ScienceClaw - A research-native agent platform launched by Zhongke Zidong Taichu

What is ScienceClaw

ScienceClaw is a research-native agent platform launched by Zhongke Zidong Taichu (incubated by the Institute of Automation, Chinese Academy of Sciences). Built on the Zidong Taichu multimodal large model, it covers fields such as life sciences, materials, chemistry, physics, and astronomy. The newly upgraded AutoProject engine pushes AI from executing individual research tasks to autonomously undertaking complete research projects. It can autonomously complete project planning, task decomposition, long-horizon execution, evidence verification, and fault-tolerant repair, ultimately accumulating reusable research assets such as papers, data, models, and code, and supporting human-machine collaborative intervention.

Main features of ScienceClaw

• Project-level planning (Project2Task): Input a vague research idea and automatically decompose it into a structured task network and plan the research path.
• Long-horizon autonomous execution (TaskExecutor): Advances projects on-site over the long term, autonomously runs experiments, handles exceptions, and dynamically iterates without the need for high-frequency manual intervention.
• Evidence-driven verification (EviGraph): Performs full-chain evidence verification for every conclusion, and automatically traces back and repairs deviations when they are found.
• Literature search and study: Leverages a literature library of over 100 million papers across all disciplines to autonomously complete literature research and extract key findings.
• Data analysis and code execution: Uniformly schedules professional tools such as code, data analysis, and simulation to complete research computing tasks.
• Accumulation of research assets: Automatically outputs papers, datasets, models, code, and a full set of experimental records that are reusable and traceable.
• Human-machine collaboration: Researchers can intervene at any node to evaluate and correct the research path and intermediate results.
• Multi-platform access: Supports invocation via the official website, WeChat, and Feishu plugins.

Technical principles of ScienceClaw

• Multimodal large model foundation: Uses the Zidong Taichu 4.0 multimodal large model as the "brain," bringing text, images, and speech into a unified semantic space. It has multimodal reasoning, autonomous planning, and feedback correction capabilities, enabling the system to uniformly understand formulas, charts, and experimental curves in papers as well as heterogeneous research information such as code, data, and simulations.
• Hierarchical autonomous multi-agent architecture: Unlike the traditional "master-slave Agent" model, the system automatically generates a task graph around research goals and dynamically schedules specialized Agents for disciplines, code, search, data analysis, simulation, and more according to domain, tools, and execution status, forming an agent cluster capable of dynamic networking, collaborative execution, and feedback-driven replanning.
• Project2Task project planning mechanism: Conducts global modeling by integrating research goals, literature evidence, resource constraints, and task dependencies. It selects the optimal decomposition strategy through signal/noise analysis, supports four project topologies—horizontal, vertical, horizontal-then-vertical, and vertical-then-horizontal—and plans serial/parallel task relationships and cross-task asset reuse.
• TaskExecutor long-horizon execution mechanism: Breaks the linear "one call, one return" model and builds a goal-driven autonomous research loop. It globally monitors project status and continuously captures experimental feedback. When a task fails or results are abnormal, it automatically traces back for evaluation, reconstructs the task network, and reruns experiments, adapting to the nonlinear iterative nature of research.
• EviGraph evidence graph system: Organize the research process into a dynamic evidence chain of "research question → research gap → hypothesis → experiment → finding → conclusion," operating in three stages: initial graph construction, inspection and iterative repair, and manuscript generation; continuously verify cross-task logical consistency, and when deviations are detected, locate the root cause along the evidence chain and autonomously repair it, with short-term/long-term graph repositories enabling version rollback and experience accumulation.

How to use ScienceClaw

• Visit the official website: Visit the official ScienceClaw website https://scienceclaw.zidongtaichu.com/，完成账号注册登录。
• Enter a research goal: No need to break down steps or write detailed prompts; simply submit a high-level research idea or project goal.
• Confirm the project plan: The system automatically generates a task breakdown and research path, supporting multiple decomposition methods such as horizontal/vertical, which can be manually confirmed or adjusted.
• Wait for autonomous execution: The AI stays on the job long-term and autonomously completes literature research, experiment execution, data analysis, and anomaly repair, with no need for frequent intervention throughout.
• Intervene and correct course at any time: During project progress, you can view intermediate results at any time and evaluate and revise the research path, hypotheses, and data.
• Obtain research outputs: After the project is completed, obtain reusable research assets such as papers, datasets, models, code, and a full set of experimental records.
• Multi-platform access: In addition to the web interface, research tools can also be accessed anytime and anywhere through WeChat and Feishu plugins.
• Consult the user guide: The official website provides detailed user documentation to further understand advanced features and operational details.

ScienceClaw's core advantages

• Project-level autonomous research: The industry's first to achieve the leap from "doing tasks" to "doing projects," capable of undertaking complete research topics rather than isolated tasks.
• Full-chain closed-loop capability: Covers the six major stages of planning, decomposition, execution, verification, repair, and accumulation, running an entire project in one chain.
• Conclusions are credible and traceable: The self-developed EviGraph evidence graph ensures every conclusion is supported by experiments and data, increasing the traceable evidence rate by 40.74% over the best baseline.
• Long-horizon autonomous execution: Breaks through the single-call model, can stay on the job long-term, autonomously iterate through trial and error, and sustain continuous exploration over multi-month cycles.
• Autonomous fault-tolerant repair: When experiments fail or data is abnormal, automatically traces the root cause, reconstructs the task network, and reruns, without manual intervention.
• Accumulation of research assets: A single study produces a full set of assets including papers, data, models, code, and experimental records, reusable and iterable, ending the practice of "valuing papers over assets."
• Multimodal technological foundation: Built on the Zidong Taichu multimodal large model, it can uniformly understand heterogeneous research information such as formulas, charts, code, and data.
• Layered multi-agent collaboration: Dynamically schedules specialized Agents in disciplines, code, search, simulation, and more to form networks for collaboration, rather than a single Agent working alone.
• Full-discipline coverage: Already covers fields such as life sciences, materials, chemistry, physics, and astronomy, backed by a literature repository of hundreds of millions of papers across all disciplines.

Comparison of ScienceClaw with similar competitors

Application scenarios of ScienceClaw

• Full-process research on a topic: Starting from a vague idea, autonomously complete the entire project from planning and experiments to paper delivery.
• Literature review and survey: Relying on a literature database of over 100 million papers, complete literature retrieval and review organization that originally took hours in just minutes.
• AI modeling and experiments: Tasks such as strip steel surface defect recognition and YOLO modeling, including data processing, training, and iterative optimization.
• Life sciences and drug development: Supports protein research, vaccine development, and drug screening, shortening drug property prediction from one week to 10 minutes.
• Materials and chemistry research: Scenarios combining dry and wet experiments such as compound simulation, materials metallurgy, and synthesis route planning.