Multi-agent scientific research task execution method and system

Abstract

This invention discloses a multi-agent scientific research task execution method and system, relating to the fields of artificial intelligence and automation scientific research technology, to address the problems of task context loss, deviation from the original goal, and lack of scientific basis that easily occur in existing methods during scientific research task execution. The method includes: parsing the user's scientific research goal instructions to generate a sequence of task nodes; before executing any task node, performing retrieval and data matching in a three-dimensional data base, trimming the matched data to obtain a task context data package; generating system prompts, executing the task instructions of the current task node based on the systematized prompts, and obtaining the output result; performing multiple verifications on the output result and providing feedback. The multi-agent scientific research task execution method provided by this invention can avoid context loss and deviation from the original goal, and ensures the scientific validity and reliability of the scientific research task execution process through static knowledge graphs and dynamic memory graphs.

Description

Technical Field

[0001] This invention relates to the fields of artificial intelligence and automation research technology, and in particular to a method and system for executing multi-agent research tasks. Background Technology

[0002] With the rapid development of artificial intelligence, large language models, and intelligent agent technology, scientific research tasks have gradually incorporated related technologies to assist in execution. Scientific research task execution schemes generally adopt a master-slave scheduling and division-of-labor execution model, where a master intelligent agent is responsible for task breakdown and allocation, specialized sub-intelligent agents execute corresponding stages, and results are summarized and output through simple information exchange. However, in long-chain tasks, information transmission decays over time, leading to loss of task context, and intelligent agents are prone to deviating from the original goal when executing complex tasks, ultimately resulting in significant errors between the output code or text and the target.

[0003] Therefore, there is an urgent need for a multi-agent scientific research task execution method and system. Summary of the Invention

[0004] The purpose of this invention is to provide a method and system for executing multi-agent scientific research tasks, which solves the problems of task context loss, deviation from the original goal, and lack of scientific basis that are easily encountered in the execution of scientific research tasks by existing methods.

[0005] To achieve the above objectives, the present invention provides the following technical solution: In a first aspect, the present invention provides a method for executing multi-agent scientific research tasks, comprising: Obtain user research goal instructions and user cognitive profile; The user's research objective instructions are parsed to generate a sequence of task nodes; the task nodes in the sequence of task nodes include at least task definition information. Before executing any task node, a search is performed in the three-dimensional data base according to the task definition information to obtain matching data that meets the matching conditions, and the matching data is trimmed to obtain a task context data package; the three-dimensional data base includes a hierarchical heterogeneous data warehouse, a static knowledge graph and a dynamic memory graph. The system prompt word template is filled with the user's research goal instruction, the task definition information, the task context data package, and the user's cognitive profile to generate system prompt words. Based on the system prompt words, the task instruction of the current task node is executed to obtain the output result. The system prompt words include role definition, global project goal, knowledge constraints, and current task instruction. The output results are subjected to multiple verifications, and feedback processing is performed based on the verification results; the multiple verifications include at least formal compliance verification and logical consistency verification.

[0006] Optionally, the static knowledge graph includes theoretical concept nodes and standard implementation nodes; the dynamic memory graph includes operation trajectory nodes and result feedback nodes; the step of retrieving matching data that meets the matching conditions in the three-dimensional data base according to the task definition information, and then trimming the matching data to obtain the task context data package includes: By embedding the model, the task definition information is transformed into a numerical vector; Retrieve nodes from the static knowledge graph and calculate the first cosine similarity between the first semantic vector and the numerical vector; the first semantic vector is the semantic vector in the node attributes of the theoretical concept node and the standard implementation node. Retrieve nodes from the dynamic memory graph and calculate the second cosine similarity between the second semantic vector and the numerical vector; the second semantic vector is the semantic vector in the node attributes of the operation trajectory node and the result feedback node; Retrieve the metadata layer in the hierarchical heterogeneous data warehouse and calculate the third cosine similarity between the metadata header and the numerical vector; Data with cosine similarity greater than a first preset threshold are determined as matching vectors; the matching vectors include nodes in the static knowledge graph, nodes in the dynamic memory graph, and one or more in the metadata header; The matching vectors are sorted in descending order of cosine similarity to obtain a sorted sequence; The sorted sequence is truncated according to a preset length to obtain a task context data packet.

[0007] Optionally, the hierarchical heterogeneous data warehouse also stores scientific research data exchange standards; formal compliance verification of the output results includes: According to the task type of the task node, the scientific research data exchange standard is read from the hierarchical heterogeneous data warehouse; The output is deserialized using a JSON parser. If parsing is successful, the parsed result is obtained; if parsing fails, it is determined to be a format error. Check whether the parsing result contains all the necessary fields defined in the scientific research data exchange standard and whether the data type matches the preset data type in the scientific research data exchange standard.

[0008] Optionally, performing logical consistency verification on the output results includes: Extract theoretical benchmarks related to the current task node from the static knowledge graph; Verify the consistency between the logic of the code that outputs the results and the logic in the theoretical benchmark; The logical consistency of the operators and formulas in the code is verified using a thought process.

[0009] Optionally, the step of executing the task of the current task node according to the system prompt and obtaining the output results includes: The system prompts are input into the large language model to start the reasoning and action engine. Through closed-loop reasoning involving observation, thinking, and action, the output results are obtained. The output results include code or text artifacts.

[0010] Optionally, obtaining a user cognitive profile includes: Select a user profile template based on the user's research goals and areas of expertise; Based on the user profile template, the user's scientific research level characteristics and technical preference characteristics are extracted from the historical interaction data in the dynamic memory graph to generate a user cognitive profile. The number of new nodes in the dynamic memory graph is monitored, and the user cognitive profile is updated when the number of new nodes reaches a preset threshold.

[0011] Optionally, the task node sequence is generated based on a directed acyclic graph; feedback processing based on the verification results includes: When either formal compliance verification or logical consistency verification fails, a structured verification report is generated; the structured verification report includes the error type and correction suggestions. If the error type is a soft error, update the task context data packet and re-execute the task of the current task node based on the updated task context data packet; If the error type is a hard error, then the directed acyclic graph is corrected or a node is inserted.

[0012] Optionally, before sorting the matching vectors according to their cosine similarity from largest to smallest to obtain a sorted sequence, the method further includes: Determine whether the third cosine similarity is greater than the second preset threshold. When the third cosine similarity is greater than the second preset threshold, determine the data volume corresponding to the third cosine similarity as the matching vector; the second preset threshold is greater than the first preset threshold.

[0013] Compared with existing technologies, the present invention provides a multi-agent scientific research task execution method, which includes acquiring user scientific research target instructions and user cognitive profiles; parsing user scientific research target instructions to generate a task node sequence; before executing any task node, searching in a three-dimensional data base according to task definition information to obtain matching data that meets the matching conditions, and trimming the matching data to obtain a task context data packet; filling the user scientific research target instructions, task definition information, task context data packet, and user cognitive profile into a system prompt word template to generate system prompt words; and executing the task instructions of the current task node based on the system prompt words to obtain the output results. This application encapsulates the task context according to the task information when executing each task node, which can avoid the situation where information decays and task context is lost in long-link tasks. Furthermore, the system prompt words generated by filling the system prompt word template with task definition information, task context data packet, and user cognitive profile include role definitions, global project goals, knowledge constraints, and current task instructions, which can ensure that each task execution does not deviate from the global project goals, thereby improving the accuracy of the task output results. Furthermore, performing formal compliance and logical consistency verification on the output results after each task execution can avoid the problem of research results lacking theoretical basis. In addition, the three-dimensional data foundation includes a static knowledge graph storing scientific truths and a dynamic memory knowledge graph storing interactive memories. Therefore, the task context information generated based on this has theoretical support and historical data support from users, thereby ensuring the scientific rigor and reliability of the research task execution process.

[0014] Secondly, the present invention provides a multi-agent scientific research task execution system, applied to the aforementioned multi-agent scientific research task execution method, the system comprising at least: The main control agent is used to receive user research goal instructions and user cognitive profiles, and send the user research goal instructions to the dynamic orchestration engine; A dynamic orchestration engine is used to parse user research objective instructions and generate a sequence of task nodes; the task nodes in the sequence of task nodes include at least task definition information. The semantic trimming and routing module is used to retrieve matching data that meets the matching conditions in the three-dimensional data base according to the task definition information before executing any task node, and trim the matching data to obtain a task context data package; the three-dimensional data base includes a hierarchical heterogeneous data warehouse, a static knowledge graph and a dynamic memory graph. The multi-agent collaborative execution module is used to fill the system prompt word template with the research goal instruction of the user, the task definition information, the task context data packet, and the user cognitive profile, generate system prompt words, execute the task instruction of the current task node based on the system prompt words, and obtain the output result; the system prompt words include role definition, global project goal, knowledge constraints, and current task instruction. The closed-loop verification module is used to perform multiple verifications on the output results and to perform feedback processing based on the verification results; the multiple verifications include at least formal compliance verification and logical consistency verification.

[0015] Optionally, the system may also include: The state manager drives the system to transition between signal resolution state, pipelined dispatch state, asynchronous blocking state, and topology evolution state. A hierarchical heterogeneous data warehouse module is used to encapsulate the output of each task node into a modular storage unit; the modular storage unit is physically divided into a metadata header and a data body; the metadata header is used to store index information; the index information includes a natural language summary, semantic vector, and resource list extracted from the output based on a summary generation template; the data body is used to store the output. The static knowledge graph construction module is used to construct theoretical concept nodes and standard implementation nodes, and to determine the feature overlap between the theoretical concept nodes and the standard implementation nodes through alignment analysis. Strong constraint edges are established between theoretical concept nodes and standard implementation nodes whose feature overlap exceeds the overlap threshold to form a static knowledge graph. The dynamic memory graph generation module is used to generate a dynamic memory graph consisting of operation trajectory nodes and result feedback nodes based on the execution and feedback of task nodes.

[0016] Compared with the prior art, the beneficial effects of the multi-agent scientific research task execution system provided by the present invention are the same as the beneficial effects of the multi-agent scientific research task execution method described in the above technical solution, and will not be repeated here. Attached Figure Description

[0017] The accompanying drawings, which are included to provide a further understanding of the invention and form part of this invention, illustrate exemplary embodiments of the invention and are used to explain the invention, but do not constitute an undue limitation of the invention. In the drawings: Figure 1 A flowchart of a multi-agent scientific research task execution method provided by the present invention; Figure 2 This invention provides a system block diagram for executing multi-agent scientific research tasks. Detailed Implementation

[0018] To facilitate a clear description of the technical solutions in the embodiments of the present invention, the terms "first" and "second" are used to distinguish identical or similar items with essentially the same function and effect. For example, the first threshold and the second threshold are merely used to distinguish different thresholds and do not limit their order. Those skilled in the art will understand that the terms "first" and "second" do not limit the quantity or execution order, and that the terms "first" and "second" are not necessarily different.

[0019] It should be noted that in this invention, the terms "exemplary" or "for example" are used to indicate examples, illustrations, or descriptions. Any embodiment or design described as "exemplary" or "for example" in this invention should not be construed as being more preferred or advantageous than other embodiments or designs. Specifically, the use of terms such as "exemplary" or "for example" is intended to present the relevant concepts in a concrete manner.

[0020] In this invention, "at least one" refers to one or more, and "more than one" refers to two or more. "And / or" describes the relationship between related objects, indicating that three relationships can exist. For example, A and / or B can represent: A alone, A and B simultaneously, or B alone, where A and B can be singular or plural. The character " / " generally indicates that the preceding and following related objects are in an "or" relationship. "At least one of the following" or similar expressions refer to any combination of these items, including any combination of single or plural items. For example, at least one of a, b, or c can represent: a, b, c, a combination of a and b, a combination of a and c, a combination of b and c, or a, b, and c, where a, b, and c can be single or multiple.

[0021] Existing scientific research task execution methods suffer from the following shortcomings: Loss of task context, leading to information attenuation over time in long-chain tasks. Logical divergence (illusion) in long-chain tasks: agents easily deviate from the original goal when executing complex steps. Lack of theoretical basis in generated code: generated scientific code often lacks supporting scientific principles or formulas. Lack of personalized service adaptation: the system cannot dynamically adjust interaction strategies and coding style based on the user's research level or technology stack preferences, such as PyTorch / TensorFlow.

[0022] To address the aforementioned problems, this invention provides a method and system for executing multi-agent scientific research tasks, which will be described below with reference to the accompanying drawings.

[0023] See Figure 1 The present invention provides a multi-agent scientific research task execution method comprising the following steps: Step 100: Obtain user research goal instructions and user cognitive profile.

[0024] User cognitive profiles are generated based on dynamic memory graphs and reflect a user's research level and technical preferences. Research level characteristics include error frequency and task complexity, while technical preferences include commonly used algorithm libraries and naming styles. User cognitive profiles are stored as independent structured objects.

[0025] Step 200: Parse the user's research objective instructions to generate a task node sequence; the task nodes in the task node sequence include at least task definition information.

[0026] Specifically, the master control agent parses the user's research objective instructions, constructs a task dependency model based on a directed acyclic graph (DAG), and sends this DAG task dependency model to the user for confirmation and locking via a user interface, generating an immutable task node sequence. Each task node in the task node sequence includes the following four data fields: Task definition information: This is used to record the natural language description of the task objective of this node, as well as specific constraints injected through the semantic trimming and routing module, i.e., the task context data packet. The data packet may include programming specifications, prohibited libraries, etc.

[0027] Execution result data: This is only used to record the ID of the specific artifact stored in the metadata header in the hierarchical data warehouse.

[0028] Status indicators: These record the real-time execution status of task nodes, such as pending, executing, under review, completed, suspended, and circuit breaker, for the runtime state manager to monitor progress. Pending indicates incomplete prerequisites and is in the blocking pool; executing indicates it has been distributed to a sub-agent and is currently being computed; under review indicates the sub-agent has submitted it and is running closed-loop verification; completed indicates verification passed, the artifact has been archived, and downstream nodes can unlock it; suspended indicates user intervention or insufficient system resources causing a pause; and circuit breaker indicates the retries have been exhausted and verification still failed, triggering graph reconstruction.

[0029] Verification Log: Used to record the structured verification report generated by the closed-loop verification module and targeted correction suggestions. The structured verification report includes pass or fail markers.

[0030] Step 300: Before executing any task node, search the three-dimensional data base according to the task definition information to obtain matching data that meets the matching conditions, and trim the matching data to obtain the task context data packet.

[0031] The three-dimensional data foundation comprises a hierarchical heterogeneous data warehouse, a static knowledge graph, and a dynamic memory graph. The hierarchical heterogeneous data warehouse employs a two-layer encapsulation structure, encapsulating the output of each task node into an independent modular storage unit. This modular storage unit is physically divided into a metadata header and a data body. The metadata header corresponds to a lightweight metadata layer, storing index information with a volume smaller than a storage volume threshold; this threshold can be set as needed. The index information includes a natural language summary, semantic vectors, and a resource list extracted from the output based on a summary generation template. The data structure corresponds to the full content layer, used to store the output results.

[0032] Static knowledge graphs are used to provide scientific theoretical support and constrain the accuracy of code generation. A static knowledge graph includes theoretical concept nodes, standard implementation nodes, and strong constraint edges between theoretical concept nodes and standard implementation nodes. Theoretical concept nodes store natural language descriptions of scientific principles and LaTeX representations of mathematical formulas. Standard implementation nodes store verified function-level code snippets or algorithm templates. The steps for establishing strong constraint edges are as follows: Alignment analysis of theory and code is performed using an offline ETL computing engine. Specifically, mathematical symbol features from theoretical concept nodes and variable naming features from standard implementation nodes are extracted, and feature overlap is calculated. Only when the overlap exceeds an overlap threshold is a strong constraint edge established between the two types of nodes, forming a static constraint network that guides theoretical time. The data sources for constructing the static knowledge graph are reviewed scientific literature and high-reputation open-source code repositories. The node content in the static knowledge graph remains read-only during system operation and does not change with the execution of user tasks; it is only updated periodically through synchronization with external knowledge bases.

[0033] Dynamic memory graphs are used to record system execution trajectories, enable experience backtracking, and build user knowledge profiles. The data source for constructing the dynamic memory graph is real-time logs and interaction data during task execution. The dynamic memory graph includes operation trajectory nodes, result feedback nodes, temporal edges connecting operation trajectory nodes, and causal edges connecting operation trajectory nodes and result feedback nodes. Operation trajectory nodes record the agent's decision-making process, tool call parameters, and environmental state snapshots; result feedback nodes record the execution results of each task, error stack information, and user-generated correction data. The dynamic memory graph generation steps are as follows: a runtime listener captures the task execution flow in real time; whenever a task node completes, a corresponding operation trajectory node is generated, and temporal edges are established according to the execution order; when errors occur or corrections are made, causal edges are established.

[0034] In practical applications, the transmission and reception process of the task context data packet is as follows: The master agent serializes the assembled task context data packet into a JSON-formatted task payload, which is then pushed to a specific channel in the Redis message queue. The "containerized runtime" of the child agents executing the task node maintains a blocking listener on the Redis channel. Once a new message is captured, the system immediately deserializes the JSON object and extracts the task context data packet.

[0035] Step 400: Fill the system prompt word template with the user's research goal instruction, the task definition information, the task context data package, and the user's cognitive profile to generate system prompt words. Execute the task instruction of the current task node based on the system prompt words to obtain the output result. The system prompt words include role definition, global project goal, knowledge constraints, and current task instruction.

[0036] The tasks at each task node are executed by sub-agents. A sub-agent is an AI entity with independent reasoning capabilities that focuses on completing a specific sub-task.

[0037] The system prompt template includes slots for role definition, global project goal, knowledge constraint, and current task. The system prompt generation process involves: filling the user's cognitive profile into the role definition slot, filling the user's research goal instruction into the global project goal slot, filling the task context data package into the knowledge constraint slot, and filling the node task goals from the task definition information into the current task slot. This ultimately generates a very long string containing complete background information, constraints, and action goals; this process is called dynamic injection. Knowledge constraints can include scientific formulas and standard code templates retrieved from static knowledge graphs, historical error correction records and successful experiences retrieved from dynamic memory graphs, and index pointers or complete data contents of artifacts retrieved from hierarchical heterogeneous data warehouses.

[0038] Step 500: Perform multiple verifications on the output results and provide feedback based on the verification results.

[0039] Multiple verifications include formal compliance verification, logical consistency verification, and target alignment verification. When all three verifications (formal compliance, logical consistency, and target alignment) are successful, the task for the next node is executed.

[0040] Figure 1The aforementioned method encapsulates the task context based on task information before executing each task node, avoiding information attenuation and task context loss during long-chain tasks. Furthermore, it fills the system prompt word template with task definition information, task context data packets, and user cognitive profiles to generate system prompt words. These prompt words include role definitions, global project goals, knowledge constraints, and current task instructions, ensuring that each task execution does not deviate from the global project goals, thereby improving the accuracy of task output results. Additionally, after each task execution, the output results undergo formal compliance and logical consistency verification, preventing the problem of research results lacking theoretical basis.

[0041] based on Figure 1 In addition to the method described herein, this specification also provides some specific implementation methods of the method, which will be described below.

[0042] As an optional approach, obtaining the user cognitive profile in step 100 can be implemented based on S11-S13.

[0043] S11: Select a user profile template based on the user's research goal instruction.

[0044] User profile templates are all pre-set. For example, if a user is a computer science researcher, their characteristics include an interest in coding style, algorithm frameworks, computing resources, and open-source licenses; the corresponding user profile template in the computer field can include technical preference characteristics such as technology stack preference, coding habits, issues of interest, and research methods, as well as research level characteristics such as error frequency and task complexity. If a user is a history researcher, their characteristics include an interest in dates, historical sources, citation formats, and archival retrieval; the corresponding user profile template in the history field can include technical preference characteristics such as historical source preference, writing habits, and issues of interest, as well as research level characteristics such as error frequency and task complexity.

[0045] S12: Extract the user's scientific research level characteristics and / or technology preference characteristics from the historical interaction data in the dynamic memory graph according to the user profile template, and generate a user cognitive profile; For example, the technical preference features extracted from the user profile template in the computer field are as follows: Technology stack preference: PyTorch over TensorFlow; Coding habits: Variable naming habits, experts usually have fewer comments, while novices have more; Whether to enforce type hints; Whether to use try-except; Focus on issues: Quantitative technology; Research methods: Preference to write experimental code first and then write papers; Preference to the latest preprints; Use of Weights & Biases.

[0046] In practical applications, historical interaction data and user profile templates can be directly input into a large language model to complete feature extraction and user cognitive profile generation.

[0047] S13: Monitor the number of new nodes in the dynamic memory graph. When the number of new nodes reaches a preset threshold, update the user cognitive profile.

[0048] In practical applications, user cognitive profile updates can be triggered based on counters, and preset quantity thresholds can be set according to needs.

[0049] The above steps record every scientific research interaction of the user through a dynamic memory graph, extract a high-dimensional user cognitive profile from it using user profile templates, and inject the user cognitive profile as prior knowledge into the intelligent agent in subsequent tasks. That is, the system prompts are generated based on the user cognitive profile, realizing personalized task planning and code generation, and solving the problem that general intelligent agents cannot understand the user's personalized scientific research habits.

[0050] As an alternative approach, step 300 can be implemented based on S31-S38.

[0051] S31: The task definition information is transformed into a numerical vector through an embedding model; Embedding models are AI models that convert unstructured data such as text, images, and audio into numerical vectors that computers can understand. The core goal of embedding models is that semantically similar content results in vectors with short distances, while semantically different content results in vectors with long distances.

[0052] S32: Retrieve nodes in the static knowledge graph and calculate the first cosine similarity between the first semantic vector and the numerical vector.

[0053] The first semantic vector is the semantic vector in the node attributes of the theoretical concept node and the standard implementation node.

[0054] The first, second, and third cosine similarities can all be calculated using an encoder that integrates BERT or Transformer architectures.

[0055] S33: Retrieve nodes in the dynamic memory graph and calculate the second cosine similarity between the second semantic vector and the numerical vector.

[0056] The second semantic vector is the semantic vector in the node attributes of the operation trajectory node and the result feedback node. The second semantic vector is a user correction record that has occurred or been marked as high value.

[0057] S34: Retrieve the metadata layer in the hierarchical heterogeneous data warehouse and calculate the third cosine similarity between the metadata header and the numerical vector.

[0058] S35: The data corresponding to the cosine similarity greater than the first preset threshold is determined as the matching vector; the matching vector includes one or more nodes in the static knowledge graph, nodes in the dynamic memory graph, and metadata header.

[0059] S36: Determine whether the third cosine similarity is greater than the second preset threshold. When the third cosine similarity is greater than the second preset threshold, determine the data volume corresponding to the third cosine similarity as the matching vector; the second preset threshold is greater than the first preset threshold. The second preset threshold and the first preset threshold can be set according to requirements. For example, the first preset threshold is 0.75 and the second preset threshold is 0.9.

[0060] This step employs a hierarchical heterogeneous data warehouse with a two-layer encapsulation structure. During retrieval, only the metadata header is loaded for vector matching. Only when the third cosine similarity exceeds a second preset threshold, or when the controlling agent explicitly requests access to a specific data body, does it initiate an HTTP GET request to the content layer based on the index pointer in the metadata header, read the corresponding data body, and decode it into text. This significantly reduces transmission overhead and, to some extent, solves the context window limitation problem.

[0061] S37: Sort the matching vectors in descending order of cosine similarity to obtain a sorted sequence; S38: Truncate the sorted sequence according to a preset length to obtain a task context data packet.

[0062] Specifically, the data is truncated from the Nth ranked data in the permutation sequence, and the first N ranked data are assembled into a task context data packet. The task context data packet is less than or equal to a preset length. This preset length is a context window limit, which can be set according to actual needs. For example, the context window limit is set to a maximum context length of 2000. After obtaining the task context data packet, it is stored in the task definition information of the task node.

[0063] Optionally, a dynamic Top-K truncation algorithm can be used. Starting with the data with the highest cosine similarity in the sorted sequence, the text length of each data point is accumulated. If the accumulated value is less than a preset length, the data is retained; if the accumulated value exceeds the preset length, the data and all subsequent data are discarded, thus pruning the sorted sequence. The value of K can be 5. This process ensures that only the most semantically relevant data proceeds to the next stage.

[0064] As an alternative approach, step 400, "execute the task instruction of the current task node based on the systematized prompt words and obtain the output result", can be implemented based on S41.

[0065] S41: Input the system prompts into the large language model, start the reasoning and action engine, and obtain the output results through closed-loop reasoning of observation, thinking, and action; the output results include code or text artifacts. Update the status flag of the task node after the task is completed.

[0066] In practical applications, the sub-agent calls the core `run()` method to start the engine, using the generated system prompt as the first message in the dialogue history, thus forcibly setting the short-term memory space of the LLM tool. The engine sends this message to the LLM API. Based on the injected system prompt, the LLM performs its first round of thought chain reasoning, planning an action such as "calling the `read_file` tool." The engine captures the LLM's tool call request, schedules the "standardized tool interface layer" to execute the actual operation, appends the execution result to the dialogue history, and sends it back to the LLM. This process repeats until the task is completed.

[0067] As an optional approach, the hierarchical heterogeneous data warehouse also stores a scientific research data exchange standard. This standard is a set of structured, machine-readable, cross-platform data formats and metadata specifications. Its core is a unified data structure, fields, semantics, and encoding rules, enabling different systems, tools, and institutions to exchange, parse, and reuse scientific research data without ambiguity. This standard can be customized according to requirements.

[0068] Performing formal compliance verification on the output results involves checking whether the output results have complete data fields and correct data types. This step can be implemented based on S511-S513: S511: Read the scientific research data exchange standard from the hierarchical heterogeneous data warehouse according to the task type of the task node.

[0069] S512: Use a JSON parser to deserialize and parse the output result. If the parsing is successful, the parsing result is obtained; if the parsing fails, it is determined to be a format error, and the verification result is output.

[0070] S513: Check whether the parsed result contains all the necessary fields defined in the scientific research data exchange standard and whether the data type matches the preset data type in the scientific research data exchange standard. If all the necessary fields exist and match the preset data type, the formal verification passes, and the verification result is output.

[0071] For example, for the "model evaluation" task, the scientific data exchange standard stipulates that the output must include three fields: accuracy (floating-point), loss (floating-point), and confusion_matrix (two-dimensional array). It is necessary to verify whether the output contains these three necessary fields.

[0072] As an optional approach, the output results are subjected to logical consistency verification, which aims to verify whether the code of the output results conforms to the theory and whether the conclusions are verifiable. Since unstructured logic is involved, this step is performed by the review agent using natural language reasoning tools, specifically, it can be implemented based on S521-S23.

[0073] S521: Extract theoretical benchmarks related to the current task node from the static knowledge graph.

[0074] For example, extract the LaTe mathematical formula for the "Softmax function": As a theoretical benchmark.

[0075] S522: Verify the consistency between the logic of the code that outputs the result and the logic in the theoretical benchmark.

[0076] Specifically, construct a comparison prompt containing "code to be reviewed" and "theoretical benchmark". The instruction format is "Please check line by line whether the code logic strictly corresponds to the order of operations and the meaning of variables in the above mathematical formula".

[0077] S523: Verify the logical consistency of the operators and formulas in the code using a thought chain.

[0078] Specifically, the process involves examining the intelligent agent's thought chain reasoning and comparing the operators in the code with the symbols in the formula.

[0079] Compliance Judgment Criteria: Logical Consistency: The review agent did not detect any missing operators, such as forgetting to calculate the sum of denominators or reversing the logical order, and output a clear PASS mark. Accurate Source Tracing: The review agent detected that the third-party library or dataset ID referenced in the code could be found in the system's data warehouse index, and no "illusion" generated false paths appeared.

[0080] The target alignment verification specifically involves examining the node definitions in the original task node sequence loaded by the agent and comparing them with the global project target, and then comparing the actual output of the sub-agent. If the actual output is correct in code but deviates from the preset global project target of the current node, for example, if a data loader is required but model training logic is written instead, it is judged as target drift, forcibly rejected, and a soft retry is triggered.

[0081] As an optional approach, feedback processing based on the verification results can be implemented based on S531-S533.

[0082] S531: When any of the formal compliance verification and logical consistency verification fails, a structured verification report is generated. This report includes the error type and suggested corrections. When all three verifications—formal compliance, logical consistency, and target alignment—pass, the generated report includes a verification pass flag. This report is then transmitted to the master control agent. Error types include formal errors and logical errors. The master control agent evaluates the error type and decides whether to trigger a soft retry or a hard circuit breaker.

[0083] S532: If the error type is a soft error, update the task context data packet and re-execute the task of the current task node according to the updated task context data packet, i.e., trigger a retry; when the error type is a formal error, it is judged as a soft error. When the error type is a logical error, it is judged as a hard error.

[0084] S533: If the error type is a hard error, then the directed acyclic graph is reconstructed by path correction or node insertion.

[0085] As an optional approach, after performing multiple verifications on the output results, the verification results and the output results need to be stored in a hierarchical heterogeneous data warehouse. This can be implemented based on S541-S542: S541: Extract key information from the output results based on the summary generation template and generate a natural language summary.

[0086] In practical applications, this step can be written collaboratively by the archiving agent and the review agent. When a task node is completed, the archiving agent automatically reads the output of that task node, extracts key information based on a pre-made summary generation template, and generates a natural language summary. After verification, it is stored in a hierarchical heterogeneous data warehouse.

[0087] S542: Store the natural language summary in the metadata header of the hierarchical heterogeneous data warehouse, and store the output result in the data body of the hierarchical heterogeneous data warehouse.

[0088] In the above steps, the storage mechanism of this application includes a static knowledge graph that stores scientific truths and a dynamic memory graph that stores interactive memories. After the task is completed, multiple verifications are introduced, which not only check the code format, but also force the code logic to be aligned with the mathematical formulas and theories in the static knowledge graph, thereby solving the problem of lack of theoretical basis in AI scientific research.

[0089] As an optional approach, a 3D data foundation needs to be constructed before performing step 100. Specifically: Upon system startup, the Docker container cluster is initialized first. A hierarchical heterogeneous data warehouse is built: MinIO (content layer) and Milvus (metadata layer) are started, establishing an index association between "metadata header" and "data body". A static knowledge graph is built: Neo4j is started, ETL scripts are run to parse papers and code repositories, establishing cross-modal alignment edges between "theory" and "implementation". A dynamic memory graph is built: the dynamic graph structure is initialized, and log listeners and user profiling analysis processes are started.

[0090] The embodiments of the present invention can divide functional modules according to the above method examples. For example, each function can be divided into its own functional module, or two or more functions can be integrated into one processing module. The integrated module can be implemented in hardware or as a software functional module. It should be noted that the module division in the embodiments of the present invention is illustrative and only represents one logical functional division; other division methods may be used in actual implementation.

[0091] like Figure 2 As shown, this invention also provides a multi-agent scientific research task execution system comprising: a master control agent 201, a dynamic orchestration engine 202, a state manager 203, a multi-agent collaborative execution module 204, a closed-loop verification module 205, a semantic trimming and routing module 206, a hierarchical heterogeneous data warehouse module 207, a static knowledge graph construction module 208, and a dynamic memory graph generation module 209. The master control agent 201 transmits data with the dynamic orchestration engine 202, the state manager 203, and the sub-agents 1-n in the multi-agent collaborative execution module via a message bus. The multi-agent collaborative execution module 204 transmits data with the hierarchical heterogeneous data warehouse module 207, the static knowledge graph construction module 208, and the dynamic memory graph generation module 209 through the semantic trimming and routing module 206. Specifically: The main control agent 201 is used to receive user research goal instructions and user cognitive profiles, and send the user research goal instructions to the dynamic orchestration engine.

[0092] In practical applications, the master control agent calls LLM to decompose the thought chain. After the dynamic orchestration engine generates a directed acyclic graph containing task nodes such as data download, preprocessing, and training, it sends the directed acyclic graph to the user interface for confirmation, locks the Plan as an immutable baseline blueprint, and writes it into the state manager.

[0093] The dynamic orchestration engine 202 is used to parse user research objective instructions and generate a sequence of task nodes; the task nodes in the sequence of task nodes include at least task definition information.

[0094] The semantic trimming and routing module 206 is used to retrieve matching data that meets the matching conditions in the three-dimensional data base according to the task definition information before executing any task node, and trim the matching data to obtain a task context data packet.

[0095] The multi-agent collaborative execution module 204 is used to fill the user's research goal instruction, the task definition information, the task context data packet, and the user's cognitive profile into the system prompt word template, generate system prompt words, execute the task instruction of the current task node based on the system prompt words, and obtain the output result; the system prompt words include role definition, global project goal, knowledge constraints, and current task instruction.

[0096] In practical applications, the execution architecture of the multi-agent collaborative execution module adopts a "containerized execution architecture based on standardized interfaces." The executing agents are built upon a general agent runtime unit, including a context loader, a standardized tool interface layer, and inference and action loop engines. The communication bus uses a Redis list or publish-subscribe pattern. Communication protocol: The master agent and the executing agents communicate non-blockingly via standardized JSON messages on the message bus, achieving complete decoupling between command issuance and task execution.

[0097] The closed-loop verification module 205 is used to perform multiple verifications on the output results; the multiple verifications include at least formal compliance verification and logical consistency verification, and feedback processing is performed based on the verification results.

[0098] In practical applications, the closed-loop verification module 205 is used for quality control and is configured with a unified acceptance protocol based on the review agent to perform formal compliance verification, logical consistency verification, and target alignment verification.

[0099] State Manager 203 is used to drive the system to transition between signal resolution state, pipelined distribution state, asynchronous blocking state, and topology evolution state; technical implementation: based on Redis Event Loop to achieve four-state transition.

[0100] The specific logic description of state transition is as follows: The process jumps from PENDING to EXECUTING, condition: the state of all preceding dependencies becomes COMPLETED. Action: DagManager identifies the node as having an in-degree of 0, extracts the task definition and constraints, and distributes it to the Worker Agent.

[0101] EXECUTING jumps to REVIEWING, condition: The Worker Agent completes the task and feeds back ACTION_DONE in the message queue. Action: The status manager submits the node output to the closed-loop verification module and starts the Reviewer Agent.

[0102] REVIEWING jumps to COMPLETED, condition: The verification module returns PASS. Action: The node status is locked and the blocking of subsequent dependent nodes is removed.

[0103] REVIEWING changes to PENDING, condition: The verification module returns FAIL_SOFT and retry_count < max_retries. Action: retry_count is incremented by 1, the fix_suggestions in the validation_log are injected into the next Prompt, and the status is reset to PENDING waiting for redistribution.

[0104] REVIEWING / EXECUTING changes to CIRCUIT_BROKEN, condition: retry_count >= max_retries or the verification module returns FAIL_HARD. Action: Stop the node and trigger the "graph reconstruction" process (Global Replan) of the dynamic orchestration engine.

[0105] ANY STATE changes to SUSPENDED, condition: Receive the user's manual suspension instruction (UserInterrupt). Action: Save the current context snapshot and stop all related Agent actions. Hierarchical heterogeneous data warehouse module 207: Used to encapsulate the output result of each task node into a modular storage unit; the modular storage unit is physically divided into a metadata header and a data body; the metadata header is used to store index information; the index information includes a natural language summary, a semantic vector, and a resource list extracted from the output result based on a summary generation template; the data body is used to store the output result; Static knowledge graph construction module 208: Used to construct theoretical concept nodes and standard implementation nodes, and determine the feature overlap degree between the theoretical concept nodes and the standard implementation nodes through alignment analysis, and establish strong constraint edges between the theoretical concept nodes and the standard implementation nodes with a feature overlap degree exceeding the overlap threshold to form a static knowledge graph; specifically, the alignment analysis algorithm extracts the mathematical symbol features in the theoretical concept nodes and the variable naming features in the standard implementation nodes, and calculates the feature overlap degree.

[0106] Dynamic memory map generation module 209: Used to generate a dynamic memory map consisting of operation trajectory nodes and result feedback nodes based on the execution and feedback of task nodes.

[0107] Optionally, the three-dimensional data foundation includes a hierarchical heterogeneous data warehouse, a static knowledge graph, and a dynamic memory graph; the semantic trimming and routing module 206 may specifically include: The numerical vector conversion unit is used to convert the task definition information into numerical vectors through the embedding model; The first retrieval unit is used to retrieve nodes in the static knowledge graph and calculate the first cosine similarity between the first semantic vector and the numerical vector; the first semantic vector is the semantic vector in the node attributes of the theoretical concept node and the standard implementation node. The second retrieval unit is used to retrieve nodes in the dynamic memory graph and calculate the second cosine similarity between the second semantic vector and the numerical vector; the second semantic vector is the semantic vector in the node attributes of the operation trajectory node and the result feedback node; The third retrieval unit is used to retrieve the metadata layer in the hierarchical heterogeneous data warehouse and calculate the third cosine similarity between the metadata header and the numerical vector. The matching vector determination unit is used to determine the data corresponding to the cosine similarity greater than a first preset threshold as the matching vector; the matching vector includes one or more nodes in the static knowledge graph, nodes in the dynamic memory graph, and metadata header; A sorting unit is used to sort the matching vectors from largest to smallest according to their cosine similarity to obtain a sorted sequence; The trimming unit is used to truncate the sorted sequence according to a preset length to obtain a task context data packet.

[0108] Optionally, the hierarchical heterogeneous data warehouse also stores scientific research data exchange standards; the task node sequence is generated based on a directed acyclic graph; the closed-loop verification module 205 further includes: The format verification unit is used to read the scientific research data exchange standard in the hierarchical heterogeneous data warehouse according to the task type of the task node; use a JSON parser to deserialize and parse the output result; if the parsing is successful, the parsing result is obtained; if the parsing fails, it is determined to be a format error; check whether all the necessary fields defined in the scientific research data exchange standard exist in the parsing result and whether the data type matches the preset data type in the scientific research data exchange standard.

[0109] The logical consistency verification unit is used to extract theoretical benchmarks related to the current task node from the static knowledge graph; verify the consistency between the logic of the code in the output result and the logic in the theoretical benchmark; and verify the logical consistency of the operators and formulas in the code through the thought chain.

[0110] The feedback unit is used to generate a structured verification report when either formal compliance verification or logical consistency verification fails. The structured verification report includes the error type and correction suggestions. If the error type is a soft error, the task context data packet is updated and the task of the current task node is re-executed according to the updated task context data packet. If the error type is a hard error, the path is corrected or a node is inserted into the directed acyclic graph.

[0111] Optionally, the multi-agent cooperative execution module 204 can be specifically used for: The system prompts are input into the large language model to start the reasoning and action engine. Through closed-loop reasoning involving observation, thinking, and action, the output results are obtained. The output results include code or text artifacts.

[0112] Optionally, the master control agent 201 may include: The user cognitive profile generation unit is used to select a user profile template based on the user's research goal instruction. Based on the user profile template, the user's scientific research level characteristics and technical preference characteristics are extracted from the historical interaction data in the dynamic memory graph to generate a user cognitive profile. The number of new nodes in the dynamic memory graph is monitored, and the user cognitive profile is updated when the number of new nodes reaches a preset threshold.

[0113] Optionally, before sorting the matching vectors according to their cosine similarity from largest to smallest to obtain a sorted sequence, the method further includes: Determine whether the third cosine similarity is greater than the second preset threshold. When the third cosine similarity is greater than the second preset threshold, determine the data volume corresponding to the third cosine similarity as the matching vector; the second preset threshold is greater than the first preset threshold.

[0114] Figure 2 The aforementioned system generates and maintains task lists through a dynamic orchestration engine; provides precise scientific concept constraints using a static knowledge graph; automatically evolves user cognitive profiles based on memory graphs; and performs closed-loop verification of execution results through a dual-check mechanism. This achieves automation and standardization of the entire scientific research process, effectively improving execution accuracy, logical coherence, and user interaction experience.

[0115] The above mainly describes the solutions provided by the embodiments of the present invention from the perspective of the interaction between various modules. It is understood that, in order to achieve the above functions, it includes corresponding hardware structures and / or software modules for executing each function. Those skilled in the art should readily recognize that, in conjunction with the units and algorithm steps of the various examples described in the embodiments disclosed herein, the present invention can be implemented in hardware or a combination of hardware and computer software. Whether a function is executed in hardware or by computer software driving hardware depends on the specific application and design constraints of the technical solution. Those skilled in the art can use different methods to implement the described functions for each specific application, but such implementation should not be considered beyond the scope of the present invention.

[0116] In the above embodiments, implementation can be achieved entirely or partially through software, hardware, firmware, or any combination thereof. When implemented using software, it can be implemented entirely or partially in the form of a computer program product. The computer program product includes one or more computer programs or instructions. When the computer program or instructions are loaded and executed on a computer, the processes or functions described in the embodiments of the present invention are performed entirely or partially. The computer can be a general-purpose computer, a special-purpose computer, a computer network, a terminal, a user equipment, or other programmable device. The computer program or instructions can be stored in a computer-readable storage medium or transferred from one computer-readable storage medium to another. For example, the computer program or instructions can be transferred from one website, computer, server, or data center to another website, computer, server, or data center via wired or wireless means. The computer-readable storage medium can be any available medium that a computer can access or a data storage device such as a server or data center that integrates one or more available media. The available medium can be a magnetic medium, such as a floppy disk, hard disk, or magnetic tape; it can also be an optical medium, such as a digital video disc (DVD); or it can be a semiconductor medium, such as a solid-state drive (SSD).

[0117] Although the invention has been described herein in conjunction with various embodiments, those skilled in the art will understand and implement other variations of the disclosed embodiments by reviewing the accompanying drawings, the disclosure, and the appended claims in carrying out the claimed invention. In the claims, the word "comprising" does not exclude other components or steps, and "a" or "an" does not exclude a plurality. A single processor or other unit can implement several functions listed in the claims. While different dependent claims may recite certain measures, this does not mean that these measures cannot be combined to produce good results.

[0118] Although the invention has been described in conjunction with specific features and embodiments, it is obvious that various modifications and combinations can be made therein without departing from the spirit and scope of the invention. Accordingly, this specification and drawings are merely exemplary descriptions of the invention as defined by the appended claims, and are considered to cover any and all modifications, variations, combinations, or equivalents within the scope of the invention. Clearly, those skilled in the art can make various alterations and modifications to the invention without departing from its spirit and scope. Thus, if such modifications and modifications of the invention fall within the scope of the claims and their equivalents, the invention is also intended to include such modifications and modifications.

Claims

1. A method for executing multi-agent scientific research tasks, characterized in that, include: Obtain user research goal instructions and user cognitive profile; Parse the user's research objective instructions and generate a sequence of task nodes; The task nodes in the task node sequence include at least task definition information; Before executing any task node, a search is performed in the three-dimensional data base according to the task definition information to obtain matching data that meets the matching conditions, and the matching data is trimmed to obtain a task context data package; the three-dimensional data base includes a hierarchical heterogeneous data warehouse, a static knowledge graph and a dynamic memory graph. The user's research goal instruction, the task definition information, the task context data package, and the user's cognitive profile are filled into the system prompt word template to generate system prompt words. Based on the system prompt words, the task instruction of the current task node is executed to obtain the output result. The system prompts include role definition, global project goals, knowledge constraints, and current task instructions; The output results are subjected to multiple verifications, and feedback processing is performed based on the verification results; the multiple verifications include at least formal compliance verification and logical consistency verification.

2. The multi-agent scientific research task execution method according to claim 1, characterized in that, The static knowledge graph includes theoretical concept nodes and standard implementation nodes; the dynamic memory graph includes operation trajectory nodes and result feedback nodes; based on the task definition information, a search is performed in the three-dimensional data base to obtain matching data that meets the matching conditions, and the matching data is trimmed to obtain a task context data package including: By embedding the model, the task definition information is transformed into a numerical vector; Retrieve nodes from the static knowledge graph and calculate the first cosine similarity between the first semantic vector and the numerical vector; the first semantic vector is the semantic vector in the node attributes of the theoretical concept node and the standard implementation node. Retrieve nodes from the dynamic memory graph and calculate the second cosine similarity between the second semantic vector and the numerical vector; the second semantic vector is the semantic vector in the node attributes of the operation trajectory node and the result feedback node; Retrieve the metadata layer in the hierarchical heterogeneous data warehouse and calculate the third cosine similarity between the metadata header and the numerical vector; Data with cosine similarity greater than a first preset threshold are determined as matching vectors; the matching vectors include nodes in the static knowledge graph, nodes in the dynamic memory graph, and one or more in the metadata header; The matching vectors are sorted in descending order of cosine similarity to obtain a sorted sequence; The sorted sequence is truncated according to a preset length to obtain a task context data packet.

3. The multi-agent scientific research task execution method according to claim 2, characterized in that, The hierarchical heterogeneous data warehouse also stores scientific research data exchange standards; formal compliance verification of the output results includes: According to the task type of the task node, the scientific research data exchange standard is read from the hierarchical heterogeneous data warehouse; The output is deserialized using a JSON parser. If parsing is successful, the parsed result is obtained; if parsing fails, it is determined to be a format error. Check whether the parsing result contains all the necessary fields defined in the scientific research data exchange standard and whether the data type matches the preset data type in the scientific research data exchange standard.

4. The multi-agent scientific research task execution method according to claim 2, characterized in that, The logical consistency verification of the output results includes: Extract theoretical benchmarks related to the current task node from the static knowledge graph; Verify the consistency between the logic of the code that outputs the results and the logic in the theoretical benchmark; The logical consistency of the operators and formulas in the code is verified using a thought process.

5. The multi-agent scientific research task execution method according to claim 1, characterized in that, The process of executing the task of the current task node based on the system prompts and obtaining the output results includes: The system prompts are input into the large language model to start the reasoning and action engine. Through closed-loop reasoning involving observation, thinking, and action, the output results are obtained. The output results include code or text artifacts.

6. The multi-agent scientific research task execution method according to claim 1, characterized in that, Obtaining a user cognitive profile includes: Select a user profile template based on the user's research goals and areas of expertise; Based on the user profile template, the user's scientific research level characteristics and technical preference characteristics are extracted from the historical interaction data in the dynamic memory graph to generate a user cognitive profile. The number of new nodes in the dynamic memory graph is monitored, and the user cognitive profile is updated when the number of new nodes reaches a preset threshold.

7. The multi-agent scientific research task execution method according to claim 1, characterized in that, The task node sequence is generated based on a directed acyclic graph; feedback processing based on the verification results includes: When either formal compliance verification or logical consistency verification fails, a structured verification report is generated; the structured verification report includes the error type and correction suggestions. If the error type is a soft error, update the task context data packet and re-execute the task of the current task node based on the updated task context data packet; If the error type is a hard error, then the directed acyclic graph is corrected or a node is inserted.

8. The multi-agent scientific research task execution method according to claim 2, characterized in that, The step of sorting the matching vectors according to their cosine similarity from largest to smallest to obtain a sorted sequence includes, prior to: Determine whether the third cosine similarity is greater than the second preset threshold. When the third cosine similarity is greater than the second preset threshold, determine the data volume corresponding to the third cosine similarity as the matching vector; the second preset threshold is greater than the first preset threshold.

9. A multi-agent scientific research task execution system, applied to the multi-agent scientific research task execution method according to any one of claims 1-8, characterized in that, The system includes at least: The main control agent is used to receive user research goal instructions and user cognitive profiles, and send the user research goal instructions to the dynamic orchestration engine; A dynamic orchestration engine is used to parse user research objective instructions and generate a sequence of task nodes; the task nodes in the sequence of task nodes include at least task definition information. The semantic trimming and routing module is used to retrieve matching data that meets the matching conditions in the three-dimensional data base according to the task definition information before executing any task node, and trim the matching data to obtain a task context data package; the three-dimensional data base includes a hierarchical heterogeneous data warehouse, a static knowledge graph and a dynamic memory graph. The multi-agent collaborative execution module is used to fill the user's research goal instruction, the task definition information, the task context data package, and the user's cognitive profile into the system prompt word template, generate system prompt words, execute the task instruction of the current task node based on the system prompt words, and obtain the output result; the system prompt words include role definition, global project goal, knowledge constraints, and current task instruction. The closed-loop verification module is used to perform multiple verifications on the output results and to perform feedback processing based on the verification results; the multiple verifications include at least formal compliance verification and logical consistency verification.

10. The multi-agent scientific research task execution system according to claim 9, further comprising: The state manager drives the system to transition between signal resolution state, pipelined dispatch state, asynchronous blocking state, and topology evolution state. A hierarchical heterogeneous data warehouse module is used to encapsulate the output of each task node into a modular storage unit; the modular storage unit is physically divided into a metadata header and a data body; the metadata header is used to store index information. The index information includes a natural language summary, semantic vector, and resource list extracted from the output results based on the summary generation template; the data body is used to store the output results. The static knowledge graph construction module is used to construct theoretical concept nodes and standard implementation nodes, and to determine the feature overlap between the theoretical concept nodes and the standard implementation nodes through alignment analysis. Strong constraint edges are established between theoretical concept nodes and standard implementation nodes whose feature overlap exceeds the overlap threshold to form a static knowledge graph. The dynamic memory graph generation module is used to generate a dynamic memory graph consisting of operation trajectory nodes and result feedback nodes based on the execution and feedback of task nodes.

Images