DeepScientist – A Fully Automated AI Scientist System Launched by Westlake University

What is DeepScientist？

DeepScientist is an autonomous research system developed by Westlake University and other institutions, enabling full-process automation from hypothesis generation to experimental validation, result analysis, and paper writing. Through a multi-agent architecture and reinforcement learning strategies, the AI can continuously explore, verify, and advance scientific frontiers without human intervention. The system is built around a closed-loop “Hypothesize–Verify–Analyze” process, combined with Bayesian optimization, treating scientific discovery as a problem of finding optimal solutions in a vast methodological space. DeepScientist also features a Findings Memory, which records and reuses historical experimental results to enhance research efficiency and innovation.

Key Features of DeepScientist

1. Automated Hypothesis Generation: The system can autonomously propose new research hypotheses or improvements based on existing knowledge and experimental results, enabling automated innovation.

2. Experiment Design and Execution: It converts hypotheses into executable experimental workflows or code, automatically builds experimental environments, runs models, collects data, and validates results.

3. Result Analysis and Report Generation: After experiments, the system automatically analyzes results, summarizes patterns, and generates academic reports or draft papers, supporting direct research output.

4. Findings Memory Management: Records all experiments, hypotheses, and results to form a reusable knowledge base, assisting future research decisions.

5. Autonomous Optimization and Learning: Uses Bayesian optimization and other methods to balance exploration and exploitation in research, continuously improving efficiency and outcomes.

6. Multi-Agent Collaboration: Multiple specialized AI agents (for hypothesis generation, code execution, result analysis, etc.) work collaboratively to create a safe, controllable, and modular research workflow.

7. Safety and Verification Mechanisms: Uses sandboxed and containerized execution strategies to ensure experiment safety, result reliability, and automatic verification of conclusions.

Technical Principles of DeepScientist

• Scientific Discovery as an Optimization Problem: Treats research innovation as finding optimal solutions in a vast methodological space, using Bayesian optimization and surrogate models to efficiently evaluate and filter hypotheses.

• Closed-Loop Research Process: A “Hypothesize–Verify–Analyze” three-stage cycle continuously generates, tests, and refines scientific hypotheses, forming a self-driven research iteration system.

• Multi-Agent Architecture: Composed of specialized agents responsible for strategy planning, code implementation, result analysis, and report generation, collaboratively completing the full research workflow.

• Findings Memory System: Maintains a long-term knowledge repository storing past experiments, hypotheses, and results to guide new research exploration and optimization.

• Hierarchical Verification and Surrogate Evaluation: Uses a two-tier validation strategy (low-fidelity and high-fidelity) where surrogate models first assess potential before running computationally intensive real experiments, saving resources.

• Containerized and Sandboxed Execution Environment: Runs code and experiments in isolated, secure environments to prevent conflicts and errors, ensuring reproducibility and trustworthy results.

• Automated Result Re-Validation: The system independently re-executes verification after experiments to prevent false positives and ensure the reliability and verifiability of scientific conclusions.

Project Links

• Official Website: https://ai-researcher.net

• GitHub Repository: https://github.com/ResearAI/DeepScientist

• arXiv Technical Paper: https://arxiv.org/pdf/2509.26603

Application Scenarios of DeepScientist

1. AI Algorithm Research: Can autonomously explore model architectures, optimization strategies, and training methods, advancing AI reasoning efficiency, interpretability, and robustness.

2. Automated Scientific Innovation: Automatically generates and validates new hypotheses in fields like machine learning, computer vision, and natural language processing, accelerating research iteration.

3. Experimental Science Assistance: Applicable to disciplines like physics, chemistry, and biology that require extensive experimental validation, automatically screening potential discoveries through virtual experiments and data analysis.

4. Agent System Optimization: Improves strategies and communication mechanisms in multi-agent collaboration or reinforcement learning tasks, optimizing system performance.

5. Research Workflow Automation: Helps research teams automate the entire process from conception to report generation, enhancing efficiency and output quality.

6. Academic Paper Generation and Peer Review Simulation: Automatically drafts papers based on experimental results and performs self-checks and quality evaluation using AI review modules.