Method and system for controllably and automatically generating creativity evaluation situation based on evolution optimization

Abstract

The invention discloses a creativity evaluation situation controllable automatic generation method and system based on evolution optimization, and belongs to the technical field of intelligent education. According to the method, a closed-loop process of hierarchical constraint planning-Monte Carlo sentence level generation-evolution variation screening-simulation evaluation feedback iteration is adopted, a situation outline is planned firstly, then a seed text is generated, diversity and quality are optimized through Map-Elites, and a high-quality text is screened in combination with virtual evaluation feedback iteration. The system comprises a user request module, a data loading module, a model generation module and an interaction module. The problems that existing situation generation depends on experts, efficiency is low and controllability is insufficient are solved, the assessment situation with the complete structure and high adaptability is automatically generated, development cost is reduced, and large-scale standardized assessment is assisted.

Description

Technical Field

[0001] This invention belongs to the field of intelligent education technology and relates to a method and system for controllable automatic generation of creativity assessment scenarios based on evolutionary optimization. Background Technology

[0002] Creativity assessment is an important topic in the field of education, and researchers in related fields have proposed numerous methods and indicators for creativity assessment. [1] In contextualized assessment, constructing complex assessment scenarios that closely resemble real-world application scenarios is of great significance for efficient and effective creativity assessment. [2] .

[0003] In traditional methods, creativity assessments primarily rely on expert evaluation, mainly using questionnaires. Users complete the questionnaires in written form, data is collected manually, and experts score each questionnaire to determine the user's creative potential. [3] This approach, which focuses on abstract textual propositions without constructing concrete scenarios related to real life, leads to a disconnect between the assessment scenario and the actual application of creativity. It lacks the "sense of immersion" and "sense of participation" required for an immersive experience. This not only easily leads to perfunctory or false answers but also directly weakens the authenticity and validity of the assessment results, making it difficult to accurately reflect the true level of creativity of the test takers. [4] .

[0004] Existing contextualized creativity assessments primarily generate two types of scenarios: the first is expert-led scenario design, exemplified by PISA. [5] Its core logic relies on experts' deep understanding of creativity to tailor-make targeted assessment scenarios. This approach ensures that the scenarios accurately stimulate creativity, and that the assessment process is scientific and rigorous. [6] Meanwhile, its development process strictly follows the standardized procedures of psychological testing, ensuring the reliability and validity of the assessment tool. However, this approach has significant limitations. Scenario design heavily relies on expert experience and expertise, resulting in long generation cycles, low efficiency, and difficulty in quickly adapting to the batch needs of large-scale assessments. Furthermore, it lacks adaptability when dealing with flexible scenarios such as diverse assessment topics and differentiated assessment subjects. The second approach involves autonomously generating assessments using large models, such as the SS_GEN method. [7-9]Its core lies in leveraging the autonomous generation capabilities of a large language model. Through precise prompting word design, such as constructing contextual text examples, clarifying contextual text structure specifications, and defining application scenario boundaries, the model is guided to directly output complete assessment scenarios. This approach significantly simplifies the scenario generation process, greatly improves scenario output efficiency by utilizing the model's scalable generation advantages, and can cover a wider range of scenario themes based on the model's generalization capabilities, further expanding the application scenarios of contextualized assessment. However, this approach currently still faces challenges such as insufficient controllability of scenario generation, and the generated text failing to fully meet the rigorous requirements of contextualized creativity assessment in terms of structural integrity, logical consistency, and suitability to assessment objectives. Summary of the Invention

[0005] To address the shortcomings of existing technologies, the present invention aims to provide a controllable and automatic generation method and system for creativity assessment scenarios based on evolutionary optimization. This method utilizes a large language model and evolutionary optimization algorithms to solve the problems of scarce and complex design of creativity assessment scenarios in the field of education.

[0006] This invention first uses a topic-based hierarchical constraint planning module to perform structured hierarchical planning of elements and constraints for a given topic, outputting a situation outline that includes the field scale, key constraints, and challenge cues. Based on this, a Monte Carlo tree search is used to perform sentence-level progressive generation, and a multi-dimensional quality assessment is conducted to obtain a set of structurally complete and logically coherent initial seed texts. Subsequently, a diversity-quality joint optimization based on Map-Elites is introduced to perform constrained mutation and screening of the seed texts, forming an elite archive covering multiple challenge dimensions. Finally, a two-stage feedback mechanism for simulated assessment (pre-test feasibility and readability assessment, post-test group response validity test) is used to perform closed-loop iterative optimization of the generation process, retaining high-quality situation texts that can effectively induce creative divergent thinking and problem reconstruction.

[0007] This invention mainly comprises two parts: context seed data generation and context diversity-quality evolution. The context seed data generation part includes steps 1-6, and the context diversity-quality evolution part includes steps 7-10, specifically including the following steps:

[0008] Step 1: Hierarchical Planning with Theme Constraints

[0009] Given topic With predefined rule sets Given a pre-trained language model DeepSeek-V3.1 Depth of Reasoning With unfolded width Under the condition of the current context tree Select the optimal scenario hyperlink to be expanded. Specifically, based on the topic... Initialize the supertree for the root node:

[0010] ,

[0011] At each layer depth Above, map the current supertree to a set of hyperlinks:

[0012] ,

[0013] If the number of superchains Then, a large model is used to score each hyperlink and retain the top-scoring hyperlinks. strip:

[0014] ,

[0015] This will allow subsequent steps to select the optimal hyperlink. Further planning and refinement will be carried out.

[0016] Step 2: Selection and refinement of decomposable nodes

[0017] From the optimal hyperlink The process involves extracting and expanding the set of divisible nodes from the candidate hyperlinks using a larger model.

[0018] ,

[0019] Then, relevant rules are selected based on the rule set, and child nodes are generated:

[0020] , .

[0021] Attach the generated child nodes to the supertree to complete this round of refinement:

[0022] .

[0023] Step 3: Synthesis and Selection of Situational Assessment Outline

[0024] At the preset reasoning depth With unfolded width Next, repeat steps 1 and 2 in a loop to process the context hypertree. The candidate hyperlinks in the selection process undergo multiple rounds of "selection-expansion" operations until all leaf nodes can no longer be refined or a preset depth threshold is reached. The loop terminates when the time is right, and the language model is used. A comprehensive evaluation is performed on each candidate hyperlink in the supertree, and the hyperlink with the highest score is selected. As the final scenario planning outline.

[0025] Step 4: Monte Carlo Tree Search Modeling and Node Selection

[0026] By constructing a Monte Carlo tree search mechanism, sentence-level generation of contextual text is performed, and the generated contextual text sequence is modeled as a state. Model the candidate sentences or fragments as actions And maintain an access count for each state and state-action pair. , and valuation Node selection uses a heuristic function of UCB form:

[0027] ,

[0028] In the expansion phase, candidate sentences are generated from the large language model and a comprehensive score is calculated. Update incrementally during the backtracking phase:

[0029] ,

[0030] This allows for a balance between value assessment and exploration efficiency during the search process.

[0031] Step 5: Candidate Sentence Generation and Multidimensional Quality Scoring

[0032] The text generation model is invoked during the expansion phase of the search. The candidate set is evaluated using a scorer to obtain the comprehensive value of the candidate actions, which is then used as the basis for action selection and ranking in Monte Carlo search.

[0033] Step 6: Backtracking Update and Adaptive Unfolding

[0034] During the backtracking phase of the Monte Carlo tree search, the comprehensive value of the candidate sentences is traced along the search path. Backhaul layer by layer, and average value based on path. With preset threshold Based on the comparison results, the subsequent sampling unfolding depth is adaptively adjusted. :when Increase the number of expansion steps when the value is higher to enhance the exploration of high-value branches. Reduce the number of unrolling steps at lower values ​​to reduce sampling overhead on low-value branches.

[0035] Step 7: Diversity - Quality Elite Profile Construction

[0036] The seed context text obtained by MCTS is subjected to Map-Elites behavior space mapping and fitness evaluation mechanism, utilizing feature mapping function. Extracting behavioral feature vectors , It refers to a numerical representation composed of multiple quantifiable dimensions. In this invention, The text is characterized using three complementary contextual style dimensions, and LLM provides three normalized scores for each dimension. It is a three-dimensional numerical vector. Based on the fitness function... Calculate the fitness value and store or update only the elite samples with the highest fitness in the corresponding behavior grid, thereby maintaining the diversity coverage of the behavior space while ensuring text quality.

[0037] Step 8: Contextual Variation and Evolution

[0038] Based on the set of mutation operators (Including operations such as expanding, deleting, and replacing) on ​​the selected archive sample in Archive. Mutate to generate new samples And calculate its fitness. If the target behavior cell is empty or ( If the current sample is in that cell, then insert or replace it in the Archive according to the archive replacement rules. By replacing existing samples, this process continuously improves the "optimal representative" of each behavioral grid while satisfying constraints such as thematic consistency, structural integrity, and readability, thereby enhancing the diversity of contextual text in the behavioral space while ensuring fitness quality.

[0039] Step 9: Simulate testing and verify evolution

[0040] By invoking a large language model and combining it with a pre-set "virtual student" prompt template, batch simulations and scoring are performed on candidate context texts in the Archive. Scores on creative divergent thinking indicators such as novelty, practicality, and fluency are collected, and the mean and distribution characteristics are statistically analyzed to form an evaluation index vector. This evaluation index vector serves as an external feedback signal, used to correct the fitness function and update evolutionary strategy parameters, thereby verifying and improving the effectiveness of candidate context texts in stimulating creative divergent thinking responses.

[0041] Step 10: Closed-loop iteration and convergence output

[0042] A closed-loop iterative optimization mechanism is constructed based on virtual evaluation feedback: the set of candidate scenarios output by the Map-Elites module is recorded as the initial candidate pool. In the During the round of iteration, the candidate pool is... Each scenario Calculate the average creativity score using the virtual assessment module. And combined with a preset elimination threshold With maximum number of iterations A screening process is conducted. The elimination threshold is set at 65 points to determine whether the generated scenarios meet the usable lower limit in terms of evaluation effectiveness. This threshold is calibrated and determined by psychology experts based on the generated scenarios and their corresponding simulated responses. This ensures that the scenarios can consistently elicit scoreable and discriminative answers while also considering computational resource consumption and avoiding excessive iteration on obviously low-quality samples. The maximum number of iterations is set to 3 to control the closed-loop optimization cost and avoid diminishing marginal returns, allowing the system to terminate iteration promptly after completing necessary corrections for weak samples.

[0043] , ;

[0044] when or The iteration will terminate at the time. The final set of high-quality scenarios; otherwise, the elimination set. New candidates are generated using mutation operators, and the candidate pool is updated. This process is repeated, and the text generation and evolution process is continuously optimized based on simulation evaluation feedback. Contextual texts that are difficult to stimulate creative responses are gradually eliminated, while high-quality contextual texts are retained.

[0045] Based on the above, the present invention also proposes a controllable automatic generation system for creativity assessment scenarios based on evolutionary optimization, the system including a user request module, a data loading module, a model generation module and a user interaction module;

[0046] The user request module is responsible for processing user-sent requests and the subject data contained in the request message body.

[0047] The data loading module is used to load data according to the standardized topic representation. The rules and background knowledge related to the topic are retrieved and loaded from the rule base, knowledge base and corpus, and then preprocessed as topic constraint data for subsequent generation.

[0048] The model generation module is used to call the large language model to perform the above-mentioned hierarchical task planning, Monte Carlo tree search generation, behavior space mapping and evolutionary screening, and virtual evaluation feedback iterative optimization under the joint constraints of the topic information and topic constraint data, so as to generate a set of complex contextual texts that match the current topic.

[0049] The user interaction module displays the generated scenario to the user, who then takes a creativity assessment and answers related questions within that scenario. Optionally, the module can also receive user feedback to optimize the subsequent scenario generation effect.

[0050] The method and system for controllable automatic generation of creativity assessment scenarios based on evolutionary optimization disclosed in this invention have the following beneficial effects:

[0051] This invention constructs an automatically controllable framework for generating creativity assessment scenarios based on evolutionary optimization. Compared to general story generation or pure text-based large-scale model solutions lacking assessment attributes, this framework, guided by themes and constraints, can stably and controllably generate problem scenarios for creativity assessment, ensuring structural integrity, logical consistency, and scoreability, while effectively stimulating participants' creative divergent thinking. Compared to traditional methods relying on expert manual development, this invention significantly shortens the development cycle of creativity assessment scenarios, reduces time development costs, and facilitates rapid customization and large-scale supply for different educational stages, subjects, and populations, thereby achieving efficient and effective creativity assessment. This invention aims to shift the development process of creativity assessment scenarios from expert manual development to automated generation, thereby significantly reducing the time cost invested by experts and shortening the development cycle. In the traditional process, experts need to first determine the assessment objectives and plan the distribution of assessment clues, then write long text scenarios and undergo multiple rounds of review and revision to ensure structural integrity, logical consistency, and scoreability. Subsequently, small-scale trial testing or peer review is often required to screen out unstable samples and make adjustments. Therefore, its costs are mainly reflected in the man-hours invested by experts, the number of iterations, and the development cycle from topic to usable assessment questions. This cycle usually takes several weeks, and if real-person testing and feedback loops are included, it often extends to a month or even longer. Attached Figure Description

[0052] 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.

[0053] Figure 1 This is a flowchart of a method for controllable automatic generation of creativity assessment scenarios based on evolutionary optimization, according to the present invention.

[0054] Figure 2 This is a flowchart of a controllable automatic generation system for creativity assessment scenarios based on evolutionary optimization, according to the present invention. Detailed Implementation

[0055] The invention will be further described in detail below with reference to the specific embodiments and accompanying drawings. Except for the contents specifically mentioned below, the processes, conditions, and experimental methods for implementing the invention are all common knowledge and general knowledge in the art, and the invention does not have any particular limitations.

[0056] This invention proposes a controllable automatic generation method for creativity assessment scenarios based on evolutionary optimization. The method employs a closed-loop process of "hierarchical constraint programming—Monte Carlo sentence-level generation—evolutionary mutation screening—simulated assessment feedback iteration" to perform structured modeling and finely controlled generation of scenario texts, producing stimulus scenario texts with a solid knowledge base, clear structure, and multi-dimensional challenges. The method automatically plans an assessment point structure outline matching a given theme, including field scale, key constraints, and potential challenges. Based on this outline, Monte Carlo tree search is used to generate seed data for scenario texts sentence by sentence. To balance the diversity and quality of generated scenario texts to meet the needs of different assessment groups, a "diversity-quality" elite profile is constructed. The seed texts undergo constrained mutation and screening to form diverse and high-quality scenario texts. The assessment attributes and validity are verified through simulated assessments with a virtual subject group, ensuring effective stimulation of subjects' creative divergent thinking. This method can assist professional creativity assessment teams in quickly developing assessment scenario questions to meet the large-scale needs of different groups, achieving efficient and effective automatic generation of creativity assessment scenario texts. The present invention also discloses a system for implementing the above-mentioned method for automatically and controllably generating complex contextual texts for creativity assessment.

[0057] This invention uses the questions and thematic constraints of creativity assessment as input to construct an automatically controllable generation model of contextual text with assessment attributes. Employing a two-stage process of "contextual seed generation" and "diversity-quality evolution," it produces contextual texts that better stimulate students' creative and divergent thinking, enhancing their reading engagement and willingness to express themselves, and ultimately allowing their true creative and divergent thinking abilities to be fully revealed.

[0058] The following provides a further explanation of the specific implementation of automatic generation of contextual text for creativity assessment. The invention will be further explained using the "AI Partner" creativity task context.

[0059] Implement a method for automatically and controllably generating complex-situation texts for creativity assessment using the following steps:

[0060] The method of the present invention includes the following steps:

[0061] Step 1: Based on the specified topic and predefined rules Key points for hierarchical planning scenario assessment, from multiple candidate hyperlinks Choose the optimal one Further planning is needed;

[0062] Step 2: From Extract decomposable nodes Further expansion is needed; these nodes correspond to contextual assessment points that require further refinement. A language model will be used. To select the most suitable node for expansion ;

[0063] Step 3: Repeat the first two steps, performing multiple rounds of "selection-expansion" operations on the hyperlinks in the context hypertree until all leaf nodes can no longer be refined or reach the preset depth, thus constructing a complete context outline structure. This is then processed using a language model. Select the optimal hyperlink And use it as the final planning outline;

[0064] Step 4: Based on the planning outline To constrain the search, construct the state space for Monte Carlo Tree Search (MCTS). and action space Initialize and maintain node access counts. State-Action Count Value estimation and exploration coefficient .

[0065] Step 5: In the expanded search phase, invoke the text generation model. Regarding the current state Generate candidate set And a scorer is used to evaluate the quality of the candidate set.

[0066] Step 6: In the backtracking phase, the quality scores of candidate sentences are aligned with the search trajectory. Return and update the corresponding and And adaptively adjust the subsequent deployment depth based on the average path value. To increase exploration on high-value pathways;

[0067] Step 7: Apply the feature mapping function to the seed context text obtained by MCTS. Extracting behavioral feature vectors According to the fitness function Calculate its fitness value and map the text and its features to the Map-Elites archive structure, storing or updating the elite sample with the best fitness in the corresponding action cell;

[0068] Step 8: Based on the set of mutation operators (Including operations such as expansion, deletion, and replacement) Mutate the selected archive sample to generate a new sample and calculate its fitness. If the sample is superior to the existing sample in the cell, it is inserted or replaced in Archive according to the Archive replacement rules, thereby improving the diversity of contextual texts while ensuring quality.

[0069] Step 9: Construct a student modeling and simulated answering module, using candidate scenario texts from the files as input stimulus materials, collect simulated users' answers and scores on creativity indicators such as novelty, practicality, and fluency, statistically analyze their mean and distribution characteristics, form an evaluation index vector, and use it as an external feedback signal to feed back to the fitness function and evolution process.

[0070] Step 10: Continuously optimize the text generation and evolution process based on simulation evaluation feedback, gradually eliminate situational texts that are difficult to stimulate creative responses, and retain high-quality situational texts.

[0071] Example

[0072] The following provides a further explanation of the specific implementation of automatic generation of contextual text for creativity assessment. The invention will be further explained using the "AI Partner" creativity task context.

[0073] Implement a method for automatically and controllably generating complex-situation texts for creativity assessment using the following steps:

[0074] Step 1: Based on the specified topic and predefined rules From multiple candidate hyperlinks Choose the optimal one Further planning

[0075] Given the "AI Partner" topic and theme, and pre-defined scaffolding rules defined by experts for the assessment scenario, the assessment framework for the "AI Partner" topic is planned. First, the hypertree is initialized with the "AI Partner" topic and theme as the root node: And given a pre-trained language model Depth of Reasoning With unfolded width Under the condition of the current context tree Select the optimal scenario hyperlink to be expanded. At each depth... Above, map the current supertree to a set of hyperlinks: , If the number of superchains Then, a large model is used to score each hyperlink and retain the top-scoring hyperlinks. strip: This allows subsequent steps to select the optimal hyperlink. Further planning and refinement will be carried out.

[0076] Step 2: From Extract decomposable nodes Further expansion is needed; these nodes correspond to contextual assessment points that require further refinement. A language model will be used. To select the most suitable node for expansion

[0077] From the optimal hyperlink The process involves extracting and expanding the set of divisible nodes from the candidate hyperlinks using a larger model. Then, based on the rule set, relevant rules are selected and child nodes are generated:

[0078] , .

[0079] Finally, the child node is attached to the supertree:

[0080] .

[0081] Step 3: Repeat the first two steps, performing multiple rounds of "selection-expansion" operations on the hyperlinks in the context hypertree until all leaf nodes can no longer be refined or reach the preset depth, thus constructing a complete context outline structure. This is then processed using a language model. Select the optimal hyperlink And use it as the final planning outline;

[0082] Step 4: Based on the planning outline To constrain the search, construct the state space for Monte Carlo Tree Search (MCTS). and action space Initialize and maintain node access counts. State-Action Count Value estimation and exploration coefficient

[0083] Sentence-level generation of contextual text is performed by constructing a Monte Carlo tree search mechanism: the generated contextual text sequence is modeled as a state. Model the candidate sentences or fragments as actions And maintain an access count for each state and state-action pair. , and valuation Node selection uses a heuristic function of the form PUCT / UCB.

[0084] ,

[0085] In the expansion phase, candidate sentences are generated from the large language model and a comprehensive score is calculated. Update incrementally during the backtracking phase:

[0086] ,

[0087] This allows for a balance between value assessment and exploration efficiency during the search process.

[0088] Step 5: In the expanded search phase, invoke the text generation model. Regarding the current state Generate candidate set And a scorer is used to evaluate the quality of the candidate set.

[0089] The text generation model is invoked during the expansion phase of the search. Regarding the current state Generate candidate set The candidate set is evaluated using a scorer to obtain the comprehensive value of the candidate actions, which is then used as the basis for action selection and ranking in Monte Carlo search.

[0090] Step 6: In the backtracking phase, the quality scores of candidate sentences are aligned with the search trajectory. Return and update the corresponding and And adaptively adjust the subsequent deployment depth based on the average path value. To increase exploration on high-value pathways

[0091] During the backtracking phase of the Monte Carlo tree search, the comprehensive value of the candidate sentences is traced along the search path. Backhaul layer by layer, and average value based on path. With preset threshold Based on the comparison results, the subsequent sampling unfolding depth is adaptively adjusted. :when Increase the number of expansion steps when the value is higher to enhance the exploration of high-value branches. Reduce the number of unrolling steps at lower values ​​to reduce sampling overhead on low-value branches.

[0092] Step 7: Apply the feature mapping function to the seed context text obtained by MCTS. Extracting behavioral feature vectors According to the fitness function Calculate the fitness value and map the text and its features into the Map-Elites archive structure, storing or updating the currently best-fit elite sample in the corresponding action cell.

[0093] The seed context text obtained by MCTS is subjected to Map-Elites behavior space mapping and fitness evaluation mechanism, utilizing feature mapping function. Extracting behavioral feature vectors According to the fitness function Calculate the fitness value and store or update only the elite samples with the highest fitness in the corresponding behavior grid, thereby maintaining the diversity coverage of the behavior space while ensuring text quality.

[0094] Step 8: Based on the set of mutation operators (Including operations such as expansion, deletion, and replacement) Mutate the selected archive sample to generate a new sample and calculate its fitness. If the sample is superior to the existing sample in the corresponding cell, it is inserted or replaced in the Archive according to the archive replacement rules, thereby improving the diversity of contextual texts while ensuring quality.

[0095] Based on the set of mutation operators (Including operations such as expanding, deleting, and replacing) on ​​the selected archive sample in Archive. Mutate to generate new samples And calculate its fitness. If the target behavior cell is empty or ( If the current sample is in that cell, then insert or replace it in the Archive according to the archive replacement rules. By replacing existing samples, the diversity of contextual text in the behavioral space can be improved while ensuring fitness quality.

[0096] Step 9: Construct a student modeling and simulated answering module. Use candidate scenario texts from the student files as input stimulus materials. Collect simulated user responses and scores on creativity indicators such as novelty, practicality, and fluency. Statistically analyze the mean and distribution characteristics to form an evaluation index vector, which serves as external feedback signals to the fitness function and evolutionary process.

[0097] By invoking a large language model and combining it with a preset "virtual student" prompt template, batch simulations and scoring are performed on candidate context texts in Archive. Scores on creativity indicators such as novelty, practicality, and fluency are collected, and the mean and distribution characteristics are statistically analyzed to form an evaluation index vector. The evaluation index vector serves as an external feedback signal to modify the fitness function or update the evolutionary strategy parameters, thereby verifying and improving the effectiveness of candidate context texts in stimulating creative responses.

[0098] Step 10: Continuously optimize the text generation and evolution process based on simulation assessment feedback, gradually eliminating situational texts that are difficult to stimulate creative responses, and retaining high-quality situational texts.

[0099] A closed-loop iterative optimization mechanism is constructed based on virtual evaluation feedback: the set of candidate scenarios output by the Map-Elites module is recorded as the initial candidate pool. In the During the round of iteration, the candidate pool is... Each scenario Calculate the average creativity score using the virtual assessment module. And combined with a preset elimination threshold With maximum number of iterations Filter:

[0100] , ;

[0101] when or The iteration will terminate at the time. The final set of high-quality scenarios; otherwise, the elimination set. New candidates are generated using mutation operators, and the candidate pool is updated. This process is repeated, and the text generation and evolution process is continuously optimized based on simulation evaluation feedback. Contextual texts that are difficult to stimulate creative responses are gradually eliminated, while high-quality contextual texts are retained.

[0102] References

[0103] [1] Guilford, JP (1967). Creativity: Yesterday, today and tomorrow. Journal of Creative Behavior, l(1), 3-14.

[0104] [2] Patterson, JD, Pronchick, J., Panchanadikar, R.et al. (2025).CAP: The creativity assessment platform for online testing and automatedscoring. Behav Res 57, 264.

[0105] [3] The standard definition of creativity. (2012). CreativityReasearch Journal 24.1(2012):92-96.

[0106] [4] Abdulla, AM, Paek, SH, Cramond, B., & Runco, MA (2020). Problem finding and creativity: A meta-analytic review. Psychology ofAesthetics, Creativity, and the Arts, 14(1), 3

[0107] [5] Grey, S., & Morris, P. (2024). Capturing the spark: PISA, twenty-first century skills and the reconstruction of creativity. Globalisation,Societies and Education, 22(2), 156-171.

[0108] [6] Kapoor, H., & Tagat, A. (2025). The Importance of ContextAwareness in PISA Creativity Data. The Journal of Creative Behavior, 59(4),e70066.

[0109] [7] Feng, Y., Song, M., Wang, J., Chen, Z., Bi, G., Huang, M., ... &Yu, J. (2025, April). SS-GEN: A Social Story Generation Framework with LargeLanguage Models. In Proceedings of the AAAI Conference on ArtificialIntelligence (Vol. 39, No. 2, pp. 1300-1308).

[0110] [8] Bai, Y., Zhang, J., Lv, X., Zheng, L., Zhu, S., Hou, L., ... &Li, J. (2024). Longwriter: Unleashing 10,000+ word generation from longcontext llms. arXiv preprint arXiv:2408.07055.

[0111] [9] Laverghetta Jr, A., Luchini, S., Linell, A., Reiter-Palmon, R., &Beaty, R. (2024). The creative psychometric item generator: a framework foritem generation and validation using large language models. arXiv preprintarXiv:2409.00202.

[0112] Unless otherwise defined, all technical and scientific terms used in this invention have the same meaning as commonly understood by one of ordinary skill in the art to which this invention pertains. The terminology used in this specification is for the purpose of describing particular embodiments only and is not intended to be limiting of the invention.

[0113] As used in this invention, the term "comprising" is an open-ended expression, meaning it includes the contents specified in this invention but does not exclude other aspects.

[0114] As used in this invention, the term "and / or" includes any one or more of the related listed items and all combinations thereof.

[0115] The scope of protection of this invention is not limited to the above embodiments. Any variations and advantages that can be conceived by those skilled in the art without departing from the spirit and scope of the inventive concept are included in this invention and are protected by the appended claims.

Claims

1. A method for controllable automatic generation of creativity assessment scenarios based on evolutionary optimization, characterized in that, Includes the following steps: Step 1: Based on the specified topic With predefined rule sets Hierarchical planning of evaluation points, and selection from multiple candidate hyperlinks Choose the optimal one Further refinement; Step 2: From Extract decomposable nodes To expand upon this, these nodes correspond to contextual assessment points that still need further refinement, using a language model. To select the most suitable node for expansion ; Step 3: Repeat Steps 1 and 2, performing multiple rounds of "selection-expansion" operations on the hyperlinks in the context hypertree until all leaf nodes can no longer be refined or reach the preset depth. This completes the context outline structure, which is then analyzed using a language model. Select the optimal hyperlink And use it as the final planning outline; Step 4: Develop a planning outline To constrain the process, construct the state space for Monte Carlo tree search. and action space and initialize the search parameters; Step 5: Invoke the text generation model during the expansion phase. Regarding the current state Generate candidate set And a scorer is used to evaluate the quality of the candidate set; Step 6: In the backtracking phase, the quality scores of candidate sentences are aligned with the search trajectory. Return the data, update the search parameters, and adjust the subsequent unfolding depth based on the average value of the path; Step 7: Extract behavioral feature vectors from the seed context text obtained by MCTS. It also calculates fitness values, maps text and its features to the Map-Elites archive structure, and stores or updates the best-fit elite samples in the corresponding behavior grid. Step 8: Based on the set of mutation operators The selected file sample is mutated to generate a new sample and its fitness is calculated. If the new sample is better than the existing sample in the corresponding behavior cell, it is inserted or replaced according to the rules to improve the diversity of contextual text. Step 9: Construct a student modeling and simulated answering module, using candidate scenario texts from the files as stimulus input, collect simulated user responses and scores on creativity and divergent thinking indicators, form an evaluation indicator vector, and feed it back to the fitness function and evolution process; Step 10: Continuously optimize the contextual text generation and evolution process based on simulation assessment feedback, gradually eliminate contextual texts that are difficult to stimulate creative and divergent thinking, and retain high-quality contextual texts.

2. The method as described in claim 1, characterized in that, In step 1, given the topic With predefined rule sets In pre-trained language models Depth of Reasoning With unfolded width Selecting the optimal hyperlink under the given conditions; including: by topic Initialize the supertree for the root node: ， At each layer depth Above, map the current supertree to a set of hyperlinks: ， If the number of superchains Then, a large model is used to score each hyperlink and retain the top-scoring hyperlinks. strip: ， Select the optimal hyperlink from these. Further planning and refinement will be carried out.

3. The method as described in claim 1, characterized in that, In step 2, from the optimal hyperlink Extract decomposable nodes and expand them: First, the large model selects a set of separable nodes from the candidate hyperlinks: ， Then, relevant rules are selected based on the rule set, and child nodes are generated: ， 。 Finally, the child node is attached to the supertree: 。 4. The method as described in claim 1, characterized in that, In step 3, the preset inference depth threshold is: Through Execute steps 1 and 2 repeatedly within the scope of the context hypertree. The candidate hyperlinks in the algorithm undergo multiple rounds of "selection-expansion" operations; the loop terminates when all leaf nodes cannot be further refined or a preset depth threshold K is reached, and the language model is then utilized. A comprehensive evaluation is performed on each candidate hyperlink in the supertree, and the hyperlink with the highest score is selected. As the final scenario planning outline.

5. The method as described in claim 1, characterized in that, In step 4, the search parameters include node visit count. State-Action Count Value estimation and exploration coefficient The sentence-level generation logic of Monte Carlo tree search is as follows: model the generated contextual text sequence as a state. Model the candidate sentences or fragments as actions Node selection uses a heuristic function of UCB form: ， In the expansion phase, candidate sentences are generated from the large language model and a comprehensive score is calculated. Update incrementally during the backtracking phase: ， This balances value assessment with exploration efficiency.

6. The method as described in claim 1, characterized in that, In step 6, the specific method for adjusting the subsequent expansion depth is as follows: The comprehensive value of the candidate sentences is aligned with the search trajectory. After each layer is sent back, the average value of the comparison path is calculated. With preset threshold ;when Increase the number of expansion steps when the value is higher to enhance the exploration of high-value branches. Reduce the number of unfolding steps at lower values ​​to reduce sampling overhead on low-value branches; the unfolding depth is determined by parameters. Characterization.

7. The method as described in claim 1, characterized in that, In step 7, through the feature mapping function Extracting behavioral feature vectors Through fitness function Calculate its fitness value; The storage rules of the Map-Elites archive structure are as follows: only the elite samples with the highest fitness are stored or updated in the corresponding behavior grid, so as to maintain the diversity coverage of the behavior space while ensuring the quality of the text.

8. The method as described in claim 1, characterized in that, In step 8, the set of mutation operators This includes operations such as expansion, deletion, and replacement; the logic for mutation and replacement is as follows: [The text abruptly shifts to a different topic] ...selected archive sample in Archive... Generate new samples by mutation Calculate its fitness If the target behavior cell is empty or If the current sample is in this cell, then insert or replace it in the Archive according to the archive replacement rules. Replace existing samples to improve the diversity of contextual texts in the behavioral space while ensuring fitness quality.

9. The method as described in claim 1, characterized in that, In step 9, the specific method for simulating answering and scoring is as follows: A large language model is invoked, combined with a preset "virtual student" prompt template, to perform batch simulated answering and scoring of candidate context texts in the Archive; the creative divergent thinking indicators include novelty, practicality, and fluency; after collecting scores for each indicator, the mean and distribution characteristics are statistically analyzed to form an evaluation indicator vector; the evaluation indicator vector serves as an external feedback signal, used to correct the fitness function or update the evolutionary strategy parameters, thereby improving the effectiveness of candidate context texts in stimulating creative divergent thinking responses; And / or, In step 10, the specific logic of the closed-loop iterative optimization is as follows: the set of candidate scenarios output by the Map-Elites module is recorded as the initial candidate pool. ; in the During the round of iteration, the candidate pool is... Each scenario Calculate the average creativity score using the virtual assessment module. Combined with preset elimination thresholds With maximum number of iterations The retained set is obtained by filtering: ， ； when or The iteration will terminate at the time. The final set of high-quality scenarios; otherwise, the elimination set. Apply the mutation operator to generate new candidates and update the candidate pool, repeating the above process.

10. A controllable automatic generation system for creativity assessment scenarios based on evolutionary optimization, characterized in that, The system implements the method as described in any one of claims 1-9, the system comprising: a user request module, a data loading module, a model generation module, and a user interaction module; wherein, The user request module is responsible for processing user-sent requests and the subject data contained in the request message body. The data loading module is used to load data based on standardized topic representations. The rules and background knowledge related to the topic are retrieved and loaded from the rule base, knowledge base and corpus, and then preprocessed as topic constraint data for subsequent generation. The model generation module is used to call a large language model to perform hierarchical task planning, Monte Carlo tree search generation, behavior space mapping and evolutionary filtering, and virtual evaluation feedback iterative optimization under the joint constraints of the topic information and topic constraint data, so as to generate a set of complex contextual texts that match the current topic. The user interaction module displays the generated scenario to the user, who then takes a creativity assessment and answers related questions within that scenario. Optionally, the module can also receive user feedback to optimize the subsequent scenario generation effect.

Images