Contextualized interactive evaluation method and system for creativity diagnosis and excitation and application

Abstract

The invention discloses a situational interactive evaluation method for creativity diagnosis and motivation, which performs creativity evaluation by executing two-stage tasks of supervision fine tuning and near-end strategy optimization, and comprises the following steps: pre-establishing an evaluation index system, initializing a dialogue-level reward calculation mechanism, and establishing an evaluation result; updating language model parameters by utilizing the pre-collected interaction data through a supervision fine tuning mode, and generating an initial dialogue strategy; real user behaviors are simulated through a user simulator, candidate guide strategies are generated, the current round of interaction is added into temporary session history, one or more rounds of user dialogue tracks are obtained through sampling, multi-round perception rewards are calculated, and model parameters are updated through a near-end strategy optimization algorithm; repeating the interaction optimization model until convergence to obtain a self-adaptive interaction strategy; and performing multi-round interaction with the real user based on the trained and optimized language model to realize creativity evaluation of the real user. The invention further discloses a system for implementing the method, and the system has a wide application prospect.

Description

Technical Field

[0001] This invention belongs to the field of intelligent education and relates to a contextualized interactive assessment method, system, and application for creativity diagnosis and stimulation. Background Technology

[0002] In the field of education, creativity assessment is an important topic. Researchers in related fields have proposed numerous methods and indicators for creativity assessment. In the interactive and contextualized assessment process, adaptive guidance based on user feedback is of great significance for efficient and effective creativity assessment.

[0003] Traditionally, creativity assessments primarily rely on expert evaluation, mainly through questionnaires. Users complete the questionnaires in written form, data is collected manually, and experts score each questionnaire to determine the user's creativity potential. This method is highly subjective; different opinions and interpretations from individual experts can lead to different results. Furthermore, in large-scale scenarios, it is time-consuming and labor-intensive, requiring significant time for scoring and data analysis when dealing with large data samples.

[0004] Existing online creativity assessments mainly take two forms: one uses an online platform to integrate traditional questionnaires, such as CAP. [1] The first type, the Creativity Assessment Platform, collects data through online questionnaires and uses technologies such as deep learning to achieve automated scoring and measurement. This approach overcomes spatial limitations, greatly reducing the time and expert resources required by traditional questionnaires and minimizing the subjectivity issues caused by human expert scoring. The second type uses interactive, contextualized assessment models, driven by large language models to facilitate human-computer interaction. Most of these models employ static generation or single-round question-and-answer mechanisms, such as Prometheus. [2] A 13B large language model for specific assessments was trained to score any given response based on user-provided rating criteria. Compared to questionnaires, this format offers novelty and flexibility, allowing for better user subjectivity and interactivity. These methods largely address the difficulties of traditional questionnaire-based creativity assessments, reducing the need for traditional human, expert, and time resources, and improving the efficiency and flexibility of creativity assessment. However, problems remain, including a lack of responsiveness and guidance to multiple rounds of student feedback, weak interactivity, low user participation, and an inability to fully stimulate users' creative potential. In the context of open-ended questions like creativity, the model cannot continuously guide users to think critically and answer in depth, ultimately affecting the effectiveness and accuracy of the assessment. Summary of the Invention

[0005] To address the shortcomings of existing technologies, the purpose of this invention is to provide a contextualized interactive assessment method, system, and application for creativity diagnosis and stimulation, utilizing large language models and neural network models to solve creativity assessment problems in the field of education.

[0006] In the model construction process of this invention, the dialogue-level reward calculation mechanism is first initialized based on the established evaluation index system. In the first stage, language model parameters are updated using supervised fine-tuning with pre-collected interaction data, providing an initial dialogue strategy for the second stage optimization. In the second stage, a user simulator is introduced to simulate real user behavior. Based on the set context and task objectives, and using historical conversation records and the current input from the user simulator, the model generates candidate guidance strategies. Monte Carlo sampling is performed by alternating responses generated by the user simulator and the model to obtain several possible subsequent dialogue trajectories. Dialogue-level rewards and multi-turn perceptual rewards are calculated, and the model parameters are updated using a proximal policy optimization algorithm. Through multiple conversation training sessions, the model is fully converged. The resulting contextualized assessment method stimulates user enthusiasm and implicit genuine creativity during the assessment process, enhancing users' willingness to express themselves and their interactive experience.

[0007] The model training of this invention mainly includes two parts: supervised fine-tuning and proximal policy optimization. The supervised fine-tuning part includes steps 1.1-1.2, and the proximal policy optimization part includes steps 2.1-2.6, specifically including the following steps:

[0008] Step 1.1: Initialize dialogue-level rewards based on the established evaluation index system. Computerized system;

[0009] Specifically,

[0010] An evaluation index system was established, comprising three stages: identifying challenges, selecting challenges, and proposing solutions. Corresponding evaluation indicators were established for each stage, and the specific definitions of the indicators are shown in Table 1 below.

[0011] Table 1 Evaluation Indicators for the Project Phase

[0012]

[0013]

[0014] Initialize the conversational reward calculation mechanism, including rewards for stimulating creativity. Internal interaction quality reward Penalty for leaking questions Interactive guidance rewards and interaction efficiency rewards This comprehensively reflects the dialogue's performance in promoting creativity, maintaining content quality, and achieving goals efficiently. Overall reward score. The calculation formula can be expressed as:

[0015] ,

[0016] in, Indicates interaction history. Expressing task objectives and rewarding creativity It is a reward used to assess the creativity demonstrated in a user's response during a conversation, an intrinsic reward for the quality of interaction. It is an interactive reward used to evaluate the quality of AI assistant interactions with users in conversations, and a penalty for missing questions. It's a penalty system designed to address potential leaks of answers by AI assistants during conversations, along with interactive guidance rewards. It is a reward used to evaluate whether an AI assistant can successfully guide the user back to the intended topic when the user deviates from the main theme; it is an interaction efficiency reward. It is a reward used to measure the efficiency of completing an interaction;

[0017] Step 1.2: In the first stage, using the pre-collected interaction data, the language model parameters, including those based on large language models such as Qwen3-8B, are updated using supervised fine-tuning to provide an initial dialogue strategy for the second stage optimization.

[0018] Specifically,

[0019] Supervised fine-tuning is used to train a large language model with an interactive dataset. Given the input, minimize the following negative log-likelihood objective:

[0020] ,

[0021] in, For context input, To guide the output towards the goal. For model parameters, For the model in parameters Given the input Generate target output in time The probability of;

[0022] Step 2.1: In the second stage, a user simulator is introduced to simulate real user behavior based on session history. Probability distribution of generating real user behavior Output a response that conforms to the user's language style and typical behavior. ;

[0023] Specifically,

[0024] Each round of the session consists of user input and model output, defined as: ,in, Indicates the first The user input wheel, Indicates the first A guidance strategy for round-based model generation is implemented. A user simulator is introduced to simulate real user behavior, addressing the issues of high computational cost for multi-round forward dialogue sampling during training and the high cost and lack of scalability of directly using real user dialogue. Specifically, in the first round... In the round-robin interaction, the user simulator Based on the given conversation history Generate possible user responses And use it as input to the current user simulator, through probability distribution Simulate real user behavior to generate multi-turn dialogue responses that conform to the user's language style and typical behavior;

[0025] Step 2.2: Based on the set context and task objectives, and using historical conversation records... and the current input of the user simulator Model-generated candidate guidance strategies ;

[0026] Specifically,

[0027] The task objective is as follows: In a given context, the dialogue between the user and the model must proceed according to a prescribed problem-solving process. This involves first identifying the potential challenges inherent in the context; then selecting the most critical or urgent challenge from among several challenges as the core issue; next, proposing and explaining solutions around this challenge, and conducting a systematic evaluation of the solutions from multiple dimensions, including feasibility and risk; and finally, based on the evaluation results, making targeted improvements and optimizations to the original solutions, thereby forming more creative and improved solutions.

[0028] Historical session records are defined as:

[0029] ,

[0030] in Indicates the preceding Turn all the conversation history, Indicates the first The current input of the user simulator, the model Receive historical sessions As input, and generate the first Round-robin candidate guidance strategy , can be represented as:

[0031] ;

[0032] Step 2.3: Add this round of interaction to the temporary session history for subsequent sampling and reward calculation;

[0033] Specifically,

[0034] This round of interaction Stored in temporary session history The record format is as follows:

[0035] ;

[0036] Step 2.4: Distribution from forward dialogue Starting with Monte Carlo sampling, responses are generated alternately using a user simulator and a model to obtain several possible subsequent dialogue trajectories. At the maximum number of rounds or the end of the session, a dialogue-level reward is calculated. ;

[0037] Specifically,

[0038] Based on a forward sampling strategy, multiple possible dialogue trajectories are generated using the Monte Carlo method. Each trajectory is generated based on the session history. Forward dialogue The probability is expressed as Then, the corresponding dialogue-level reward is calculated for each trajectory. This is used to evaluate the model's response quality in the current dialogue. This is achieved by limiting the maximum number of forward sampling rounds. This significantly reduces computational overhead while maintaining sufficient contextual information, thereby ensuring the scalability and efficiency of multi-round contextual dialogue training.

[0039] Step 2.5: Dialogue-based rewards and forward dialogue distribution Calculate multi-round perception rewards Select the optimal bootloader Update the main session history and use a proximate strategy (PPO, a reinforcement learning method that balances training stability and performance improvement by limiting the policy update magnitude) to optimize the algorithm to update model parameters;

[0040] The multi-round perception reward The calculation is as follows:

[0041] ,

[0042] in, The trajectory set generated by forward sampling, based on the obtained Select the optimal bootloader Update the main session history. The model parameters are updated based on Proximal Policy Optimization (PPO) as the output of the reward function. The optimization objective is:

[0043] ,

[0044] in, , Expressing expectations, For the dominant function, For the cropping operation, To limit the hyperparameters of the cropping range, For action, For state, For the new dialogue strategy, This is an old dialogue strategy;

[0045] Step 2.6: Repeat steps 2.2-2.5 for multiple rounds and multiple sessions of training until the model training converges and the final diagnostic and activation interaction strategy is obtained;

[0046] By repeatedly performing session generation, reward evaluation, and policy optimization until the model converges, an adaptive interaction policy is obtained.

[0047] This invention also proposes a creativity contextual interactive assessment system based on the above method. The system includes a user request module, a data processing module, a model training module, and a user interaction module.

[0048] The user request module is responsible for processing user-sent requests and the interactive data contained in the request message body.

[0049] The data processing module loads user-uploaded interactive data into memory or loads interactive data from the database into memory based on specified dataset information, and preprocesses the data.

[0050] The model training module trains the language model’s dynamic guidance capability based on preprocessed offline data, using supervised fine-tuning and proximal optimization strategies to obtain an adaptive interaction strategy.

[0051] The user interaction module uses a trained language model to engage in dialogue with the user. Through multiple rounds of interaction and dynamic adjustment of the guided interaction content, it enhances the user's willingness to express themselves and their interactive experience during the assessment process, thereby stimulating and assessing the user's creative potential.

[0052] The present invention also provides the application of the above-described method or system in the assessment of creativity intelligence, etc.

[0053] The beneficial effects of this invention include: This invention constructs a language model that supports multi-turn interaction and can dynamically adjust the content guiding the interaction. Compared with most existing interactive contextual assessment models that use static generation or single-turn question-and-answer mechanisms, it can better enhance students' willingness to express themselves and their interactive experience during the assessment process, help them express their implicit true level of creativity, enhance users' willingness to express themselves and their interactive experience during the assessment process, and achieve efficient, effective, and scalable creativity diagnosis and stimulation. Attached Figure Description

[0054] 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.

[0055] Figure 1 This is a flowchart of the contextualized assessment method of the present invention.

[0056] Figure 2 This is a system architecture diagram for the contextualized evaluation of this invention. Detailed Implementation

[0057] The present 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 present invention are all common knowledge and general knowledge in the art, and the present invention does not have any particular limitations.

[0058] This invention proposes a contextualized interactive assessment method for creativity diagnosis and stimulation. It constructs and trains a language model supporting multi-turn interactions and dynamically adjusting guidance strategies based on collected contextual interaction data. This model is trained using supervised fine-tuning and proximal optimization strategies to achieve automatic guided generation of creative scenarios. Dynamic guidance through multi-turn interactions aims to stimulate creativity, thereby accurately assessing user creativity performance. The method employs a two-stage training paradigm combining supervised fine-tuning and proximal strategy optimization algorithms. It constructs a model capable of dynamically adjusting guidance strategies and defines a dialogue-level reward function encompassing multiple dimensions, including creativity stimulation, intrinsic interaction quality, missing question penalty, interaction guidance, and efficiency. During training, a user simulator is introduced to simulate real user behavior, and Monte Carlo sampling is used to evaluate multi-turn dialogue trajectories under different guidance strategies, thereby calculating multi-turn perceived rewards. This strengthens the model's ability to guide and stimulate user creativity in multi-turn interactions. Through adaptive interaction strategies, it can stimulate user enthusiasm and implicit true creativity levels during the assessment process, enhancing user expression and interactive experience, achieving efficient, effective, and scalable creativity diagnosis and stimulation. In large-scale scenarios, this invention enables efficient and accurate automatic generation of creative contexts, assisting enterprises, schools, and other organizations in conducting creative context interactions and stimulating creative potential. This invention also provides a system for implementing the aforementioned contextualized interactive assessment method for creativity diagnosis and stimulation.

[0059] This invention uses interactive data from creative contexts as input to construct a language model that supports multi-turn interactions and can dynamically adjust and guide interactive content. Through two phases of training—supervised fine-tuning and proximal policy optimization—the model can better enhance students' willingness to express themselves and their interactive experience during the assessment process, thus helping them express their implicit, true level of creativity.

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

[0061] Step 1: Establish an evaluation index system in advance, initialize the dialogue-level reward calculation mechanism, update the language model parameters through supervised fine-tuning using pre-collected interaction data, and generate an initial dialogue strategy.

[0062] Step 2: Simulate real user behavior through a user simulator, generate candidate guidance strategies, add the current round of interaction to the temporary session history, sample one or more rounds of user dialogue trajectory, calculate multi-round perception rewards, and use the near-end strategy optimization algorithm to update model parameters.

[0063] Step 3: Repeat the interaction optimization model until convergence to obtain the adaptive interaction strategy;

[0064] Step 4: Based on the trained and optimized language model, conduct multiple rounds of interaction with real users to achieve real user creativity assessment.

[0065] Step one further includes the following:

[0066] Step 1.1: Initialize dialogue-level rewards based on the established evaluation index system. Computerized system;

[0067] Step 1.2: Using the pre-collected interaction data, update the model parameters of the language model using supervised fine-tuning to provide an initial strategy for the optimization in Step 2;

[0068] Step two further includes the following:

[0069] Step 2.1: In step two, a user simulator is introduced to simulate real user behavior based on session history. Generation probability distribution Output a response that conforms to the user's language style and typical behavior. ;

[0070] Step 2.2: Based on the set context and task objectives, and using historical conversation records... and the current input of the user simulator Model-generated candidate guidance strategies ;

[0071] Step 2.3: Add this round of interaction to the temporary session history for subsequent sampling and reward calculation;

[0072] Step 2.4: Distribution from forward dialogue Starting with Monte Carlo sampling, responses are generated alternately using a user simulator and a model to obtain several possible subsequent dialogue trajectories. At the maximum number of rounds or the end of the session, a dialogue-level reward is calculated. ;

[0073] Step 2.5: Dialogue-based rewards and forward dialogue distribution Calculate multi-round perception rewards Select the optimal bootloader Update the main session history and use the near-end strategy optimization algorithm to update the model parameters;

[0074] Step 2.6: Repeat steps 2.2-2.5 for multiple rounds and multiple sessions of training until the model training converges and the final diagnostic and stimulation interaction strategy is obtained.

[0075] An interactive creativity assessment system based on the above method is provided in this invention, such as... Figure 2As shown, it includes a user request module, a data processing module, a model training module, and a user interaction module;

[0076] The user request module is responsible for processing user-sent requests and the interactive data contained in the request message body.

[0077] The data processing module loads the user-uploaded interactive data into memory and preprocesses the data;

[0078] The model training module trains the language model’s dynamic guidance capability based on preprocessed offline data, using supervised fine-tuning and proximal optimization strategies to obtain an adaptive interaction strategy.

[0079] The user interaction module uses a trained language model to engage in dialogue with the user. Through multiple rounds of interaction and dynamic adjustment of the guided interaction content, it enhances the user's willingness to express themselves and their interactive experience during the assessment process, thereby stimulating and assessing the user's creative potential.

[0080] Example

[0081] The following provides further explanation of the specific implementation of the creativity assessment. This invention is further illustrated through the creativity task of "identifying and solving the core challenges of the Ola Kai project."

[0082] Implement a contextualized, interactive assessment method for creativity diagnosis and stimulation using the following steps: Figure 1 As shown:

[0083] Steps a and b constitute the first stage of supervised fine-tuning. An evaluation index system is established in advance, the dialogue-level reward calculation mechanism is initialized, and the language model parameters are updated through supervised fine-tuning using pre-collected interaction data to generate an initial dialogue strategy.

[0084] Step ch is the second stage of near-end strategy optimization. It simulates real user behavior through a user simulator, generates candidate guidance strategies, adds the current round of interaction to the temporary session history, samples one or more rounds of user dialogue trajectory, calculates multi-round perceived rewards, and updates model parameters using the near-end strategy optimization algorithm.

[0085] Step a: Initialize dialogue-level rewards based on the established evaluation index system. Computerized

[0086] Based on an evaluation index system designed by creativity experts, the process is divided into three stages: identifying challenges, selecting challenges, and proposing solutions. For each stage, creativity experts establish corresponding indicators. Based on the expert-designed indicator system, a dialogue-level reward calculation mechanism is initialized, encompassing rewards for stimulating creativity. Internal interaction quality reward Penalty for leaking questions Interactive guidance rewards and interaction efficiency rewards This comprehensively reflects the dialogue's performance in promoting creativity, maintaining content quality, and efficiently achieving its goals. Overall Award The calculation formula can be expressed as:

[0087] ,

[0088] in, Indicates interaction history. The task objectives are indicated, and the various reward items are defined as follows: Creativity Stimulation Reward Based on changes in user creativity scores, an evaluation is conducted according to creativity dimensions (diversity, originality, problem identification ability, problem analysis ability, and problem-solving ability), with values ​​ranging from [0,1]. For example, if a user's score changes from "only identifying 3 types of challenges" to "supplementing hidden challenges such as economic feasibility and cross-boundary execution," then... Improvement; Intrinsic Interaction Quality Rewards Based on the fluency and logic of the dialogue, the value range is [0,1]. For example, if the model guides the user to gradually delve into the analysis of "the relationship between microplastic pollution and regulatory loopholes," then... Approaching 1; Penalty for Leaking Questions This is used to penalize users when the model reveals answers that should have been completed by the user during interaction. The value range is [0,1]. For example, directly stating "The core challenge of the Ola Kai project is the lack of regulations" will result in a penalty; Interaction guidance reward. This is used to evaluate the model's effectiveness in guiding users back to the main task when they deviate from the topic. The scoring range is [0,1]. If a user mentions "irrelevant marine tourism development," the model guides the user with the question, "Will tourism development affect the efficiency of plastic pollution cleanup?" Approaching 1; Interaction efficiency reward This is used to measure the efficiency of completing an interaction, defined as follows: This aims to encourage students to produce more content to enhance cognitive engagement, while also encouraging models to generate more concise responses to improve interaction efficiency; among them, This refers to the number of tokens corresponding to the content output by students during the interaction process. This refers to the number of tokens corresponding to the content output by the system during interaction. Tokens are the basic unit when large models process text. This is an adjustment coefficient used to control the weighting relationship between the number of tokens output by students and the number of tokens output by the system. (Except for...) In addition, the reward or penalty scores are automatically calculated using a large language model driven by prompt words. In this embodiment, the prompt words for different reward or penalty items are as follows:

[0089] Creativity inspires rewards The system uses creativity prompts to enable large language models to evaluate the creativity of users' responses in dialogues. It mainly includes five parts: diversity, originality, problem discovery, problem analysis, and problem solving, and scores them on a scale of 0, 0.5, and 1. The output is a JSON file containing the thought process and the final score.

[0090] Intrinsic Interaction Quality Reward The quality of AI assistant interaction with users in conversations is evaluated using interactive prompts. It mainly includes three parts: engagement, clarity of response, and ability to understand user needs. The scores are 0, 0.5, and 1, and the output is a JSON file containing the thought process and interactivity scores.

[0091] Penalty for leaking questions The algorithm uses clue words to detect the possibility of the AI ​​assistant leaking answers during conversations, scoring them on a scale of 0, 0.5, and 1, and outputting a JSON file containing the thought process and the penalty score for leaking answers.

[0092] Interactive guidance rewards The AI ​​assistant is evaluated using interactive prompts to determine whether it can successfully guide the user back to the intended topic when the user deviates from the topic. The score is given on a scale of 0, 0.5, and 1, and the output is a JSON file containing the thought process and the interactive prompt score.

[0093] Step b: In the first stage, using pre-collected interaction data, the language model parameters are updated using supervised fine-tuning to provide an initial strategy for optimization in the second stage.

[0094] The language model is trained using supervised fine-tuning with an interactive dataset. Given the input, minimize the following negative log-likelihood objective:

[0095] ,

[0096] in, For context input, To guide the output towards the goal. For model parameters, For the model in parameters Given the input Generate target output in time The probability of;

[0097] Step c: In the second stage, a user simulator is introduced to simulate real user behavior based on session history. Generation probability distribution Output a response that conforms to the user's language style and typical behavior.

[0098] Each round of the session consists of user input and model output, defined as:

[0099] ,

[0100] in, Indicates the first The user input wheel, Indicates the first The guiding strategy generated by the wheel model.

[0101] A user simulator is introduced to simulate real user behavior, addressing the issues of high computational cost for multi-turn forward dialogue sampling during training and the high cost and lack of scalability of directly using real user dialogue. Specifically, in the... In the round-robin interaction, the user simulator Based on the given conversation history Generate possible user responses The user's input will be used as the input for that round, and the result will be processed through a probability distribution. Simulating real user behavior generates multi-turn dialogue responses that match the user's language style and typical behaviors. Taking the second round of data in this task as an example, the previous round's historical conversation showed that the user was concerned about "the continued existence of marine pollution, especially microplastic pollution." The simulator generates... The output includes phrases like, "Your point is very insightful. Ocean pollution, especially the persistence of microplastics, is indeed a multi-layered and far-reaching global environmental crisis," mimicking past user behavior.

[0102] Step d: Based on the set context and task objectives, and using historical conversation records... and the current input of the user simulator Model-generated candidate guidance strategies

[0103] The task objective is as follows: In a given context, the dialogue between the user and the model must proceed according to a prescribed problem-solving process. This involves first identifying the potential challenges inherent in the context; then selecting the most critical or urgent challenge from among several challenges as the core issue; next, proposing and explaining solutions around this challenge, and conducting a systematic evaluation of the solutions from multiple dimensions, including feasibility and risk; and finally, based on the evaluation results, making targeted improvements and optimizations to the original solutions, thereby forming more creative and improved solutions.

[0104] Historical session records are defined as:

[0105] ,

[0106] in Indicates the preceding Turn all the conversation history, Indicates the first The current input from the user simulator, taking the data from the second round of this task as an example. Enter "I agree with your point. In a deeper discussion, I believe the persistent marine pollution (especially microplastic pollution) is a core issue because it directly threatens marine ecosystems and can cause long-term environmental impacts. Weak regulatory enforcement in industrialized countries means that companies lack legal or economic incentives to reduce plastic waste, making it difficult to address the problem at its root. Finally, the effective management and conversion of collected plastic waste into fuel presents a significant logistical and technological challenge, impacting the project's scalability and sustainability. Which of these issues do you consider the most critical and the one that needs the highest priority?" Receive historical sessions And generate a second round of candidate guidance strategies. , respectively,

[0107] Candidate Guide "The three challenges you raised are indeed interconnected, but prioritizing them depends on our perspective. If we're looking for immediate results, then addressing microplastic pollution may be the most urgent, as its damage to ecosystems is irreversible. However, for systemic change, we may first need to address the lack of regulation—stricter policies can incentivize industry innovation and reduce waste at its source."

[0108] While the technological challenges are complex, the prospects are vast if breakthroughs can be achieved and scaled up in converting plastics into fuel. This could revolutionize waste management and create economic impetus for cleanup efforts. How do you view the interplay of these challenges over time? Would you prioritize one to drive progress in the others?

[0109] Alternatively, we could explore a hybrid approach—for example, combining public advocacy to drive regulations with R&D investment. In your opinion, what kind of solution would be the most effective starting point?

[0110] Candidate Guide "Your analysis is spot on! Each challenge is deeply interconnected, but prioritizing them depends on the project's short-term goals. Addressing regulatory weaknesses can drive systemic change by creating incentives for industries to reduce plastic emissions, but this requires political action. Meanwhile, tackling microplastic pollution may require innovative cleanup technologies combined with public awareness campaigns."

[0111] However, if the goal is to achieve rapid progress, then improving waste-to-fuel technologies or licensing is considered a core challenge, as it directly guarantees the project's operational success and demonstrates tangible results. In this context, how would you weigh the urgency of policy change against technological innovation? Will one create conditions for the other?

[0112] (You've done an excellent job analyzing these complex problems!) Would you be willing to discuss potential solutions for any of them?

[0113] and candidate guidance "All three challenges are deeply intertwined, but addressing weak regulation could be a fundamental step. Stricter policies would incentivize industry to cut plastic production, thereby reducing the amount of waste entering the ocean and easing pressure on microplastic pollution and waste management systems. This regulatory shift could also foster innovation in recycling and fuel conversion technologies."

[0114] Would you like to explore potential strategies for lobbying governments or incentivizing business transformation? Alternatively, we could examine how technological advancements in waste-to-fuel processes can circumvent some of these obstacles. In which direction would you like to explore this next?

[0115] Step e: This round of interaction Add to temporary session history Used for subsequent sampling and reward calculation

[0116] This round of interaction Stored in temporary session history The record format is as follows:

[0117] ;

[0118] Guided by candidates For example, ={(“User”:“Based on the text, I believe the most significant challenges facing the Ola Kai project are the persistent marine pollution (especially microplastic pollution), weak regulations in industrialized countries, and the difficulty in effectively managing and converting collected plastic waste into fuel. What are your thoughts on these challenges?”,

[0119] Assistant: "Your analysis of the challenges facing the Ola Kai project is very insightful! The persistent marine pollution (particularly microplastics) and the weak regulatory framework are indeed major obstacles. In addition, the technical challenges of efficiently converting collected plastics into fuel are also key factors affecting the project's scalability and sustainability." Which challenge would you like to explore further? For example, how can stricter regulations or innovative waste management technologies address these issues? Or can we collaboratively devise creative solutions to tackle microplastic pollution at its source? Please let me know how I can help deepen your thinking!

[0120] (User: "I agree with your point. In a deeper discussion, I believe the persistent marine pollution (especially microplastic pollution) is a core issue because it directly threatens marine ecosystems and can cause long-term environmental impacts. Weak regulatory oversight in industrialized countries means that companies lack legal or economic incentives to reduce plastic waste, making it difficult to address the problem at its root. Finally, effectively managing and converting collected plastic waste into fuel presents significant logistical and technological challenges, impacting the project's scale and sustainability. Which of these issues do you consider the most critical and the one that needs the highest priority?")

[0121] The assistant stated: "All three challenges are deeply intertwined, but addressing weak regulation could be a fundamental step. Stricter policies would incentivize industry to cut plastic production, thereby reducing the amount of waste entering the ocean and easing pressure on microplastic pollution and waste management systems. This regulatory shift could also foster innovation in recycling and fuel conversion technologies." Would you like to explore potential strategies for lobbying governments or incentivizing corporate transformation? Alternatively, we could examine how technological advancements in waste-to-fuel processes can circumvent some of these obstacles. In which direction would you like to explore this next?

[0122] Step f: Distribution from the forward dialogue Starting with Monte Carlo sampling, responses are generated alternately using a user simulator and a model to obtain several possible subsequent dialogue trajectories. At the maximum number of rounds or the end of the session, a dialogue-level reward is calculated.

[0123] Based on a forward sampling strategy, multiple possible dialogue trajectories are generated using the Monte Carlo method. Each trajectory is generated based on the session history. Forward dialogue The probability is expressed as Guided by the candidate model in this example For example, its three forward trajectories are as follows:

[0124] User: "Based on the text, I believe the most significant challenges facing the Ola Kai project are the persistent marine pollution (particularly microplastic pollution), weak regulations in industrialized countries, and the difficulty in effectively managing and converting collected plastic waste into fuel. What are your thoughts on these challenges?"

[0125] Assistant: "Your analysis of the challenges facing the Ola Kai project is very insightful! The persistent marine pollution (particularly microplastics) and the weak regulatory framework are indeed major obstacles. In addition, the technical challenges of efficiently converting collected plastics into fuel are also key factors affecting the project's scalability and sustainability." Which challenge would you like to explore further? For example, how can stricter regulations or innovative waste management technologies address these issues? Or can we collaboratively devise creative solutions to tackle microplastic pollution at its source? Please let me know how I can help deepen your thinking!

[0126] User: "I agree with your point. Delving deeper, I believe the persistent marine pollution (especially microplastic pollution) is a core issue because it directly threatens marine ecosystems and can cause long-term environmental impacts. Weak regulatory oversight in industrialized countries means that companies lack legal or economic incentives to reduce plastic waste, making it difficult to address the problem at its root. Finally, effectively managing and converting collected plastic waste into fuel presents significant logistical and technological challenges, impacting the project's scalability and sustainability. Which of these issues do you consider the most critical and the one that needs the highest priority?"

[0127] "Assistant": "All three challenges are deeply intertwined, but prioritizing them depends on our perspective. If we're looking for immediate results, addressing microplastic pollution may be the most urgent, as its damage to ecosystems is irreversible. However, for systemic change, we may first need to address the lack of regulation—stricter policies can incentivize industry innovation and reduce waste at its source." While the technological challenges are complex, the prospects are vast if breakthroughs can be achieved and scaled up in converting plastics into fuel. This could revolutionize waste management and create economic impetus for cleanup efforts. How do you view the interplay of these challenges over time? Would you prioritize one to drive progress in the others? Alternatively, we could explore a hybrid approach—for example, combining public advocacy to drive regulations with R&D investment. In your opinion, what kind of solution would be the most effective starting point?

[0128] User: "I understand your point about the interconnectedness of these challenges. After some consideration, I believe the most critical and immediate challenge is the weak regulation in industrialized countries. By implementing stricter policies, we can reduce the amount of plastic waste entering the ocean, which will help alleviate the persistent problem of microplastic pollution and the logistical challenges of managing and converting plastic waste. Stricter regulations can also drive innovation in waste management and cleanup technologies. What do you think about starting with this aspect of regulation?"

[0129] "Assistant": "Focusing on stronger regulation is a strategic option—it addresses the root cause by creating systemic change. Industrialized nations have enormous influence, and their policies can set global standards to reduce plastic inflows into the ocean." For this method to be effective, we need to: 1. Data-driven advocacy Emphasizing economic, health, and environmental costs to drive policy change. 2. Corporate participation : Use corporate social responsibility to put pressure on key industries and support regulation. 3. pilot projects Showcase success stories (such as Ola Kai's work) to build political will for stricter laws. However, this is a slow process. Can we combine regulatory efforts with urgent technological developments (such as advanced plastic-to-fuel conversion methods) or grassroots cleanup campaigns to create an immediate impact while awaiting policy changes? How can we balance these approaches? Your emphasis on regulation aligns very well with the perspective of "systemic change." Would you like to develop a phased strategy focusing on this?

[0130] User: "I agree that a multi-pronged approach is crucial. While focusing on stronger regulation is important for long-term systemic change, we should also combine it with immediate action. For example, investing in advanced plastic-to-fuel conversion technologies and organizing more frequent and effective cleanup campaigns can help reduce current pollution levels and demonstrate the feasibility and benefits of such initiatives. This dual approach can provide momentum and support for regulatory change. What do you think of this balanced strategy?"

[0131] "Assistant": "This balancing strategy is very pragmatic and is likely the most effective way forward. Here's how the components complement each other:" ### Regulatory Focus (Long-Term) - Promote Extended Producer Responsibility (EPR) laws that require manufacturers to be responsible for waste. - Advocate for a ban on single-use plastics and incentivize circular economy models. - Harmonize global standards using international treaties (such as the UN Plastics Treaty). ### Technology and cleanup operations (short-term) - Expand Ola Kai’s work by utilizing emerging pyrolysis or catalytic conversion technologies to improve the efficiency of plastic-to-fuel conversion. - Collaborate with local communities and NGOs to conduct rapid cleanup operations (e.g., beach cleanup + river interception systems). - Pilot the "pollution-to-profit" model, allowing waste collectors to earn income by recycling / fueling materials. ### Synergistic effect Successful cleanup campaigns can be used to lobby the government. Technological advancements reduce costs, making stricter policies economically feasible. - Public support from corporate partnerships can accelerate the regulatory timeline. Would you like to elaborate on any specific aspects of this strategy? For example, which technological or policy levers are most promising in terms of near-term impact? Note: I've added some structure to highlight potential paths—I'd be happy to adjust it to suit your priorities. --- Do you need resources on demonstration policies, case studies of successful technology implementations, or grassroots mobilization strategies? },

[0132] User: "Based on the text, I believe the most significant challenges facing the Ola Kai project are the persistent marine pollution (particularly microplastic pollution), weak regulations in industrialized countries, and the difficulty in effectively managing and converting collected plastic waste into fuel. What are your thoughts on these challenges?"

[0133] Assistant: "Your analysis of the challenges facing the Ola Kai project is very insightful! The persistent marine pollution (particularly microplastics) and the weak regulatory framework are indeed major obstacles. In addition, the technical challenges of efficiently converting collected plastics into fuel are also key factors affecting the project's scalability and sustainability." Which challenge would you like to explore further? For example, how can stricter regulations or innovative waste management technologies address these issues? Or can we collaboratively devise creative solutions to tackle microplastic pollution at its source? Please let me know how I can help deepen your thinking!

[0134] User: "I agree with your point. Delving deeper, I believe the persistent marine pollution (especially microplastic pollution) is a core issue because it directly threatens marine ecosystems and can cause long-term environmental impacts. Weak regulatory oversight in industrialized countries means that companies lack legal or economic incentives to reduce plastic waste, making it difficult to address the problem at its root. Finally, effectively managing and converting collected plastic waste into fuel presents significant logistical and technological challenges, impacting the project's scalability and sustainability. Which of these issues do you consider the most critical and the one that needs the highest priority?"

[0135] "Assistant": "All three challenges are deeply intertwined, but prioritizing them depends on our perspective. If we're looking for immediate results, addressing microplastic pollution may be the most urgent, as its damage to ecosystems is irreversible. However, for systemic change, we may first need to address the lack of regulation—stricter policies can incentivize industry innovation and reduce waste at its source." While the technological challenges are complex, the prospects are vast if breakthroughs can be achieved and scaled up in converting plastics into fuel. This could revolutionize waste management and create economic impetus for cleanup efforts. How do you view the interplay of these challenges over time? Would you prioritize one to drive progress in the others? Alternatively, we could explore a hybrid approach—for example, combining public advocacy to drive regulations with R&D investment. In your opinion, what kind of solution would be the most effective starting point?

[0136] "User": "I understand the interconnectedness of these challenges. After careful consideration, I believe that addressing the regulatory gap is the most critical challenge. Stricter regulations can encourage improved waste management practices and reduce the amount of plastic entering the ocean from the outset, which will alleviate the pressure on the Ola Kai project and other similar initiatives." In summary, the key challenges include: 1. The persistence of marine pollution (especially microplastics). This directly threatens the marine ecosystem and is an urgent problem that needs to be addressed. 2. Weak regulation in industrialized countries The lack of strong legal and economic incentives for companies to reduce plastic waste. 3. The difficulty of effectively managing collected plastic waste and converting it into fuel It affects the scalability and sustainability of the project. In my opinion, addressing the regulatory gaps seems like a reasonable starting point. What do you think? [[End of conversation]]

[0137] User: "Based on the text, I believe the most significant challenges facing the Ola Kai project are the persistent marine pollution (particularly microplastic pollution), weak regulations in industrialized countries, and the difficulty in effectively managing and converting collected plastic waste into fuel. What are your thoughts on these challenges?"

[0138] Assistant: "Your analysis of the challenges facing the Ola Kai project is very insightful! The persistent presence of marine pollution (particularly microplastics) and the weak regulatory framework are indeed major obstacles. In addition, the technical challenges of efficiently converting collected plastics into fuel are also key factors affecting the project's scalability and sustainability." Which challenge would you like to explore further? For example, how can stricter regulations or innovative waste management technologies address these issues? Or can we collaboratively devise creative solutions to tackle microplastic pollution at its source? Please let me know how I can help deepen your thinking!

[0139] User: "I agree with your point. Delving deeper, I believe the persistent marine pollution (especially microplastic pollution) is a core issue because it directly threatens marine ecosystems and can cause long-term environmental impacts. Weak regulatory oversight in industrialized countries means that companies lack legal or economic incentives to reduce plastic waste, making it difficult to address the problem at its root. Finally, effectively managing and converting collected plastic waste into fuel presents significant logistical and technological challenges, impacting the project's scalability and sustainability. Which of these issues do you consider the most critical and the one that needs the highest priority?"

[0140] "Assistant": "All three challenges are deeply intertwined, but prioritizing them depends on our perspective. If we're looking for immediate results, addressing microplastic pollution may be the most urgent, as its damage to ecosystems is irreversible. However, for systemic change, we may first need to address the lack of regulation—stricter policies can incentivize industry innovation and reduce waste at its source." While the technological challenges are complex, the prospects are vast if breakthroughs can be achieved and scaled up in converting plastics into fuel. This could revolutionize waste management and create economic impetus for cleanup efforts. How do you view the interplay of these challenges over time? Would you prioritize one to drive progress in the others? Alternatively, we could explore a hybrid approach—for example, combining public advocacy to drive regulations with R&D investment. In your opinion, what kind of solution would be the most effective starting point?

[0141] User: "I believe the most critical and urgent challenge is the weak regulation in industrialized countries. While microplastic pollution is extremely harmful and urgently needs to be addressed, stricter regulations can create more sustainable long-term solutions by reducing the amount of plastic waste entering the ocean. However, I also recognize the value of a hybrid approach, which simultaneously promotes regulatory reform and invests in R&D for better plastic-to-fuel technologies. This way, we can address the problem at its source while also improving our ability to manage existing pollution. What are your thoughts on this hybrid approach?"

[0142] "Assistant": "A hybrid approach—simultaneously pushing for stricter regulations (to curb future pollution) and developing technologies (to treat existing waste)—is a highly strategic choice. Here are some specific ways to implement it:" 1. Regulatory recommendations: - Promote Extended Producer Responsibility (EPR) legislation that requires manufacturers to bear the costs of plastic recycling or safe disposal. - Provide subsidies and incentives for industries that adopt a circular economy model. - Enforce single-use plastic bans in major polluting countries. 2. Technological Innovation: - Fund research into pyrolysis or other chemical recycling to improve the efficiency of converting plastics into fuels. - Develop AI-driven sorting systems to optimize waste management processes. - Explore biodegradable alternatives to reduce reliance on traditional plastics. 3. Public-private partnerships: - Collaborate with NGOs, governments, and businesses to scale up cleanup operations (e.g., the "Ocean Cleanup" project). - Establish a "marine-friendly" product certification system to guide consumer demand. This dual-track strategy encompasses both prevention and governance. The key lies in synergy—how can policy support technology implementation, and vice versa? Would you like to add specific policy or technological directions to this framework? For example, imposing a "carbon tax" on plastic producers could fund cleanup and research and development; or using blockchain to trace the flow of plastic waste to ensure compliance. These ideas could further enhance the effectiveness of hybrid approaches. Would you like to develop a specific five-year plan for such an initiative? I'd be happy to assist with the details! (Note: I have avoided repeating previous points, but have implied them within the framework. Please let me know if you would prefer a more concise reply!)

[0143] "User": "I believe the core challenge is the weak regulation in industrialized countries, because this could create a more sustainable long-term solution by reducing the amount of plastic waste entering the ocean. However, I also agree on the value of a hybrid approach, which simultaneously promotes regulatory reform and technological innovation. Specific challenges include: 1) the persistent presence of marine pollution (especially microplastics); 2) weak regulation in industrialized countries; and 3) the difficulty of effectively managing collected plastic waste and converting it into fuel. Let's focus on developing a comprehensive five-year plan to address these issues. How can this plan be structured to be most effective?"

[0144] "Assistant": "Developing an effective five-year plan to address the challenges of the Ola Kai project requires a combination of policy advocacy, technological innovation, and grassroots mobilization. The following is a phased approach:" ### Phase 1: Infrastructure Development (Years 1-2) Target: We accumulate momentum through publicity, pilot projects, and alliance building. - Policy recommendations: - Launch a global initiative to encourage G20 countries to develop strict post-consumer plastics emission reduction regulations. - Collaborate with think tanks to draft a model for the Extended Producer Responsibility (EPR) system. - Technology pilot: - Fund 3-5 pilot projects to test advanced plastic-to-fuel technologies in areas with severe plastic pollution. - Develop an artificial intelligence and robot-assisted microplastic waste sorting system. - Public participation: - Organize cleanup activities in coastal communities to attract media attention and raise public awareness. - Collaborate with brands to promote the "zero waste" initiative. ### Phase Two: Policy and Technology Promotion (Years 2-4) Target: Reach binding agreements and scale up the application of mature technologies. - Regulatory Breakthrough: - Strive to get at least three major polluting countries to commit to mandatory biodegradable packaging or plastic taxes. - Cooperate with international organizations (such as the International Maritime Organization and the United Nations Environment Programme) to improve marine protection treaties. - Technology Implementation: - Increase the processing capacity of the successful pilot plant by 10 times. - Promote "smart trash cans" equipped with IoT sensors to track waste flow in real time. - Behavior change: - Collaborate with schools to launch the "Plastic-Free Youth" education program. - Incentivize consumer participation through blockchain-certified recycling rebates. ### Phase Three: Global Integration (Years 4-5) Target: Transform the results into long-term structural changes. - Policy standardization: - Promote a "Global Plastics Treaty" modeled after the Paris Agreement. - Advocate for the WTO to impose sanctions on countries that export high levels of plastics. - Ecological restoration: - Deploy large autonomous navigation cleaning drones in the ocean circulation zone. - Restore coastal mangroves to naturally filter microplastics. - Sustainable Model: - Certify "plastic neutral" companies and provide green bond incentives for eco-innovators. ### Key success metrics: - The percentage of G20 countries that have adopted EPR laws. - The number of tons of plastic removed from the ocean each year. - The number of large-scale plastic-to-fuel conversion facilities in operation. - The degree of public willingness to shift to a plastic-free lifestyle. ### Risks and Responses: - Industry obstacles: Pressure companies to comply with regulations through shareholder activism. - Technical bottlenecks: Distribute R&D investment and explore multiple transformation technologies. - Funding shortage: Mobilize philanthropy and carbon offset income streams. Do you wish to focus on a specific area (such as policy or technology) or explore a particular regional strategy (such as targeting Southeast Asia)? This program is flexible – please let us know how you would like to further customize it! (Note: This plan clearly embodies a hybrid approach and clearly defines the phases.) --- Please let me know if you require a more detailed version of any part! ”}.

[0145] Subsequently, the corresponding dialogue-level reward is calculated for each trajectory. This is used to evaluate the model's response quality in the current dialogue. This is achieved by limiting the maximum number of forward sampling rounds. This significantly reduces computational overhead while preserving sufficient contextual information, thereby ensuring the scalability and efficiency of multi-round contextual dialogue training. In this embodiment, the maximum number of forward sampling rounds is limited. Candidate model guidance , , The dialogue-level rewards are (43.55, 21.85, 272.22), (-219.86, -173.48, 12.42), and (-227.98, 0.38, 84.34).

[0146] Step g: Dialogue-based rewards and forward dialogue distribution Calculate multi-round perception rewards Select the optimal bootloader Update the main session history and use a near-end strategy optimization algorithm to update the model parameters.

[0147] Finally, multi-round perception rewards The calculation is as follows:

[0148] ,

[0149] in, The trajectory set generated by forward sampling, based on the obtained Select the optimal bootloader Update the main session history. In this embodiment, candidate model guidance... , , The multi-round perception rewards were 112.54, -126.97, and -47.75 respectively, therefore, the answer is... Update the session history.

[0150] by As the output of the reward function, the model parameters are updated based on Proximal Policy Optimization (PPO, a reinforcement learning method that balances training stability and performance improvement by limiting the magnitude of policy updates). The optimization objective is:

[0151] ,

[0152] in, , Expressing expectations, For the dominant function, For the cropping operation, To limit the hyperparameters of the cropping range, For action, For state, For the new dialogue strategy, This is an old dialogue strategy;

[0153] Step h: Repeat step dg for multiple rounds and multiple sessions of training until the model training converges, obtaining the final diagnostic and interaction stimulation strategy.

[0154] Repeat the process of session generation, reward evaluation, and policy optimization until the model converges, resulting in a language model with dynamic guidance capabilities. For example, when a user suggests, "Microplastics are too small; collection technologies are useless," The prompt asks: "Do you think that 'source control (such as banning microplastic products)' can reduce the amount that needs to be collected? Does the Ola Kai project in the document, 'Focusing on Microplastics,' mention any shortcomings in source control?"

[0155] References

[0156] [1] Patterson, J.D., Pronchick, J., Panchanadikar, R.et al. (2025). CAP: The creativity assessment platform for online testing and automated scoring. Behav Res 57, 264.

[0157] [2] Kim, S., Shin, J., Cho, Y., Jang, J., Longpre, S., Lee, H., Yun, S., Shin, S., Kim, S., Thorne, J., & Seo, M. (2023). Prometheus: Inducing Fine-grained Evaluation Capability in Language Models. ArXiv, abs / 2310.08491.

[0158] [3] Guilford, J. P. (1967). Creativity: Yesterday, today and tomorrow. Journal of Creative Behavior, l(1), 3-14.

[0159] [4] Runco, Mark A., and Garrett J. Jaeger. The standard definition of creativity. (2012). Creativity Reasearch Journal 24.1(2012):92-96.

[0160] [5] Abdulla, A. M., Paek, S. H., Cramond, B., & Runco, M. A. (2020). Problem finding and creativity: A meta-analytic review. Psychology of Aesthetics, Creativity, and the Arts, 14(1), 3

[0161] [6]Sasson Lazovsky, G., Raz, T., & Kenett, Y. N. (2025). The art ofcreative inquiry—from question asking to prompt engineering. The Journal ofCreative Behavior, 59(1), e671.

[0162] [7]Raz, T., Reiter-Palmon, R., & Kenett, Y. N. (2023). The role ofasking more complex questions in creative thinking. Psychology of Aesthetics,Creativity, and the Arts.

[0163] [8]Wróblewska, A., Korbin, M., Kenett, Y. N., Dan, D., Ganzha, M., &Paprzycki, M. (2024). Applying text mining to analyze human question askingin creativity research. In International Conference on Big Data Analytics(pp. 44-73). Cham: Springer Nature Switzerland.

[0164] [9]Plucker, Jonathan A., Meihua Qian, and Stephanie L. Schmalensee.(2014). Is what you see what you really get? Comparison of scoring techniquesin the assessment of realworld divergent thinking. Creativity ResearchJournal 26.2: 135-143.

[0165]

[10] Reiter-Palmon, R., Forthmann, B., & Barbot, B. (2019). Scoringdivergent thinking tests: A review and systematic framework. Psychology ofAesthetics, Creativity, and the Arts, 13(2), 144.

[0166]

[11] Michael D.Mumford, Tristan Mclntosh. (2017). Creative ThinkingProcesses: The Past and the Future. Jonrnal of Creative Behavior. Volume 51,Issue 4.

[0167]

[12] Michael D.Mumford, Deborah L.Whetzel, Roni Reiter-Palmon. (1997).Thinking Creatively at Work: Organization Influences on Creative ProblemSolving. Jonrnal of Creative Behavior. Volume 31, Issue 1.

[0168]

[13] Cropley, D. (2010). Functional Creativity: Products and theGeneration of Effective Novelty. In J. C. Kaufman and R. J. Sternberg (Eds.),Cambridge Handbook of Creativity, Chapter 16 (Pp. 301-317). New York:Cambridge University Press.

[0169]

[14] Simonton, DK (2012). Taking the US Patent Office Criteria Seriously: A Quantitative Three-Criterion Creativity Definition and ItsImplications. Creativity Research Journal, 24(2–3), 97–106.

[0170] 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 this invention are included in this invention and are protected by the appended claims.

Claims

1. A contextualized interactive assessment method for diagnosing and stimulating creativity, characterized in that, The method assesses creativity through a two-stage task involving supervised fine-tuning and proximal policy optimization, including: Step 1: Establish an evaluation index system in advance, initialize the dialogue-level reward calculation mechanism, update the language model parameters through supervised fine-tuning using pre-collected interaction data, and generate an initial dialogue strategy. Step 2: Simulate real user behavior through a user simulator, generate candidate guidance strategies, add the current round of interaction to the temporary session history, sample one or more rounds of user dialogue trajectory, calculate multi-round perception rewards, and use the near-end strategy optimization algorithm to update model parameters. Step 3: Repeat the interaction optimization model until convergence to obtain the adaptive interaction strategy; Step 4: Based on the trained and optimized language model, conduct multiple rounds of interaction with real users to achieve real user creativity assessment.

2. The method as described in claim 1, characterized in that, In step one, the evaluation index system is established through multi-level indicators. The first-level indicators include: identifying challenges, selecting challenges, and proposing solutions. The second-level indicators corresponding to identifying challenges include fluency, novelty, flexibility, complexity, real-world insight, and applicability. The second-level indicators corresponding to selecting challenges include fluency, novelty, and flexibility. The second-level indicators corresponding to proposing solutions include fluency, novelty, flexibility, feasibility, effectiveness, and relevance. The dialogue-level reward calculation mechanism includes: creativity stimulation reward, intrinsic interaction quality reward, question omission penalty, interaction guidance reward, and interaction efficiency reward; Overall reward = Creativity stimulation reward + Intrinsic interaction quality reward - Question omission penalty + Interaction guidance reward + Interaction efficiency reward.

3. The method as described in claim 1, characterized in that, In step one, a large language model is trained using supervised fine-tuning with the pre-collected interactive dataset as input, and the objective is to minimize the negative log-likelihood. ， in, For context input, To guide the output towards the goal. For model parameters, For interactive datasets, For the model in parameters Given the input Generate target output in time The probability of.

4. The method as described in claim 1, characterized in that, In step two, each round of conversation consists of user input and model output. A user simulator is introduced to simulate real user behavior and generate multi-round dialogue responses that conform to the user's language style and typical behavior. In the In the round-robin interaction, the user simulator Based on the given conversation history Generate possible user responses And use it as input to the current user simulator, through probability distribution Simulate real user behavior.

5. The method as described in claim 1, characterized in that, In step two, based on the set context and task objectives, and using historical conversation records... and the current input of the user simulator Model-generated candidate guidance strategy ; Historical session records are defined as: , in, Indicates the preceding Turn all the conversation history, Indicates the first The current input of the user simulator, the model Receive historical sessions As input, and generate the first Round-robin candidate guidance strategy , can be represented as: 。 6. The method as described in claim 1, characterized in that, In step two, the current interaction is added to the temporary session history, represented as: ， in, This indicates the current round of interaction. Indicates the history of temporary sessions.

7. The method as described in claim 1, characterized in that, In step two, the forward dialogue distribution... Starting with Monte Carlo sampling, responses are generated alternately using a user simulator and a model to obtain several possible subsequent dialogue trajectories. At the maximum number of rounds or the end of the session, a dialogue-level reward is calculated. Each trajectory is generated based on the session history. Forward dialogue The probability is expressed as Limit the maximum number of forward sampling rounds. .

8. The method as described in claim 1, characterized in that, In step two, based on dialogue-level rewards and forward dialogue distribution Calculate multi-round perception rewards Select the optimal bootloader Update the main session history and use a near-end strategy optimization algorithm to update the model parameters; among which, The multi-round perception reward The calculation is as follows: ， in, The trajectory set generated by forward sampling, based on the obtained Select the optimal bootloader Update the main session history; by As the output of the reward function, the model parameters are updated based on the near-end policy optimization. The optimization objective is: ， in, , Expressing expectations, For the dominant function, For the cropping operation, To limit the hyperparameters of the cropping range, For action, For state, For the new dialogue strategy, This is an old dialogue strategy.

9. An evaluation system for implementing the method as described in any one of claims 1-8, characterized in that, The evaluation system includes: a user request module, a data processing module, a model training module, and a user interaction module; wherein, The user request module processes the requests sent by the user and the interactive data contained in the request message body; The data processing module loads user-uploaded interactive data into memory or loads interactive data from the database into memory based on specified dataset information, and preprocesses the data. The model training module trains the language model’s dynamic guidance capability based on preprocessed offline data, using supervised fine-tuning and proximal optimization strategies to obtain an adaptive interaction strategy. The user interaction module utilizes a trained language model to stimulate and assess users' creative potential through multiple rounds of interaction.

10. The method as described in any one of claims 1-8, or the system as described in claim 9, in the application of creativity intelligence assessment.

Images