Intelligent teaching system based on knowledge graph and multi-agent cooperation

Abstract

This invention discloses a smart teaching system based on knowledge graphs and multi-agent collaboration, relating to the subject of artificial intelligence reasoning technology. The system includes: a multimodal graph dataset and lightweight knowledge graph construction module, which constructs a multimodal computer science graph benchmark dataset and generates a lightweight computer science subject knowledge graph; a knowledge graph enhanced pre-inference module, which outputs a standardized preprocessing package; a CS-MMA collaborative architecture module, including an adaptation agent unit, a basic agent unit, an advanced agent unit, a professional agent unit, and a verification agent unit. The adaptation agent unit determines the task level and generates driving instructions, the three agent units perform differentiated reasoning respectively, and the verification agent unit verifies the execution results; and a system verification module for performance verification and iterative optimization. This invention achieves highly accurate analysis and personalized solutions for computer science graph-based teaching tasks.

Description

Technical Field

[0001] This invention relates to the field of artificial intelligence reasoning technology, specifically to a smart teaching system based on knowledge graphs and multi-agent collaboration. Background Technology

[0002] The deep integration of artificial intelligence technology and education has driven the rapid development of smart education, and the construction of an intelligent education service system has become an important direction for the modernization of education. Currently, smart education has made significant progress in the popularization of online learning platforms and the application of AI-assisted teaching tools, effectively improving teaching efficiency and the balance of resource allocation. In the field of computer science education, due to the highly specialized, logical, and practical nature of its knowledge system, students often face multiple challenges, including understanding complex concepts, debugging code, and solidifying theoretical foundations. Traditional teaching models, in large-class settings, struggle to provide timely and differentiated guidance tailored to individual differences, especially during self-study after class, where learning difficulties cannot be quickly addressed or accurately resolved, resulting in a significant teaching gap.

[0003] In current technological development, AI education in computer science exhibits three characteristics: evaluation benchmarks have shifted from general tasks to assessments of professional knowledge points; model capabilities have expanded from text reasoning to multimodal graph understanding; and teaching systems are becoming increasingly collaborative and professional, with a growing emphasis on the practical application of technology. However, these technologies have limitations in practical applications: 1) Current benchmarks are mostly designed for pure text reasoning or general image recognition, lacking systematic evaluation resources for computer science-specific graphs. The ability to understand professional graphs such as CPU data paths, protocol stack layers, and process state transitions is difficult to effectively evaluate, resulting in a lack of clear objectives for model optimization. 2) Tests show that mainstream models perform reasonably well on computer science multiple-choice questions, but their accuracy drops significantly when faced with comprehensive questions that require associating multiple graphs and maintaining long-chain states. The coordination mechanism between visual information extraction and logical reasoning is imperfect, leading to the phenomenon of "understanding the graph but not being able to deduce the reasoning." 3) Existing solutions mostly adopt a single-model architecture or simply connect multiple models in parallel, resulting in loose organization of professional knowledge, which easily leads to state loss or logical contradictions during complex reasoning processes. The system can complete basic question and answer tasks, but it performs poorly in cross-knowledge point association, in-depth principle explanation, and multi-step consistency maintenance. There is also a prominent contradiction between computing power consumption and response timeliness. Summary of the Invention

[0004] The technical solution of this invention is used to solve the problem of insufficient accuracy in reasoning about computer-related graphical knowledge.

[0005] This invention solves the above-mentioned technical problems through the following technical means: a smart teaching system based on knowledge graphs and multi-agent collaboration, comprising:

[0006] The module for constructing a multimodal graph dataset and a lightweight knowledge graph is used to build the MMCS-Bench multimodal computer science subject graph benchmark dataset, which includes the subject and major graph of each question, the corresponding question, the answer, and the explanation. Based on the MMCS-Bench multimodal computer science subject graph benchmark dataset, a lightweight computer science subject knowledge graph is generated. ; The knowledge graph-enhanced pre-inference module, connected to the multimodal graph dataset and lightweight knowledge graph construction module, is used to receive user-input tasks to be processed. Computer science diagrams and question text Through cross-modal semantic fusion units, knowledge graph fusion reasoning units, preprocessing result standardization units, and model optimization and tuning units, a standardized preprocessing package is output. The standardized preprocessing package Encapsulated with: cross-modal fusion semantic text vectors Preliminary reasoning results Knowledge gap label And the illustrated sub-map ; The CS-MMA (Computer Science - Multimodal Multi-Agent) collaborative architecture module, connected to the knowledge graph augmented pre-inference module, includes an adaptation agent unit, a basic agent unit, an advanced agent unit, a specialized agent unit, and a verification agent unit, used to receive the standardized preprocessing package. And output the final reasoning result; The adaptation agent unit includes a task complexity scoring model, used for knowledge gap labeling. and preliminary reasoning results The task level is quantitatively scored, and the task level is determined as simple, medium or complex. The driving instructions are then generated and output. The basic agent unit, advanced agent unit, and professional agent unit are respectively used to receive driving instructions for simple, medium, and complex tasks and perform differentiated reasoning. The verification Agent unit is used to receive the output results of each Agent unit and perform verification. The system verification module, connected to the CS-MMA collaborative architecture module, is used to verify and iteratively optimize system performance.

[0007] Furthermore, the multimodal graph dataset and lightweight knowledge graph construction module includes: The data collection and preprocessing unit is used to select visual problems that cannot be solved without explaining the relevant charts from computer science teaching resources as initial candidate samples. For the initial candidate samples, the unit captures the corresponding professional illustrations and identifies them according to a naming protocol. The unit manually aligns the professional illustrations with the corresponding question text, options, answers, and explanation text to establish a mapping relationship. Based on the mapping relationship, the unit uses multimodal large model processing to generate structured data containing the professional illustrations, question text, options, answers, and explanations. The structured data is then proofread and output. The structured sample construction unit is used to remove duplicate samples and filter low-quality data by semantic similarity calculation of the structured data; the multiple-choice sample is converted into fill-in-the-blank sample without options by a large model and the uniqueness of the answer is ensured by manual verification; the subject, major illustration, corresponding question, answer and analysis of each processed question are packaged into a unified task sample to form a multimodal computer science illustration benchmark dataset MMCS-Bench covering the four major subjects of computer science. The lightweight knowledge graph unit extracts three core knowledge elements from the task samples: entities, intra-subject entity relationships, and cross-subject general relationships. Based on the subject tags of the task samples, the knowledge elements are organized into the corresponding subject knowledge graph using the Neo4j graph database. Redundant node merging and association path simplification operations are performed on the subject knowledge graph to obtain a lightweight subject knowledge graph. These lightweight knowledge graphs are then integrated to form a lightweight computer science subject knowledge graph. And a large language model was used to analyze the lightweight computer science subject knowledge graph. Semantic transformation is performed to obtain a lightweight computer science subject knowledge graph. Natural language description.

[0008] Furthermore, the cross-modal semantic fusion unit includes: Integrating visual language models, large language models, and cross-modal deep fusion functions; employing visual language models for the task at hand. Computer science diagrams Semantic parsing is performed to generate a natural language description of the diagram; a large language model is used to analyze the question text. Perform semantic understanding to generate a natural language description of the problem text; invoke a lightweight computer science subject knowledge graph. Computer science diagrams Entity and Lightweight Computer Science Discipline Knowledge Graph Match entities in the graph to construct a graphical subgraph containing entities and relationships. ; The illustrated subgraph is analyzed using a large language model. Perform semantic transformation to generate a graphical sub-map. The natural language description; the cross-modal semantic fusion unit has a built-in cross-modal deep fusion function, which unifies the semantic granularity and expression paradigm of the three types of natural language descriptions through the semantic adaptation module built into the cross-modal deep fusion function, establishes the correspondence between the three through logical association matching, and integrates the three into a unified cross-modal fused semantic text through a deep fusion algorithm. .

[0009] Furthermore, the knowledge graph fusion reasoning unit is used to process the task to be processed. Cross-modal fusion semantic text Lightweight Computer Science Discipline Knowledge Graph The natural language descriptions are combined and input into a large model. Through the attention mechanism built into the large model, cross-modal semantic text is fused. The natural language description in the text calls upon a lightweight computer science subject knowledge graph. Entities, intra-subject entity relationships, and cross-subject general relationships Complete the entity extraction and basic logic derivation shown in the diagram, and output the task to be processed. Preliminary reasoning results ; The preliminary reasoning results are calculated by using the knowledge gap annotation function built into the knowledge graph fusion reasoning unit. With the illustration of the sub-map The semantic matching degree is used to determine the initial inference result when the matching degree is lower than a preset threshold. The missing computer science knowledge in the text, and the generation of knowledge gap tags. .

[0010] Furthermore, the adaptation agent unit, built based on the large model Gemini-2.5-Flash, is used to receive tasks to be processed. With standardized preprocessing packages From standardized preprocessing packages Preliminary reasoning results were obtained from the analysis. Knowledge Gap Labels ; The task complexity scoring model built into the adapter agent unit is invoked based on the task to be processed. Preliminary reasoning results Knowledge gap label and the aforementioned lightweight computer science subject knowledge graph The task complexity score is quantitatively evaluated from four dimensions: subject specialization, logical hierarchy, answer output requirements, and diagrammatic complexity. The comprehensive score is obtained by weighted summation. The overall task complexity score is compared with a preset threshold range to determine the task to be processed. The task levels include simple, medium, and complex. Real-time acquisition of the real-time status sets of basic agent units, advanced agent units, and professional agent units within the system. ; Based on the determined task level, combined with the real-time status set This generates driver instructions for driving Agent units that match the task level.

[0011] Furthermore, the basic Agent unit, built upon the large model Qwen3-VL-30B-A3B-Instruct, is used to receive driving instructions and tasks to be processed from the Agent unit. and standardized preprocessing packages And based on standardized preprocessing packages Cross-modal fusion semantic text encapsulated in Computer science diagrams Entity recognition is performed by calling the illustration sub-map in the standardized preprocessing package. The identified entities are compared with the illustrated sub-map. The entities and relationships are compared to verify the existence of the entities and the clarity of the relationships; If entity or relation missing information is found during the verification process, consider using a lightweight computer science subject knowledge graph. The corresponding data is retrieved and completed; the results of recognition, matching and completion are integrated to generate and output the reasoning results.

[0012] Furthermore, the advanced Agent unit, built upon the large model Claude-Sonnet-4-5-20250929-thinking, is used to receive driving instructions and tasks to be processed from the adapted Agent unit. and standardized preprocessing packages ; Cross-modal fusion semantic text based on standardized preprocessing packages With the illustration of the sub-map For computer science diagrams Perform logical link parsing, including: executing the entity in the diagram - the diagram subgraph. Entity anchoring in the graph involves calling a graph neural network (GNN) to analyze the subgraph. Multi-hop relationship propagation is performed to construct the initial logical link, and the illustrated sub-graph is generated through the large model Qwen3-VL-8B. The relationships in the code are transformed into natural language logical descriptions, combined with standardized preprocessing packages. Verify the logical syntax and explicit rules of the initial logical link; Based on the results of the logical link parsing and the knowledge gap tags in the standardized preprocessing package From a lightweight computer science discipline knowledge graph The core application layer knowledge points are retrieved in the targeted search, and the preliminary reasoning results are analyzed. Supplement knowledge; Preliminary reasoning results The process involves step-by-step decomposition and verification, breaking it down into single-step reasoning units, entity association links, and subject knowledge reference segments. Logical consistency verification and fusion completion are then performed based on explicit logical breakpoint types, which include one or more of the following: incomplete fields, inconsistencies in reasoning steps within a single subject, incorrect concept references, conflicts between diagram annotations and text descriptions, and matching gaps between knowledge graph entities and reasoning entities. The results of integrating logical link analysis, knowledge supplementation, and fusion completion are used to generate and output the reasoning results.

[0013] Furthermore, the specialized agent unit is built on a dual-model architecture, including the GPT-5 model for handling multiple-choice and fill-in-the-blank questions, and the GLM-4.6V model for handling comprehensive application questions. The specialized agent unit receives driving instructions and tasks to be processed from the adapted agent unit. and standardized preprocessing packages ; Based on the task format determination result, the appropriate large model is selected and invoked, and the invoked large model is based on the cross-modal fused semantic text. Computer science diagrams A full-dimensional analysis is performed, which covers entity nesting relationships, control signal annotation, and system state mapping. Based on knowledge gap tagging From a lightweight computer science discipline knowledge graph The system retrieves deep-seated fundamental principles and, based on cross-subject common relationships, retrieves related cross-subject knowledge to refine the preliminary reasoning results. To supplement; Based on the aforementioned deep-seated fundamental knowledge, cross-subject related knowledge, and comprehensive analysis results, long-chain nested logical reasoning and system-level consistency verification are performed; simultaneously, preliminary reasoning results are identified. The implicit logical conflicts in the process include one or more of the following: cross-subject contradictions, inconsistencies in mechanisms, and unreasonable frameworks. Based on the deep-seated principles and using cross-subject related knowledge as a logical bridge, the conflicts in the preliminary reasoning results are resolved to ensure the logical coherence of multi-sub-problems and multi-step reasoning. It integrates the results of comprehensive analysis, knowledge supplementation, long-chain nested logical reasoning, and conflict resolution to generate and output the reasoning results.

[0014] Furthermore, the verification agent unit, built based on the large model Gemini-3-Pro-Preview, is used to receive the inference results and standardized preprocessing packages output by the upstream functional agent unit. The preset quantization thresholds that match each functional Agent unit are combined with the standardized preprocessing package. Knowledge gap tags Illustration of sub-map and lightweight computer science subject knowledge graph The system performs item-by-item verification of the inference results. If the verification results reach the corresponding quantization accuracy threshold, the final inference result is directly output. If the verification results do not reach the corresponding quantization accuracy threshold, the system traces the source of the substandard inference results and generates correction suggestions. At the same time, the verification results are synchronized to the adaptation agent unit and the upstream functional agent unit to achieve cross-layer bidirectional feedback. The verification results are also sent back to the knowledge graph enhancement pre-inference module and the model optimization and tuning unit to achieve cross-modal fusion semantic text optimization and model tuning linkage. The corrected and integrated content is transformed into the final inference result with the same content structure as the output content of the corresponding functional agent unit.

[0015] Furthermore, the system verification module is used to execute a hierarchical verification process: automating tests are performed based on the multimodal computer science subject knowledge graph benchmark dataset MMCS-Bench to batch verify the overall system performance. The verification indicators for the overall system performance include answer accuracy, logical completeness of explanations, and utilization rate of diagrammatic information. Next, real-world testing is conducted based on actual teaching cases for each subject, collecting feedback from actual usage in teaching scenarios. Finally, based on the automated test data and user feedback from the real-world testing, the system model, the CS-MMA collaborative architecture module, and the lightweight computer science subject knowledge graph are evaluated. Perform iterative optimization.

[0016] The advantages of this invention are: This invention addresses the fragmentation and redundant knowledge representation issues in traditional teaching systems by constructing the multimodal computer science diagrammatic benchmark dataset MMCS-Bench and a lightweight computer science subject knowledge graph. It enhances the alignment and fusion of cross-modal semantic text through a knowledge graph-enhanced pre-inference module, outputting standardized preprocessing packages to improve the accuracy of knowledge associations. The CS-MMA collaborative architecture module uses task complexity grading (simple / medium / complex) to drive differentiated inference for basic, advanced, and professional agents. Combined with quantitative accuracy verification and a two-way feedback mechanism for the validation agents, it achieves dynamic adaptation of tasks and a closed-loop result verification, ensuring the rigor of the inference logic. Finally, automated testing and experimental feedback through the system verification module continuously improve system performance. Overall, this invention significantly enhances the analytical depth and inference accuracy of computer science diagrammatic teaching tasks. Attached Figure Description

[0017] Figure 1 This is a schematic diagram of the modules of the intelligent teaching system based on knowledge graph and multi-agent collaboration according to Embodiment 1 of the present invention; Figure 2 This is a quantity distribution diagram of subject areas and task types in Embodiment 1 of the present invention; Figure 3 This is a distribution diagram of the number of sub-problems in the comprehensive problem set of Embodiment 1 of the present invention; Figure 4 This is a heatmap of the reasoning ability of Embodiment 1 of the present invention; Figure 5 This is a line graph showing the performance of different large models in different subjects for multiple-choice and fill-in-the-blank questions in Embodiment 1 of the present invention; Figure 6 These are line graphs showing the performance of different large models in different subject-specific integrated problems according to Embodiment 1 of the present invention; Figure 7A This is a bar chart showing the fine-grained inference performance of the large models Qwen3-VL-30B-A3B-Instruct, Doubao-1.5-Vision-Pro, Qwen-VL-max, Claude-Sonnet-4-5-20250929-thinking, and Gemini-2.5-Flash in Embodiment 1 of the present invention. Figure 7B This is a bar chart showing the fine-grained inference performance of the large models GPT-5, Gemini-3-Pro-Preview, GLM-4.6V, the large model combination, and CS-MMA in Embodiment 1 of the present invention. Detailed Implementation

[0018] To make the objectives, technical solutions, and advantages of the embodiments of the present invention clearer, the technical solutions of the embodiments of the present invention will be clearly and completely described below in conjunction with the embodiments of the present invention. Obviously, the described embodiments are only some embodiments of the present invention, not all embodiments. Based on the embodiments of the present invention, all other embodiments obtained by those skilled in the art without creative effort are within the scope of protection of the present invention.

[0019] Example 1 This embodiment uses the teaching assistance of four core computer science subjects—Data Structures and Algorithms (DSA), Computer Organization (CO), Computer Network (CN), and Operating System (OS)—as an application scenario to provide a detailed explanation of the technical solution of the intelligent teaching system based on knowledge graphs and multi-agent collaboration of the present invention.

[0020] This system adopts a distributed architecture, deployed in layers: hardware, software, and application. Each layer collaborates to ensure stable system operation. The hardware layer deploys graphics workstations and supporting data acquisition equipment. The graphics workstations are equipped with NVIDIA A100 GPUs and 128GB of memory to meet the computing power requirements for multimodal data processing, model training, and inference. The software layer is built on a Python 3.9 development environment, integrating PyTorch 2.0 for model training, Neo4j for knowledge graph construction, LangChain for agent-based collaborative scheduling, and FastAPI for interface development, providing underlying development and operational support for each module. The application layer provides teaching assistance services through a web client developed with Vue3 and a mobile client adapted for both Android and iOS systems. Data linkage and functional collaboration between functional modules are achieved through RESTful APIs. The system is deployed to a cloud server using Docker containers, supporting up to 100 concurrent accesses. At the same time, a system-level closed-loop iteration mechanism is established to continuously collect user feedback from teaching scenarios. The system model is incrementally fine-tuned every quarter based on the feedback, and the lightweight computer science subject knowledge graph is updated to continuously optimize the overall system performance and knowledge coverage.

[0021] like Figure 1 As shown, the intelligent teaching system based on knowledge graphs and multi-agent collaboration includes: The module for constructing a multimodal graph dataset and a lightweight knowledge graph is used to build the MMCS-Bench multimodal computer science subject graph benchmark dataset, which includes the subject and major graph of each question, the corresponding question, the answer, and the explanation. Based on the MMCS-Bench multimodal computer science subject graph benchmark dataset, a lightweight computer science subject knowledge graph is generated. .

[0022] The data collection and preprocessing unit is used to select visual problems from computer science teaching resources that cannot be solved without interpreting the relevant diagrams as initial candidate samples. These resources include past years' 408 postgraduate entrance examination questions for computer science in China, authoritative textbooks for the four core computer science courses, and formal mock exam materials to ensure the data's professional authority and depth. During the data acquisition phase, high-precision scanning equipment is used to digitize paper materials, with a scanning resolution set at 300 dpi or higher to generate PDF files containing text layers. OCR technology is used to ensure accurate text extraction. After digitization, computer science students manually screen the data. The screening criteria are strictly limited to visual problems that cannot be solved without interpreting the relevant diagrams. These problems cover types such as CPU data path diagrams and instruction pipeline spacetime diagrams from Computer Organization, binary trees and hash representations from Data Structures, TCP / IP protocol stack hierarchy diagrams and routing representations from Computer Networks, and process state transition diagrams and page table mapping diagrams from Operating Systems. Each problem requires independent judgment by two students, and only after both students reach a consensus can it be confirmed as an initial candidate sample. For each initial candidate sample, a corresponding professional illustration is captured using professional screenshotting or image processing tools (such as Adobe Photoshop). The capture process follows strict quality standards, requiring an image resolution of no less than 1920×1080 pixels and an effective chart area occupying more than 60% of the total image area. A naming protocol with a four-level coding structure including subject, year, question type, and question number is used to uniformly identify the captured illustrations and their corresponding questions. Using this identifier as an index, computer science students manually align the identified captured illustrations with the corresponding question text, options, standard answers, and explanation text, establishing a one-to-one mapping relationship between chart elements and text content. The alignment results are stored in a relational database. The aligned image and text data is then input into the large model Gemini, using structured JSON. The schema definition and LaTeX formula retain instruction prompts, automatically extracting and generating structured data that integrates professional illustrations, question text, option information, standard answers, and explanations. The generated structured data undergoes two rounds of validation to ensure quality: the first round is automatic validation, checking field completeness and JSON format validity; the second round is manual proofreading, verifying the academic standardization of LaTeX formulas, the consistency of fields with the preset schema, the tightness of the integration of explanation text and illustrations, and the correctness of answers. This results in high-quality, structured multimodal exam question data for downstream modules.

[0023] The structured sample construction unit further processes the structured data output from the data collection and preprocessing unit to form the multimodal computer science diagramming benchmark dataset MMCS-Bench. Semantic similarity analysis is performed on the structured data to identify and remove duplicate samples. The Sentence-BERT model is used to embed and encode all question texts and calculate cosine similarity. A 95% similarity threshold is set for multiple-choice questions and a 92% similarity threshold for comprehensive questions to identify duplicate candidates. Redundant samples are deleted after manual review and confirmation. Low-quality samples with insufficient image resolution, low chart / figure area proportions, mismatched question text and diagram labels, or ambiguous answers are also filtered out. To further enhance the robustness of the dataset, multiple-choice questions were converted into fill-in-the-blank questions without options. This conversion was not performed on all multiple-choice questions, but rather based on the following specific selection criteria: only multiple-choice questions whose stems constitute independent, complete questions with unique answers were selected (e.g., questions testing specific concepts, numerical values, or calculations); questions where the stem and options are logically highly coupled and cannot exist independently of the options were excluded (e.g., "Which of the following is correct?" or "True or False" type multiple-choice questions). For the selected qualified samples, the option information was removed using the Gemini model, and the question text was rewritten into a fill-in-the-blank format directly asking for the answer. The conversion results were manually verified to ensure the uniqueness and accuracy of the answers. Finally, all data underwent review by domain experts to verify logical correctness, text and image alignment accuracy, and the accuracy of subject labels for each question. The final approved questions, along with their subject, illustration, corresponding question, answer, and explanation, were packaged into a unified task sample, forming the multimodal computer science illustration benchmark dataset MMCS-Bench, covering four major subjects: data structures, computer organization, computer networks, and operating systems. This dataset contains 901 task samples, each including the subject, illustration, corresponding question, and standard answer. Figure 2 As shown, the task sample formats include multiple-choice questions, fill-in-the-blank questions, and comprehensive questions.

[0024] Multiple choice questions (MCQ): These questions require the model to extract visual information from a single professional illustration and select the correct answer based on subject knowledge. They are used to evaluate the model's local reasoning ability in a fixed visual scene.

[0025] Fill-in-the-blank (FBQ): Derived from MCQ by removing answer options, the question stem and corresponding illustration are the same as MCQ, but the model is required to output the answer directly instead of selecting from the options. It is used to evaluate the robustness of the model's reasoning without option hints, and each FBQ corresponds to a single illustration.

[0026] Comprehensive Problems (CMP): Each CMP describes a complex system scenario supported by 1-3 graphs, accompanied by 1 to 6+ interdependent comprehensive sub-problems. It is used to evaluate the model's system-level logical consistency reasoning ability, meaning the answers to subsequent sub-problems depend on intermediate states obtained from preceding steps. For example... Figure 3 The graph shows the distribution of subproblems within a Comprehensive Problem (CMP), where the y-axis represents the number of subproblems in each comprehensive problem. "6+ subproblems" in the graph indicates a comprehensive problem containing 6 or more subproblems.

[0027] Lightweight knowledge graph units are used to extract entities (computer science concepts), intra-subject entity relationships, and cross-subject general relationships from task samples to construct a lightweight computer science subject knowledge graph. This study employs a BERT-BiLSTM-CRF-based named entity recognition model combined with a computer science and technology terminology dictionary. From 901 task samples containing the stated subjects, professional diagrams, corresponding questions, answers, and explanations, three core knowledge elements are extracted: entities (such as "register," "data bus," "binary tree," and "TCP / IP protocol stack"), intra-subject entity relationships (such as "PC→address bus→memory," "root node→child node→leaf node"), and cross-subject general relationships (such as "data transmission," "control dependency," and "timing synchronization"). Based on the subject tags of the task samples, the knowledge elements are organized into the corresponding subject knowledge graph using the Neo4j graph database. Redundant node merging (such as merging "general-purpose register set" with "GPRs") and association path simplification operations (such as deleting redundant relay nodes) are performed on the subject knowledge graph to reduce its size, resulting in a lightweight subject knowledge graph. These four lightweight subject knowledge graphs are then integrated to form a lightweight computer science professional subject knowledge graph. And the large language model GPT-4o is used to analyze the lightweight computer science subject knowledge graph. Semantic transformation is performed to obtain a lightweight computer science subject knowledge graph. Natural language description.

[0028] The knowledge graph-enhanced pre-inference module, connected to the multimodal graph dataset and lightweight knowledge graph construction module, is used to receive user-input tasks for processing. Computer science diagrams and question text Through cross-modal semantic fusion units, knowledge graph fusion reasoning units, preprocessing result standardization units, and model optimization and tuning units, a standardized preprocessing package is output. .

[0029] Specifically, this includes: a cross-modal semantic fusion unit that integrates a visual language model, a large language model, and a cross-modal deep fusion function; and the use of the visual language model Qwen-VL-max for the task being processed. Computer science diagrams (e.g., a "pipeline CPU data path diagram") undergoes semantic parsing to extract the structural logic and relationships of entities (such as registers and buses) in the diagram, generating a natural language description of the diagram; the large language model GPT-4o is then used to analyze the question text. (e.g., analyzing the execution process of the load instruction) performs semantic understanding to generate a natural language description of the problem text; simultaneously, it invokes a lightweight computer science subject knowledge graph. Computer science diagrams All entities appearing in the graph (i.e., graphical computer science concepts, such as "PC register," "address bus," "memory," and "ALU" in the "pipeline CPU data path diagram") are included in the lightweight computer science subject knowledge graph. Match the corresponding entities (ensuring conceptual consistency) and link these entities with the lightweight computer science subject knowledge graph. The relationships within (such as "PC register - output → address bus" and "address bus - connection → memory") are constructed into a diagrammatic subgraph. The sub-graph was analyzed using the large language model GPT-4o. Semantic transformation is performed to obtain a natural language description of the illustrated sub-map, which accurately maps the topological relationships between hardware entities (such as "unidirectional data flow") and professional semantics (such as "PC register output address is directly connected to memory address port via address bus").

[0030] The cross-modal semantic fusion unit incorporates a cross-modal deep fusion function. This function achieves semantic alignment, logical association, and deep fusion of three types of natural language descriptions. The cross-modal deep fusion function includes a semantic adaptation module, logical association matching, and a deep fusion algorithm. The semantic adaptation module unifies the semantic granularity (refined to the "entity-relationship" level) and expression paradigm (e.g., all using a "subject-action-object" structure) of the natural language descriptions of the graph, the question, and the graph subgraph. Logical association matching establishes correspondences between the three types of natural language descriptions (e.g., recognizing the equivalence between "PC → address bus" in the graph's natural language description and "PC register output to address bus" in the question's natural language description, and associating "analyze load instruction" in the question's natural language description with "memory access phase" in the graph's natural language description and "memory read signal" in the graph subgraph's natural language description). Based on the results of logical association matching, the deep fusion algorithm integrates the three types of semantics, outputting a unified cross-modal fused semantic text. .

[0031] Knowledge graph fusion reasoning unit, used to process tasks Cross-modal fusion semantic text Lightweight Computer Science Discipline Knowledge Graph The natural language descriptions are combined and input into the large model. In this embodiment, the Qwen3-VL-8B-Instruct large model is selected as the base model. Through the attention mechanism built into the large model, cross-modal fusion semantic text is associated. The natural language description in the text calls upon a lightweight computer science subject knowledge graph. Extract entities, intra-subject entity relationships, and cross-subject general relationships from the diagram, complete the "entity extraction (such as the ALU and register modules in the 5th stage of the pipeline)" and "basic logic derivation (such as the flow logic of the load instruction in each stage)", and output the task to be processed. Preliminary reasoning results The completeness of the reasoning results is evaluated using the knowledge gap annotation function built into the knowledge graph fusion reasoning unit. The specific operation of this function is as follows: The preliminary reasoning results are... Natural language description and illustration sub-map The natural language descriptions are input into the BGE-M3 model to generate dense vectors. The semantic matching degree is obtained by calculating the cosine similarity between the two vectors, with the value ranging from [0,1]. When the matching degree is lower than the preset threshold of 0.75, the preliminary inference result is located. The system lacks relevant computer science knowledge and generates knowledge gap tags. This knowledge gap label This includes missing knowledge types (e.g., "risk type"), related entities (e.g., "pipeline pause"), locations (e.g., "logic chain step 3"), and search keywords (e.g., "data risk - forwarding mechanism").

[0032] Preprocessing result normalization unit, used to fuse cross-modal semantic text Preliminary reasoning results Knowledge gap label And the illustrated sub-map The package is standardized preprocessed to provide a unified input for subsequent CS-MMA collaborative architecture modules.

[0033] The model optimization and tuning unit is used to compress the Qwen3-VL-8B-Instruct model to 200M using static knowledge distillation technology, enabling lightweight deployment. LoRA / QLoRA (4-bit NF4 quantization) technology is used for static parameter compression of the model, with the low-rank matrix rank r=16 and scaling factor α=16, reducing memory usage without changing the original pre-trained weights. Finally, TensorRT is used to accelerate inference processing of the optimized model, reducing inference latency and meeting the real-time response requirements of teaching scenarios.

[0034] The CS-MMA collaborative architecture module, connected to the knowledge graph enhanced pre-inference module, includes an adaptation agent unit, a basic agent unit, an advanced agent unit, a specialized agent unit, and a verification agent unit, used to receive standardized preprocessing packages. It then outputs the final reasoning result.

[0035] Specifically, this includes: an adapted Agent unit, built on the lightweight large model Gemini-2.5-Flash, used to receive tasks to be processed. With standardized preprocessing packages From standardized preprocessing packages Preliminary reasoning results were obtained from the analysis. Knowledge Gap Labels ; Invoke the task complexity scoring model built into the adapter agent unit, based on the task Preliminary reasoning results Knowledge gap label and lightweight computer science subject knowledge graph A quantitative score is generated from four dimensions: subject specialization, logical hierarchy, answer output requirements, and diagrammatic complexity. The comprehensive task complexity score is obtained by summing the scores with equal weights.

[0036] The core grading rules for the four scoring indicators are as follows: Subject specialization (maximum score 25 points): Based on entity coverage (the proportion of entities in the diagram in the lightweight computer science subject knowledge graph). The proportion of total entities), relationship depth (the longest path layer between entities), and knowledge gap tags. The core criteria are the degree of knowledge gaps (the proportion of knowledge that needs to be supplemented to the total task requirements) and the number of cross-subject sub-related connections: 21-25 points: Entity coverage of 4.5% or higher, relationship depth of 3 levels or higher. The marked knowledge gap is ≥15%, and it involves the integration principle across 3 or more sub-subjects. The sub-subject integration principle refers to the principle that involves the relationship between two or more subjects.

[0037] 16-20 points: Entity coverage is 3%-4.5%, relationship depth is 2 levels. The marked knowledge gap is 10%-15%, and involves the integration principles of 1-2 sub-subjects.

[0038] 11-15 points: Entity coverage is 1.5%-3%, relationship depth is 1 level. The marked knowledge gap is 5%-10%, and there is no cross-subject correlation.

[0039] 0-10 points: Entity coverage is below 0%-1.5%, and the relationship depth is 1 level. The marked knowledge gap is 0%-5%, and there is no cross-subject correlation.

[0040] Logical level (25 points in total): Based on the preliminary reasoning results Logical consistency (whether there is a conflict) is determined based on the number of reasoning steps required to complete the task (the number of steps that need to be broken down to complete the task) and the depth of logical nesting (the number of logical layers including multiple branches / loops): 21-25 points: There are obvious logical conflicts (such as causal contradictions between steps), the reasoning steps are ≥6, and the number of nested / multi-branched intersections is ≥3.

[0041] 16-20 points: There are a few logical inconsistencies (such as the order of non-core steps being reversed), the reasoning steps are 4-5, and the number of nested logical layers is 1-2.

[0042] 11-15 points: The logic is basically the same (only the details differ), the reasoning steps are 2-3, and there is no logical nesting. 0-10 points: The logic is completely consistent, requiring only one reasoning step (such as direct calculation), with no logical nesting.

[0043] Answer output requirements (25 points in total): Based on knowledge gap tags The need to supplement missing knowledge (the proportion of knowledge to be supplemented to the total task requirements) and preliminary reasoning results The core criteria are completeness (the percentage of existing reasoning that covers the task requirements) and depth of output content (the level of detail in conclusions / steps / principles). 21-25 points: The marked knowledge gap needs to be filled in ≥15% Completeness ≤70%, must output "conclusion, full-step derivation, in-depth principle analysis, accurate diagram annotation, and cross-knowledge point expansion" (such as explaining data transmission logic by combining "TCP congestion control" and "routing table structure"), and must systematically integrate professional knowledge.

[0044] 16-20 points: The knowledge gaps identified in the annotations require supplementation at a rate of 10%-15%. The completeness should be 70%-80%, and the output should include "conclusion, step breakdown and explanation of core principles", and the derivation basis should be explained in relation to relevant knowledge points.

[0045] 11-15 points: The knowledge gaps identified in the annotations require supplementation at a rate of 5%-10%. The completeness is 80%-90%, and the output is "conclusion and breakdown of basic steps", without the need for explanation of the principles.

[0046] 0-10 points: The knowledge gap filling requirement is 0%-5%. Completeness is 90%-100%, only core conclusions / values / keywords need to be output, no explanation of the underlying principles is required.

[0047] Graph Association Complexity (25 points maximum): The structural complexity of the graph corresponding to the task (number of entities in the graph, types of associations) and the graph's subgraphs. The mapping relationship between entities in the diagram and entities in the illustration (the ratio of entities in the illustrated sub-graph matching entities in the illustration) is used as the criterion for judgment: 21-25 points: The number of entities in the diagram is ≥11, including 3 or more types of cross-entity group (comprising multiple entities and modular structural units with clear functional boundaries) associations (by identifying multi-level nested connections, control signals / state transitions in the diagram, and cross-entity group collaborative logic). Less than 70% of the entities in the diagram are mapped to the entities shown in the illustration.

[0048] 16-20 points: The diagram shows 6-10 entities, including 1-2 types of cross-entity group associations. 70%-80% of the entities in the diagram are accurately mapped to the entities shown in the illustration.

[0049] 11-15 points: The diagram shows 2-5 entities, including only those related to the same entity group. 80%-90% of the entities in the diagram are accurately mapped to the entities shown in the diagram.

[0050] 0-10 points: The diagram shows a single entity with no valid relationships. 90% or more of the entities in the diagram are accurately mapped to the entities in the illustration.

[0051] Task levels are determined based on thresholds: 0-30 points are easy, 31-70 points are medium, and 71-100 points are complex.

[0052] Real-time acquisition of real-time status sets of basic agent units, advanced agent units, and professional agent units within the system. (Including load status and computing power usage). Driver instructions are generated based on task level: simple tasks are assigned to basic agent units, medium tasks to advanced agent units, and complex tasks to specialized agent units. To improve efficiency, a dynamic load fine-tuning strategy is introduced. This strategy monitors the overall load level of various agent units in real time. When the load of a large model in an advanced agent unit exceeds the 80% threshold, tasks with complexity ratings between "simple" and "medium" are scheduled to basic agent units with lower loads, and tasks with complexity ratings between "medium" and "complex" are scheduled to specialized agent units with lower loads, thus distributing the load. To ensure fairness and symmetry in load balancing, when the load of a basic or specialized agent unit exceeds the threshold, fine-tuning is also triggered, scheduling tasks in its queue with complexity ratings adjacent to the medium task boundary to advanced agent units with relatively lower loads, thereby avoiding overload of a single node. Driver instructions directly drive agent units matched to the task level, ensuring collaborative efficiency.

[0053] The basic Agent unit, built upon the large model Qwen3-VL-30B-A3B-Instruct, is used to receive driver instructions and tasks to be processed from the adapted Agent unit. and its standardized preprocessing package And based on standardized preprocessing packages Cross-modal fusion semantic text encapsulated in Computer science diagrams After entity recognition is completed, the unit calls the standardized preprocessing package. The diagram sub-map encapsulated in the middle The entities in the identified diagram and the diagram sub-maps Entity matching is performed on entities within the text. This process includes: fusing semantic text across modalities. Extracted natural language descriptions of the obtained illustrations, and use them to identify computer science illustrations. The entities in the graph (such as ALUs, linked list nodes) and their corresponding subgraphs are shown in the diagram. The entities (such as "arithmetic logic unit") and relationships (such as "PC register → address bus → memory") in the graph are compared; based on the above comparison, the consistency between the identification results and the illustrated subgraph is verified. This includes: verifying the entities in the graph within the illustrated subgraph. Does the diagram contain corresponding entity definitions (entity existence)? Verify the relationships between entities in the diagram's subgraph. Is there a clear definition (relationship explicitness)? If, during entity matching, one or more of the following are identified: entity missing or relationship missing, then directly from the lightweight computer science subject knowledge graph... In the process, the corresponding entity definition or relationship is retrieved for completion, the results of recognition, matching and completion are integrated, and the reasoning result is generated and output.

[0054] The advanced Agent unit, built on the medium-sized large model Claude-Sonnet-4-5-20250929-thinking, is used to receive driving instructions and tasks to be processed from the adapted Agent unit. and standardized preprocessing packages Its core work revolves around analyzing the logical links shown in the diagram, supplementing knowledge, and refining the reasoning results.

[0055] This unit performs logical link parsing. It is based on a standardized preprocessing package. Cross-modal fusion semantic text encapsulated in With the illustration of the sub-map Treating and handling tasks Computer science diagrams The logical link parsing process includes: First, performing entity anchoring from the entity in the diagram to the entity in the diagram's subgraph, based on cross-modal fusion semantic text. Extracted natural language descriptions of the obtained illustrations, and use them to identify computer science illustrations. The entities in the diagram, and the sub-graphs shown. Semantic matching is performed on entities in the graph to obtain a set of matching entities in the graph and entities in the graph subgraph; then, logical link mining of the matching entities is performed, and a graph neural network (GNN) is called to process the graph subgraph. Multi-hop relationship propagation is performed to mine the topological relationships and directed data flows between entities, and an initial logical link is constructed. Next, relational semantic generation of the graph subgraph is executed, and the graph subgraph is processed using the large model Qwen3-VL-8B. The relationships in the code are transformed into natural language logical descriptions. Finally, a standardized preprocessing package is used. The logical syntax and explicit rules (such as field format and logical constraints) of the initial logical link are verified against the aforementioned natural language logical description, and finally the relationships and basic operational logic between entities in the diagram are identified.

[0056] Based on the relationships between entities and the basic operational logic obtained from the logical link parsing diagram, combined with standardized preprocessing packages... Knowledge gap tags encapsulated in the middle From a lightweight computer science discipline knowledge graph Targeted retrieval of core application layer knowledge points, and standardized preprocessing packages Preliminary inference results encapsulated in the middle Supplementing knowledge. This includes identifying knowledge gap tags. Includes: Types of missing knowledge: Clear preliminary reasoning results The categories of knowledge to be supplemented (such as core course principles, multi-step logic chains, cross-module related logic, practical application scenarios and functional meanings) define the scope of knowledge retrieval; related entities: locate the entity nodes of the diagrammatic subgraph related to the knowledge gap (such as ALU, TCP protocol stack, page table entries, etc.) to anchor the related objects for knowledge supplementation; location information: mark the knowledge gap in the diagram. Initial logical link or preliminary reasoning results The system provides precise location information for data completion; it also offers a lightweight computer science subject knowledge graph for keyword search. The search entry point enables precise matching of knowledge points. Core application layer knowledge points include: core course principles covering pipelining hazards, hash collisions, TCP congestion control, virtual memory management, etc.; multi-step logic chains covering the complete data path process, OS seven-layer model call chain, page table traversal process, etc.; cross-module correlation logic covering the relationship between ALU and pipeline, protocol stack layer linkage, process scheduling and interruption, etc.; practical application scenarios and functional meanings covering the meaning of page table entries, routing table structure, red-black tree characteristics, etc.

[0057] Based on the relationships between entities and the basic operational logic obtained from the aforementioned logical link analysis, the preliminary reasoning results are... The process involves breaking down and verifying the initial reasoning results step by step. The logical link is broken down hierarchically. Based on the relationships between entities and the basic operational logic shown in the diagram, it is divided into single-step reasoning units (e.g., given attribute A → deduce conclusion B; entity X satisfies rule Y → determine result Z), entity relationship links (referring to connecting scattered single steps into a complete path based on the relationships between entities and the basic operational logic shown in the diagram), and subject knowledge reference segments (lightweight computer science subject knowledge graph). The results are categorized into three types: (content fragments being called and matched), and (explicit rule judgments and logical consistency verification based on breakpoint features). This involves verifying field completeness, consistency of reasoning steps within a single subject, accuracy of concept citations, matching of diagram annotations and text descriptions, and matching of knowledge gap tags with entities. The location and type of explicit logical breakpoints are then identified. If explicit logical breakpoints are identified during the verification process, the lightweight computer science subject knowledge graph is used as the basis for further analysis. Retrieve relevant knowledge for logical connection; combine the retrieved relevant knowledge with the preliminary reasoning results. To integrate and complete.

[0058] The explicit logical breakpoint types include one or more of the following: incomplete fields, inconsistencies in reasoning steps within a single subject, incorrect concept references, conflicts between diagram annotations and text descriptions, and matching breaks between knowledge graph entities and reasoning entities.

[0059] Specifically, incomplete fields: This type of breakpoint is characterized by... Some fields only present partial content, lacking complete field definitions, constraints, or technical details (e.g., only describing "hash collision resolution" but not clarifying the core logic of "chaining"). Completing this requires clarifying the knowledge category of the incomplete fields based on the underlying operational logic, starting from a lightweight computer science subject knowledge graph. The standard format, complete definition, and technical constraints of this field are retrieved. Subsequent analysis of the preliminary inference results... The fields in the document are completed, corrected, and standardized to ensure that they meet the requirements of computer science teaching content.

[0060] Inconsistent reasoning steps within a single subject: This type of breakpoint is characterized by logical sequence errors, process conflicts, or violations of the subject's underlying operational rules in the reasoning steps within the same subject (e.g., in CPU pipelined reasoning, placing "pipeline hazard handling" after "instruction execution" violates the pipeline execution sequence). Completion requires relying on the initial logical link topology from the logical link analysis, calling a graph neural network (GNN) to backtrack the relationships between entities in the diagram, and utilizing a lightweight computer science subject knowledge graph. KG The standard multi-step logic chain of the subject is retrieved (such as page table traversal process and complete data path process); subsequently, contradictory steps are reordered and logically reconstructed, erroneous reasoning steps are replaced, logical conflicts between steps are eliminated, and the reasoning chain returns to a self-consistent state consistent with the basic operational logic.

[0061] Conceptual reference error: This type of breakpoint is characterized by... There are conceptual reference errors and incorrect use of technical terminology (e.g., incorrectly referring to "heap (priority queue)" as "ordinary array" to describe the underlying implementation of OS process scheduling). Completing this requires locating the application scenarios of the incorrect concepts based on the relationships between entities in the diagram, and drawing upon a lightweight computer science subject knowledge graph. The standard definition, core characteristics, and applicable rules of this concept are retrieved from the database. Subsequent replacements... To correct any erroneous concepts, we will supplement the explanations of the underlying principles, common pitfalls, and application boundaries of each concept, and rectify any deviations in the use of these concepts to ensure their professionalism and accuracy.

[0062] Conflicts between diagram annotations and text descriptions: This type of breakpoint characteristic is found in computer science diagrams. entity annotation content and The textual logical descriptions are inconsistent (e.g., in the diagram, the ALU's input ports are labeled "instruction flow and data flow," but the text description only mentions "instruction flow"). Completing this requires cross-modal fusion semantic text that relies on logical link parsing. Re-anchor the entity nodes in the diagram, from the lightweight computer science subject knowledge graph The standard annotation rules and corresponding text description for the entity are retrieved. Subsequently, bidirectional alignment corrections are performed on the illustration annotations or text content to unify the semantic expression of the entity with the illustration presentation, achieving consistency between the semantics of the text and the illustration, and eliminating conflicting breakpoints.

[0063] Discontinuity in matching knowledge graph entities with inference entities: This type of discontinuity is characterized by lightweight computer science subject knowledge graphs. The entity nodes in the data, and the preliminary reasoning results In the inference entities, a valid correspondence cannot be established (e.g., the inference link mentions "routing table structure," but the lightweight computer science subject knowledge graph). (No corresponding entity is associated with this). Completion requires a diagrammatic subgraph with logical link parsing. For reference, the semantic matching degree between the reasoned entities and the graph entities is re-verified, and the entity relationships are updated based on the relationships between entities in the graph. Subsequent work will focus on lightweight computer science subject knowledge graphs. The system retrieves relevant knowledge for matching, completes the correspondence between entities, eliminates gaps in entity matching, and ensures the semantic connectivity between inference entities and the subject knowledge graph.

[0064] Based on the results of logical link parsing, knowledge supplementation, and fusion completion, it generates and outputs reasoning results that include principle explanations, step breakdowns, basic error-prone tips, and key link annotations in diagrams. Balancing reasoning accuracy and response efficiency, it eliminates the need for in-depth principle analysis and cross-subject knowledge connections, meeting the needs of routine teaching support such as after-class homework tutoring and in-depth explanations of core knowledge points. It possesses professional reasoning and content generation capabilities supported by moderate computing power.

[0065] The professional agent unit is built on a dual-model architecture, including the GPT-5 model for handling multiple-choice and fill-in-the-blank questions, and the GLM-4.6V model for handling comprehensive application questions.

[0066] The professional agent unit receives driving instructions and tasks to be processed from the adapted agent unit. and standardized preprocessing packages Based on the task format determination result, the appropriate model is selected and invoked. The invoked model then fuses semantic text across modalities. Computer science diagrams This involves a comprehensive analysis encompassing entity nesting relationships, control signal annotations, and system state mapping. Entity nesting relationships refer to the inclusion, subordination, and hierarchical relationships between entities in the diagram. For example, in computer organization, the CPU contains an ALU (Arithmetic Logic Unit), which in turn contains adders. In data structures, the root node of a binary tree contains a left subtree and a right subtree, and each subtree contains leaf nodes. Control signal annotations refer to the signal identifiers (usually arrows, enable signals, clock signals, etc.) used for instruction transmission, state switching, and data flow in the diagram. For example, in computer networks, router routing signals and packet forwarding instructions. System state mapping refers to deriving the system's operating state at different times based on input signals and entity states in the diagram, and mapping abstract logic to physical diagrams. For example, in computer organization, a state machine switches from an "idle state" to a "running state" based on input conditions.

[0067] Based on knowledge gap tagging From a lightweight computer science discipline knowledge graph The process involves retrieving deep-seated fundamental principles, which form the underlying operational logic and basic theoretical rules supporting the reasoning in the questions. These principles include, but are not limited to: within a single subject, the underlying mechanisms for resolving hash collisions in data structures (chaining, open addressing); and the hazard mechanism of CPU instruction pipelines in computer organization principles. Across subjects, the process flow transitions from data structures to operating systems, from the operating system's process scheduling queue to the underlying implementation of priority queues (heaps) in data structures. The process also extends from computer networks to operating systems, from TCP congestion control to operating system flow control and buffer management. Based on the general relationships across subjects, related cross-subject knowledge is retrieved to refine the preliminary reasoning results. Further details are provided. Based on in-depth fundamental knowledge, cross-subject related knowledge, and comprehensive analysis results, long-chain nested logical reasoning and system-level consistency verification are performed: Long-chain nested logical reasoning, based on comprehensive analysis results and in-depth fundamental knowledge, extends single-step reasoning into a complete logical chain covering the dynamic interaction sequence of entities; System-level consistency verification, based on cross-subject related knowledge, performs cross-subject global verification to verify the compatibility of multi-subject knowledge and the reachability of system state.

[0068] Simultaneously identify the preliminary reasoning results The implicit logical conflicts within the framework are addressed by using deep-seated theoretical knowledge as a foundation and cross-disciplinary knowledge as a logical bridge to refine the preliminary reasoning results. Conflict resolution is carried out to ensure the logical coherence of multi-sub-problems and multi-step reasoning.

[0069] Among them, implicit logical conflicts include three types: cross-subject contradictions, inconsistencies in mechanisms, and unreasonable frameworks. Specifically, cross-disciplinary contradictions refer to conflicts in the technical logic or operational rules of different subjects during the reasoning process, violating the underlying physical or logical facts of the computer system. Simply put, it involves incorrectly applying knowledge from subject A to subject B, resulting in their fundamental incompatibility. Solution: Based on deep-seated theoretical knowledge, and combined with interdisciplinary connections built from related cross-subject knowledge, accurately pinpoint the initial reasoning result. In the case of contradictions that violate the underlying physical or logical facts of computer systems (such as incorrectly using ordinary arrays to implement process scheduling priority queues), we directly discard erroneous associations that cause conflicts between technical logic or operating rules, and reconstruct associations that conform to the underlying physical or logical facts of computer systems based on cross-subject general relationships, thereby eliminating cross-subject contradictions at their root.

[0070] Inconsistent mechanisms refer to logical flaws in the reasoning chain within a single subject or entity group, causing the system's operation to be inconsistent. Although the syntax of the statements is grammatically correct, crucial processing flows or state transitions are omitted when executing specific hardware or software mechanisms. Solution: Address the initial reasoning results... Logical flaws in the reasoning chain (such as the handling of omissions and hazards in CPU pipeline reasoning) can be addressed without reconstructing the entire framework. Instead, the logical flaws can be located, and the corresponding deep-level principles and cross-disciplinary knowledge can be retrieved to directly fill the gaps and repair the initial reasoning results in situ. To restore its internal consistency.

[0071] An unreasonable framework refers to implicit contradictions between multi-branch reasoning paths, or an overall reasoning chain that contradicts the actual scenario illustrated in computer science diagrams. Simply put, it's not a localized error or logical flaw, but a deviation in the overall planning direction (e.g., designing a "single-threaded serial architecture" reasoning scheme for a "high-concurrency multi-threaded scheduling" scenario). Solution: Address the initial reasoning results... The implicit contradictions between different branch reasoning paths (e.g., for the same state machine, one path deduces it as being in a "running state," while another path deduces it as being in a "dormant state") contradict the actual scenario illustrated in computer science diagrams and are preliminary reasoning results that cannot be locally repaired. The overall chain (such as adapting a single-threaded architecture to high-concurrency scheduling) is traced back to the source node of the conflict. Based on deep principles and combined with cross-disciplinary related knowledge, the reasoning direction is replanned, and a new preliminary reasoning result that conforms to the actual scenario illustrated in computer science diagrams is reconstructed. The reasoning path eliminates implicit contradictions and errors in the overall reasoning chain from a macro perspective.

[0072] Based on the retrieved deep-seated fundamental knowledge, the results of long-chain nested logical reasoning and system-level consistency verification, and the preliminary reasoning results. The elimination of implicit logical conflicts results in a complete teaching reasoning outcome that includes in-depth analysis of underlying principles, detailed derivation of all steps, multi-dimensional illustrations, cross-knowledge point connections, and system-level risk warnings.

[0073] The Validation Agent Unit, built upon the large model Gemini-3-Pro-Preview, receives inference results and standardized preprocessing packages from the Basic Agent Unit, Advanced Agent Unit, and Specialized Agent Unit. The reasoning output of agents at all levels is quantitatively verified in all dimensions to ensure the professionalism, accuracy and logical coherence of the reasoning results, while strengthening the rigor of long-chain logical reasoning.

[0074] The verification agent unit is configured with preset quantization accuracy thresholds matching each functional agent unit, requiring an accuracy rate of ≥T% for graphical analysis, ≥X% for inference logic, ≥K% for knowledge matching, and ≥E% for solution content coverage. The verification process and rules are fully adapted to the requirements of computer science education: receiving the inference results output by upstream functional agent units (basic agent unit, advanced agent unit, and professional agent unit), and combining them with standardized preprocessing packages. Knowledge gap tags Illustration of sub-map and lightweight computer science subject knowledge graph The reasoning results are checked item by item. If the check results reach the corresponding quantization accuracy threshold, the final reasoning result is directly output.

[0075] If the verification result does not reach the corresponding quantification accuracy threshold, the error source of the substandard inference result is traced and a precise correction suggestion is generated. At the same time, the verification result is synchronized to the adaptation agent unit and the upstream functional agent unit to achieve cross-layer bidirectional feedback. The verification result is also sent back to the cross-modal semantic extraction and fusion unit and the model optimization and tuning unit of the knowledge graph enhanced pre-inference module to achieve linkage between feature encoding optimization and model tuning. Finally, the corrected and integrated content is transformed into the final inference result with the same content structure as the output content of the corresponding functional agent unit.

[0076] The system verification module, connected to the CS-MMA collaborative architecture module, is used for comprehensive verification and iterative optimization of the system's overall performance, inference accuracy, and teaching adaptability. The system verification module executes a layered verification process combining automated testing and real-world testing: Automated testing is performed based on the multimodal computer science knowledge graph benchmark dataset MMCS-Bench to batch verify the overall system performance, including three core indicators: answer accuracy, explanation logic completeness, and utilization of graph information; Real-world testing is then conducted based on real teaching cases for each subject, collecting feedback from actual usage in teaching scenarios; Finally, based on the automated testing data and real-world user feedback, the system model, the CS-MMA collaborative architecture module, and the lightweight computer science knowledge graph are evaluated. The content undergoes targeted iterative optimization. This further improves the system's performance and its adaptability to instruction, while simultaneously embedding validated optimization rules and parameters into the system, forming a closed loop of "validation-feedback-iteration-optimization." Optimization rules refer to the logic, strategies, and judgment criteria summarized from validation that guide the system's optimization. Optimization parameters refer to the configurable numbers, thresholds, and switches determined after validation to control system behavior. Effective improvement methods and optimal configurations summarized during validation are fixed into the system, allowing it to utilize optimized experience.

[0077] The specific implementation process is as follows: First, automated testing is performed, and three core quantitative evaluation dimensions are set: answer accuracy (number of correct tasks / total number of tasks × 100%), explanation logic completeness (number of knowledge points covered / number of knowledge points in the standard answer × 100%), and illustration information utilization rate (number of extracted visual features / total number of key features × 100%). The multimodal computer science illustration benchmark dataset MMCS-Bench is used as test data to verify the core inference accuracy of the system.

[0078] Further on-site testing and verification were conducted, with 200 computer science and technology students and 10 professional teachers participating in the testing to comprehensively collect feedback on actual use in teaching scenarios. The specific collection content included: (1) core quality feedback on the output results of each functional agent (accuracy of answer, completeness of explanation logic, utilization rate of diagram information, with a focus on the cross-subject reasoning of professional agent units and the accuracy of logical link analysis of advanced agent units); (2) feedback on the rationality of hierarchical scheduling and reasoning output of upstream functional agent units (such as whether the complexity classification of the adapting agent is accurate, the effectiveness of the second-level response of the basic agent, the adaptability of the dual-model switching answer of the professional agent, and the accuracy of error tracing and correction suggestions of the verification agent); (3) feedback on system response speed and ease of operation; the collection method adopted a combination of automatic recording of usage logs, teacher and student questionnaires and random interviews to ensure that the feedback was comprehensive and targeted.

[0079] Experimental verification To verify the technical effectiveness of the knowledge graph-based and multi-agent collaborative intelligent teaching system (CS-MMA), the MMCS-Bench dataset, a multimodal computer science subject graph benchmark dataset, was used as a unified evaluation benchmark. Three sets of comparative experiments were designed: performance comparison with mainstream large models, ablation experiments with large models, and quantitative testing of inference efficiency. All calls to large models were made via API. The MMCS-Bench dataset contains 901 task samples covering four major subjects: computer organization, data structures, computer networks, and operating systems. It employs three task formats and a five-dimensional inference ability classification system. All experiments were conducted under zero-sample conditions, with the temperature parameter fixed at 0.

[0080] (1) Division of reasoning dimensions To rigorously quantify the model's inference requirements, this invention employs a fine-grained classification method for each question in the dataset, dividing them into five complementary inference capability dimensions. The same question may involve multiple capabilities simultaneously, requiring the model to progressively advance from basic structural understanding to higher-order system-level logical reasoning. These capability dimensions are summarized using an inference capability matrix. Figure 4 As shown.

[0081] The accuracy rate of the corresponding reasoning ability = the total score of the large model on questions with the corresponding reasoning ability label / the total number of questions with the corresponding reasoning ability label.

[0082] Structural reasoning capability is used to understand spatial layout, hierarchical organization, and the relationships between nodes and links. This capability has a 97.6% coverage rate in Data Structures Analysis (DSA), where tasks such as tree traversal and graph structure analysis are very common, making it suitable for typical task scenarios such as tree traversal and graph structure analysis.

[0083] Timing reasoning capability is used to trace the execution order of timing events or protocol execution sequences. This capability has a coverage rate of 48.3% in Computer Networks (CN) for network handshake process analysis, and a coverage rate of 50.7% in Computer Organization (CO) for pipeline timing and instruction scheduling. This capability is significantly demonstrated in relevant scenarios in both subjects.

[0084] State mapping reasoning capability is used to maintain and transform the state of an abstract logical space, such as finite state machine transformations or mapping from virtual addresses to physical addresses. This capability has a coverage rate of 82.2% in operating systems (OS) and 73.2% in computer networks (CN). In these subjects, the state of dynamic systems must be continuously updated and maintained.

[0085] Computational reasoning ability is used to extract relevant numerical information from charts and graphs and combine them with subject-specific formulas to complete logical operations or numerical calculations, such as throughput calculation and Hamming distance calculation.

[0086] System-level consistent reasoning capability is used to ensure that a model maintains consistency of its internal state during multi-step continuous reasoning. This capability has a coverage rate of 85.7% in Computer Organization and 70.2% in Operating Systems. It is particularly relevant in synthesis problems involving dependent states that require tracking dynamic changes in hardware resources.

[0087] (2) Scoring mechanism In the evaluation scheme design, multiple-choice and fill-in-the-blank questions adopt a unified automated evaluation protocol. For multiple-choice questions, rule-based scoring is applied using regular expressions to verify exact matches with basic facts, awarding 1 point for each correct match. For fill-in-the-blank questions, due to the semantic variations in the answers, exact string matching is insufficient; therefore, GPT-4o-mini is used as an automatic judge to evaluate the scientific semantic equivalence between the model output and the reference answer. The maximum score for each fill-in-the-blank question is 1 point, with scores evenly distributed across multiple blanks in a single item. Multiple-choice question score rate = total multiple-choice score / total number of multiple-choice questions; Fill-in-the-blank question score rate = total fill-in-the-blank score / total number of fill-in-the-blank questions.

[0088] The comprehensive question assesses system-level reasoning through interdependent sub-problems requiring long-term context state maintenance. To score these complex responses, DeepSeek-V3.2 with a keypoint-based quantitative scoring mechanism is employed. The reference solution is decomposed into discrete keypoints, allowing the scoring model to evaluate semantic coverage and accumulate partial credit, thereby mitigating subjective bias. For the overall score, a sample average normalization strategy is used, with a maximum score of 1 point per question. For comprehensive questions containing N sub-problems, each sub-problem is assigned a weight of 1 / N, effectively distinguishing expert-level reasoning ability. The specific scoring rules for the comprehensive question are as follows: Basic scoring system: This problem has N subproblems, each worth 10 points.

[0089] Scoring Point Breakdown Logic: Observe the standard answer. If the answer contains a semicolon (“:” or “;”), it means that the answer consists of multiple key scoring points; the score for each scoring point = 10 points / the number of scoring points for that question. For example, if the standard answer is “A:B”, then A accounts for 5 points and B accounts for 5 points.

[0090] Scoring criteria: Students' answers do not need to be word-for-word identical to the standard answer; semantic matching is sufficient for scoring. Students must include all scoring points to receive full marks (10 points). If only some scoring points are included, the scores for those parts will be added together.

[0091] Final score = Total score / (10) N).

[0092] Comprehensive question score rate = Total comprehensive question score / Total number of comprehensive questions; The percentage of correct answers on all comprehensive and major questions = the total number of correct answers on all comprehensive and major questions / the total number of comprehensive and major questions; The percentage of correct answers to all sub-questions in the comprehensive question = the total number of correct answers to all sub-questions in the comprehensive question / the total number of comprehensive questions.

[0093] Experiment 1: Performance Comparison of CS-MMA and Mainstream Large Models In the performance evaluation experiments of mainstream large-scale models, 13 models spanning from 7B to ultra-large parameter scales were selected, including the open-source models DeepSeek-VL-7B-Chat, Qwen3-VL-8B-Instruct, Qwen3-VL-235B-A22B-Instruct, Qwen3-VL-30B-A3B-Instruct, and GLM-4.6V, as well as proprietary models Gemini-3-Pro-Preview, Gemini-2.5-Flash, Claude-Sonnet-4-5-20250929, Claude-Sonnet-4-5-20250929-thinking, GPT-5, GPT-4o, Doubao-1.5-Vision-Pro, and Qwen-VL-max. Zero-sample testing was conducted on these models under the same evaluation conditions as the CS-MMA model of this invention.

[0094] The experimental results are shown in Tables 1-3. The overall scores of the 13 mainstream large-scale models ranged from 17.66% to 69.02%, with the optimal model, Gemini-3-Pro-Preview, achieving a score of 69.02%. In contrast, the overall score of the CS-MMA model presented in this invention reached 81.02%, a 12 percentage point improvement over the optimal mainstream model. Figures 5-6As shown, in terms of task format, fill-in-the-blank questions are derived from multiple-choice questions by removing answer options, and option removal reduces model performance. Removing answer options in fill-in-the-blank tasks leads to a significant and consistent performance drop across all models. On average, compared to multiple-choice questions, large models experience a 10%-20% decrease in scores on fill-in-the-blank questions. For example, GPT-5 saw a 18.65% decrease, from 84.39% to 65.74%. While CS-MMA's performance also slightly decreased after option removal, its score dropped by approximately 9% compared to mainstream large models, indicating that CS-MMA provides more accurate knowledge retrieval. Across all subjects, Computer Organization remains the most robust without answer options (e.g., Claude-Sonnet-4-5-20250929 achieves 87.08% in this subject), while Operating Systems and Computer Networks suffer greater declines due to their greater reliance on accurate system state tracking. The overall score rate of mainstream large-scale models decreased by 25% to 45% compared to multiple-choice questions. Although Claude-Sonnet-4-5-20250929-thinking improved by 8.99% (77.78% vs. 68.79%) in multiple-choice questions compared to Claude-Sonnet-4-5-20250929, it actually decreased by 3.36% in overall questions. Compared to other large-scale models, CS-MMA showed a significant improvement in the overall question accuracy rate and the overall sub-question accuracy rate, indicating that CS-MMA has a significantly improved ability to maintain logical consistency in long and interdependent reasoning chains. In terms of subject dimensions, the mainstream large-scale models achieved the best score of 92.59% on multiple-choice questions in Computer Organization Principles, but only 24.58% on comprehensive questions in Operating Systems. In Data Structures, Gemini-3-Pro-Preview's comprehensive question score plummeted from 86.79% to 31.18% compared to the multiple-choice score. In contrast, the CS-MMA of this invention outperformed the best mainstream large-scale models in both multiple-choice and comprehensive questions across all four major subjects of Computer Science. Figure 7A and Figure 7B As shown, in terms of fine-grained reasoning ability, the large model GLM-4.6V achieved an accuracy of 59.2% in the system-level consistency dimension, while the CS-MMA of this invention outperforms mainstream large models in all five dimensions of reasoning ability.

[0095] Table 1. Score Rate of Large Models, Large Model Combinations, and CS-MMA in Multiple-Choice Questions in Computer Science Disciplines

[0096] Table 2. Score Rates of Large Models, Large Model Combinations, and CS-MMA in Comprehensive Questions of Computer Science Disciplines

[0097] Table 3. Score Rate of Large Models, Large Model Combinations, and CS-MMA in Computer Science Disciplines

[0098] Experiment 2: Ablation Experiment Combining CS-MMA and Large Model Four mainstream high-performance large-scale models—Qwen3-VL-8B-Instruct, GLM-4.6V, GPT-5, and Gemini-3-Pro-Preview—were simply superimposed to form a large-scale model combination. This combination only synthesized the answers from the four models as the final answer; the remaining evaluation conditions and procedures remained completely consistent with the 13 mainstream large-scale models and CS-MMA, as shown in Tables 1-3. The overall score rate of the large-scale model combination was 75.58%, an improvement of 6.56% compared to the best single large-scale model, Gemini-3-Pro-Preview, but the performance improvement was not significant. The overall score rate of CS-MMA was 81.02%, and the score rates for multiple-choice questions, comprehensive questions, and the average score in each subject were significantly higher than those of the mainstream large-scale models and the large-scale model combination. Figure 5 and Figure 6 As shown, from the task format analysis, CS-MMA's score rate dropped by about 9% after option removal, far lower than the average drop of 10%-20% for mainstream large models. The drop in the combined large model score rate was consistent with that of a single large model, indicating no improvement in the reliance on option discrimination. For comprehensive questions, CS-MMA achieved significant improvements in both the overall correct score rate for comprehensive questions and the overall correct score rate for sub-questions. While the combined large model score rate slightly improved compared to the optimal single large model GLM-4.6V, it lagged significantly behind CS-MMA. Figure 7A and Figure 7B As shown, from the perspective of fine-grained reasoning ability analysis, CS-MMA is in a leading position in all five dimensions of structure, timing, state mapping, computation, and system consistency, especially in the state mapping dimension, where it has made significant improvements. However, the performance of large model combinations in each dimension is only slightly better than the best single large model, and it has not broken through the capability bottleneck.

[0099] Experiment 3: Reasoning Efficiency Test In addition to inference accuracy, this invention also quantitatively tests the inference efficiency of 13 mainstream large models, combinations of large models and CS-MMA. Taking all 901 task samples in the MMCS-Bench dataset as the unified inference object, under the same hardware deployment environment (graphics workstation configured with NVIDIA A100 GPU and 128 GB memory), the actual time taken by each model to complete the inference of all questions is statistically analyzed. Test results show significant differences in the time taken by the 13 mainstream large-scale models to complete all questions. Specifically, Deepseek-VL-7B-Chat took 3.0 hours, Qwen3-VL-8B-Instruct took 3.2 hours, Qwen3-VL-30B-A3B-Instruct took 3.8 hours, Qwen3-VL-235B-A22B-Instruct took 4.4 hours, Qwen-VL-max took 3.5 hours, Doubao-1.5-Vision-Pro took 3.4 hours, GLM-4.6V took 3.6 hours, and Gemini-2.5-F took [the remaining time is missing from the original text]. The inference time was 3.3h for lash, 3.7h for Claude-Sonnet-4-5-20250929, 3.9h for Claude-Sonnet-4-5-20250929-thinking, 3.8h for GPT-4o, 4.1h for GPT-5, and 4.1h for Gemini-3-Pro-Preview, with the overall time ranging from 3.0h to 4.4h. Inference time was positively correlated with the model's computing power, with heavy large models having significantly longer inference times than light large models. Among them, the Qwen3 series models gradually increased inference time as the parameter scale increased, and the overall inference time was relatively long. Large model combinations require driving four large models to work in parallel, resulting in an extremely large workload for processing a single task and a significantly longer reasoning time for the entire set of questions compared to a single heavy-duty large model. This leads to a severe waste of computing power and extremely poor reasoning efficiency. In contrast, the CS-MMA proposed in this invention completes the reasoning of all 901 task samples in just 2 hours and 23 minutes, demonstrating a significant improvement in reasoning efficiency compared to 13 mainstream large models. These results show that CS-MMA maintains high reasoning accuracy while also possessing superior reasoning efficiency, enabling rapid response in teaching scenarios and making it more suitable for the time-sensitive teaching support needs of computer science majors, such as classroom Q&A and after-class homework tutoring.

[0100] Based on the combined results of the three experiments, although the large model combination integrates four CS-MMA core base models, it has not achieved a performance breakthrough and suffers from wasted computing power and low efficiency. In contrast, CS-MMA, through a collaborative mode of multi-agent hierarchical collaboration, dynamic scheduling, and precise matching, has achieved significant improvements in inference accuracy, task format robustness, fine-grained inference capability, and inference efficiency, verifying the core technical value of this collaborative mode compared to the simple superposition of multiple models.

[0101] The above embodiments are only used to illustrate the technical solutions of the present invention, and are not intended to limit it. Although the present invention has been described in detail with reference to the foregoing embodiments, those skilled in the art should understand that modifications can still be made to the technical solutions described in the foregoing embodiments, or equivalent substitutions can be made to some of the technical features. Such modifications or substitutions do not cause the essence of the corresponding technical solutions to deviate from the spirit and scope of the technical solutions of the embodiments of the present invention.

Claims

1. A smart teaching system based on knowledge graphs and multi-agent collaboration, characterized in that: include: The module for constructing a multimodal graph dataset and a lightweight knowledge graph is used to build the MMCS-Bench multimodal computer science subject graph benchmark dataset, which includes the subject and major graph of each question, the corresponding question, the answer, and the explanation. Based on the MMCS-Bench multimodal computer science subject graph benchmark dataset, a lightweight computer science subject knowledge graph is generated. ; The knowledge graph-enhanced pre-inference module, connected to the multimodal graph dataset and lightweight knowledge graph construction module, is used to receive user-input tasks to be processed. Computer science diagrams and question text Through cross-modal semantic fusion units, knowledge graph fusion reasoning units, preprocessing result standardization units, and model optimization and tuning units, a standardized preprocessing package is output. The standardized preprocessing package Encapsulated with: cross-modal fusion semantic text vectors Preliminary reasoning results Knowledge gap label and the illustrated sub-map ; The CS-MMA collaborative architecture module, connected to the knowledge graph enhanced pre-inference module, includes an adaptation agent unit, a basic agent unit, an advanced agent unit, a professional agent unit, and a verification agent unit, used to receive the standardized preprocessing package. And output the final reasoning result; The adaptation agent unit includes a task complexity scoring model, used for knowledge gap labeling. and preliminary reasoning results The task level is quantitatively scored, and the task level is determined as simple, medium or complex. The driving instructions are then generated and output. The basic agent unit, advanced agent unit, and professional agent unit are respectively used to receive driving instructions for simple, medium, and complex tasks and perform differentiated reasoning. The verification Agent unit is used to receive the output results of each Agent unit and perform verification. The system verification module, connected to the CS-MMA collaborative architecture module, is used to verify and iteratively optimize system performance.

2. The intelligent teaching system based on knowledge graphs and multi-agent collaboration according to claim 1, characterized in that, The multimodal graph dataset and lightweight knowledge graph construction module include: The data collection and preprocessing unit is used to select visual problems that cannot be solved without explaining the relevant charts from computer science teaching resources as initial candidate samples. For the initial candidate samples, the unit captures the corresponding professional illustrations and identifies them according to a naming protocol. The unit manually aligns the professional illustrations with the corresponding question text, options, answers, and explanation text to establish a mapping relationship. Based on the mapping relationship, the unit uses multimodal large model processing to generate structured data containing the professional illustrations, question text, options, answers, and explanations. The structured data is then proofread and output. The structured sample construction unit is used to remove duplicate samples and filter low-quality data by semantic similarity calculation of the structured data; the multiple-choice sample is converted into fill-in-the-blank sample without options by a large model and the uniqueness of the answer is ensured by manual verification; the subject, major illustration, corresponding question, answer and analysis of each processed question are packaged into a unified task sample to form a multimodal computer science illustration benchmark dataset MMCS-Bench covering the four major subjects of computer science. The lightweight knowledge graph unit extracts three core knowledge elements from the task samples: entities, intra-subject entity relationships, and cross-subject general relationships. Based on the subject tags of the task samples, the knowledge elements are organized into the corresponding subject knowledge graph using the Neo4j graph database. Redundant node merging and association path simplification operations are performed on the subject knowledge graph to obtain a lightweight subject knowledge graph. These lightweight knowledge graphs are then integrated to form a lightweight computer science subject knowledge graph. And a large language model was used to analyze the lightweight computer science subject knowledge graph. Semantic transformation is performed to obtain a lightweight computer science subject knowledge graph. Natural language description.

3. The intelligent teaching system based on knowledge graphs and multi-agent collaboration according to claim 1, characterized in that, The cross-modal semantic fusion unit includes: Integrating visual language models, large language models, and cross-modal deep fusion functions; employing visual language models for the task at hand. Computer science diagrams Semantic parsing is performed to generate a natural language description of the diagram; a large language model is used to analyze the question text. Perform semantic understanding to generate a natural language description of the problem text; invoke a lightweight computer science subject knowledge graph. Computer science diagrams Entity and Lightweight Computer Science Discipline Knowledge Graph Match entities in the graph to construct a graphical subgraph containing entities and relationships. ; The illustrated subgraph is analyzed using a large language model. Perform semantic transformation to generate a graphical sub-map. The natural language description; the cross-modal semantic fusion unit has a built-in cross-modal deep fusion function, which unifies the semantic granularity and expression paradigm of the three types of natural language descriptions through the semantic adaptation module built into the cross-modal deep fusion function, establishes the correspondence between the three through logical association matching, and integrates the three into a unified cross-modal fused semantic text through a deep fusion algorithm. .

4. The intelligent teaching system based on knowledge graphs and multi-agent collaboration according to claim 1, characterized in that, The knowledge graph fusion reasoning unit is used to process the task to be processed. Cross-modal fusion semantic text Lightweight Computer Science Discipline Knowledge Graph The natural language descriptions are combined and input into a large model. Through the attention mechanism built into the large model, cross-modal semantic text is fused. The natural language description in the text calls upon a lightweight computer science subject knowledge graph. Based on the entities, intra-subject entity relationships, and cross-subject general relationships shown in the diagram, complete the entity extraction and basic logic derivation, and output the task to be processed. Preliminary reasoning results ; The preliminary reasoning results are calculated by using the knowledge gap annotation function built into the knowledge graph fusion reasoning unit. With the illustration of the sub-map The semantic matching degree is used to determine the initial inference result when the matching degree is lower than a preset threshold. The missing computer science knowledge in the text, and the generation of knowledge gap tags. .

5. The intelligent teaching system based on knowledge graphs and multi-agent collaboration according to claim 1, characterized in that, The adaptation agent unit, built on the large model Gemini-2.5-Flash, is used to receive tasks to be processed. With standardized preprocessing packages From standardized preprocessing packages Preliminary reasoning results were obtained from the analysis. Knowledge Gap Labels ; The task complexity scoring model built into the adapter agent unit is invoked based on the task to be processed. Preliminary reasoning results Knowledge gap label and the aforementioned lightweight computer science subject knowledge graph The task complexity score is quantitatively evaluated from four dimensions: subject specialization, logical hierarchy, answer output requirements, and diagrammatic complexity. The comprehensive score is obtained by weighted summation. The overall task complexity score is compared with a preset threshold range to determine the task to be processed. The task levels include simple, medium, and complex. Real-time acquisition of the real-time status sets of basic agent units, advanced agent units, and professional agent units within the system. ; Based on the determined task level, combined with the real-time status set This generates driver instructions for driving Agent units that match the task level.

6. The intelligent teaching system based on knowledge graphs and multi-agent collaboration according to claim 1, characterized in that, The basic Agent unit, built upon the large model Qwen3-VL-30B-A3B-Instruct, is used to receive driver instructions and tasks to be processed from the Agent unit. and standardized preprocessing packages And based on standardized preprocessing packages Cross-modal fusion semantic text encapsulated in Computer science diagrams Entity recognition is performed by calling the illustration sub-map in the standardized preprocessing package. The identified entities are compared with the illustrated sub-map. The entities and relationships are compared to verify the existence of the entities and the clarity of the relationships; If entity or relation missing information is found during the verification process, consider using a lightweight computer science subject knowledge graph. The corresponding data is retrieved and completed; the results of recognition, matching and completion are integrated to generate and output the reasoning results.

7. The intelligent teaching system based on knowledge graphs and multi-agent collaboration according to claim 1, characterized in that, The advanced agent unit, built on the large model Claude-Sonnet-4-5-20250929-thinking, is used to receive driving instructions and tasks to be processed from the adapted agent unit. and standardized preprocessing packages ; Cross-modal fusion semantic text based on standardized preprocessing packages With the illustration of the sub-map For computer science diagrams Perform logical link parsing, including: executing the entity in the diagram - the diagram subgraph. Entity anchoring in the graph involves calling a graph neural network (GNN) to analyze the subgraph. Multi-hop relationship propagation is performed to construct the initial logical link, and the illustrated sub-graph is generated through the large model Qwen3-VL-8B. The relationships in the code are transformed into natural language logical descriptions, combined with standardized preprocessing packages. Verify the logical syntax and explicit rules of the initial logical link; Based on the results of the logical link parsing and the knowledge gap tags in the standardized preprocessing package From a lightweight computer science discipline knowledge graph The core application layer knowledge points are retrieved in the targeted search, and the preliminary reasoning results are analyzed. Supplement knowledge; Preliminary reasoning results The process involves step-by-step decomposition and verification, breaking it down into single-step reasoning units, entity association links, and subject knowledge reference segments. Logical consistency verification and fusion completion are then performed based on explicit logical breakpoint types, which include one or more of the following: incomplete fields, inconsistencies in reasoning steps within a single subject, incorrect concept references, conflicts between diagram annotations and text descriptions, and matching gaps between knowledge graph entities and reasoning entities. The results of integrating logical link analysis, knowledge supplementation, and fusion completion are used to generate and output the reasoning results.

8. The intelligent teaching system based on knowledge graphs and multi-agent collaboration according to claim 1, characterized in that, The specialized agent unit is built on a dual-model architecture, including the GPT-5 model for handling multiple-choice and fill-in-the-blank questions, and the GLM-4.6V model for handling comprehensive application questions. The specialized agent unit receives driving instructions and tasks to be processed from the adapted agent unit. and standardized preprocessing packages ; Based on the task format determination result, the appropriate large model is selected and invoked, and the invoked large model is based on the cross-modal fused semantic text. Computer science diagrams A full-dimensional analysis is performed, which covers entity nesting relationships, control signal annotation, and system state mapping. Based on knowledge gap tagging From a lightweight computer science discipline knowledge graph The system retrieves deep-seated fundamental principles and, based on cross-subject common relationships, retrieves related cross-subject knowledge to refine the preliminary reasoning results. To supplement; Based on the aforementioned deep-seated fundamental knowledge, cross-subject related knowledge, and comprehensive analysis results, long-chain nested logical reasoning and system-level consistency verification are performed; simultaneously, preliminary reasoning results are identified. The implicit logical conflicts in the process include one or more of the following: cross-subject contradictions, inconsistencies in mechanisms, and unreasonable frameworks. Based on the deep-seated principles and using cross-subject related knowledge as a logical bridge, the conflicts in the preliminary reasoning results are resolved to ensure the logical coherence of multi-sub-problems and multi-step reasoning. It integrates the results of comprehensive analysis, knowledge supplementation, long-chain nested logical reasoning, and conflict resolution to generate and output the reasoning results.

9. The intelligent teaching system based on knowledge graphs and multi-agent collaboration according to claim 1, characterized in that, The verification agent unit, built on the large model Gemini-3-Pro-Preview, is used to receive the inference results and standardized preprocessing packages output by the upstream functional agent unit. The preset quantization thresholds that match each functional agent unit are combined with the standardized preprocessing package. Knowledge gap tags Illustration of sub-map and lightweight computer science subject knowledge graph The reasoning results are checked item by item; If the verification result reaches the corresponding quantization accuracy threshold, the final inference result will be output directly. If the verification result does not reach the corresponding quantitative accuracy threshold, the error source of the substandard inference result is traced and a correction suggestion is generated. At the same time, the verification result is synchronized to the adaptation agent unit and the upstream functional agent unit to achieve cross-layer bidirectional feedback. The verification result is also sent back to the knowledge graph enhancement pre-inference module and the model optimization and tuning unit to achieve cross-modal fusion semantic text optimization and model tuning linkage. The corrected and integrated content is transformed into the final inference result with the same content structure as the output content of the corresponding functional agent unit.

10. The intelligent teaching system based on knowledge graphs and multi-agent collaboration according to claim 1, characterized in that, The system verification module is used to execute a hierarchical verification process: Automated testing is performed based on the Multimodal Computer Science Subject Graph Benchmark dataset MMCS-Bench to batch verify the overall system performance. The verification indicators for the overall system performance include answer accuracy, logical completeness of explanations, and utilization rate of graphical information. Next, real-world testing is conducted based on actual teaching cases for each subject, collecting feedback from practical use in teaching scenarios. Finally, based on the automated testing indicator data and user feedback from the real-world testing, the system model, the CS-MMA collaborative architecture module, and the lightweight computer science subject knowledge graph are evaluated. Perform iterative optimization.

Images