Multi-agent collaborative chemical knowledge explicit distillation method based on cognitive apprentice mechanism
By employing a multi-agent collaborative explicit distillation method for chemical knowledge, the challenge of extracting chemical knowledge from large language models was solved, enabling the generation and dynamic updating of structured knowledge and improving the accuracy and cross-scenario applicability of chemical knowledge.
Patent Information
- Application Number
- CN202610177415.9
- Authority / Receiving Office
- CN · China
- Patent Type
- Applications(China)
- Current Assignee / Owner
- Filing Date
- 2026-02-07
- Publication Date
- 2026-05-15
AI Technical Summary
Existing technologies struggle to efficiently and accurately extract structured, reusable chemical expertise from large language models, particularly in chemical reasoning and complex question-answering tasks where generalization is insufficient. Furthermore, knowledge base construction is statically lagging and updates are inefficient, failing to adapt to the dynamic evolution of the chemical field.
A multi-agent collaborative explicit distillation method for chemical knowledge based on a cognitive apprenticeship mechanism is adopted. Through the collaborative work of senior students, teaching assistants and teachers, a structured list of chemical knowledge is generated. The accuracy and completeness of the knowledge are ensured by a two-dimensional arbitration evaluation mechanism, and finally stored in a pre-set knowledge base.
It significantly improves the interpretability and disciplinary rigor of chemical knowledge, enhances the accuracy, completeness, and cross-scenario reusability of knowledge, adapts to the dynamic evolution of chemical knowledge, and constructs a high-quality, dynamic, and scalable knowledge base.
Smart Images

Figure CN122047387A_ABST
Abstract
Description
Technical Field
[0001] This invention relates to the fields of artificial intelligence and machine learning technology, and more specifically, to a multi-agent collaborative chemical knowledge explicit distillation method and system based on a cognitive apprenticeship mechanism. Background Technology
[0002] With the rapid development of Large Language Models (LLMs) technology, they have demonstrated significant advantages in knowledge processing and solving complex problems, and are widely regarded as a parameterized carrier of massive amounts of knowledge. However, in the field of artificial intelligence, how to efficiently and accurately extract structured, reusable, and specific subject-matter expertise from these black-box models has become one of the key challenges that urgently needs to be overcome.
[0003] While large language models demonstrate some knowledge application capabilities during reasoning, their lack of transparency in internal mechanisms makes it difficult to directly obtain the complete explicit knowledge points supporting their problem-solving logic. "Explicit knowledge" here refers to knowledge units clearly expressed in natural language or symbolic form, possessing well-defined semantic boundaries and logical structures, such as definitions, rules, formulas, and typical examples. This type of knowledge can be verified, reused, and combined independently of specific problems. Currently, the main technical approaches used for knowledge extraction and model optimization include traditional knowledge distillation and knowledge base construction, but both have significant limitations.
[0004] I. Inherent defects of traditional knowledge distillation technology:
[0005] Traditional knowledge distillation (KD) uses the "soft labels" output by the teacher model as a supervision signal to guide the student model to fit the output distribution of the teacher model, aiming to achieve the transfer and inheritance of "hidden knowledge." This technical approach is essentially an implicit, black-box knowledge transfer model, and has three major shortcomings:
[0006] 1. Lack of knowledge structure representation ability: The distillation process only focuses on fitting the numerical distribution of the model output layer, without explicitly modeling the decision logic, semantic association, reasoning path and knowledge hierarchy relationship within the teacher model, thus failing to form structured and decomposable knowledge units.
[0007] 2. Prone to superficial imitation and generalization failure: In tasks requiring multi-step logical deduction, such as chemical reasoning and complex question answering, student models can only statistically simulate the output patterns of teacher models, failing to inherit traceable and verifiable reasoning bases and core principles. When faced with out-of-distribution (OOD) data, multi-hop reasoning scenarios, or changing boundary conditions, the model's generalization ability and output reliability decrease significantly.
[0008] 3. Lack of support for knowledge intervention and cross-scenario application: Due to the lack of symbolic and structured expression of knowledge itself, traditional knowledge distillation cannot achieve knowledge auditing, error attribution, conflict resolution, and cross-task transfer, making it difficult to meet the application needs of fields with high rigor and high transparency requirements such as medicine, chemistry, and finance.
[0009] II. Technical Bottlenecks of Traditional Knowledge Base Construction Methods:
[0010] Traditional knowledge base construction mainly relies on external static text resources such as encyclopedias and professional classics, and uses technologies such as entity linking, relation extraction, and rule annotation to input knowledge into the database. Its core bottleneck is:
[0011] 1. Static lag and inefficient updating: External knowledge resources have long update cycles and high iteration costs, making it impossible to adapt in real time to emerging concepts in the field of chemistry, dynamically evolving reaction mechanisms, and cross-domain knowledge connections, and making it difficult to cover cutting-edge knowledge and dynamic logic.
[0012] 2. The problem of sparse and fragmented knowledge coverage: A large amount of implicit subject knowledge, cross-knowledge point reasoning chains, and multi-scale relational logic are not explicitly recorded, and the knowledge base has serious problems of fragmentation and silos; at the same time, the cost of manual annotation, rule maintenance and conflict verification is extremely high, making it difficult to achieve large-scale, high-quality knowledge construction.
[0013] 3. Failure to activate the parameterized knowledge potential of large language models: Existing technologies mostly treat large language models as passive text generation or question-answering interfaces, without establishing a systematic transformation process from the model's "parameterized implicit knowledge" to "symbolic explicit knowledge". Furthermore, there is a lack of consistency verification, conflict resolution, and quality control mechanisms for model-generated knowledge, which fails to fully explore the knowledge value within the model.
[0014] III. Fundamental flaws in the cue-guided reasoning extraction method:
[0015] In recent years, scholars have attempted to guide large language models to explicitly output the reasoning process through prompting strategies such as chain-of-thought, self-explanation, and program-of-thought, in order to achieve knowledge extraction. However, these methods still belong to the paradigm of single agent and single generation, and have unavoidable technical shortcomings:
[0016] 1. Consistency and accuracy of knowledge are difficult to guarantee: Reasoning and knowledge generation rely heavily on the wording, structure and guiding logic of the prompt template, lack multi-dimensional verification and conflict resolution mechanisms, and are prone to generating logical contradictions, factual errors or knowledge content with ambiguous boundaries, resulting in prominent model "illusion" problems;
[0017] 2. Knowledge is strongly coupled with scenarios and has poor transferability: The extracted "knowledge" is often deeply bound to specific questions, scenario assumptions, and numerical special cases. It is difficult to extract general principles, core rules, and standardized knowledge units that are detached from specific scenarios, making it impossible to achieve reuse across questions and scenarios.
[0018] 3. Lack of multi-perspective verification and role-based division of labor: The single-agent generation model cannot verify the accuracy, completeness, and teaching applicability of knowledge from different dimensions, making it difficult to distinguish between model illusion and real knowledge, and even more difficult to assess the disciplinary rigor and logical integrity of knowledge.
[0019] The aforementioned technical shortcomings are particularly pronounced in the field of chemistry. The chemical knowledge system possesses strong structure and multi-scale characterization (macroscopic experimental phenomena – microscopic reaction mechanisms – symbolic formulas), and is subject to strict physicochemical constraints (such as mass conservation, charge balance, thermodynamic feasibility, and reaction boundary conditions). This demands extremely high levels of rigor, logic, symbolic representation, and hierarchical coherence in the knowledge. For example, regarding problems related to "calculating the electromotive force of a galvanic cell," it is necessary not only to master the expression of the Nernst equation but also to connect core knowledge points such as the standard electrode potential of redox couples, the influence of ion activity, non-standard state correction rules, and reaction spontaneity criteria. Furthermore, there are strict logical dependencies and applicable boundaries between these knowledge points. If traditional knowledge distillation and single-agent prompt extraction methods are used, only surface reasoning steps can be fitted, failing to explicitly model the logical connections and constraints between concepts. This easily generates knowledge fragments that are "formally correct but mechanistically incorrect" (such as incorrectly applying the standard electrode potential to systems of arbitrary concentrations, violating the applicable premise of the Nernst equation), thus causing serious misguidance in chemical teaching, experimental assistance, and problem-solving scenarios. Meanwhile, chemical knowledge exists in a large amount of the form of symbolic formulas, reaction equations, and material structural formulas. Traditional distillation methods based on pure text distribution fitting are difficult to effectively capture the semantic connotation and constraint rules of symbolic knowledge, further exacerbating the risk of knowledge distortion and logical breakage.
[0020] There are currently no effective solutions to the problems in the relevant technologies. Summary of the Invention
[0021] To address the problems in related technologies, this invention proposes a multi-agent collaborative chemical knowledge explicit distillation method and system based on a cognitive apprenticeship mechanism, in order to overcome the aforementioned technical problems existing in the prior art.
[0022] Therefore, the specific technical solution adopted by the present invention is as follows:
[0023] According to one aspect of the present invention, a multi-agent cooperative chemical knowledge explicit distillation method based on a cognitive apprenticeship mechanism is provided, the method comprising:
[0024] S1. An explicit knowledge distillation framework, including senior student agents, teaching assistant agents, and teacher agents, is used to process the collected chemistry questions and generate a knowledge list.
[0025] S2. The knowledge list is evaluated using a two-dimensional arbitration evaluation mechanism, and the decision-making committee determines whether the evaluation result is qualified. If it is, the knowledge list is transmitted to the learner decision-maker; if not, the tutoring teacher agent is triggered to generate a supplementary knowledge list and transmit it to the learner decision-maker.
[0026] S3. The learner decision-maker verifies whether the received result meets the preset knowledge quality standard. If yes, the received result is stored in the preset knowledge base; otherwise, the explicit knowledge distillation framework is triggered to carry out the next round of explicit distillation process.
[0027] Preferably, an explicit knowledge distillation framework comprising senior student agents, teaching assistant agents, and teacher agents is used to process the collected chemistry questions and generate a knowledge list including:
[0028] S11. Collect chemical questions and their corresponding standard answers, and use the collected results as input for explicit knowledge distillation;
[0029] S12. Based on the preset multi-agent coordination framework, define senior student agent, teaching assistant agent and teacher agent according to role, goal and background to construct an explicit knowledge distillation framework.
[0030] S13. Use the explicit knowledge distillation framework to perform knowledge distillation on the chemical problem, and verify the knowledge distillation results in conjunction with the standard answer to output a knowledge list.
[0031] Preferably, an explicit knowledge distillation framework is used to perform knowledge distillation on the chemical problem, and the results of the knowledge distillation are verified in conjunction with the standard answer. The output knowledge list includes:
[0032] S131. Utilize the senior student agent in the explicit knowledge distillation framework to receive chemistry questions, and trigger the preset large language model based on the received results;
[0033] S132. The chemical problem is analyzed by triggering the preset large language model, and an initial list of knowledge point titles is identified and generated based on the analysis results.
[0034] S133. Input the initial knowledge point title list and the corresponding standard answers into the teaching assistant intelligent agent simultaneously, use the teaching assistant intelligent agent to calibrate the initial knowledge point title list, and output the knowledge point title list.
[0035] S134. The teacher agent receives the list of knowledge point titles and performs a structured transformation of the list of knowledge point titles according to the preset hierarchical rules to generate a knowledge list.
[0036] Preferably, the initial list of knowledge point titles and the corresponding standard answers are simultaneously input into the teaching assistant agent. The teaching assistant agent then calibrates the initial list of knowledge point titles and outputs a list of knowledge point titles including:
[0037] S1331. The teaching assistant intelligent agent synchronously receives the initial knowledge point title list and the corresponding standard answers, and filters the initial knowledge point title list with the standard answers as a reference.
[0038] S1332. Adjust the granularity of the filtered initial knowledge point title list to obtain a candidate knowledge point title list with uniform granularity.
[0039] S1333. Based on the reasoning chain in the standard answer, complete the list of candidate knowledge point titles, standardize the terminology of the completed list of candidate knowledge point titles, and output the list of knowledge point titles.
[0040] Preferably, the teacher agent receives a list of knowledge point titles and performs a structured transformation of the list according to preset hierarchical rules to generate a knowledge list including:
[0041] S1341. Based on the received list of knowledge point titles, the teacher agent writes the definition, principle and logical relationship for each knowledge point title according to the preset textbook standards, and completes the construction of the definition and principle layer;
[0042] S1342. Based on the knowledge definition and principle layer construction results, supplement the formulas, parameters, case templates, misconceptions and constraints corresponding to each knowledge point title to complete the data and application layer construction.
[0043] S1343. By integrating the construction results of the definition and principle layer, data and application layer, the knowledge point titles and the content of each layer are structurally integrated to generate a knowledge list.
[0044] Preferably, a two-dimensional arbitration evaluation mechanism is used to evaluate the knowledge list, and a decision-making committee determines whether the evaluation result is satisfactory. If so, the knowledge list is transmitted to the learner decision-maker; if not, the tutoring teacher agent is triggered to generate a supplementary knowledge list and transmit it to the learner decision-maker, including:
[0045] S21. Based on the preset multi-agent coordination framework, define the tutoring teacher agent, learner decision-maker, accuracy and completeness arbitrator, and objectivity and portability arbitrator according to roles, goals and background.
[0046] S22. Establish a two-dimensional arbitration evaluation mechanism based on arbitrators with accuracy and completeness, and arbitrators with objectivity and portability, and use the two-dimensional arbitration evaluation mechanism to evaluate the knowledge list and generate an evaluation report.
[0047] S23. Based on the pre-set professional division of labor model and multi-dimensional indicator constraints, establish a decision-making committee using a review mechanism;
[0048] S24. The decision-making committee judges the evaluation report according to the preset adjudication rules and confirms whether the evaluation report is qualified. If the evaluation report is qualified, the knowledge list is transmitted to the learner decision-maker. If the evaluation report is unqualified, the tutoring teacher agent is triggered to generate a supplementary knowledge list and transmit it to the learner decision-maker.
[0049] Preferably, a two-dimensional arbitration evaluation mechanism is established based on arbitrators of accuracy and completeness, and arbitrators of objectivity and portability. This two-dimensional arbitration evaluation mechanism is then used to evaluate the knowledge list, generating an evaluation report including:
[0050] S221. Based on the definition of arbitrators with accuracy and completeness, and arbitrators with objectivity and portability, establish a two-dimensional arbitration evaluation mechanism;
[0051] S222. Use a two-dimensional arbitration evaluation mechanism to evaluate the knowledge content and knowledge expression of the knowledge list, and generate evaluation scores and corresponding evaluation suggestions based on the evaluation results.
[0052] S223. Combine the assessment scores with the corresponding assessment recommendations to obtain an assessment report.
[0053] Preferably, the input to the tutoring teacher agent includes an initial list of knowledge point titles, a knowledge list, and an assessment report.
[0054] Preferably, the learner decision-maker verifies whether the received result meets the preset knowledge quality standard. If so, the received result is stored in the preset knowledge base; otherwise, the explicit knowledge distillation framework is triggered to perform the next round of explicit distillation, including:
[0055] S31. The learner decision-maker receives the knowledge list and the knowledge point supplement list, answers the questions based on the corresponding knowledge points, and uses the answer as a feedback signal to generate a confidence score for the feedback signal.
[0056] S32. Verify whether the confidence score meets the preset knowledge quality standard. If it meets the preset knowledge quality standard, store the received result and the corresponding knowledge point into the preset knowledge base. If it does not meet the preset knowledge quality standard, trigger the explicit knowledge distillation framework to carry out the next round of explicit distillation process.
[0057] According to another aspect of the invention, a multi-agent cooperative chemical knowledge explicit distillation system based on a cognitive apprenticeship mechanism is also provided, the system comprising:
[0058] The knowledge list generation module is used to process the collected chemistry questions and generate a knowledge list using an explicit knowledge distillation framework that includes senior student agents, teaching assistant agents, and teacher agents.
[0059] The assessment and decision-making module is used to evaluate the knowledge list using a two-dimensional arbitration assessment mechanism and determine whether the assessment result is qualified through the decision-making committee. If it is qualified, the knowledge list is transmitted to the learner decision-maker; if not, the tutoring teacher agent is triggered to generate a supplementary knowledge list and transmit it to the learner decision-maker.
[0060] The verification and storage module is used to verify whether the received results meet the preset knowledge quality standards through the learner decision-maker. If they do, the received results are stored in the preset knowledge base; otherwise, the explicit knowledge distillation framework is triggered to carry out the next round of explicit distillation process.
[0061] The beneficial effects of this invention are as follows:
[0062] 1. This invention adopts an explicit knowledge distillation framework that integrates senior students, teaching assistants, and teacher agents. It simulates the hierarchical processing of chemical problems by cognitive apprentices, transforming the implicit parameter knowledge of large models into explicit structured knowledge units with clear semantics and logic. This breaks through the limitations of traditional black-box distillation, explicitly models the logical connections and symbolic expressions of chemical knowledge, and greatly improves the interpretability of knowledge and the rigor of the discipline.
[0063] 2. This invention constructs a two-dimensional arbitration assessment and decision-making committee with a two-layer quality control mechanism. It comprehensively assesses the knowledge list from the dimensions of content and expression, identifies and supplements unqualified knowledge, effectively identifies model illusions, knowledge gaps and logical contradictions, adapts to chemically constrained specifications, and significantly improves the accuracy, completeness and cross-scenario reusability of knowledge.
[0064] 3. This invention establishes a closed-loop verification and iterative distillation process through a learner decision-maker. Qualified knowledge is stored in the database, while unqualified knowledge is restarted. This solves the drawbacks of static and solidified traditional knowledge extraction, adapts to the dynamic evolution of chemical knowledge, and builds a dynamic and scalable knowledge base to provide high-quality knowledge support for multi-scenario applications. Attached Figure Description
[0065] To more clearly illustrate the technical solutions in the embodiments of the present invention or the prior art, the drawings used in the embodiments 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.
[0066] Figure 1 This is a flowchart of a multi-agent collaborative chemical knowledge explicit distillation method based on a cognitive apprenticeship mechanism according to an embodiment of the present invention;
[0067] Figure 2 This is a knowledge hierarchy diagram of a multi-agent collaborative chemical knowledge explicit distillation method based on a cognitive apprenticeship mechanism according to an embodiment of the present invention.
[0068] In the picture:
[0069] 1. Knowledge list generation module; 2. Evaluation and decision-making module; 3. Verification and data entry module. Detailed Implementation
[0070] To further illustrate the various embodiments, the present invention provides accompanying drawings, which are part of the disclosure of the present invention. These drawings are mainly used to illustrate the embodiments and can be used in conjunction with the relevant descriptions in the specification to explain the operating principles of the embodiments. With reference to these drawings, those skilled in the art should be able to understand other possible implementation methods and the advantages of the present invention.
[0071] According to embodiments of the present invention, a multi-agent collaborative chemical knowledge explicit distillation method and system based on a cognitive apprenticeship mechanism are provided.
[0072] The present invention will now be further described in conjunction with the accompanying drawings and specific embodiments, such as... Figures 1-2 As shown, according to an embodiment of the present invention, a multi-agent cooperative chemical knowledge explicit distillation method based on a cognitive apprenticeship mechanism includes:
[0073] S1. An explicit knowledge distillation framework, including senior student agents, teaching assistant agents, and teacher agents, is used to process the collected chemistry questions and generate a knowledge list.
[0074] It should be further explained that the present invention aims to solve the core problem of existing knowledge distillation methods, which are difficult to extract the most appropriate explicit knowledge from large language models under the condition of known standard chemical question-answer pairs (Q,A); the so-called "most appropriate" means that the extracted knowledge unit simultaneously satisfies: accuracy (strictly conforms to chemical principles), completeness (covers all key elements required to solve the problem), decontextualization (strips away the question-specific information and retains universal rules) and verifiability (supports logical verification independent of the original question).
[0075] Furthermore, this invention is inspired by cognitive apprenticeship, an educational psychology model that externalizes implicit professional thinking through mechanisms such as expert demonstration, tiered guidance, and reflective correction. This invention transforms this theory into a multi-agent collaborative explicit chemical knowledge distillation framework. This framework uses QA pairs as the sole monitoring signal and, through a role-based workflow, systematically transforms the parametric memory within the large language model into structured knowledge units composed of titles, definitions and principles, and key data / formulas. The entire process strictly isolates early interference from answers in knowledge generation, introducing a two-dimensional arbitration and closed-loop feedback mechanism to effectively suppress factual errors, logical paradoxes, and subjective biases. This invention not only overcomes the limitations of traditional distillation, which focuses solely on model performance transfer while neglecting knowledge reusability, but also constructs a traceable, quality-controlled, and logically transparent chemical knowledge infrastructure, providing solid support for lightweight model training, dynamic knowledge base construction, and highly reliable intelligent chemical education.
[0076] In this optional embodiment, an explicit knowledge distillation framework comprising senior student agents, teaching assistant agents, and teacher agents is used to process the collected chemistry questions and generate a knowledge list including:
[0077] S11. Collect chemical questions and their corresponding standard answers, and use the collected results as input for explicit knowledge distillation.
[0078] S12. Based on the preset multi-agent coordination framework, define senior student agent, teaching assistant agent and teacher agent according to role, goal and background to construct an explicit knowledge distillation framework.
[0079] It should be further explained that the core of the explicit knowledge distillation framework consists of three agents: Senior Agent, Teacher Assistant Agent, and Teacher Agent. These agents distill the required knowledge units from known standard chemistry question-and-answer pairs to obtain a structured natural language knowledge list. Specifically, based on a pre-defined multi-agent coordination framework (Crew AI), the three agents are defined through three dimensions: role, goal, and context.
[0080] 1. Senior student's AI agent:
[0081] Role: Chemistry Problem Analyst (Senior); Goal: To carefully analyze chemistry problems and generate a complete and accurate list of knowledge points needed to solve them; Background: An experienced chemistry educator and mentor, skilled at in-depth analysis of problems, accurately identifying every concept, formula, reaction principle, or calculation method involved, providing a comprehensive and well-organized knowledge list that serves as a reliable knowledge foundation for other agents.
[0082] 2. Teaching Assistant Intelligent Agent:
[0083] Role: Knowledge List Reviewer (Teaching Assistant); Objective: Review the knowledge list submitted by students, correct any issues, and deliver a complete and integrated version; Background: A rigorous reviewer of chemistry knowledge, adept at identifying omissions, redundancies, and logical inconsistencies, and integrating the corrections into a coherent and clear knowledge list; Refinement work ensures that downstream agents can carry out subsequent tasks based on reliable information.
[0084] 3. Teacher's Intelligent Agent:
[0085] Role: Structured Knowledge Writer (Teacher); Goal: To transform the integrated knowledge point list into accurate, well-formatted, and easy-to-use knowledge entries; Background: Senior chemistry teacher, skilled at extracting complex content into clear and easy-to-understand summaries; concise and accurate expression, closely aligned with problem-solving needs, avoiding redundant embellishment while fully preserving key details, ensuring that other intelligent agents can efficiently utilize this knowledge.
[0086] S13. Use the explicit knowledge distillation framework to perform knowledge distillation on the chemical problem, and verify the knowledge distillation results in conjunction with the standard answer to output a knowledge list.
[0087] In this optional embodiment, an explicit knowledge distillation framework is used to perform knowledge distillation on the chemical problem, and the results of the knowledge distillation are verified in conjunction with the standard answer. The output knowledge list includes:
[0088] S131. Utilize the senior student agent in the explicit knowledge distillation framework to receive chemistry questions, and trigger the preset large language model based on the received results.
[0089] S132. The chemical problem is analyzed by triggering the preset large language model, and an initial list of knowledge point titles is generated based on the analysis results.
[0090] It should be further explained that the senior student agent receives raw chemistry questions without standard answers. It activates the reasoning ability of general large language models (such as the Qwen series and DeepSeek series) solely based on the question text, and autonomously identifies and outputs a set of chemistry knowledge points highly relevant to the question, such as redox reactions, Le Chatelier's principle, and the naming rules for coordination compounds, while maintaining semantic relevance. The core objective at this stage is to construct the first layer of the knowledge hierarchy, that is, to output only a list of knowledge point titles, strictly prohibiting the generation of any explanatory content, formula derivations, or suggestive statements that may be related to the answer. The reason for adopting the design of no-answer input and title-level output is to fundamentally cut off the reverse inducement of the answer to the knowledge extraction process, ensuring that the identified knowledge points are entirely derived from a deep analysis of the question's semantics, context, and implicit constraints, rather than a post-hoc rationalization of known results. The final output list of knowledge point titles will serve as the input basis for subsequent teaching assistant agents, providing structured anchors for precise alignment in the next stage while preserving ample room for exploration.
[0091] S133. Input the initial knowledge point title list and the corresponding standard answers into the teaching assistant agent simultaneously, use the teaching assistant agent to calibrate the initial knowledge point title list, and output the knowledge point title list.
[0092] In this optional embodiment, the initial list of knowledge point titles and the corresponding standard answers are synchronously input into the teaching assistant agent. The teaching assistant agent then calibrates the initial list of knowledge point titles and outputs a list of knowledge point titles including:
[0093] S1331. The teaching assistant intelligent agent synchronously receives the initial knowledge point title list and the corresponding standard answers, and filters the initial knowledge point title list with the standard answers as a reference.
[0094] S1332. Adjust the granularity of the filtered initial knowledge point title list to obtain a candidate knowledge point title list with uniform granularity.
[0095] S1333. Based on the reasoning chain in the standard answer, complete the list of candidate knowledge point titles, standardize the terminology of the completed list of candidate knowledge point titles, and output the list of knowledge point titles.
[0096] It should be further explained that after receiving the preliminary, rough, and potentially redundant or generalized list of knowledge point titles output by the senior student agent, the teaching assistant agent enters the knowledge alignment stage. This stage introduces standard answers as contextual reference signals for the first time, but their role is limited to calibration, not generation. That is, they do not directly restate the answers, but rather use the correct problem-solving logic contained in the answers as anchor points to reassess the relevance of the previously identified knowledge points, remove redundancies, adjust granularity, and supplement as necessary. Specifically, the teaching assistant agent performs the following operations:
[0097] Relevance Filtering: Eliminate semantically related knowledge points that do not substantially contribute to the solution of this problem, such as listing crystal field theory when the question involves acid-base titration; Granularity Optimization: Refine overly broad titles, such as refining chemical equilibrium to weak acid ionization equilibrium or the Henderson-Hasselbalch equation for buffer solutions, or merge overly fragmented items, such as integrating oxidation number calculation and half-reaction balancing into redox reaction balancing rules; Missing Information Completion: Based on the reasoning chain implicit in the standard answer, supplement key links that senior students' intelligence agents may have missed, for example, if the answer relies on activity coefficient correction, the ionic strength effect in non-ideal solutions needs to be explicitly added; Terminology Standardization: Use IUPAC or curriculum standard terminology uniformly to ensure subject consistency in knowledge expression.
[0098] This step achieves precise convergence from the potential knowledge set driven by the problem to the necessary knowledge set aligned with the task. It retains the breadth of the previous divergent exploration while enhancing the relevance of knowledge and the relevance of teaching through the constraint of the answer. Crucially, the standard answer is introduced only at this stage, strictly avoiding its contamination effect in the initial knowledge recognition stage. This effectively prevents the model from developing confirmation bias or constructing conforming interpretations due to prior knowledge of the answer, ensuring the objectivity and transferability of the final distilled knowledge. The complete list of knowledge point titles output at this stage will serve as the structural skeleton for the teacher agent to build complete knowledge content, providing a clear, focused, and non-redundant guiding framework for the subsequent generation of definitions, principles, formulas, and typical examples.
[0099] S134. The teacher agent receives the list of knowledge point titles and performs a structured transformation of the list of knowledge point titles according to the preset hierarchical rules to generate a knowledge list.
[0100] In this optional embodiment, the teacher agent receives a list of knowledge point titles and performs a structured transformation on the list according to preset hierarchical rules to generate a knowledge list including:
[0101] S1341. Based on the received list of knowledge point titles, the teacher agent writes the definition, principle and logical relationship for each knowledge point title according to the preset textbook standards, and completes the construction of the definition and principle layer;
[0102] S1342. Based on the knowledge definition and principle layer construction results, supplement the formulas, parameters, case templates, misconceptions and constraints corresponding to each knowledge point title to complete the data and application layer construction.
[0103] S1343. By integrating the construction results of the definition and principle layer, data and application layer, the knowledge point titles and the content of each layer are structurally integrated to generate a knowledge list.
[0104] It should be added that, for the complete list of knowledge point titles output by the teaching assistant agent, the teacher agent is responsible for constructing a complete explicit knowledge expression, specifically completing the second layer (definition and principle layer) and the third layer (key data and application layer) of the knowledge hierarchy. This process strictly follows the three principles of decontextualization, universality, and structuring to ensure that the generated content does not depend on the specific values, experimental conditions, or answer directions of the original questions.
[0105] In the second layer, the teacher agent provides the following for each knowledge point:
[0106] Precise definition: Use standard expressions from IUPAC or mainstream textbooks to clearly define the boundaries of concepts. For example, Le Chatelier's principle describes the response trend of a closed system to external disturbances in equilibrium, rather than the rate or path. Core principle: Clarify its physical / chemical nature and applicable premises. For example, the Nernst equation is only applicable to electrochemical equilibrium, and when the activity coefficient is assumed to be 1, it should be noted that it is an approximation. Logical relationship: Explain the dependence or difference with other knowledge points when necessary. For example, the naming rules for coordination compounds are based on the oxidation state of the central metal, which is different from the systematic nomenclature of covalent molecules.
[0107] In the third layer, further supplementation is provided:
[0108] Key formulas and parameters: including standard form, variable meanings and units, such as This formula is the Nernst Equation, used to calculate the actual electrode potential of an electrochemical cell under non-standard conditions. ,in, It is the standard electrode potential. It is the gas constant (8.314 J·mol⁻¹·K⁻¹). It is absolute temperature (unit: K). It is the number of electrons transferred during the reaction. It is the Faraday constant (approximately 96485 C·mol⁻¹), while This is the reaction quotient, representing the ratio of the current concentrations (or partial pressures) of reactants and products; this equation shows that the battery potential changes with conditions such as concentration and temperature, when all substances are in standard conditions ( )hour, Typical examples: Select representative, general cases that deviate from the original problem's context, such as using a Zn / Cu²⁺ galvanic cell to illustrate electromotive force calculations, rather than reusing specific redox pairs from the original problem; Common pitfalls and limitations: Proactively point out common mistakes, such as neglecting the self-ionization of water in buffer solution pH calculations only when [HA]≫10⁻ 6 The equation holds true under M conditions, or has applicable boundaries, such as the ideal gas law failing under high pressure.
[0109] The entire explanation process completely hides the standard answers to the original questions, using only the titles of knowledge points as input. This ensures that the generated content focuses on the universality of the knowledge itself, rather than a retrospective explanation of a specific problem-solving path. This answer-less paraphrasing mechanism fundamentally avoids the erroneous generalization of question-specific information (such as the specific temperature, concentration, or products of a reaction) into general rules, thus guaranteeing the portability of distillation knowledge. That is, the same knowledge unit can be reliably applied to similar but different instances of problems, supporting cross-question transfer and combinatorial reasoning. The final output is a set of clearly structured, semantically complete, and independently verifiable natural language knowledge units, which form the basic carrier for subsequent arbitration assessment and student application, and also provide high-quality atomized materials for building a dynamic and updatable chemical structured knowledge base.
[0110] Furthermore, through the design of question input, knowledge extraction (without answer), knowledge refinement (with answer), and knowledge explanation (repetition without answer), a high degree of relevance (knowledge points closely related to the question) and high objectivity (avoiding hints about the answer) are cleverly achieved.
[0111] S2. The knowledge list is evaluated using a two-dimensional arbitration evaluation mechanism, and the decision-making committee determines whether the evaluation result is qualified. If it is, the knowledge list is transmitted to the learner decision-maker; if not, the tutoring teacher agent is triggered to generate a supplementary knowledge list and transmit it to the learner decision-maker.
[0112] It should be further explained that, to ensure the quality of knowledge distillation, the system introduces a two-dimensional arbitration evaluation mechanism to ensure that the knowledge is the most appropriate set required. This mechanism mainly consists of three arbitrators plus a tutor, specifically structured as follows: Also based on a pre-defined multi-agent coordination framework, defined by roles, goals, and context:
[0113] 1. Tutor's AI Agent:
[0114] Role: Knowledge Supplement Tutor; Goal: To supplement and improve the knowledge explanations based on the review committee's feedback, ensuring that all identified knowledge gaps or problems are properly addressed; Background: An expert tutor specializing in optimizing chemistry education content. When the review committee points out omissions, ambiguities, or deficiencies in the knowledge presentation, the tutor carefully analyzes the feedback, accurately supplements the relevant content, maintaining scientific accuracy while improving the clarity and completeness of the presentation, and strictly adheres to the structured format and quality standards of the original knowledge entries to ensure the consistency and reliability of the overall content.
[0115] 2. Student Agent:
[0116] After knowledge is verified through arbitration, the learner's decision-maker performs specific tasks based on the explicit knowledge, such as answering questions, completing blanks, and explaining phenomena, and uses the task execution results as the final feedback signal; if the task fails or the confidence level is low, the knowledge backtracking and correction process can be triggered to form a closed-loop optimization.
[0117] 3. Accuracy & Completeness Arbiter:
[0118] Role: Content Accuracy and Completeness Arbitrator (CAC); Goal: To rigorously review the entire workflow from the perspective of content accuracy and completeness; Background: Specializes in auditing the accuracy and completeness of knowledge content in complex problem-solving processes, focusing on whether the content of each knowledge point is factually accurate, precisely expressed, logically consistent, and has sufficient detail and completeness to effectively support problem-solving, clearly record the problems found, and provide practical improvement suggestions.
[0119] 4. Objectivity & Portability Arbiter:
[0120] Role: Content Objectivity Arbitrator (CO); Goal: To rigorously review the entire workflow from the perspective of content objectivity; Background: Specializes in auditing the objectivity of knowledge content in complex problem-solving processes, focusing on: whether the description of each knowledge point is sufficiently objective, whether it avoids subjective judgments, personal opinions, or biased wording, to ensure good knowledge portability, prevent answer leakage, clearly record problems found, and provide practical improvement suggestions.
[0121] 4. Decision-making Committee (Committee):
[0122] Role: Chairperson of the Arbitration Committee; Objective: To synthesize the opinions of all arbitrators and form a coherent and consistent final assessment and decision; Background: An experienced committee chairperson skilled in coordinating and integrating different review opinions, able to systematically compare the findings of all arbitrators, extract consensus, clarify differences, and make a clear, decisive, and well-founded final ruling on this basis, ensuring that the entire knowledge review process is efficient, fair, and the conclusions are reliable.
[0123] In this optional embodiment, a two-dimensional arbitration evaluation mechanism is used to evaluate the knowledge list, and a decision-making committee determines whether the evaluation result is qualified. If qualified, the knowledge list is transmitted to the learner decision-maker; if not, the tutoring teacher agent is triggered to generate a supplementary knowledge point list and transmit it to the learner decision-maker, including:
[0124] S21. Based on the preset multi-agent coordination framework, define the tutoring teacher agent, learner decision-maker, accuracy and completeness arbitrator, and objectivity and portability arbitrator according to roles, goals, and background.
[0125] S22. Establish a two-dimensional arbitration evaluation mechanism based on arbitrators of accuracy and completeness, and arbitrators of objectivity and portability, and use the two-dimensional arbitration evaluation mechanism to evaluate the knowledge list and generate an evaluation report.
[0126] In this optional embodiment, a two-dimensional arbitration evaluation mechanism is established based on arbitrators of accuracy and completeness, and arbitrators of objectivity and portability. This two-dimensional arbitration evaluation mechanism is then used to evaluate the knowledge list and generate an evaluation report, including:
[0127] S221. Based on the definition of arbitrators with accuracy and completeness, and arbitrators with objectivity and portability, establish a two-dimensional arbitration evaluation mechanism;
[0128] S222. Use a two-dimensional arbitration evaluation mechanism to evaluate the knowledge content and knowledge expression of the knowledge list, and generate evaluation scores and corresponding evaluation suggestions based on the evaluation results.
[0129] S223. Combine the assessment scores with the corresponding assessment recommendations to obtain an assessment report.
[0130] It should be further explained that the Accuracy & Completeness Arbiter is used to assess whether the knowledge content conforms to the facts of chemistry and whether it covers key elements (such as reaction conditions, limiting assumptions, etc.), as detailed below:
[0131] Accuracy: Are the knowledge points factually accurate? Are there any conceptual errors? Are definitions, formulas, and key data presented correctly and precisely? Completeness: Is the content of each knowledge point sufficiently detailed and complete? Does it provide enough necessary information to support general applications, such as definitions, key data, and formulas? Logical Rigor: Is the explanation logically consistent? Are the cited formulas or concepts used correctly? Omissions and Problems: Are there any knowledge points that are too vague, lack key details, or contain factual errors? Please provide specific suggestions on how to correct inaccuracies and enhance the completeness of specific knowledge points.
[0132] The Objectivity & Portability Arbiter is used to examine whether the knowledge representation is neutral and free from subjective bias, and to determine whether it is applicable to similar problems, rather than just suitable for the current question, as follows:
[0133] Objectivity and Portability: Is the description of the knowledge points written in an objective and factual manner, applicable to similar questions? Does it avoid subjective language, personal opinions, or value judgments, thus avoiding being tied to a specific question? Prevention of Answer Leakage: Does the description avoid revealing the correct answer or problem-solving strategy? Are there any hints, clues, question-specific references, or implicit guidance that make the answer obvious or directly point to the correct option? Neutrality: Is the language neutral and unbiased? Are there any emotional expressions, evaluative statements, or persuasive language that should be removed? Factual Presentation: Are the facts presented as facts without unnecessary emphasis or explanation? Is the tone appropriate for educational content that should be reusable? Generalization Ability: Can these knowledge points be understood and applied to other similar questions without modification, or do they contain question-specific references that limit their portability? Please provide specific suggestions on how to make the description more objective and portable, including examples of language that may reveal or imply the answer, such as question-specific references, direct mentions of the correct option, or implicit problem-solving strategies, and propose objective alternative expressions.
[0134] S23. Based on the pre-set specialized division of labor model and multi-dimensional indicator constraints, establish a decision-making committee using a review mechanism.
[0135] It should be further explained that, by combining the scores and suggestions from both arbitrators, and using preset rules, if any dimension falls below a threshold, the decision is rejected or a weighted fusion strategy is employed to make a final ruling. This review mechanism draws on the LLM as Judge concept, but does not rely on a single model for judgment. Instead, it achieves more reliable knowledge quality control through specialized division of labor and multi-dimensional indicator constraints. Specifically, it includes: consensus points; points of disagreement (write None if there is no disagreement); supplementary suggestions (write None if no supplementation is needed); final ruling: passed / requires supplementation / modification; if passed, the process can proceed to the solution stage; if supplementation / modification is required, please clearly indicate the highest priority item that must be corrected first; remarks (optional), used to explain observations or provide warnings.
[0136] S24. The decision-making committee judges the evaluation report according to the preset adjudication rules and confirms whether the evaluation report is qualified. If the evaluation report is qualified, the knowledge list is transmitted to the learner decision-maker. If the evaluation report is unqualified, the tutoring teacher agent is triggered to generate a supplementary knowledge list and transmit it to the learner decision-maker.
[0137] In this optional embodiment, the input to the tutoring teacher agent includes an initial list of knowledge point titles, a knowledge list, and an assessment report.
[0138] It should be further explained that the Tutor Agent (temperature 0.7, because the tutoring is executed after the review committee's rejection, so the model needs to widely incorporate the review committee's opinions) is activated when the knowledge distillation result is deemed unqualified by the decision committee. This agent receives three types of input: the initial list of knowledge point titles generated by the senior student agent; the complete knowledge content processed by the teaching assistant agent and the teacher agent, i.e., the knowledge list, including titles, definitions, formulas, examples, etc.; and the structured feedback provided by the two-dimensional arbitrator, i.e., the evaluation report, such as insufficient accuracy (failure to mention the thermodynamic feasibility constraints of the reaction) or weak portability (examples overly relying on the original question values). Based on the above information, the Tutor Agent performs targeted knowledge reconstruction. Its core task is not simple regeneration, but supervised iterative correction.
[0139] For knowledge points with logical conflicts or ambiguous expressions, the tutoring agent can trigger local re-reasoning, invoking the LLM to regenerate the content of that submodule under controlled prompts, and then resubmit it for arbitration: A complete knowledge list is republished, covering all gaps identified by the committee, with each item using a title and definition / key data / function format, totaling three sentences; existing knowledge points are updated or new items are added based on committee feedback; if a knowledge point needs to be replaced, the original title should be retained; if new content is added, a new title should be used; any outdated or conflicting content is deleted to ensure the final list fully reflects the latest guidance. The requirements are as follows: No additional comments, case studies, tables, or lists are allowed; only concise statements are permitted. The final answer should not be revealed or the correct option implied. The purpose should describe the general application of the knowledge point, not a specific problem. This process can be iterative, typically with a maximum number of retries set, such as 2–3 rounds, until the knowledge unit passes committee review. By introducing a tutoring agent, the system achieves a paradigm shift from single-generation to feedback-driven refinement, significantly improving the robustness and fault tolerance of knowledge distillation, ensuring that the final output knowledge is not only correct but also possesses instructional clarity and generalizability.
[0140] S3. The learner decision-maker verifies whether the received result meets the preset knowledge quality standard. If yes, the received result is stored in the preset knowledge base; otherwise, the explicit knowledge distillation framework is triggered to carry out the next round of explicit distillation process.
[0141] In this optional embodiment, the learner decision-maker verifies whether the received result meets the preset knowledge quality standard. If yes, the received result is stored in the preset knowledge base; otherwise, the explicit knowledge distillation framework is triggered to perform the next round of explicit distillation, including:
[0142] S31. The learner decision-maker receives the knowledge list and the knowledge point supplement list, answers the questions based on the corresponding knowledge points, and uses the answer as a feedback signal to generate a confidence score for the feedback signal.
[0143] S32. Verify whether the confidence score meets the preset knowledge quality standard. If it meets the preset knowledge quality standard, store the received result and the corresponding knowledge point into the preset knowledge base. If it does not meet the preset knowledge quality standard, trigger the explicit knowledge distillation framework to carry out the next round of explicit distillation process.
[0144] The workflow of this invention is based on explicit chemical knowledge distillation, which is a multi-agent collaborative verification process used to extract, verify, and refine a high-quality, transferable list of knowledge points from an NMR spectrum selection question. The entire process is strictly executed according to the following steps:
[0145] 1. Question Input and Preliminary Knowledge Recognition (Senior Agent):
[0146] Input question: Determine how many signals the compound (CH3)3CCH2CHO has in ¹H NMR and ¹³C NMR, respectively; Senior agent (model: deepseek-r1-250528, temperature 0.0) analyzes the question and lists 20 preliminary knowledge topics, covering NMR basics, symmetry, functional group characteristics, etc.
[0147] 2. Knowledge Title Review and Simplification (Teacher Assistant):
[0148] The Teacher Assistant (using the same model) reviews the Senior's list against the known correct answers (three ¹H signals and four ¹³C signals); removes redundant or non-core entries, such as splitting patterns and integral principles, and retains 15 necessary and non-overlapping knowledge titles.
[0149] 3. Knowledge Point Content Filling (Teacher):
[0150] Teacher (model: deepseek-r1-250528, temperature 0.7) writes three sentences for each title: Definition; Key data / formula; Role (emphasizing generality, not specific to any particular title).
[0151] 4. Two-dimensional arbitration assessment:
[0152] (1) Accuracy & Completeness Arbiter:
[0153] Three errors / missing information were found: functional group identification misused infrared (IR) data and should be replaced with NMR chemical shifts; quaternary carbon detection incorrectly described DEPT spectra, as quaternary carbons are not visible in any DEPT; and the molecular symmetry definition omitted the inversion operation and lacked examples.
[0154] (2) Objectivity & Portability Arbiter:
[0155] The above issues will reduce the reusability of knowledge or mislead problem-solving; it is emphasized that all content must be limited to the NMR field, avoid implying answers, and use neutral language.
[0156] 5. Committee Chair's Decision:
[0157] Model: deepseek-v3 (0°); Decision: Needs supplementation; Three high-priority modifications are specified: replace IR data for functional group identification with NMR chemical shifts; revise the description of DEPT for quaternary carbon detection; improve the definition of molecular symmetry and add examples such as benzene.
[0158] 6. Knowledge supplementation and updating (Tutor):
[0159] Tutor (model: deepseek-r1-250528) rewrote all 15 knowledge points based on the ruling, ensuring: data accuracy (e.g., quaternary carbon is not visible in DEPT); complete content (e.g., adding typical examples such as ethanol and benzene); domain focus (using only NMR data); and objective and transferable language.
[0160] 7. Final Solution:
[0161] The Solver (model: qwen3-8b) performs reasoning based on the revised knowledge base: the molecule contains three proton environments (aldehyde H, methylene CH2, tert-butyl 9H) → 3 ¹H signals; it contains four carbon environments (aldehyde C, CH2, quaternary carbon C, three equivalent methyl C) → 4 ¹³C signals; the output answer is: option C (corresponding to three ¹H signals and four ¹³C signals); consistent with the standard answer, verification successful.
[0162] 8. Knowledge Archiving: The final 15 high-quality knowledge points are saved in structured JSON format to TAKD.json. This process embodies a high-quality knowledge distillation paradigm of multi-agent division of labor and cooperation, explicit knowledge construction, double verification, and continuous iteration, which is particularly suitable for highly rigorous fields such as chemistry.
[0163] Using this framework, effective knowledge was extracted from 773 questions out of 1244 questions in Chembench (a dataset of chemistry questions), totaling 8951 pieces of knowledge, with an extraction rate of 62.14%. Furthermore, the top 100 questions from Chembench were selected and compared with some established benchmarks. These included the pass rate of the two-dimensional arbitrator, the accuracy of the student arbitrator before and after adding a retry mechanism, and the measurement of the accuracy of the new student arbitrator model to assess the portability of the knowledge itself. SetterKD, or direct distillation, also known as the question setter's perspective, directly allows the model to accept the questions and answers and provide the required knowledge. Mind chain, multi-way voting, and model debate are all methods used to reduce the direct answer hints, as shown in Table 1.
[0164] Table 1. Result Table for Question 1-100
[0165] Judgment device pass rate Initial accuracy Final accuracy Retry accuracy Transplant accuracy (SetterKD) provides the answer directly. 3% 57% 62% 11.63% 87.10% (TAKD) This invention 7% 55% 62% 15.55% 87.10% (COT) Mind Chain 0% 50% 58% 16% 86% (Major) Multiple voting channels 3% 51% 59% 16.33% 92% (Debate) Model Debate 4% 55% 60% 11.11% 86.67%
[0166] Using a large model as the referee and setting up a two-dimensional arbitrator, the accuracy, completeness, portability, and objectivity of the distillation method itself were evaluated. Experiments showed that, despite the overly stringent evaluation mechanism, the present invention had a significantly higher pass rate than other methods. Further evaluation of objectivity alone revealed that the pass rate of the present invention could reach 24%. Furthermore, by statistically analyzing the accuracy before and after adding the retry mechanism, the effectiveness of distilling explicit knowledge using different methods was deeply evaluated. The present invention showed significant advantages over thought chains, multi-way voting, and multi-model debate. Moreover, after retrying, it directly approached the method with obvious answer hints (SetterKD), even though the prompts emphasized the generality of knowledge. This further demonstrated the overall superiority of the present invention's framework. Subsequently, the model of the student agent was changed (from Qwen3-8B to DeepSeek-R1-0528) to evaluate the transferability of knowledge, and the present invention still maintained a good accuracy rate.
[0167] In an example of this invention, the observed signals in the ¹H NMR and ¹³C NMR spectra of compound (CH3)3CCH2CHO (3,3-dimethylbutyraldehyde) were specifically: five ¹H signals and four ¹³C signals; five ¹H signals and six ¹³C signals; three ¹H signals and four ¹³C signals; and three ¹H signals and six ¹³C signals. The performance of SetterKD and TAKD was compared as follows:
[0168] I. Key points of TAKD:
[0169] 1. Knowledge Point Title: Analysis of Organic Molecular Structures Used in Nuclear Magnetic Resonance (NMR) Spectroscopy
[0170] Definition: By analyzing chemical shifts, coupling splitting modes, and signal integrals in nuclear magnetic resonance (NMR) spectra, the connection modes and functional groups of molecules can be determined; Key data: None; Function: Enables chemists to infer the complete structure of unknown organic molecules by correlating spectral data.
[0171] 2. Knowledge Point Title: The Principle of Chemical Equivalence in Nuclear Magnetic Resonance:
[0172] Definition: If atoms are in exactly the same electronic environment, they will have the same resonance frequency in the nuclear magnetic resonance spectrum, which is called chemical equivalence; Key data: None; Function: Used to predict the number of different signals in the nuclear magnetic resonance spectrum and simplify the structure analysis process.
[0173] 3. Knowledge Point Title: ¹Environmental Assessment of Hydrogen Atoms in HNMR:
[0174] Definition: By evaluating the chemical shift, splitting mode, and integral value of a proton, the symmetry of its neighboring atoms and molecules can be inferred; Key data: The chemical shift of a proton typically ranges from 0 to 14 ppm; Function: Used to identify proton types and their local electronic environments, and to assist in the characterization of functional groups.
[0175] 4. Knowledge Point Title: Environmental Assessment of Carbon Atoms in ¹³CNMR:
[0176] Definition: By analyzing the chemical shift and signal multiplicity of carbon, we can determine its hybridization state, functional group type, and molecular symmetry; Key data: The chemical shift range of carbon is usually 0–220 ppm; Function: To reveal the diversity of the carbon skeleton and distinguish structural types such as alkyl, carbonyl, or aromatic systems.
[0177] 5. Knowledge Point Title: The Influence of Symmetry on Atomic Equivalence:
[0178] Definition: Molecular symmetry operations (such as rotation and mirror reflection) can place different atoms in equivalent positions, thereby merging them into a single signal in the NMR spectrum; Key data: None; Function: Reduces spectral complexity and helps identify symmetrical structural segments (such as cyclic structures or alkyl chains).
[0179] 6. Key Point Title: NMR Characteristics of Aldehyde Functional Groups:
[0180] Definition: Due to the electron withdrawing effect, the carbonyl carbon and formyl protons of the aldehyde group exhibit a significant low-field shift in the NMR spectrum; Key data: Aldehyde protons: δ9–10 ppm; Carbonyl carbon: δ190–205 ppm; Function: These signals can serve as diagnostic markers of the presence of the aldehyde group for spectral analysis.
[0181] 7. Knowledge Point Title: Equivalence Analysis in Symmetrical Alkyl Groups (e.g., tert-butyl):
[0182] Definition: Symmetrical alkyl groups, due to rapid rotation or molecular symmetry, have chemically equivalent protons or carbon atoms, producing only a single signal in the NMR spectrum; Key data: tert-butyl protons exhibit a single peak at δ~1.3 ppm; Function: To interpret the simplified spectral patterns in symmetrical substituents, aiding in group identification.
[0183] 8. Knowledge Point Title: The Proton and Carbon Environment of Methylene (–CH2–):
[0184] Definition: The chemical shift of the methylene group varies due to the influence of adjacent electron-withdrawing or electron-donating groups; Key data: Methylene protons: δ0.5ppm (alkyl) to δ4–5ppm (allylic / benzyl); Carbon: δ10–50ppm; Function: Helps to distinguish different types of methylene groups (e.g., esters vs. ethers) and achieve accurate structural assignment.
[0185] 9. Knowledge Point Title: Detection of Quaternary Carbon in ¹³ CNMR:
[0186] Definition: Quaternary carbon atoms are not bonded to hydrogen atoms and show no signal in DEPT spectra, but are visible in broadband decoupled ¹³ CNMR spectra; Key data: Quaternary carbon atoms show no signal in both DEPT-135 and DEPT-90 spectra; Role: Identifying proton-free carbon atoms is crucial for constructing molecular skeletons.
[0187] 10. Knowledge Point Title: Systematic Counting Methods for the Number of Nuclear Magnetic Resonance Signals:
[0188] Definition: By assessing molecular symmetry and chemical environment, the system identifies equivalent sets of atoms, thereby determining the number of different NMR signals; Key data: None; Function: Predicts the minimum number of signals in NMR spectra, providing a basis for preliminary structure analysis.
[0189] II. SetterKD Knowledge Points:
[0190] 1. Knowledge Point Title: Proton Nuclear Magnetic Resonance (¹HNMR) Signal Counting:
[0191] Definition: ¹H NMR shows that the number of different signals corresponds to the number of chemically inequivalent hydrogen atom environments in a molecule; Key data: None; Function: Can be used to predict spectral complexity and assist in the structural analysis of organic compounds.
[0192] 2. Knowledge Point Title: ¹Chemical Equivalence in HNMR:
[0193] Definition: Due to molecular symmetry or rapid rotation, hydrogen atoms are chemically equivalent if they are in the same magnetic environment; Key data: None; Function: Determines the amount of signal, simplifying spectral analysis by reducing the amount of observable signal.
[0194] 3. Knowledge Point Title: NMR Characteristics of Tert-Butyl Groups:
[0195] Definition: The (CH3)3C– group, due to free rotation, makes the three methyl groups equivalent, producing only one signal in ¹H NMR; Key data: It usually appears as a single 9H peak with a chemical shift of about 1.0 ppm; Function: It provides diagnostic signal patterns of branched alkane structures in NMR spectra.
[0196] 4. Key Point Title: The Methylene Proton Adjacent to the Carbonyl Group:
[0197] Definition: The CH2 group directly attached to the aldehyde group is significantly deshielded due to the electron-withdrawing effect of the carbonyl group; Key data: It usually appears as a 2H triplet with a chemical shift of about 2.4 ppm; Function: It helps to locate the position of functional groups in the molecule through characteristic chemical shifts.
[0198] 5. Knowledge Point Title: Nuclear Magnetic Resonance Characteristics of Aldehyde Protons:
[0199] Definition: The formyl proton (CHO) appears at a low field due to the strong deshielding effect of the carbonyl group; Key data: It usually appears as a 1H singlet with a chemical shift of about 9–10 ppm; Function: It provides clear evidence for the presence of aldehyde functional groups in organic molecules.
[0200] 6. Knowledge Point Title: Carbon-13 Nuclear Magnetic Resonance (¹³CNMR) Signal Counting:
[0201] Definition: The number of different signals in ¹³ CNMR corresponds to the number of chemically inequivalent carbon atoms; Key data: None; Function: Reveals molecular symmetry and helps determine the carbon skeleton of organic compounds.
[0202] 7. Knowledge Point Title: Detection of Quaternary Carbon:
[0203] Definition: Carbons without hydrogen atoms (such as the central carbon in (CH3)3C–) produce a weak but identifiable signal in ¹³ CNMR; Key data: Aliphatic quaternary carbons typically appear in 30–50 ppm; Role: Crucial for the resolution of complete structures, as these carbons are not visible in ¹H NMR spectra.
[0204] 8. Knowledge Point Title: Chemical Shift of Carbonyl Carbon:
[0205] Definition: Aldehyde carbon exhibits a characteristic shift at low fields due to a strong deshielding effect; Key data: Typically appears in 190–205 ppm; Function: Provides conclusive evidence for carbonyl functional groups in ¹³ CNMR spectra.
[0206] 9. Key Point Title: The Role of Symmetry in NMR Analysis:
[0207] Definition: Molecular symmetry reduces the number of observable NMR signals by making multiple atomic nuclei equivalent; Key data: None; Function: Reduces the number of signals while preserving structural information, simplifying spectral analysis.
[0208] 10. Knowledge Point Title: Properties of Active Nuclei in NMR:
[0209] Definition: Both ¹H and ¹³C have nuclear spin (I=½), and therefore can be detected by NMR spectroscopy; Key data: The natural abundance of ¹H is approximately 99.98%, and that of ¹³C is approximately 1.1%; Comparing different answers to the same knowledge point, it was found that SetterKD's answer had a clear bias towards the answer, while TAKD did not have this problem.
[0210] III. Comparison Results:
[0211] 1. First point: NMR characteristics of tert-butyl:
[0212] According to TAKD, symmetrical alkyl groups, due to rapid rotation or molecular symmetry, have chemically equivalent protons or carbon atoms, producing only a single signal in the NMR spectrum. Tert-butyl protons exhibit a single peak at δ~1.3 ppm.
[0213] According to SetterKD, the (CH3)3C– group results in three methyl groups being equivalent due to free rotation, producing only one signal in ¹H NMR. This typically manifests as a single 9H peak with a chemical shift of approximately 1.0 ppm.
[0214] Hint: SetterKD directly states that only one signal is generated and provides integral information about the 9H single peak, telling the reader that it contributes 1 signal and has 9 hydrogens.
[0215] 2. Second point: The methylene proton adjacent to the carbonyl group (the most obvious clue):
[0216] According to TAKD, the chemical shift of the methylene group varies due to the influence of adjacent electron-withdrawing or electron-donating groups. Methylene protons: δ0.5 ppm (alkyl) to δ4–5 ppm (allylic / benzyl).
[0217] According to SetterKD, the CH2 group directly attached to the aldehyde group is significantly deshielded due to the electron-withdrawing effect of the carbonyl group. This typically manifests as a 2H triplet with a chemical shift of approximately 2.4 ppm.
[0218] Hint: SetterKD explicitly states that CH2 is directly linked to the aldehyde group, which is a description of the specific molecule (CH3)3CCH2CHO in the question; at the same time, it gives a 2H triplet, directly telling the reader that this is 1 signal and 2 hydrogens.
[0219] 3. Thirdly: NMR characteristics of aldehyde matrix particles:
[0220] According to TAKD: The carbonyl carbon and formyl protons of the aldehyde group exhibit a significant low-field shift in the NMR spectrum; aldehyde protons: δ9–10 ppm; carbonyl carbon: δ190–205 ppm.
[0221] According to SetterKD, the formyl proton (CHO) appears at a low field due to the strong deshielding effect of the carbonyl group. It usually appears as a single 1H peak with a chemical shift of about 9–10 ppm.
[0222] Hint: SetterKD explicitly states a single peak at 1H, directly telling the reader that this is a signal.
[0223] 4. Fourth Article: Detection of Quaternary Carbon:
[0224] According to TAKD: Quaternary carbon atoms are not connected to hydrogen atoms and have no signal in DEPT spectrum, but are visible in broadband decoupled ¹³ CNMR spectrum; quaternary carbons have no signal in both DEPT-135 and DEPT-90 spectra.
[0225] According to SetterKD, carbons without hydrogen atoms (such as the central carbon in (CH3)3C–) produce a weak but identifiable signal in ¹³ CNMR; aliphatic quaternary carbons typically appear at 30–50 ppm.
[0226] Hint: SetterKD directly names the central carbon in (CH3)3C–, analyzing the specific structure of the molecule in question.
[0227] 5. Article 5: Chemical shift of the carbonyl carbon:
[0228] TAKD approach: Incorporate the information of the carbonyl carbon into the NMR characteristics of the aldehyde functional group.
[0229] The SetterKD approach involves listing the chemical shift of the carbonyl carbon as a separate knowledge point, demonstrating that it provides conclusive evidence for the carbonyl functional group in ¹³C NMR spectra.
[0230] Hint: Listing this point separately implies that carbonyl carbon needs to be counted as a separate signal.
[0231] IV. Summary:
[0232] If the question asks for the number of signals in the ¹H NMR spectrum of (CH3)3CCH2CHO, the answer can be obtained directly by addition based on the knowledge of Setter-KD:
[0233] The tert-butyl group contributes 1 signal (the knowledge mentions a 9H singlet), the methylene group contributes 1 signal (the knowledge mentions a 2H triplet), and the aldehyde group contributes 1 signal (the knowledge mentions a 1H singlet). The answer is 1 + 1 + 1 = 3 signals. Students do not need to understand the principles of NMR or do any reasoning; they only need to count how many times the "one signal" appears in the knowledge points and add them up to get the answer. This is why SetterKD provides hints about the answer.
[0234] According to another embodiment of the present invention, a multi-agent cooperative chemical knowledge explicit distillation system based on a cognitive apprenticeship mechanism is also provided, the system comprising:
[0235] The knowledge list generation module 1 is used to process the collected chemistry questions and generate a knowledge list using an explicit knowledge distillation framework that includes senior student agents, teaching assistant agents, and teacher agents.
[0236] The assessment and decision-making module 2 is used to evaluate the knowledge list using a two-dimensional arbitration assessment mechanism and determine whether the assessment result is qualified through the decision-making committee. If it is qualified, the knowledge list is transmitted to the learner decision-maker; if not, the tutoring teacher agent is triggered to generate a supplementary knowledge list and transmit it to the learner decision-maker.
[0237] The verification and storage module 3 is used to verify whether the received results meet the preset knowledge quality standards through the learner decision-maker. If they do, the received results are stored in the preset knowledge base; otherwise, the explicit knowledge distillation framework is triggered to carry out the next round of explicit distillation process.
[0238] The above description is only a preferred embodiment of the present invention and is not intended to limit the present invention. Any modifications, equivalent substitutions, improvements, etc., made within the spirit and principles of the present invention should be included within the protection scope of the present invention.
Claims
1. A multi-agent collaborative chemical knowledge explicit distillation method based on a cognitive apprenticeship mechanism, characterized in that, The method includes: S1. An explicit knowledge distillation framework, including senior student agents, teaching assistant agents, and teacher agents, is used to process the collected chemistry questions and generate a knowledge list. S2. The knowledge list is evaluated using a two-dimensional arbitration evaluation mechanism, and the decision-making committee determines whether the evaluation result is qualified. If it is, the knowledge list is transmitted to the learner decision-maker; if not, the tutoring teacher agent is triggered to generate a supplementary knowledge list and transmit it to the learner decision-maker. S3. The learner decision-maker verifies whether the received result meets the preset knowledge quality standard. If yes, the received result is stored in the preset knowledge base; otherwise, the explicit knowledge distillation framework is triggered to carry out the next round of explicit distillation process.
2. The multi-agent collaborative chemical knowledge explicit distillation method based on a cognitive apprenticeship mechanism according to claim 1, characterized in that, The method employs an explicit knowledge distillation framework comprising senior student agents, teaching assistant agents, and teacher agents to process the collected chemistry questions and generate a knowledge list including: S11. Collect chemical questions and their corresponding standard answers, and use the collected results as input for explicit knowledge distillation; S12. Based on the preset multi-agent coordination framework, define senior student agent, teaching assistant agent and teacher agent according to role, goal and background to construct an explicit knowledge distillation framework. S13. Use the explicit knowledge distillation framework to perform knowledge distillation on the chemical problem, and verify the knowledge distillation results in conjunction with the standard answer to output a knowledge list.
3. The multi-agent collaborative chemical knowledge explicit distillation method based on a cognitive apprenticeship mechanism according to claim 2, characterized in that, The explicit knowledge distillation framework is used to perform knowledge distillation on chemical problems, and the results are verified by combining the standard answer. The output knowledge list includes: S131. Utilize the senior student agent in the explicit knowledge distillation framework to receive chemistry questions, and trigger the preset large language model based on the received results; S132. The chemical problem is analyzed by triggering a preset large language model, and an initial list of knowledge point titles is identified and generated based on the analysis results. S133. Input the initial knowledge point title list and the corresponding standard answers into the teaching assistant intelligent agent simultaneously, use the teaching assistant intelligent agent to calibrate the initial knowledge point title list, and output the knowledge point title list. S134. The teacher agent receives the list of knowledge point titles and performs a structured transformation of the list of knowledge point titles according to the preset hierarchical rules to generate a knowledge list.
4. The multi-agent collaborative chemical knowledge explicit distillation method based on a cognitive apprenticeship mechanism according to claim 3, characterized in that, The process involves synchronously inputting the initial knowledge point title list and its corresponding standard answers into the teaching assistant AI, using the teaching assistant AI to calibrate the initial knowledge point title list, and outputting the knowledge point title list, which includes: S1331. The teaching assistant intelligent agent synchronously receives the initial knowledge point title list and the corresponding standard answers, and filters the initial knowledge point title list with the standard answers as a reference. S1332. Adjust the granularity of the filtered initial knowledge point title list to obtain a candidate knowledge point title list with uniform granularity. S1333. Based on the reasoning chain in the standard answer, complete the list of candidate knowledge point titles, standardize the terminology of the completed list of candidate knowledge point titles, and output the list of knowledge point titles.
5. The multi-agent collaborative chemical knowledge explicit distillation method based on a cognitive apprenticeship mechanism according to claim 4, characterized in that, The teacher agent receives a list of knowledge point titles and performs a structured transformation of the list according to preset hierarchical rules to generate a knowledge list including: S1341. Based on the received list of knowledge point titles, the teacher agent writes the definition, principle, and logical relationship for each knowledge point title according to the preset textbook standards, and completes the construction of the definition and principle layer. S1342. Based on the knowledge definition and principle layer construction results, supplement the formulas, parameters, case templates, misconceptions and constraints corresponding to each knowledge point title to complete the data and application layer construction. S1343. By integrating the construction results of the definition and principle layer, data and application layer, the knowledge point titles and the content of each layer are structurally integrated to generate a knowledge list.
6. The multi-agent collaborative chemical knowledge explicit distillation method based on a cognitive apprenticeship mechanism according to claim 1, characterized in that, The knowledge list is evaluated using a two-dimensional arbitration evaluation mechanism, and the evaluation results are determined by a decision-making committee. If the results are satisfactory, the knowledge list is transmitted to the learner's decision-making device. If not, the tutoring teacher agent is triggered to generate a supplementary knowledge list and transmit it to the learner decision-maker, including: S21. Based on the preset multi-agent coordination framework, define the tutoring teacher agent, learner decision-maker, accuracy and completeness arbitrator, and objectivity and portability arbitrator according to roles, goals and background. S22. Establish a two-dimensional arbitration evaluation mechanism based on arbitrators with accuracy and completeness, and arbitrators with objectivity and portability, and use the two-dimensional arbitration evaluation mechanism to evaluate the knowledge list and generate an evaluation report. S23. Based on the pre-set professional division of labor model and multi-dimensional indicator constraints, establish a decision-making committee using a review mechanism; S24. The decision-making committee judges the evaluation report according to the preset adjudication rules and confirms whether the evaluation report is qualified. If the evaluation report is qualified, the knowledge list is transmitted to the learner decision-maker. If the evaluation report is unqualified, the tutoring teacher agent is triggered to generate a supplementary knowledge list and transmit it to the learner decision-maker.
7. The multi-agent collaborative chemical knowledge explicit distillation method based on a cognitive apprenticeship mechanism according to claim 6, characterized in that, The establishment of a two-dimensional arbitration evaluation mechanism based on arbitrators of accuracy and completeness, and arbitrators of objectivity and portability, and the use of this mechanism to evaluate the knowledge list and generate an evaluation report includes: S221. Based on the definition of arbitrators with accuracy and completeness, and arbitrators with objectivity and portability, establish a two-dimensional arbitration evaluation mechanism; S222. Use a two-dimensional arbitration evaluation mechanism to evaluate the knowledge content and knowledge expression of the knowledge list, and generate evaluation scores and corresponding evaluation suggestions based on the evaluation results. S223. Combine the assessment scores with the corresponding assessment recommendations to obtain an assessment report.
8. The multi-agent collaborative chemical knowledge explicit distillation method based on a cognitive apprenticeship mechanism according to claim 7, characterized in that, The inputs to the tutoring teacher agent include an initial list of knowledge point titles, a knowledge list, and an assessment report.
9. The multi-agent collaborative chemical knowledge explicit distillation method based on a cognitive apprenticeship mechanism according to claim 1, characterized in that, The learner decision-maker verifies whether the received result meets the preset knowledge quality standard. If so, the received result is stored in the preset knowledge base. If not, the explicit knowledge distillation framework will be triggered to proceed to the next round of explicit distillation, including: S31. The learner decision-maker receives the knowledge list and the knowledge point supplement list, answers the questions based on the corresponding knowledge points, and uses the answer as a feedback signal to generate a confidence score for the feedback signal. S32. Verify whether the confidence score meets the preset knowledge quality standard. If it meets the preset knowledge quality standard, store the received result and the corresponding knowledge point into the preset knowledge base. If it does not meet the preset knowledge quality standard, trigger the explicit knowledge distillation framework to carry out the next round of explicit distillation process.
10. A multi-agent collaborative chemical knowledge explicit distillation system based on a cognitive apprenticeship mechanism, used to implement the multi-agent collaborative chemical knowledge explicit distillation method based on a cognitive apprenticeship mechanism as described in any one of claims 1-9, characterized in that, The system includes: The knowledge list generation module is used to process the collected chemistry questions and generate a knowledge list using an explicit knowledge distillation framework that includes senior student agents, teaching assistant agents, and teacher agents. The assessment and decision-making module is used to evaluate the knowledge list using a two-dimensional arbitration assessment mechanism and determine whether the assessment result is qualified through the decision-making committee. If it is qualified, the knowledge list is transmitted to the learner decision-maker; if not, the tutoring teacher agent is triggered to generate a supplementary knowledge list and transmit it to the learner decision-maker. The verification and storage module is used to verify whether the received results meet the preset knowledge quality standards through the learner decision-maker. If they do, the received results are stored in the preset knowledge base; otherwise, the explicit knowledge distillation framework is triggered to carry out the next round of explicit distillation process.
