Large and small model collaborative multi-agent heuristic learning guiding method and equipment

Abstract

The invention relates to the technical field of intelligent education, in particular to a large and small model collaborative multi-agent heuristic learning guiding method and device.According to the method, through the closed-loop process of obtaining student input, sensing, calculating and analyzing, generating a control instruction, generating a heuristic reply and conducting double evaluation and verification, a large-scale model and a small-scale model are obtained; the technical problems that an existing intelligent learning guiding system is poor in controllability, high in computing power cost and insufficient in emotion and cognition splitting and specialty are solved by combining heuristic learning guiding state machine logic, a small parameter model of special data set fine tuning and edge deployment optimization. According to the method, low-delay and high-controllability heuristic learning guidance with subject specialty and emotional care is realized, network disconnection scenes can be flexibly adapted, and personalized and immersive learning guidance experience is provided for students.

Description

Technical Field

[0001] This invention relates to the field of intelligent education technology, specifically to a multi-agent heuristic teaching method and device that coordinates large and small models. Background Technology

[0002] With the rapid development of generative artificial intelligence technology, intelligent education systems have gradually shifted from the traditional "resource recommendation" model to a "generative tutoring" model centered on natural language dialogue. Large language models, with their powerful semantic understanding and content generation capabilities, have shown broad application prospects in the field of intelligent tutoring; however, when directly applied to serious educational scenarios, they still face many technical bottlenecks.

[0003] Existing intelligent learning guidance methods generally suffer from the following shortcomings: First, the teaching process lacks controllability and structured constraints, making it prone to logical leaps or answer leaks, and hindering the implementation of gradual, heuristic guidance. Second, there is a significant contradiction between computational cost and response latency; the entire process relies on large-parameter models, resulting in high inference costs and long delays, while using only small-parameter models cannot guarantee the depth and professionalism of guidance. Third, there is a disconnect between emotion and cognition; most systems focus only on knowledge transmission and fail to perceive students' emotional states in real time and dynamically adjust teaching strategies, leading to insufficient student motivation. Fourth, there is insufficient subject-specific adaptability; there is a lack of dedicated datasets built based on authoritative teaching resources, making it difficult to guarantee the professionalism and accuracy of the guidance content.

[0004] Furthermore, existing methods have limitations in deployment flexibility, making them difficult to adapt to edge scenarios such as network outages, thus restricting their widespread application in different teaching environments. Therefore, there is an urgent need for an intelligent learning guidance method that combines high controllability, cost-effectiveness, emotional support, and flexible deployment capabilities to meet the actual needs of educational scenarios. Summary of the Invention

[0005] In view of this, the purpose of the present invention is to provide a multi-agent heuristic teaching method and device for large and small model collaboration, so as to overcome the problems existing in the current technology.

[0006] To achieve the above objectives, the present invention adopts the following technical solution: On the one hand, this application provides a multi-agent heuristic teaching method for collaboration between large and small models, including: Step S1: Obtain natural language data input by the student; wherein, the natural language data includes: text data and voice data; Step S2: Perform sentiment state analysis and learning memory retrieval on the natural language data to obtain sentiment vectors and historical dialogue context data; Step S3: Based on the emotion vector, the historical dialogue context data, and the current teaching state, structured control instructions are generated from a large parameter language model based on the heuristic guided learning state machine logic. Step S4: Generate a heuristic natural language response from a small-parameter model fine-tuned with a dedicated dataset based on the structured control instructions and the historical dialogue context data; Step S5: Perform a dual evaluation mechanism to verify the heuristic natural language response. If the verification passes, output the heuristic natural language response to the student. If it fails, return to step S4 to regenerate the heuristic natural language response.

[0007] Furthermore, in the method described above, step S1 includes: Receive natural language data submitted by students through text input boxes or voice acquisition devices; The speech data in the natural language data is converted into text format using ASR technology; The natural language data is preprocessed to remove redundant characters and correct semantic ambiguities.

[0008] Furthermore, in the method described above, step S2 includes: Obtain the textual and behavioral features of the natural language data; wherein, the behavioral features include: response delay time and the number of retries for the current question; The text features and the behavior features are input into the DistilBERT classification model to extract the semantic vector of the text features and to normalize the behavior features and map them into behavior vectors. By concatenating the semantic vector and the behavioral vector, the sentiment vector is output through the Softmax function; Extract the dialogue context data from the most recent K rounds using a sliding window queue with a capacity of K=10 rounds; The natural language data is transformed into query vectors using the bge-m3 Embedding model; Based on the query vector, an approximate nearest neighbor search is performed in the historical database to recall three historical interaction records that are most relevant to the topic of the query vector; wherein, the historical interaction records include historical average skill scores; The dialogue context data is integrated with the historical interaction records to output historical dialogue context data.

[0009] Furthermore, the heuristic guided learning state machine logic described above includes: a concrete perception stage, a cognitive conflict stage, a reasoning construction stage, a transfer and application stage, and an inductive summarization stage. Step S3 includes: If the natural language data is initial question data, then a large-parameter language model generates structured control instructions based on the sentiment vector, the historical dialogue context data, and the current teaching state, so that the small-parameter model generates a heuristic natural language response corresponding to the initial question; wherein, the structured control instructions include: target stage field, action type field, skill focus field, sentiment strategy field, and constraint condition field; If the natural language data is a response data of a heuristic natural language response to the initial question, then proceed to the concrete perception stage; Obtain the knowledge points from the initial question data, and retrieve real-life scenarios or virtual experimental phenomena based on the knowledge points from the initial question data; Detect the number of matching "attribute-direction of change" pairs in the response data; When the number of matches is less than a first preset threshold, a structured control instruction is generated so that the small parameter model generates a single-dimensional guide for the unmatched attribute. When the number of matches is not less than the first preset threshold, the cognitive conflict stage is entered. Based on the real-life scenario or the virtual experiment phenomenon, a structured control instruction is generated so that the small parameter model pushes abnormal cases that contradict the student's current cognition to the student. Obtain the student's response data to the abnormal case, and use an intent recognition model to calculate the intent recognition confidence of the response data to the abnormal case; When the confidence level of the intent recognition does not exceed the second preset threshold, a structured control instruction is generated to enable the small parameter model to supplement the student with explanations of abnormal cases. When the confidence level of the intent recognition exceeds the second preset threshold, the inference construction stage is entered. Calculate the immediate skill performance score of the response data for the aforementioned anomalous cases; The student's final skill mastery score is calculated based on the immediate skill performance score and the historical average skill score. When the final skill mastery score is <0.6, a structured control instruction for the scaffolding strategy is generated so that the small parameter model can provide response data that is specific analogous to or reduces the dimensionality of the problem; When 0.6 ≤ the final skill mastery score < 0.85, a structured control instruction for generating clue prompts is generated to enable the small parameter model to generate response data that only prompts key variables; When the final skill mastery score is ≥0.85, the student enters the transfer application stage. Based on the thought chain of the CSQ dataset, the reasoning logic of the response data for the anomalous case is determined, and a new isomorphic scenario question is generated according to the reasoning logic. Obtain the students' response data to the isomorphic new scenario question, and determine the interpretation logic of the response data to the isomorphic new scenario question; The large parameter language model is used to determine whether the explanation logic correctly reuses the reasoning logic; If not, then optimize the description of the problem in the new scenario; If so, then proceed to the summarization stage; Guide students to produce a closed-loop summary of "phenomenon-principle-application"; Obtain the summary data output by the students and check whether the summary data contains the three elements of phenomenon, principle and application; If included, update the historical database and mark the mastery status of the corresponding knowledge point as "mastered"; If not included, a structured control instruction is generated to cause the small parameter model to generate question data that inquires about missing elements.

[0010] Furthermore, in the method described above, step S4 includes: Collect knowledge point data; wherein, the knowledge point data includes: knowledge point data stipulated in the "Compulsory Education Science Curriculum Standards (2022 Edition)", formal examination question data for grades 3-6 of primary school, and authoritative encyclopedia data from the National Primary and Secondary School Smart Education Platform; The knowledge point data is labeled with problem information, subject attributes, and solutions to construct the CSQ dataset; Extract triplet data sets from the CSQ dataset; wherein the triplet data sets include: the original question, the parsing thought process, and the science skill tags; The triplet data set is input into the rewritten Prompt and fed into the large parameter language model to generate "instruction-heuristic probing" training corpus; Using LoRA technology, with rank r=8, learning rate lr=2e-5 and training epoch=3, the basic language model was used as the base model. The training corpus was input, and the training objective was set to maximize the probability of generating heuristic questions and minimize the probability of generating declarative explanations. The training was carried out, and the fine-tuned basic language model was saved as a small parameter model. The structured control instructions and the historical dialogue context data are input into the small parameter model to generate a heuristic natural language response.

[0011] Furthermore, in the method described above, step S5 includes: The bge-m3 Embedding model is used to convert the heuristic natural language responses and corresponding standard answer data into vectors, respectively. Calculate the cosine similarity between the heuristic natural language response vector and the standard answer data vector; When the cosine similarity is greater than 0.7, it is determined that the answer has been directly leaked, an error code is generated, and the process returns to step S4 to regenerate the heuristic natural language response. Based on rule-based NLP parsing, it is detected whether the heuristic natural language response contains guiding keywords; If not included, return to step S4 to trigger rewriting. The rewriting instruction is "change to an open-ended question to guide students to explain their logic". Real-time monitoring of the value of the frustration dimension in the emotional vector; If the value of the frustration dimension is greater than 0.7 for two consecutive rounds, an emotion regulation signal is generated to increase the priority of emotional support. When generating the structured control instructions, the teaching strategy is downgraded from "high challenge" to "high care + low difficulty" to reduce the cognitive complexity of the problem and increase reassuring language.

[0012] Furthermore, the methods described above also include: Deploy the large-parameter language model on a cloud server; The small parameter model is compressed using INT8 quantization technology and deployed on user edge devices; The user edge device stores a subset of the CSQ dataset. When the network is disconnected, the user edge device independently generates heuristic natural language responses based on the small parameter model and the subset of the CSQ dataset, thereby realizing basic heuristic dialogue.

[0013] On the other hand, this application provides a multi-agent heuristic learning device for large-scale model collaboration, including a processor and a memory, wherein the processor is connected to the memory: The processor is used to call and execute the program stored in the memory; The memory is used to store the program, which is at least used to execute the multi-agent heuristic learning method for large-scale model collaboration described above.

[0014] The beneficial effects of this invention are as follows: High controllability of teaching: Through a five-stage heuristic learning state machine and structured control instructions, the teaching process is standardized, the rate of answer leakage is significantly reduced, and the logic and coherence of heuristic guidance are greatly improved. Optimal efficiency: The large and small models work together, with the large model responsible for complex decisions and the small model responsible for high-frequency interactions, reducing computing costs and keeping response latency within an acceptable range for real-time interactions. Highly professional: The CSQ dataset, built on authoritative teaching resources, and the targeted fine-tuning process ensure that the model-guided content aligns with the teaching syllabus, and its subject-specific accuracy and professionalism far exceed those of general models; Providing adequate emotional support: Establishing a closed-loop regulation mechanism between emotional state and teaching strategies to respond to students' emotional changes in real time, thereby enhancing their learning experience and sustained motivation; Flexible deployment: It supports hybrid deployment on cloud and edge devices, and can still provide basic instruction even when the network is offline. It is adaptable to a variety of teaching environments and has a wider range of applications. Attached Figure Description

[0015] To more clearly illustrate the technical solutions in the embodiments of the present invention or the prior art, the drawings used in the description of the embodiments or the prior art will be briefly introduced below. Obviously, the drawings described below are only some embodiments of the present invention. For those skilled in the art, other drawings can be obtained based on these drawings without creative effort.

[0016] Figure 1 This is a flowchart provided in one embodiment of the multi-agent heuristic teaching method for large and small model collaboration according to the present invention; Figure 2 This is a schematic diagram of the structure of an embodiment of a multi-agent heuristic learning device for large and small model collaboration according to the present invention. Detailed Implementation

[0017] To make the objectives, technical solutions, and advantages of this invention clearer, the technical solutions of this invention will be described in detail below. Obviously, the described embodiments are merely some embodiments of this invention, and not all embodiments. Based on the embodiments of this invention, all other implementation methods obtained by those skilled in the art without creative effort are within the scope of protection of this invention.

[0018] Figure 1 This is a flowchart provided in one embodiment of a multi-agent heuristic teaching method for large and small model collaboration according to the present invention. Please refer to... Figure 1 This embodiment may include the following steps: Step S1: Obtain the natural language data input by the student; wherein, the natural language data includes: text data and voice data; Step S2: Perform sentiment analysis and learning memory retrieval on the natural language data to obtain sentiment vectors and historical dialogue context data; Step S3: Based on the emotion vector, historical dialogue context data, and the current teaching status, structured control instructions are generated from the large parameter language model based on the heuristic guided learning state machine logic. Step S4: Generate a heuristic natural language response from a small-parameter model fine-tuned with a dedicated dataset based on structured control instructions and historical dialogue context data; Step S5: Perform a dual evaluation mechanism to verify the heuristic natural language response. If the verification passes, output the heuristic natural language response to the student. If it fails, return to step S4 to regenerate the heuristic natural language response.

[0019] Preferably, step S1 includes: Receive natural language data submitted by students through text input boxes or voice acquisition devices; Convert speech data in natural language data into text format using ASR technology; The natural language data is preprocessed to remove redundant characters and correct semantic ambiguities.

[0020] Understandably, after receiving the natural language data input by the student, the natural language data is first preprocessed at the user edge device to remove redundant characters and correct semantic ambiguities in the natural language data, so as to prevent it from affecting the subsequent heuristic natural language response.

[0021] Preferably, step S2 includes: Obtain textual and behavioral features from natural language data; among which, behavioral features include: response delay time and the number of retries for the current question; Text features and behavioral features are input into the DistilBERT classification model to extract semantic vectors from text features and normalize behavioral features into behavioral vectors. By concatenating semantic vectors and behavioral vectors, the sentiment vector is output through the Softmax function. Extract the dialogue context data from the most recent K rounds using a sliding window queue with a capacity of K=10 rounds; The bge-m3 Embedding model is used to transform natural language data into query vectors; Based on the query vector, perform an approximate nearest neighbor search in the historical database to recall three historical interaction records that are most relevant to the topic of the query vector; among them, the historical interaction records include the historical average skill score; The dialogue context data is integrated with historical interaction records to output historical dialogue context data.

[0022] It is understandable that a lightweight classification model is used to perform sentiment polarity analysis on students' natural language input and output sentiment vectors (dimensions include: confusion, frustration, focus, and challenge).

[0023] Emotional Vector V emo =[e1, e2, e3, e4], where e1 is the level of confusion, e2 is the level of frustration, e3 is the level of focus, and e4 is the level of challenge. The values ​​of e1, e2, e3, and e4 are in the range of [0,1].

[0024] Based on the VectorDB, the system uses the embedding vectors of the student's natural language input to retrieve the student's long-term memory (historical weak skills, knowledge mastery status, and historical average skill scores) and short-term memory (current dialogue context).

[0025] Preferably, the heuristic guided learning state machine logic includes: concrete perception stage, cognitive conflict stage, reasoning construction stage, transfer and application stage, and inductive summarization stage; Step S3 includes: If the natural language data is the initial question data, then the large parameter language model generates structured control instructions based on the sentiment vector, historical dialogue context data, and the current teaching state, so that the small parameter model can generate a heuristic natural language response corresponding to the initial question; the structured control instructions include: target stage field, action type field, skill focus field, sentiment strategy field, and constraint condition field; If the natural language data is the response data of the heuristic natural language response to the initial question, then proceed to the concrete perception stage; Extract the knowledge points from the initial question data, and retrieve real-life scenarios or virtual experimental phenomena based on the knowledge points in the initial question data; Detect the number of matching "attribute-direction of change" pairs in the response data; When the number of matches is less than the first preset threshold, a structured control instruction is generated so that the small parameter model can generate a single-dimensional guide for the unmatched attributes. When the number of matches is not less than the first preset threshold, the cognitive conflict stage is entered. Based on real-life scenarios or virtual experimental phenomena, structured control instructions are generated so that the small parameter model can push abnormal cases that contradict the student's current cognition to the student. Obtain student response data for anomalous cases, and use an intent recognition model to calculate the intent recognition confidence of the response data for anomalous cases; When the confidence level of intent recognition does not exceed the second preset threshold, a structured control instruction is generated to enable the small parameter model to supplement the student with explanations of anomalous cases. When the confidence level of intent recognition exceeds the second preset threshold, the inference construction stage begins. Calculate an immediate skills performance score based on response data for anomalous cases; The student's final skill mastery score is calculated based on the immediate skill performance score and the historical average skill score. When the final skill mastery score is <0.6, generate structured control instructions for the scaffolding strategy to enable the small parameter model to provide specific analogies or reduce the dimensionality of the questions; When 0.6 ≤ final skill mastery score < 0.85, generate structured control instructions to prompt clues so that the small parameter model generates response data that only prompts key variables; When the final skill mastery score is ≥0.85, the student enters the transfer and application stage. Based on the thought chain of the CSQ dataset, the reasoning logic of the response data for anomalous cases is determined, and a new isomorphic scenario question is generated according to the reasoning logic. Obtain student response data to questions in a new isomorphic scenario and determine the interpretation logic for the response data. Use a large-parameter language model to determine whether the explanation logic is correct and reuse the reasoning logic; If not, then optimize the description of the problem in the new scenario; If so, then proceed to the summarization stage; Guide students to produce a closed-loop summary of "phenomenon-principle-application"; Obtain the summary data output by students and check whether the summary data contains the three elements of phenomenon, principle and application; If included, update the historical database and mark the mastery status of the corresponding knowledge point as "mastered"; If not included, structured control instructions are generated to enable the small-parameter model to generate question data that inquires about missing features.

[0026] Understandably, based on CSQ data annotation, the following calculation formula is used to determine whether students have captured the key variables in scientific inquiry.

[0027]

[0028] Among them, K target Let K be the set of keywords corresponding to the current question in CSQ. matched The number of keywords contained in the student's answer.

[0029] Using Embedding models (such as bge-m3) to analyze student responses R student Compared with the standard thought chain T in the CSQ dataset gold Map the vectors and calculate the cosine similarity.

[0030]

[0031] Immediate skill performance score calculated based on student responses in the current round:

[0032] Where w1 + w2 = 1.

[0033] The final skill score used for state machine transition judgment (range [0,1]):

[0034] Among them, Score skill For the final skill mastery score, S history To extract the student's historical average level in this skill, β is the immediate weighting coefficient (recommended value 0.7), indicating that the system values ​​current performance more, but will also refer to historical ability to prevent misjudgment.

[0035] Preferably, step S4 includes: Collect knowledge point data; among which, knowledge point data includes: knowledge point data stipulated in the "Compulsory Education Science Curriculum Standards (2022 Edition)", data of formal examination questions for grades 3-6 of primary school, and authoritative encyclopedia data from the National Primary and Secondary School Smart Education Platform; The knowledge point data is labeled with problem information, subject attributes, and solutions to construct the CSQ dataset; Extract triplet datasets from the CSQ dataset; the triplet datasets include: the original question, the parsed thought process, and the science skill label; Input the triplet data set into the rewritten Prompt and then into the large parameter language model to generate "instruction-heuristic probing" training corpus; Using LoRA technology, with rank r=8, learning rate lr=2e-5 and training epoch=3, the basic language model was used as the base model, and the training corpus was input. The training objective was set to maximize the probability of generating heuristic questions and minimize the probability of generating declarative explanations. The fine-tuned basic language model was saved as a small parameter model. Structured control commands and historical dialogue context data are input into a small-parameter model to generate heuristic natural language responses.

[0036] Preferably, step S5 includes: The bge-m3 Embedding model is used to transform the heuristic natural language responses and corresponding standard answer data into vectors, respectively. Calculate the cosine similarity between the heuristic natural language response vector and the standard answer data vector; When the cosine similarity is greater than 0.7, it is determined that the answer has been directly leaked, an error code is generated, and the process returns to step S4 to regenerate the heuristic natural language response. Based on rule-based NLP parsing, this study detects whether heuristic natural language responses contain leading keywords. If not included, return to step S4 to trigger rewriting. The rewriting instruction is "change to an open-ended question to guide students to explain their logic". Real-time monitoring of the frustration dimension in the emotion vector; If the value of the frustration dimension is greater than 0.7 for two consecutive rounds, an emotion regulation signal is generated to increase the priority of emotional support. When generating structured control instructions, the teaching strategy is downgraded from "high challenge" to "high care + low difficulty" to reduce the cognitive complexity of the problem and increase reassuring language.

[0037] Preferred options also include: Deploy large-parameter language models on cloud servers; The small parameter model is compressed using INT8 quantization technology and deployed on user edge devices; The user edge device stores a subset of the CSQ dataset. When the network is disconnected, the user edge device independently generates heuristic natural language responses based on the small parameter model and the subset of the CSQ dataset, thus realizing basic heuristic dialogue.

[0038] In specific practice: 1. Preliminary preparations Model Deployment: The large-parameter language model (qwen2.5 72B) is deployed on Alibaba Cloud ECS cloud server; the small-parameter model (qwen2.5 7B basic model fine-tuned with CSQ dataset) adopts INT8 quantization technology and is deployed on student tablets (edge ​​devices). The tablets locally store a subset of 20,000 CSQ datasets (including knowledge points related to "buoyancy").

[0039] Dataset support: The CSQ dataset contains triplet data on the knowledge point of "buoyancy" (original question: "Why can a wooden block float on water?"; analytical thinking chain: "1. Observe the different phenomena of wooden blocks and stones placed in water; 2. Analyze the forces acting on the object (gravity, buoyancy); 3. Derive the relationship between buoyancy and the volume of water displaced by the object; 4. Summarize the application scenarios of buoyancy"; scientific skills tags: "observation of phenomena, inductive reasoning, application and transfer").

[0040] 2. Input Acquisition Students submitted their initial question via a text input box on a tablet: "Why can a wooden block float on water, but a stone sinks?" (text data); The device preprocesses the input data: removing redundant spaces, correcting semantic ambiguity, and obtaining standardized natural language data: "Why can wooden blocks float on water, but stones sink?".

[0041] 3. Emotional state analysis and learning / memory retrieval 3.1 Emotional State Analysis: Extract the text feature "Why can wooden blocks float on water, but stones sink?", and the behavioral features: response delay time 1.5 seconds (entered by the student after thinking), and 0 retries for the current question; Textual and behavioral features are input into the DistilBERT classification model to extract semantic vector H. semantic After normalizing the behavioral features, they are mapped to behavioral vectors H. behavior ; The two types of vectors are concatenated and then processed by the Softmax function to output the sentiment vector V. emo =[0.65, 0.1, 0.8, 0.75] (Confusion 0.65, Frustration 0.1, Focus 0.8, Desire for Challenge 0.75).

[0042] 3.2 Learning and Memory Retrieval: Short-term memory retrieval: Extract the context of the last 10 rounds of dialogue (empty, as it is the initial question) through a sliding window queue with a capacity of K=10 rounds. Long-term memory retrieval: The bge-m3Embedding model is used to transform text features into a query vector Vq. An approximate nearest neighbor search is performed on the tablet's local historical database to recall 3 relevant historical interaction records. 1. One month ago, a student asked, "Why can a ship float on the sea?" The historical average skill score was S. history =0.6; 2. The student made a mistake on the topic of "mass and volume of objects"; 3. Students prefer guidance based on real-life scenarios; Integrate and output historical dialogue context data H mem = Empty dialogue context, [(Question about ship buoyancy, skill score 0.6), (Incorrect knowledge of mass and volume), (Preference for everyday scenarios)]).

[0043] 4. Generate structured control instructions The natural language data was determined to be the initial problem data, and a large-parameter language model (cloud-based) was developed based on the sentiment vector V. emo Historical dialogue context data H mem Based on the current teaching status (initial stage), generate structured control instructions I. ctrl 1:

[0044] After receiving the heuristic response from the tablet, the student submitted their response data: "I tried putting a wooden block and a stone into a basin of water. The wooden block floated, while the stone sank immediately. Also, the wooden block looked lighter than the stone." The system determines that the data is a response to the initial reply and enters the concrete perception stage: Extract the knowledge point "the relationship between buoyancy and the floating and sinking of objects" from the initial question, and search for the everyday scenario "the floating and sinking phenomenon of wooden blocks, stones, and plastic bottles in a basin of water"; The "attribute-direction of change" tuples in the detected response data are: {(wood block, floating), (stone, sinking), (wood block, light), (stone, heavy)}, with N matching pairs. match =4; Set the first preset threshold to 2, because N match ≥2, entering the cognitive conflict stage, the cloud-based large model generates structured control instructions I ctrl2 :

[0045] An unusual case of a student receiving a push notification on a tablet: "If you put a very heavy block of wood (heavier than a small stone) into water along with the small stone, will the block of wood sink? Also, if you press down on the floating block of wood with your hand, what do you feel?" The submitted response data was: "The block of wood should still float, right? When I press down on the block of wood, I feel a force pushing my hand upwards." The intent recognition confidence score of this response data was calculated to be 0.85 (the student questioned the unusual case and attempted to answer it). Set the second preset threshold to 0.8. Since the confidence level is > 0.8, proceed to the inference construction stage. Calculate real-time skill performance score

[0046] S semantic The cosine similarity between the student's response and the thought process chain of the "buoyancy" parsing in the CSQ dataset is 0.75. S keyword The target keywords (buoyancy, sinking, upward force, mass) totaled 4, and students answered 3 of them correctly. keyword =0.75; .

[0047] Calculate the final skill mastery score:

[0048] Because 0.6≤Score skill <0.85, structured control commands I for generating clues and hints in cloud-based large models. ctrl 3:

[0049] The student receives the following clue: "The 'upward force' you feel is buoyancy! What factors do you think affect the magnitude of buoyancy? Think again, why does the iron block sink while the wooden block floats, even though they are of the same weight?" The student submits the following response: "Does buoyancy depend on whether the object is made of wood or on the size of the object?" Recalculate the immediate skill performance score (S) current =0.82, Final Skill Mastery Score skill =0.754, still within the range of 0.6 ≤ Score skill In the range of <0.85, the cloud generates control commands to supplement the clues, and the tablet outputs: "Great conjecture! If the wood is cut into very small pieces, will it still float? If the iron block is made into the shape of a boat, can it float on water?" The student replied, "The wooden pieces still float! A boat made of iron should float, just like a real boat." The score was then recalculated. skill =0.86 (≥0.85), proceed to the migration and application stage: Based on the thought process of the CSQ dataset, the reasoning logic is determined to be "the magnitude of buoyancy is related to the volume of water displaced by the object, but not to the material of the object, and changing the shape of the object can change the volume of water displaced". Generate a new isomorphic scenario question: "When swimming in the summer, do you feel the same buoyancy when you are fully submerged in the water and when you only submerge half of your body? Why?" The student replied: "It's different! When you're fully submerged in water, you feel a greater upward force on your body. It should be that the more your body is submerged, the more water is displaced, and the greater the buoyancy." The cloud-based big data model determined that the student's explanation was logically correct, reused the core reasoning logic, and entered the inductive summarization stage, generating control command I. ctrl 4:

[0050] 5. Generate heuristic natural language responses The tablet-based small-parameter model receives structured control commands and historical dialogue context data from each stage, and generates corresponding responses based on the fine-tuned parameters. Receive I ctrl The message then reads: "You've observed very carefully! There are many similar phenomena in life, such as plastic bottles floating on water while keys sink. Try to recall what the differences are between these floating and sinking objects, besides their weight?" Receive I ctrl 2. Generate abnormal cases and questions (as described in the above process); Receive I ctrl 3. Generate clues and hints (as described in the above process); Receive Ictrl The message then reads: "Excellent! You have grasped the core logic of buoyancy. Now, please try to summarize: What is the phenomenon of buoyancy (phenomenon)? What does the magnitude of buoyancy depend on (principle)? What other examples of using buoyancy in daily life (applications)?"

[0051] 6. Verification of the dual evaluation mechanism 6.1 Compliance Assessment: For each response generated by the tablet, the bge-m3Embedding model is used to convert the response and the standard answer of the "buoyancy" knowledge point in the CSQ dataset into vectors. The cosine similarity is less than 0.7, indicating no answer leakage. Based on rule-based NLP parsing, all responses contained guiding keywords such as "why," "what do you think," and "try," which constituted open-ended questions, and the verification passed. 6.2 Emotional Feedback Regulation: The value of the frustration dimension in the emotional vector was monitored in real time. There were no instances of frustration values ​​>0.7 in two consecutive rounds. The teaching strategy maintained "moderate challenge + encouragement". Students submitted a summary of the data: "Phenomenon: Some objects can float on the surface of a liquid, while others sink; Principle: The magnitude of buoyancy is related to the volume of liquid displaced by the object. The larger the displaced volume, the greater the buoyancy; Application: Ships use buoyancy to navigate, and life rings allow people to float on the water." The test summary data includes three elements: "phenomenon-principle-application". The local historical database is updated, and the mastery status of the "buoyancy" knowledge point is marked as "mastered". This learning session is now complete.

[0052] The present invention also provides a multi-agent heuristic learning device for large and small model collaboration, used to implement the above-described method embodiments. Figure 2 This is a schematic diagram of the structure of an embodiment of a multi-agent heuristic learning device for large and small model collaboration according to the present invention. Figure 2 As shown, the multi-agent heuristic learning device for big-small model collaboration in this embodiment includes a processor 21 and a memory 22, with the processor 21 connected to the memory 22. The processor 21 is used to call and execute the program stored in the memory 22; the memory 22 is used to store the program, which is at least used to execute the multi-agent heuristic learning method for big-small model collaboration in the above embodiments.

[0053] The specific implementation scheme of the multi-agent heuristic learning device with large and small model collaboration provided in this application can refer to the implementation scheme of the multi-agent heuristic learning method with large and small model collaboration in any of the above embodiments, and will not be repeated here.

[0054] It is understood that the same or similar parts in the above embodiments can be referred to each other, and the contents not described in detail in some embodiments can be referred to the same or similar contents in other embodiments.

[0055] It should be noted that in the description of this invention, the terms "first," "second," etc., are used for descriptive purposes only and should not be construed as indicating or implying relative importance. Furthermore, in the description of this invention, unless otherwise stated, "a plurality of" means at least two.

[0056] Any process or method description in the flowchart or otherwise herein can be understood as representing a module, segment, or portion of code comprising one or more executable instructions for implementing a particular logical function or process, and the scope of the preferred embodiments of the invention includes additional implementations in which functions may be performed not in the order shown or discussed, including substantially simultaneously or in reverse order depending on the functions involved, as will be understood by those skilled in the art to which embodiments of the invention pertain.

[0057] It should be understood that various parts of the present invention can be implemented in hardware, software, firmware, or a combination thereof. In the above embodiments, multiple steps or methods can be implemented in software or firmware stored in memory and executed by a suitable instruction execution system. For example, if implemented in hardware, as in another embodiment, it can be implemented using any one or a combination of the following techniques known in the art: discrete logic circuits having logic gates for implementing logical functions on data signals, application-specific integrated circuits (ASICs) having suitable combinational logic gates, programmable gate arrays (PGAs), field-programmable gate arrays (FPGAs), etc.

[0058] Those skilled in the art will understand that all or part of the steps of the methods described in the above embodiments can be implemented by a program instructing related hardware. The program can be stored in a computer-readable storage medium, and when executed, it includes one or a combination of the steps of the method embodiments.

[0059] Furthermore, the functional units in the various embodiments of the present invention can be integrated into a processing module, or each unit can exist physically separately, or two or more units can be integrated into a module. The integrated module can be implemented in hardware or as a software functional module. If the integrated module is implemented as a software functional module and sold or used as an independent product, it can also be stored in a computer-readable storage medium.

[0060] The storage media mentioned above can be read-only memory, disk, or optical disk, etc.

[0061] In the description of this specification, references to terms such as "one embodiment," "some embodiments," "example," "specific example," or "some examples," etc., indicate that a specific feature, structure, material, or characteristic described in connection with that embodiment or example is included in at least one embodiment or example of the invention. In this specification, the illustrative expressions of the above terms do not necessarily refer to the same embodiment or example. Furthermore, the specific features, structures, materials, or characteristics described may be combined in any suitable manner in one or more embodiments or examples.

[0062] Although embodiments of the present invention have been shown and described above, it is understood that the above embodiments are exemplary and should not be construed as limiting the present invention. Those skilled in the art can make changes, modifications, substitutions and variations to the above embodiments within the scope of the present invention.

Claims

1. A multi-agent heuristic teaching method for collaborative large and small models, characterized in that, include: Step S1: Obtain natural language data input by the student; wherein, the natural language data includes: text data and voice data; Step S2: Perform sentiment state analysis and learning memory retrieval on the natural language data to obtain sentiment vectors and historical dialogue context data; Step S3: Based on the emotion vector, the historical dialogue context data, and the current teaching state, structured control instructions are generated from a large parameter language model based on the heuristic guided learning state machine logic. Step S4: Generate a heuristic natural language response from a small-parameter model fine-tuned with a dedicated dataset based on the structured control instructions and the historical dialogue context data; Step S5: Perform a dual evaluation mechanism to verify the heuristic natural language response. If the verification passes, output the heuristic natural language response to the student. If it fails, return to step S4 to regenerate the heuristic natural language response.

2. The method according to claim 1, characterized in that, Step S1 includes: Receive natural language data submitted by students through text input boxes or voice acquisition devices; The speech data in the natural language data is converted into text format using ASR technology; The natural language data is preprocessed to remove redundant characters and correct semantic ambiguities.

3. The method according to claim 2, characterized in that, Step S2 includes: Obtain the textual and behavioral features of the natural language data; wherein, the behavioral features include: response delay time and the number of retries for the current question; The text features and the behavior features are input into the DistilBERT classification model to extract the semantic vector of the text features and to normalize the behavior features and map them into behavior vectors. By concatenating the semantic vector and the behavioral vector, the sentiment vector is output through the Softmax function; Extract the dialogue context data from the most recent K rounds using a sliding window queue with a capacity of K=10 rounds; The natural language data is transformed into query vectors using the bge-m3 Embedding model; Based on the query vector, an approximate nearest neighbor search is performed in the historical database to recall three historical interaction records that are most relevant to the topic of the query vector; wherein, the historical interaction records include historical average skill scores; The dialogue context data is integrated with the historical interaction records to output historical dialogue context data.

4. The method according to claim 3, characterized in that, The heuristic guided learning state machine logic includes: concrete perception stage, cognitive conflict stage, reasoning construction stage, transfer and application stage, and inductive summarization stage; Step S3 includes: If the natural language data is initial question data, then a large-parameter language model generates structured control instructions based on the sentiment vector, the historical dialogue context data, and the current teaching state, so that the small-parameter model generates a heuristic natural language response corresponding to the initial question; wherein, the structured control instructions include: target stage field, action type field, skill focus field, sentiment strategy field, and constraint condition field; If the natural language data is a response data of a heuristic natural language response to the initial question, then proceed to the concrete perception stage; Obtain the knowledge points from the initial question data, and retrieve real-life scenarios or virtual experimental phenomena based on the knowledge points from the initial question data; Detect the number of matching "attribute-direction of change" pairs in the response data; When the number of matches is less than a first preset threshold, a structured control instruction is generated so that the small parameter model generates a single-dimensional guide for the unmatched attribute. When the number of matches is not less than the first preset threshold, the cognitive conflict stage is entered. Based on the real-life scenario or the virtual experiment phenomenon, a structured control instruction is generated so that the small parameter model pushes abnormal cases that contradict the student's current cognition to the student. Obtain the student's response data to the abnormal case, and use an intent recognition model to calculate the intent recognition confidence of the response data to the abnormal case; When the confidence level of the intent recognition does not exceed the second preset threshold, a structured control instruction is generated to enable the small parameter model to supplement the student with explanations of abnormal cases. When the confidence level of the intent recognition exceeds the second preset threshold, the inference construction stage is entered. Calculate the immediate skill performance score of the response data for the aforementioned anomalous cases; The student's final skill mastery score is calculated based on the immediate skill performance score and the historical average skill score. When the final skill mastery score is <0.6, a structured control instruction for the scaffolding strategy is generated so that the small parameter model can provide response data that is specific analogous to or reduces the dimensionality of the problem; When 0.6 ≤ the final skill mastery score < 0.85, a structured control instruction for generating clue prompts is generated to enable the small parameter model to generate response data that only prompts key variables; When the final skill mastery score is ≥0.85, the student enters the transfer application stage. Based on the thought chain of the CSQ dataset, the reasoning logic of the response data for the anomalous case is determined, and a new isomorphic scenario question is generated according to the reasoning logic. Obtain the students' response data to the isomorphic new scenario question, and determine the interpretation logic of the response data to the isomorphic new scenario question; The large parameter language model is used to determine whether the explanation logic correctly reuses the reasoning logic; If not, then optimize the description of the problem in the new scenario; If so, then proceed to the summarization stage; Guide students to produce a closed-loop summary of "phenomenon-principle-application"; Obtain the summary data output by the students and check whether the summary data contains the three elements of phenomenon, principle and application; If included, update the historical database and mark the mastery status of the corresponding knowledge point as "mastered"; If not included, a structured control instruction is generated to cause the small parameter model to generate question data that inquires about missing elements.

5. The method according to claim 4, characterized in that, Step S4 includes: Collect knowledge point data; wherein, the knowledge point data includes: knowledge point data stipulated in the "Compulsory Education Science Curriculum Standards (2022 Edition)", formal examination question data for grades 3-6 of primary school, and authoritative encyclopedia data from the National Primary and Secondary School Smart Education Platform; The knowledge point data is labeled with problem information, subject attributes, and solutions to construct the CSQ dataset; Extract triplet data sets from the CSQ dataset; wherein the triplet data sets include: the original question, the parsing thought process, and the science skill tags; The triplet data set is input into the rewritten Prompt and fed into the large parameter language model to generate "instruction-heuristic probing" training corpus; Using LoRA technology, with rank r=8, learning rate lr=2e-5 and training epoch=3, the basic language model was used as the base model. The training corpus was input, and the training objective was set to maximize the probability of generating heuristic questions and minimize the probability of generating declarative explanations. The training was carried out, and the fine-tuned basic language model was saved as a small parameter model. The structured control instructions and the historical dialogue context data are input into the small parameter model to generate a heuristic natural language response.

6. The method according to claim 5, characterized in that, Step S5 includes: The bge-m3 Embedding model is used to convert the heuristic natural language responses and corresponding standard answer data into vectors, respectively. Calculate the cosine similarity between the heuristic natural language response vector and the standard answer data vector; When the cosine similarity is greater than 0.7, it is determined that the answer has been directly leaked, an error code is generated, and the process returns to step S4 to regenerate the heuristic natural language response. Based on rule-based NLP parsing, it is detected whether the heuristic natural language response contains guiding keywords; If not included, return to step S4 to trigger rewriting. The rewriting instruction is "change to an open-ended question to guide students to explain their logic". Real-time monitoring of the value of the frustration dimension in the emotional vector; If the value of the frustration dimension is greater than 0.7 for two consecutive rounds, an emotion regulation signal is generated to increase the priority of emotional support. When generating the structured control instructions, the teaching strategy is downgraded from "high challenge" to "high care + low difficulty" to reduce the cognitive complexity of the problem and increase reassuring language.

7. The method according to claim 6, characterized in that, Also includes: Deploy the large-parameter language model on a cloud server; The small parameter model is compressed using INT8 quantization technology and deployed on user edge devices; The user edge device stores a subset of the CSQ dataset. When the network is disconnected, the user edge device independently generates heuristic natural language responses based on the small parameter model and the subset of the CSQ dataset, thereby realizing basic heuristic dialogue.

8. A multi-agent heuristic learning device for collaborative large and small models, characterized in that, It includes a processor and a memory, wherein the processor is connected to the memory: The processor is used to call and execute the program stored in the memory; The memory is used to store the program, which is at least used to execute the multi-agent heuristic learning method for large-scale model collaboration as described in any one of claims 1-7.

Images