ScienceClaw - A research-native intelligent agent platform launched by CAS Zidong Taichu - AiBoss

What is ScienceClaw?

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

Main functions of ScienceClaw

Project-level planning (Project2Task)Input a fuzzy research idea, and it will automatically break it down into a structured task network and plan a research path.
Long-term autonomous execution (TaskExecutor): Long-term on-site project advancement, autonomously conducting experiments, handling anomalies, and dynamically iterating, without the need for frequent human intervention.
Evidence-driven verification (EviGraph)Each conclusion undergoes full-chain evidence verification, and any deviations are automatically traced back and corrected.
Literature retrieval and studyRelying on a database of hundreds of millions of documents from all disciplines, we independently completed literature surveys and extracted key findings.
Data Analysis and Code ExecutionIt unifies the use of professional tools such as scheduling code, data analysis, and simulation to complete scientific research computing tasks.
Scientific research asset accumulationAutomatically outputs papers, datasets, models, code, and a complete set of experimental records, which are reusable and traceable.
Human-machine collaborationResearchers can intervene at any point to evaluate and correct the research path and intermediate results.
Multi-terminal accessSupports calls from the official website, WeChat, and Lark plugins.

The technical principles of ScienceClaw

Multimodal large model baseUsing the Zidong Taichu 4.0 multimodal large model as the "brain", it incorporates text, images, and speech into a unified semantic space, and has the ability of multimodal reasoning, autonomous planning and feedback correction, enabling the system to uniformly understand heterogeneous scientific research information such as formulas, charts, experimental curves, code, data, and simulations in papers.
Hierarchical Autonomous Multi-Agent ArchitectureUnlike the traditional "master-slave agent" model, the system automatically generates a task graph around the research goals and dynamically schedules professional agents such as disciplines, code, search, data analysis, and simulation according to the domain, tools, and execution status, forming an intelligent agent cluster that can be dynamically networked, collaboratively executed, and replanned based on feedback.
Project2Task project planning mechanismThe system integrates research objectives, literature evidence, resource constraints, and task dependencies to perform global modeling. It selects the optimal decomposition strategy through signal/noise analysis, supports four project topologies: horizontal, vertical, horizontal-then-vertical, and vertical-then-horizontal. It also plans the parallel relationships of task strings and the reuse of cross-task assets.
TaskExecutor long-term execution mechanismBreaking away from the linear "one call, one return" model, it constructs a goal-driven autonomous scientific research cycle, globally monitors project status, continuously captures experimental feedback, and automatically backtracks and evaluates, reconstructs the task network, and reruns the experiment when a task fails or results are abnormal, thus adapting to the non-linear iterative characteristics of scientific research.
EviGraph Evidence Graph SystemThe research process is organized as a dynamic evidence chain of "research question → research gap → hypothesis → experiment → discovery → conclusion", and runs in three stages: initial map construction, inspection and iterative repair, and manuscript generation. It continuously verifies the consistency of cross-task logic, locates the root cause along the evidence chain and repairs it autonomously when deviations are detected, and realizes version rollback and experience accumulation in conjunction with short-term/long-term map libraries.

How to use ScienceClaw

Visit the official websiteVisit the ScienceClaw website https://scienceclaw.zidongtaichu.com/ and complete account registration and login.
Input research goalsThere's no need to break down the steps or write a detailed prompt; simply submit a macro-level research concept or project objective.
Confirm Project PlanningThe system automatically generates task breakdown and research paths, supporting multiple breakdown methods such as horizontal and vertical, which can be manually confirmed or adjusted.
Waiting for autonomous executionAI can autonomously complete literature research, experiment execution, data analysis, and anomaly repair on-site for extended periods without requiring frequent intervention.
Intervene and correct course at any timeIntermediate results can be viewed at any time during the project's progress, allowing for evaluation and correction of research paths, hypotheses, and data.
Obtaining scientific research resultsUpon completion of the project, you will receive reusable research assets such as papers, datasets, models, code, and a complete set of experimental records.
Multi-terminal callIn addition to the web version, research tools can also be accessed anytime, anywhere via WeChat and Lark plugins.
Consult the user guideThe official website provides detailed user documentation, which can be used to learn more about advanced features and operational details.

ScienceClaw's core advantages

Project-level independent scientific researchIt is the first in the industry to achieve the leap from "doing tasks" to "doing projects", and can undertake complete scientific research projects rather than isolated tasks.
End-to-end closed-loop capabilityIt covers six major stages: planning, decomposition, execution, verification, repair, and accumulation, and runs the entire project through a single link.
The conclusions are credible and traceable.The self-developed EviGraph evidence graph ensures that every conclusion is supported by experiments and data, and the traceable evidence rate is 40.74% higher than the best baseline.
Long-term autonomous executionBreaking away from the single-call mode, it can remain on-site for extended periods, autonomously experiment and iterate, and withstand continuous exploration over several months.
Autonomous fault-tolerant repairWhen an experiment fails or data is abnormal, it automatically traces back to the root cause, reconstructs the task network, and reruns the experiment without manual intervention.
Scientific research asset accumulationA single research output yields a complete set of assets, including papers, data, models, code, and experimental records, which are reusable and iterative, ending the trend of "emphasizing papers but neglecting assets."
Multimodal technology genesBased on the Zidong Taichu multimodal large model, it can unify the understanding of heterogeneous scientific research information such as formulas, charts, codes, and data.
Hierarchical multi-agent collaborationIt dynamically schedules professional agents such as disciplines, code, search, and simulation to form a network for collaboration, rather than a single agent working alone.
Full subject coverageIt covers fields such as life sciences, materials, chemistry, physics, and astronomy, and is backed by a database of hundreds of millions of documents from all disciplines.

ScienceClaw's Competitive Product Comparison

Comparison Dimensions	ScienceClaw (中科紫东太初)	AI Co-Scientist (Google)
Producer	Zhongke Zidong Taichu, incubated by the Institute of Automation, Chinese Academy of Sciences/Wuhan Institute of Artificial Intelligence	Google (in conjunction with Stanford and other institutions)
Model base	Zidong Taichu 4.0 Multimodal Large Model: Natively Unified Understanding of Text, Images, and Speech	Gemini 2.0 primarily focuses on text-based reasoning.
Core positioning	A project-level autonomous research system capable of independently managing complete research projects.	Virtual research collaborators focus on hypothesis generation and experimental design.
work unit	Project Level: Planning → Decomposition → Execution → Validation → Remediation → Consolidation of the Entire Process	Centered on hypothesis generation: an iterative cycle of generation → reflection → ranking → evolution.
Experiment Execution	The enterprise version can autonomously run experiments, process data, and train models. It also supports embodied execution and wet/dry experiment fusion by connecting to devices such as robotic arms.	Instead of directly conducting experiments, the hypotheses and experimental designs are output for human verification.
Multi-agent architecture	Hierarchical autonomy, dynamic networking, and scheduling of specialized agents such as discipline/code/search/simulation agents.	Six types of specialized intelligent agents—generation, reflection, ranking, evolution, proximity, and meta-censorship—are coordinated by a supervisory intelligent agent.
Conclusion verification mechanism	EviGraph's evidence graph is verifiable, traceable, and repairable across the entire chain.	Tournament Elo scoring plus self-debate optimizes hypothesis quality, primarily relying on literature evidence.
Deliverables	A complete matrix of scientific research assets including papers, datasets, models, code, and experimental records.	Research hypotheses, research review, and experimental protocol text
Subject coverage	Eight major disciplines including life sciences, materials science, chemistry, physics, and astronomy	Currently, the validation is primarily being applied in biomedical settings.
Literature Resources	Self-built database of hundreds of millions of documents across all disciplines	Primarily relies on open access literature and web search

Application Scenarios of ScienceClaw

Full-process research of the projectStarting from a vague concept, the entire project was completed independently, from planning and experimentation to the delivery of a thesis.
Literature reviewBased on a database of hundreds of millions of documents, it can complete the literature search and review process that would normally take hours in just a few minutes.
AI Modeling and ExperimentData processing, training, and iterative optimization for tasks such as strip surface defect identification and YOLO modeling.
Life Sciences and Drug DevelopmentIt supports protein research, vaccine development, and drug screening, reducing drug property prediction time from one week to 10 minutes.
Materials and Chemical ResearchApplications include dry and wet experiments such as compound simulation, materials metallurgy, and synthetic route planning.