. Currently, the policies of major top conferences regarding AI as an "author" or primary contributor are still dynamically adjusting, posing compliance risks of rejection or retraction.
• High Operating Costs: Agentic Tree Search is inherently a compute-intensive strategy, making a large number of LLM calls across multiple branches. This can lead to an exponential increase in API Token consumption, bringing non-negligible financial costs.
• Maintenance and Stability Risks: The project currently has no declared open-source license (NOASSERTION) and has accumulated dozens of unresolved Issues. Users may need to troubleshoot and fix underlying bugs themselves, lacking official real-time technical support guarantees.

Evidence Sources

• GitHub API Repository Data: https://api.github.com/repos/SakanaAI/AI-Scientist-v2 (Fetch time: 2026-03-28)
• GitHub API Release Data: https://api.github.com/repos/SakanaAI/AI-Scientist-v2/releases/latest (Fetch time: 2026-03-28)
• Project README File: https://github.com/SakanaAI/AI-Scientist-v2/blob/main/README.md (Fetch time: 2026-03-28)
• Project Homepage: https://github.com/SakanaAI/AI-Scientist-v2 (Fetch time: 2026-03-28)