An AI-guided interaction-based cognitive empowerment method and system

Abstract

The application provides a cognitive empowerment method and system based on AI guided interaction, and relates to the technical field of data processing.The method comprises the following steps: based on the perceived cognitive state, dynamically generating guided interaction content, wherein the guided interaction content comprises guided questions, which are used to guide the user to think without directly providing answers; during the multi-round interaction with the user, the change of the cognitive state of the user is continuously perceived, and the guided strategy is dynamically adjusted based on the changed cognitive state; after outputting the analytical content, a forced pause is executed, a boundary statement and / or a confirmatory question are outputted to the user, so that the final decision right is returned to the user, and independent thinking of the user is ensured.The application can significantly improve the clarity score of complex problem solving, and effectively compress the decision thinking time.

Description

Technical Field

[0001] This invention relates to the field of data processing technology, and in particular to a cognitive empowerment method and system based on AI-guided interaction. Background Technology

[0002] Currently, AI interaction systems, represented by large-scale language models, generally adopt a "passive response" interaction paradigm. In this paradigm, users must actively ask questions, and the AI ​​system generates and returns answers based on the content of those questions. However, this existing model has significant technical shortcomings.

[0003] It sets an extremely high cognitive threshold for users. The passive response mode requires users to have the ability to transform their complex and vague questions into precise and structured inquiries, that is, to master the so-called prompt word engineering skills. For the vast majority of ordinary users who do not possess this professional skill, their interaction with the AI ​​system is often inefficient, and it is difficult to obtain effective assistance from AI to solve substantial and complex problems.

[0004] Current technologies lack the ability to dynamically perceive and adapt to users' cognitive states. Users' psychological states (such as confusion, anxiety, and dependency) are constantly changing during problem-solving. Existing interaction systems cannot automatically identify these states by analyzing the linguistic features of user input, let alone dynamically adjust interaction strategies accordingly. This results in AI systems consistently responding to users in all states with a single, fixed pattern, failing to provide appropriate cognitive assistance when users most need guidance, and unable to effectively intervene when users develop dependency tendencies.

[0005] Existing technologies, when assisting in decision-making and thinking, risk eroding users' cognitive sovereignty. Mechanisms that directly provide answers can easily lead users to skip independent thinking and directly accept the AI's output. This interaction process not only lacks mechanisms to force users to pause and independently confirm at key decision points, but also lacks an honest mechanism for proactively declaring the boundaries of their own capabilities. Over time, this will lead to severe AI dependence on users, weakening their own cognitive and decision-making abilities.

[0006] In summary, the question is how to construct a novel AI interaction architecture that upgrades AI's role from a passive answer provider to an active cognitive navigator, using mechanisms such as proactive guidance, dynamic adaptation, and mandatory protection of cognitive sovereignty to address the aforementioned issues. Summary of the Invention

[0007] This invention provides a cognitive empowerment method and system based on AI-guided interaction, which can significantly improve the clarity score of solving complex problems, while effectively compressing decision-making time.

[0008] To solve the above-mentioned technical problems, the technical solution of the present invention is as follows: Firstly, a cognitive empowerment method based on AI-guided interaction, the method comprising: The problem of receiving natural language input from users; The natural language problem is analyzed to determine the user's core needs, and the user's current cognitive state is perceived by analyzing the language features of the input. Based on the perceived cognitive state, guided interactive content is dynamically generated. The guided interactive content includes guided questions to guide users to think, rather than providing answers directly. During multiple rounds of interaction with users, continuously perceive changes in the user's cognitive state and dynamically adjust the guidance strategy based on the changed cognitive state; After outputting analytical content, a forced pause is executed, and boundary statements and / or confirmatory questions are presented to the user to return the final decision-making power to the user and ensure that the user completes independent thinking.

[0009] Secondly, a cognitive empowerment system based on AI-guided interaction includes: The receiving module is used to receive natural language questions input by the user; The intent parsing and state awareness module is used to perform intent parsing on the natural language question, determine the user's core needs, and perceive the user's current cognitive state by analyzing the language features of the input. The guidance strategy dynamic generation module is used to dynamically generate guided interactive content based on the perceived cognitive state. The guided interactive content includes guided questions to guide users to think rather than directly providing answers. The multi-turn dialogue management module is used to continuously sense changes in the user's cognitive state during multiple rounds of interaction and dynamically adjust the guidance strategy based on the changed cognitive state. The Forced Pause and Sovereignty Relinquishment module is used to perform a forced pause after outputting analytical content, outputting boundary statements and / or confirmatory questions to the user, so as to return the final decision-making power to the user.

[0010] The above-described solution of the present invention has at least the following beneficial effects: By proactively generating guided questions, the system liberates users from the constraints of needing to master precise questioning skills. Through a closed loop of intent parsing and dynamic guidance based on state awareness, the system proactively helps users clarify ambiguous questions and break down complex goals. This enables users from any background to communicate in natural language and efficiently advance their thinking under the system's structured guidance. This solution significantly improves the clarity score for solving complex problems while effectively reducing decision-making time.

[0011] By continuously analyzing the linguistic features of user input, the system dynamically perceives the user's cognitive state and adjusts guidance strategies in real time based on changes in that state. This technology ensures that the AI's interaction method is not static but always precisely matched to the user's current cognitive needs. For example, when a user shifts from confusion to dependence, the system can instantly switch to a forced independent thinking mode, thus providing optimal assistance while preventing AI from misleading or providing ineffective guidance.

[0012] This invention innovatively introduces a forced pause and boundary declaration mechanism. After outputting the analysis content, the system proactively pauses and outputs a confirmatory question, forcibly returning the final analysis, judgment, and decision-making process to the user. This ensures, from a technical architecture perspective, that humans always maintain a dominant position throughout the entire decision-making loop, fundamentally resolving the ethical and technical risks of users becoming dependent on AI by directly obtaining answers. Simultaneously, by proactively declaring the analytical boundaries of the system's capabilities, it cultivates critical thinking in users and establishes a safer and more trustworthy human-computer collaboration relationship. Attached Figure Description

[0013] Figure 1 This is a flowchart illustrating a cognitive empowerment method based on AI-guided interaction, provided by an embodiment of the present invention.

[0014] Figure 2 This is a schematic diagram of a cognitive empowerment system based on AI-guided interaction provided by an embodiment of the present invention. Detailed Implementation

[0015] Exemplary embodiments of the present disclosure will now be described in more detail with reference to the accompanying drawings. While exemplary embodiments of the present disclosure are shown in the drawings, it should be understood that the present disclosure may be implemented in various forms and should not be limited to the embodiments set forth herein. Rather, these embodiments are provided so that this disclosure will be thorough and complete, and will fully convey the scope of the disclosure to those skilled in the art.

[0016] like Figure 1 As shown, the present invention provides a cognitive empowerment system based on AI-guided interaction, which comprises three layers in its architecture: a core layer, an application layer, and a defense layer. The overall system architecture will be described below to facilitate understanding of the technical carriers for the execution of steps in subsequent method embodiments.

[0017] The core layer is the core functional layer for the system to perform guided interactions, and mainly includes the following functional components: (1) Intent Resolution and Cognitive State Awareness Component: This component receives the user's question text in natural language and performs two analysis tasks on the user input in the background. The first is intent resolution, which identifies the core problem the user wants to solve from the user input, i.e., the user's core needs. The second is cognitive state awareness, which analyzes the language features of the user input text to determine the type of cognitive state the user is currently in. This component performs the above analysis silently in the background and does not directly output the analysis process or intermediate results to the user.

[0018] (2) Dynamic Guidance Strategy Generation Component: This component dynamically generates guided interactive content based on the user's core needs and current cognitive state output by the intent parsing and state awareness components. The core characteristic of the generated guided interactive content is that it is not a direct answer to the user's question, but rather structured content containing guided questions to guide the user to think independently. This component supports at least two strategy generation methods: one is to select a matching strategy template from a pre-built strategy template library based on the state-scenario mapping relationship and instantiate and populate it; the other is to generate personalized guidance strategies in real time based on the large language model according to the current dialogue context and cognitive state.

[0019] (3) Multi-turn dialogue management component: This component is responsible for maintaining the state of multi-turn interactive sessions with the user. After each round of user response, this component passes the user input of this round to the intent parsing and state awareness component for re-analysis to continuously track changes in the user's cognitive state. When a change in the user's cognitive state is detected, this component notifies the guidance strategy dynamic generation component to adjust the currently adopted guidance strategy based on the updated cognitive state.

[0020] (4) Forced Pause and Return of Sovereignty Component: After the system outputs guiding analysis content, this component performs a forced pause operation, that is, actively stops the current interactive output, discontinuously outputs subsequent content, and instead outputs boundary statements and / or confirmatory questions. The functional goal of this component is to return the initiative of analysis, judgment and final decision-making to the user, and technically ensure the user's dominant position in the decision-making loop.

[0021] The application layer provides standardized external interfaces for the various functions of the core layer. The application layer contains one or more application programming interfaces (APIs). External applications (such as mobile applications, web applications, desktop clients, instant messaging bots, etc.) send user-input natural language question text to the core layer through these interfaces and receive guided interactive content returned by the core layer. The application layer standardizes the format of input and output data, decoupling the core layer from various front-end applications and supporting access from multiple front-end formats.

[0022] The defense layer is a crucial component of the system architecture of this invention. During the normal execution of guided interactions by the system, the defense layer transparently embeds specific semantic patterns into the output guided interaction content. These semantic patterns are imperceptible to end users who are normally using the system and do not affect the functionality of the guided interactions or the user experience. However, for unauthorized third parties who attempt to functionally fit or distill the system, these semantic patterns will cause interference during their fitting or distillation process. Specifically, this manifests as: causing the fitted model to produce outputs deviating from expectations under specific types of inputs, or causing the distilled model to exhibit systematic biases in the cognitive state perception stage.

[0023] The semantic pattern embedding method of the defense layer includes, but is not limited to, at least one of the following: In the natural language expression of guided interactive content, while keeping the semantics unchanged, the sentence structure is transformed according to predetermined rules, so that different system instances with different transformation rules produce differentiated output text, thereby increasing the difficulty for the fitter to obtain consistent training samples; embedding watermark-level features that are invisible to normal users in the output text, which will leave traceable statistical features in the model parameter space when used for model training.

[0024] The defense layer also includes a data security and anti-distillation sublayer, which is used to prevent interception and bulk data collection when transmitting guided interactive content between the system and external applications.

[0025] The following examples, using multiple specific application scenarios, provide a detailed description of the cognitive empowerment method based on AI-guided interaction of the present invention. It should be noted that the cognitive state types and corresponding guidance strategies described in each embodiment are merely exemplary. In actual implementation, cognitive states may include any two or more of the following types: confused, anxious, dependent, conservative, etc., and the corresponding guidance strategies can also be adjusted and expanded according to actual needs.

[0026] Example 1: Team efficiency improvement scenario (confused cognitive state) This embodiment describes how, when a user faces a vague and complex organizational management problem (improving team efficiency), the system helps the user clarify the problem, break down the goals, and advance their thinking through guided interaction. In this embodiment, the user's initial cognitive state is perceived as confused.

[0027] Step S101: Receive user input Users input the following question in natural language text via an application installed on their mobile devices: "I want to improve team efficiency, but I don't know where to start." The application sends this natural language question text to the core layer's intent parsing and state awareness component through an API provided at the application layer. The receiving module encodes, standardizes, and verifies the integrity of the input text before forwarding it to the intent parsing and state awareness component.

[0028] Step S102: Intent parsing and state awareness The intent parsing and state awareness components perform two analysis tasks after receiving user input in the background.

[0029] The first analysis, intent parsing, involves the component performing semantic parsing on the user's input text. Using natural language processing techniques (such as semantic encoding and intent classification based on pre-trained language models), it identifies the core problem the user wants to solve, providing a starting point for improving team efficiency. Simultaneously, the component extracts key entity information: team efficiency is the target domain, and the user's lack of direction indicates a lack of a starting point for problem decomposition and a structured approach. The intent parsing results are output in structured data format, including at least a core requirement description field and a list of key entities.

[0030] The second analysis, cognitive state perception: The component performs multi-dimensional analysis of the linguistic features of the user's input text. Analysis dimensions include, but are not limited to: (a) ambiguity indicators, whether the text contains words indicating uncertainty, such as "don't know," "unclear," "maybe," "probably," etc.; (b) degree of structure, whether the user provides specific dimensions of the problem, constraints, or previously tried solutions; (c) sentence complexity, the length and logical nesting of the user's sentences; and (d) emotional characteristics, determining whether the user's expression contains emotional signals such as anxiety and frustration through sentiment analysis.

[0031] In this embodiment, through comprehensive analysis of the above-mentioned linguistic features, the system determines that the user is currently in a state of confusion. The determination criteria include: the text contains explicit expressions of confusion such as "I don't know where to start," and the ambiguity index is higher than a preset threshold; the user has not provided any specific dimensions or constraints of the problem, and the structuredness index is lower than a preset threshold; the problem statement is brief and lacks hierarchy, and the sentence complexity index is low.

[0032] Step S103: Dynamically generate guided interactive content After receiving the intent parsing results (core requirement: finding the entry point to improve team efficiency) and the cognitive state perception results (confused type), the guidance strategy dynamic generation component performs strategy matching and content generation.

[0033] For users experiencing a state of confusion, the system's pre-set guidance strategy involves structured decomposition and option limitation. The core logic of this strategy is that when users are confused, their cognitive difficulty lies in not knowing where to begin when faced with an overly broad question. Providing a direct answer at this point does not help users develop independent thinking skills. Instead, the system should proactively offer a set of structured options, breaking down the broad question into several specific and actionable sub-directions. By providing users with limited and structured choices, the system helps them focus their attention on the actionable aspects, taking the first step in thinking.

[0034] Based on the above strategy, the system dynamically generates the following guided interactive content (guided questions): Okay, let's clarify our thinking first. You want to improve team efficiency. Which of the following areas do you see as the most significant bottleneck: A. Inefficient task allocation; everyone is busy, but overall output is low; B. High cross-departmental communication costs, frequently leading to disputes and information asymmetry; C. Low team morale, lack of motivation and enthusiasm, and insufficient execution. Let's break down one aspect at a time. Which area do you think best reflects your team's actual situation? The guided interactive content is characterized by the following: First, it does not directly answer how to improve team efficiency; second, it breaks down the broad question into three specific, debatable options, each accompanied by a contextual description to aid user understanding; and third, it ends with a confirmatory question, requiring the user to make a choice before proceeding. This corresponds to the technical features of claim 1, where the guided interactive content includes guided questions to guide user thinking without directly providing answers, and claim 4, where the guided interactive content includes providing structured options.

[0035] Step S104: Continuous state awareness and dynamic policy adjustment in multi-turn interactions After receiving the guided question, the user thinks about it and replies: "It should be A, the task allocation issue is quite prominent." The multi-turn dialogue management component sends this reply text to the intent parsing and state awareness component for a new round of language feature analysis.

[0036] In this round of analysis, the system detected that the user made a clear choice (it should be A), indicating that the user has begun to find a concrete direction of thought from a state of confusion. However, the user used the uncertain modifier "should" and ended with "ba," indicating that their choice is not yet firm and still needs further guidance to solidify their judgment. Based on this, the system updated the user's cognitive state from confused to initial-stage oriented (the degree of confusion has been significantly reduced, but they have not yet entered a deep analysis state). Based on the updated state, the guidance strategy dynamic generation component dynamically adjusted to a targeted in-depth exploration strategy, no longer providing broad options, but instead breaking down the user's chosen task allocation direction at a deeper level. For example, the system further asked: In what specific aspects is the task allocation unreasonable? Is it that the task objectives are unclear? Or is the workload distribution uneven? Or is the task priority ranking chaotic? Step S105: Forced Pause and Boundary Declaration After outputting the analytical content for each round, a forced pause and automatic triggering of the sovereignty handover component occur. Specifically, the system performs the following operations: First, the system actively halts the current interactive output. After providing the guided questions or analysis, it stops continuously outputting subsequent analysis, suggestions, or conclusions, and enters a state of waiting for user response. This forced pause ensures that the system will not continue to push information before the user has completed their independent thinking, technically preserving independent thinking space for the user.

[0037] Second, add a boundary statement at the end of the guided interactive content. For example, after outputting the analysis, the system adds the following statement: "The above analysis is based on deductions from general organizational management logic. It cannot replace your final judgment based on daily observations and in-depth understanding of the team. Each team has its unique personnel composition, business background, and development stage. The above analytical framework provides a method for organizing thoughts, not a standard answer applicable to all teams. Please make independent judgments and decisions based on the actual situation of your team." This boundary statement clearly informs users that the analysis provided by the AI ​​system is based on deductions from general knowledge logic and has inherent boundaries in its analytical scope and accuracy; users need to make their final independent judgments based on their own understanding of the actual situation.

[0038] Step S106: Generate a cognitive progress report After multiple rounds of guided interaction, and after users have broken down and analyzed their team's efficiency issues in multiple layers (for example, users analyzed the task allocation problem in turn and identified three specific areas for improvement), the multi-round dialogue management component detected that the core issues had been fully broken down and triggered the generation of a cognitive progress report.

[0039] An example of a cognitive progress report would be: Through the process we just went through, we successfully broke down the initial broad goal of improving team efficiency into three verifiable and actionable specific tasks: First, optimize the weekly task allocation process and establish a clear task priority ranking mechanism; second, create a visual progress dashboard to make each member's work progress transparent and visible; third, set up weekly 15-minute stand-up meetings to synchronize progress and promptly expose bottlenecks. Compared to the initial state of not knowing where to start, you now have a clear execution path. Next, which direction would you like to begin optimizing specifically? The core elements of a cognitive progress report include: reviewing the initial problem and comparing it with the current progress; summarizing the breakdown results achieved during this interaction; listing specific actionable tasks; and returning the initiative for the next step to the user.

[0040] Example 2: New Product Pricing Scenario (Anxious Cognitive State) This embodiment describes how the system can quickly adjust its guidance strategy to match the user's immediate cognitive state when the user raises a specific business decision question in an anxious mood, and embed boundary declaration and cognitive sovereignty protection mechanisms in the process.

[0041] Step S201: Receive user input and status awareness Users can enter the following natural language question through the desktop web application: Stop talking nonsense. I need to price my new product, so just tell me, what's a suitable price? The intent parsing and state awareness components perform analysis in the background. Intent parsing results: The user's core need is to determine the price of the new product. State awareness results: Through analysis of the user's linguistic features, the system determines that the user is currently in an anxious cognitive state. The criteria for this determination include: the user used the imperative phrase "don't talk nonsense" (meaning to urge the user to stop); used the expression "just tell me" (meaning to tell me what you're asking) to guide the rejection process; the question was brief and direct, with a text length significantly shorter than the average length of normal conversation; and sentiment analysis detected negative emotional signals.

[0042] Step S202: Strategy matching and dynamic generation for anxiety-type adaptation For anxious cognitive states, the corresponding guidance strategy matched by the system is a simplified framework and direct delivery. The core logic of this strategy is that when users are anxious, their cognitive characteristics are characterized by impatience with lengthy guidance processes and a desire to quickly obtain actionable information. At this time, using slow-paced, structured guidance (such as the step-by-step breakdown and option provision method in Example 1) will not only be unacceptable to users but may also exacerbate their anxiety. After recognizing this state, the system skips the warm-up phase of step-by-step guidance and directly delivers a streamlined, actionable framework to the user. Crucially, the system still does not directly provide a specific pricing figure as an answer, because directly providing a specific figure would deprive the user of decision-making power, and such a figure may be significantly inaccurate due to a lack of specific business information about the user.

[0043] The system dynamically generates the following guided interactive content: I understand. No beating around the bush.

[0044] There is no single, universally applicable price for new products, but there are three framework methods you can use directly: Method 1: Cost-plus pricing. The calculation is: (total fixed costs divided by projected sales volume, plus variable costs per unit), then multiplied by (1 plus target profit margin). If your cost data is clear, use this method to calculate a minimum guaranteed price.

[0045] Method two: Competitive benchmarking. Identify three comparable products in the market that are similar to your product's positioning and obtain their market prices. Select the median price, or the price value at the 70th percentile of the price distribution. Newcomers to the market generally start with a price slightly lower than the benchmark price.

[0046] Method Three: Value Perception Method. Ask your first batch of target customers the following questions: Approximately how many hours of repetitive work does this product save you each day? And approximately how much extra revenue does it generate for you each month? Then follow up with: What percentage of this additional revenue would you be willing to share with us? The percentage your customer is willing to share will be your value anchor when setting your price.

[0047] To give you a more intuitive reference, let's assume that your competitors' prices are roughly between 100 and 200 yuan. Then, the suggested price ranges for the three methods are approximately: 99 yuan for the rapid sales volume strategy; 149 yuan for the stable pricing strategy; and 199 yuan and above for the brand premium strategy.

[0048] The characteristics of this guided interactive content are: (1) it maintains the essence of guidance, providing three methodological frameworks and multiple price range options, rather than a definite numerical answer; (2) it adapts to anxious states, removing the elaboration and transition and directly providing actionable content; (3) it preserves the user's decision-making space, providing multiple strategies and ranges, with the final choice still in the user's hands.

[0049] Step S203: Forced Pause, Boundary Declaration, and Confirmatory Questions After the above analytical content is output, the forced pause and sovereignty handover components execute a forced pause, the system actively stops outputting content, and does not continuously push further content, waiting for the user to process and respond. Boundary statements and confirmatory questions are also output simultaneously. Boundary Statement: The above is a calculation framework based on general business logic and pricing theory. It cannot replace your final pricing decision based on your product's actual cost structure, the purchasing power of your target customer group, and the competitive landscape of your market segment. Every product and every market has its unique variables; the framework above provides a framework, not a standard answer. After setting a price, it is recommended that you conduct test sales on a small scale first and adjust accordingly based on actual market feedback.

[0050] Confirmation question: Is this direction helpful to you? Or do you need me to further refine the calculation parameters for a specific pricing method? This confirmatory question corresponds to the technical feature used to require users to explicitly confirm their understanding or choices. Users need to explicitly confirm whether the system's guidance is helpful, or clearly indicate areas requiring further clarification.

[0051] Example 3: Career Development Decision-Making Scenario (Forced Independent Thinking Mode Switching in a Dependent Cognitive State) This embodiment highlights the core technical feature described in claim 3: when the system senses that the user is in a dependent cognitive state, it automatically switches to a forced independent thinking mode, refuses to provide answers directly, and prompts the user to complete the thinking independently through specific guiding questions.

[0052] Step S301: Triggering the perception of scene background and dependent state After the first two rounds of guided interaction (in which the system helped users outline possible career development paths: a technical expert path and a managerial path, and analyzed the required skill set and typical development trajectory for each path), in the third round of interaction, the user entered the following response: So, do you think I should choose technical or management? Which has a better future? The intent parsing and state awareness components performed linguistic feature analysis on the user's input in this round. The analysis revealed the following characteristics: (a) The user used "you" instead of "I" as the judgment subject, implying that "you think I should choose," indicating that the user was outsourcing the decision-making responsibility to the AI ​​system, shifting the decision-making subject from the user to the AI; (b) The user directly asked the system to choose between two options, specifying which was more promising, rather than expressing their own inclination or analytical logic; (c) The user had already obtained sufficient structured information about the two paths in the first two rounds, possessing the information basis to make an independent judgment, but still requested the AI ​​to make the decision on their behalf. Based on the comprehensive analysis of the above linguistic features, the system determined that the user's cognitive state had shifted from the previous initial orientation type to a dependent type.

[0053] Step S302: Automatically switch to forced independent thinking mode When the multi-turn dialogue management component detects a cognitive state update to dependency, it immediately notifies the guidance strategy dynamic generation component. The guidance strategy dynamic generation component then performs a strategy switching operation, automatically switching from the current regular guidance strategy to a forced independent thinking mode. The core rules of this mode are: the system refuses to directly provide answers or make alternative choices; the system no longer provides direct answer-oriented outputs such as "A or B, which to choose"; instead, the system generates guiding questions that encourage independent thinking, forcibly returning the thinking task to the user.

[0054] Specifically, in this mode, the system generates the following questions that force independent thinking: I notice you want me to help you make this choice. However, the answer to which path is more promising largely depends on your own judgment regarding some key issues. As a technical system, I can help you analyze information and outline the framework, but I cannot make this decision for you. I would like you to answer the following three questions independently first: First, in the past year, what moment made you feel the most accomplished? Was it solving a complex technical problem or leading the team to achieve a project goal? Second, do you enjoy delving into a problem on your own until you completely solve it, or do you enjoy communicating and coordinating with others to get things done? Third, what are your expectations for your work status five years from now? Do you want to become a recognized technical authority in a certain field, or a manager who can coordinate the overall situation? Please think carefully about these three questions first, and write down your answers. Then we can analyze together where your answers lead.

[0055] The components of this forced independent thinking guided questioning include: (a) explicitly refusing to make decisions for the user, stating that the system cannot make this decision for you; (b) providing the user with a thinking task to complete independently, consisting of three specific introspective questions that require the user to draw on their own experience and feelings; (c) requiring the user to produce independent thinking results before continuing to participate in the system, asking them to think carefully and write down their own answers; and (d) redefining the system's role as an facilitator for information analysis and organization, rather than a decision-maker.

[0056] Step S303: Force suspend execution After outputting the aforementioned questions guiding independent thinking, the forced pause and sovereignty handover components execute a forced pause operation. The system proactively stops outputting information, waiting for the user to independently complete the reflection and answers to the three introspective questions. This pause is not a technical wait (such as network latency or processing delay), but a deliberately designed cognitive pause. Its technical goal is to forcibly leave a time window for the user to independently think and produce judgments within the interaction process, ensuring that the user does not skip the independent judgment stage and directly rely on AI output.

[0057] It should be noted that the specific number and content of the three introspective questions mentioned above are merely examples. In actual implementation, the system can dynamically generate different forms and numbers of tasks to promote independent thinking based on the current dialogue topic and the user's specific context. Their common characteristic is that these tasks all require users to produce content based on their own experience, feelings, and judgments, rather than simply choosing from options provided by AI.

[0058] Step S304: Extended explanation of dependency triggering conditions The triggering conditions for the forced independent thinking mode are not limited to the scenario in the above embodiments where the user directly requests the AI ​​to make decisions on their behalf. In actual implementation, the forced independent thinking mode can be triggered when the system detects any one or more of the following signals: (1) The user did not express an independent opinion or judgment in multiple rounds of interaction, but only repeatedly asked the AI ​​for its opinion or asked the AI ​​to give a choice; (2) Users directly accept the options provided by AI in multiple rounds without making any independent screening, questioning or adjustment; (3) Users repeatedly use language patterns such as "you think you're helping me" or "you're doing it for me" in multiple rounds to shift the decision-making subject from the first person to the second person; (4) Even after the system has provided sufficient information and an analytical framework, the user still says, "I am still not sure, just tell me directly," instead of trying to make an independent judgment.

[0059] The above extended examples of triggering conditions illustrate that the technical solution of automatically switching from a dependent cognitive state to a forced independent thinking mode is not limited to a specific triggering method. Any implementation that switches strategies after detecting a tendency to shift decision-making responsibility based on user language characteristics falls within the protection scope of this invention.

[0060] Example 4: Technical Solution Evaluation Scenario (Conservative Cognitive State) This embodiment demonstrates how the system adjusts its guidance strategy to adapt to the cognitive characteristics of a user in a conservative cognitive state. Conservative users exhibit a cautious attitude towards new technologies or solutions, tend to over-collect information, and find it difficult to make decisions, resulting in a state of analytical paralysis.

[0061] Step S401: Perception of Conservative State User input: Our company wants to implement an automated testing system, but I've heard that many companies haven't seen good results after implementing it. I need to do some more research. Could you please summarize the advantages and disadvantages of automated testing so I can make a decision after gaining a comprehensive understanding?

[0062] The intent parsing and state awareness components perform linguistic feature analysis on user input. State awareness result: Determined as a conservative cognitive state. Judgment criteria include: the user mentioned hearing that many companies used it with poor results, expressing a preconceived bias towards the new solution and showing a clear preference for negative information; the user stated they needed to fully understand the information before making a decision, demonstrating an excessive pursuit of information completeness; the user actively requested a summary of advantages and disadvantages rather than seeking decision-making assistance, indicating a cognitive mode of collecting more information to postpone decision-making. Overall, the risk assessment vocabulary density, decision-postponement willingness index, and information gathering need index all exceeded preset thresholds.

[0063] Step S402: Conservative Adaptation Guidance Strategy For users exhibiting a conservative mindset, the system employs a progressive framework and minimum viable capability (MVP) guidance strategy. This strategy avoids directly persuading users to accept new solutions (which would trigger defensiveness and reinforce their conservative tendencies). Instead, it first provides a structured information framework to meet their information gathering needs, then guides them to design a minimum-cost, minimum-scope validation plan, transforming the major decision of whether to fully implement a new solution into a smaller decision of whether to try it on a very small scale. The underlying logic of this strategy is that the decision-making obstacle for conservative users lies in the perceived excessive risk. By narrowing the scope and cost of the decision, the perceived risk is reduced, thus helping users take the first step in making a decision.

[0064] The system-generated guided interactive content includes: (a) a structured information analysis framework that presents the conditions under which automated testing works well and under what conditions it does not, rather than simply listing advantages and disadvantages; (b) narrowing the user's focus from whether to implement it throughout the company to whether to try it on a small, independent module first; and (c) providing specific steps for minimum viable verification.

[0065] At the end of the content, the system performs a forced pause and outputs a boundary declaration: The above analysis is based on general observations and summaries of automated testing practices introduced in multiple industries. It cannot replace your own assessment based on your company's current code quality, team skill structure, project delivery cycle, and business quality requirements. It is recommended that you first conduct a two-week pilot verification on an independent functional module with no more than 50,000 lines of code and no more than 500 test cases. By comparing the defect detection rate and regression testing time before and after the pilot, you can obtain a basis for decision-making based on your own business environment.

[0066] Step S403: State Transition Tracking After the user responds to the above guidance, the system re-analyzes the language features based on the response to determine whether the user's cognitive state has shifted from conservative to other states. For example, if the user's response includes expressions such as "I can try it first, so I can test it on one module for two weeks," the system will detect a significant increase in the user's decision-making intention and a significant decrease in the density of risk assessment vocabulary. This will update the cognitive state from conservative to initial-stage oriented, and accordingly adjust the guidance strategy from a gradual framework to targeted in-depth exploration, guiding the user to plan more thoroughly in the direction of pilot testing.

[0067] This process embodies the technical feature of re-analyzing language features to update cognitive state based on the user's responses in multiple rounds of interaction: the update of state is not based on a single judgment in a certain round, but is a dynamic tracking that is continuously carried out throughout the entire multi-round interaction process.

[0068] Example 5: Health Management Scenario (Dynamic Switching Between Multiple States) This embodiment demonstrates how, within a single interactive session, a user's cognitive state may undergo multiple changes, and how the system continuously tracks these changes and dynamically switches the guidance strategy. This embodiment illustrates that the guidance strategy of this invention is not statically preset, but rather an adaptive system that dynamically adjusts in real time according to changes in the user's cognitive state during the dialogue.

[0069] Step S501: First Round, Confused Perception and Structured Deconstruction User's initial input: I've been feeling very tired lately, and I can't sleep well at night. I'm wondering if there's something wrong with my health. Status perception result: Confused type: The user expressed vague health concerns; the question was broad and lacked focus, containing confusing markers such as "I don't know." The system generated guided questions, breaking down the vague health concerns into analyzable dimensions such as sleep, diet, exercise, and psychological stress, helping the user focus their attention.

[0070] Step S502: Second round, state transitions to initial orientation. User's second response (after receiving the breakdown framework): Sleep is indeed a big problem. I need one or two hours to fall asleep every day, and I wake up very early in the morning. Status awareness updated, shifting from confused to initially targeted: The user has focused on the specific dimension of sleep and provided a detailed description of their symptoms (difficulty falling asleep + early awakening), indicating a preliminary ability to pinpoint the problem. The system switches to a targeted, in-depth analysis strategy to further guide the user to refine the specific patterns of their sleep problems.

[0071] Step S503: Third round, switching between dependency detection and forced independent thinking User's third response (after in-depth analysis): Doctor, just tell me what medicine to take, I don't want to analyze anymore. Status awareness update, dependency-based: The system detected multiple dependency signals from the user, referring to the system as a doctor (role misalignment, treating the AI ​​as a professional medical person), directly requesting a medication plan (outsourcing professional medical decision-making responsibility to the AI), and expressing reluctance to analyze anymore (explicitly refusing to continue independent thinking).

[0072] The system immediately switches to a forced independent thinking mode: First, it refuses to provide medication advice. As an AI system, I am not a medical professional and cannot provide medication advice. Second, it outputs a boundary statement: The following analysis is based solely on general sleep health knowledge and does not constitute medical advice. Sleep disorders have complex and diverse causes, potentially involving physiological, psychological, and environmental factors, requiring diagnosis by a professional doctor through a face-to-face consultation and necessary examinations. Third, it generates guiding questions to encourage user self-directed action. It suggests the following: Record a sleep log for one week, including your bedtime, approximate time to fall asleep, number of times you wake up during the night, time you wake up in the morning, and your daytime energy level score. Take this sleep log to consult a professional doctor. Now, can you tell me what time you usually go to bed at night and what time you need to get up in the morning? Let's start by gathering this basic information.

[0073] This embodiment demonstrates the three-stage evolution of cognitive state from confused to initially oriented to dependent during a single dialogue, as well as the complete process by which the system dynamically adjusts its guidance strategy based on real-time perception results at each stage.

[0074] The cognitive state classification system involved in this invention forms the basis for guiding the dynamic matching of strategies. The definitions, judgment criteria, and strategy mapping relationships of at least four cognitive states are described in detail below. It should be noted that the following four states are merely illustrative examples. In actual implementation, cognitive states may include any two or more of these four states, and other types of cognitive states (such as confused, confident, avoidant, curious, and skeptical states) may be added as needed for specific application scenarios. These extensions all fall within the scope of protection of this invention.

[0075] Confused type: Definition: When faced with a complex or broad problem, users lack a clear direction for thinking and the ability to break down the problem, exhibiting a cognitive state of not knowing where to begin.

[0076] Language characteristics: The input text frequently contains words indicating uncertainty (such as not knowing, not clear, clueless, and not knowing where to start); the problem is described broadly and lacks structure; the user does not provide specific dimensions of the problem, boundary conditions, or solutions already tried; the sentences are short but semantically ambiguous.

[0077] Judgment indicator combination: The ambiguity indicator is higher than the preset threshold, the structuring indicator is lower than the preset threshold, and the sentence contains confusing keywords. Corresponding guidance strategy: Structured decomposition and option limitation. The system proactively breaks down broad questions into several specific and actionable sub-directions, presenting them to the user in the form of structured options, lowering the threshold for starting thinking.

[0078] Anxious type: Definition: Users are under time pressure or are impatient, hoping to get results quickly and lacking patience for lengthy onboarding processes.

[0079] Language characteristics: The length of the input text is significantly shorter than the average length of normal conversation; imperative sentences or urging expressions are frequently used; phrases such as "Stop talking nonsense and just tell me in one sentence" are used; the system refuses or skips the structured guidance provided by the system; sentiment analysis shows that the intensity of negative emotions is higher than the preset threshold.

[0080] Judgment indicator combination: The text length is below a preset threshold, and sentiment analysis shows that the intensity of negative emotions or urgency is higher than a preset threshold, and the sentence contains keywords that urge or reject guidance. Corresponding guidance strategy: Simplified framework + direct delivery. The system skips step-by-step guidance and directly delivers a prepared actionable framework to the user, but still maintains the essence of guidance (not giving direct definitive answers), and adds boundary statements and confirmatory questions at the end.

[0081] Dependency type: Definition: Even when users have sufficient information and analytical frameworks, they still tend to delegate decision-making responsibility to AI systems, demonstrating a tendency to let AI make decisions for them.

[0082] Language characteristics: Using expressions such as "you think," "you help me choose," "you decide for me," and "you tell me what to do," shifting the decision-making subject from the first person (I) to the second person (you); directly requesting the system to make choices or judgments after the system provides an analytical framework, rather than expressing one's own inclinations and reasons for judgments; consistently failing to express independent opinions in multiple rounds of interaction, habitually accepting options provided by AI without making independent selections.

[0083] Judgment Indicator Combination: The "decision subject shift" indicator is true (detecting that the subject of the judgment / decision action is in the second person and the predicate involves decision-making or choice-related verbs), and the user has not produced independent judgment content in a consecutive preset number of rounds (e.g., 2 or 3 consecutive rounds) of interaction (i.e., the user's responses mainly consist of accepting the AI's options, requesting the AI ​​to make choices for them, or continuing to ask the AI ​​for its opinion). Corresponding Guidance Strategy: Forced Independent Thinking Mode. The system refuses to directly provide answers or alternative judgments, generates guiding questions that encourage users to produce independent thinking content, and requires users to complete an independent thinking task before the system continues to participate.

[0084] Conservative type: Definition: Users are cautious or even resistant to new solutions, methods or changes, and tend to over-collect information to maintain the status quo, which manifests as a cognitive pattern of excessive information gathering but insufficient decision-making and action.

[0085] Language characteristics: Overemphasis on negative information and risks, selectively ignoring or giving low weight to positive information; expressing a strong need for comprehensive understanding, thorough research and careful consideration; tending to collect more information rather than make decisions; sentence structure containing risk-sensitive expressions such as worry about uncertainty in case of failure.

[0086] Judgment indicator combination: Risk assessment vocabulary density exceeds a preset threshold, and the frequency of postponement intention indicators (such as the frequency of postponement expressions like "further investigation," "reconsideration," and "comprehensive understanding") exceeds a preset threshold, as does the information gathering demand indicator. Corresponding guidance strategy: Progressive framework + minimum feasible verification guidance. The system does not attempt to persuade users, but first provides a structured information framework to meet their information needs, and then guides users to break down large decisions into verification steps with the lowest cost and smallest scope, reducing the perceived risk of decision-making.

[0087] Technical implementation methods of state awareness: The perception of cognitive states can be achieved through various technical means or combinations thereof. The following describes several feasible implementation methods (which can be used individually or in combination), and these methods are provided as examples: Method 1, rule-based multidimensional feature scoring: Key linguistic features and corresponding feature extraction rules for each cognitive state are predefined (e.g., keyword matching rules, syntactic pattern matching rules, text length threshold rules, sentiment polarity threshold rules, etc.). Multidimensional feature vectors are extracted from the user input text. The matching degree between these feature vectors and the feature templates of each cognitive state is calculated (e.g., using weighted scoring or similarity metrics). The state with the highest matching degree is selected as the perception result. The advantages of this method are: transparent decision logic, strong interpretability, and no need for training data.

[0088] Method 2, based on a machine learning classification model: A text classification model is trained using a training dataset labeled with cognitive states. The model takes user-input text as input and outputs the probability distribution of each cognitive state. The state with the highest probability value is selected as the perceived result. Training data can come from manually annotated historical dialogue data, where annotators label each dialogue text with state tags based on the definition of cognitive states and linguistic feature descriptions. The model can be fine-tuned using a pre-trained language model based on the Transformer architecture, or it can be implemented using traditional machine learning classifiers (such as support vector machines, random forests, etc.) combined with text feature engineering. The advantage of this method is that it can capture complex language patterns that are difficult to cover with rule-based methods, and the classification accuracy can be continuously optimized as the amount of labeled data increases.

[0089] Method 3, Few-Shot Inference Based on a Large Language Model: The definition of cognitive states, linguistic feature descriptions, and several labeled examples are provided as prompts to a large language model. This model then infers and outputs the user's cognitive state type based on the prompts and the current user input text. The prompts may include a textual description of each cognitive state, a list of typical linguistic features, and 1 to 3 dialogue examples labeled with the state. The advantages of this method are: it allows for flexible expansion to new cognitive state types without retraining the classification model, requiring only updates to the state descriptions in the prompts; and it leverages the deep understanding of context by the large language model to handle complex situations with ambiguous boundaries.

[0090] It should be noted that the above three methods are merely illustrative examples. Those skilled in the art can choose other technical solutions that can achieve cognitive state perception functions based on actual deployment conditions (such as computing resources, availability of labeled data, and requirements for real-time state perception). These choices do not affect the scope of protection of the technical solutions of this invention.

[0091] In this invention, the generation of the guidance strategy supports at least two implementation methods, which are described in detail below.

[0092] Method 1: Matching and instantiation based on a preset strategy template library In this approach, the system pre-builds a strategy template library. The strategy template library contains multiple strategy template records. Each strategy template record contains at least the following fields: (a) the applicable cognitive state type, where a strategy template can be applied to one or more cognitive state types; (b) a core logical description of the strategy, using structured fields to define the strategy's goal (the guidance effect the strategy aims to achieve), means (the guidance method adopted by the strategy), and constraints (the rules the strategy should follow, such as not providing answers directly and requiring confirmation questions); and (c) a guidance content generation template, containing populated variable slots, which are represented by specific markers in the template, such as {user core needs}, {decomposed option list}, {domain keywords}, {structured framework content}, etc.

[0093] Once the guidance strategy dynamic generation component receives the cognitive state perception result and intent parsing result, it uses the cognitive state type as the query key to retrieve a matching strategy template from the strategy template library. After a successful retrieval, it fills the corresponding variable slots in the strategy template with key information from the intent parsing result (user core requirement description, related entity list, domain context information, etc.) to generate instantiated guided interactive content.

[0094] Taking Example 1 as an example: For users experiencing cognitive confusion, a structured strategy template with option limitations is matched. The variable slot {domain keywords} in this template is filled with team efficiency, and the {decomposition option list} is automatically generated by the system based on key entities and the domain knowledge base from the intent parsing results, resulting in three options (unreasonable task allocation, high cross-departmental communication costs, and low team morale). This instantiates personalized, guided questioning content tailored to the current user and the current problem.

[0095] Method 2: Real-time reasoning generation based on large language models In this approach, the system does not rely on a pre-defined strategy template library. Instead, it leverages the contextual understanding and reasoning capabilities of a large language model to generate personalized guidance strategies and content in real time based on the complete context of the current dialogue and the perceived cognitive state.

[0096] The specific implementation method is as follows: The system assembles the following information into a structured reasoning prompt: (a) the complete history of the current dialogue, including the user's input text in each round and the system's output text in each round, arranged in chronological order; (b) the currently perceived user cognitive state type and a detailed description of the state (including the language features, cognitive characteristics and recommended guidance strategy direction of the state); (c) the generation constraints of the guidance strategy, including: not providing answers directly, including guiding questions, attaching boundary statements after outputting analytical content, and including confirmatory questions at the end; (d) the user input text that needs to be responded to.

[0097] The reasoning prompt is submitted to the large language model, which then uses the contextual information and constraints in the prompt to generate personalized, guided interactive content that fits the current dialogue context. The generated content must satisfy all rules in the constraints.

[0098] The advantages of this approach are: First, strategy generation is not limited by the number and coverage of preset templates, and can handle complex or marginal situations not covered in the template library; Second, the guidance content can be more closely integrated with the specific context of the current dialogue and the user's personalized characteristics (such as the user's personal preferences, knowledge background, decision-making style, etc. revealed in the previous rounds); Third, the system's guidance strategy can gain new guidance capabilities and better guidance quality by updating the underlying large language model without modifying the code or strategy template library.

[0099] The two methods can be used in combination: In practice, the system can deploy both of the above strategy generation methods simultaneously and select or combine them according to the specific situation: (a) For regular interactions with clear cognitive states and typical scenarios, the first method (template matching and instantiation) is used first to obtain faster response speed and more controllable and stable output quality; (b) For situations not fully covered by the template library (such as when the confidence of template matching is lower than the preset threshold) or situations where the user context has highly personalized characteristics, the system switches to the second method (real-time reasoning of large language models) to obtain more flexible and adaptable guidance capabilities.

[0100] The switching between the two methods can be automatically controlled by matching a confidence threshold: when the overall matching confidence of the strategy template in Method 1 with the current cognitive state and intent parsing results is lower than a preset threshold (e.g., 0.7), the system automatically switches to Method 2 for strategy generation. Alternatively, the system can adopt a hybrid mode, using Method 1 to generate the basic framework of the guidance strategy, and then using Method 2 to adapt the framework to context and refine its language, thus balancing efficiency and quality.

[0101] In this invention, the forced suspension and cognitive sovereignty protection mechanism are key components for achieving the core technical goal of returning the final decision-making power to the user and ensuring that the user completes independent thinking. The following details the various implementation methods and execution modes of this mechanism.

[0102] Forced pause execution mode: The forced pause mechanism supports at least three execution modes, each of which can be flexibly selected or combined according to the current stage and content type of the conversation.

[0103] Mode 1, Pause After Analysis: After the system outputs analytical content (such as a breakdown of a problem, a comparison of multiple solutions, a deduction of causal relationships, or an outline of the information framework), a forced pause and a handover of control component trigger a pause. The system does not continue to output subsequent analysis, suggestions, or conclusions. Instead, it first outputs boundary statements and / or confirmatory questions, then actively stops outputting and enters a state awaiting user response. The purpose of this mode is to prevent the system from continuously outputting too much information, causing users to be exposed to subsequent content before they have digested the current analysis and completed independent thinking, thus passively following the system's train of thought rather than actively thinking.

[0104] Mode Two, Pause at Key Decision Points: During guided interaction, the system identifies key decision points in the dialogue (e.g., users facing A / B choices, needing to confirm their understanding of the problem, deciding on the next analytical or action direction, or the system preparing to provide a framework conclusion requiring independent user judgment). At these points, the system proactively pauses, outputting confirmatory questions and requiring the user to provide clear confirmation or a choice before proceeding to the next stage. The purpose of this mode is to ensure that users make independent judgments and choices at each key decision point, rather than passively following the system's guidance.

[0105] Mode 3, Dependency Detection-Triggered Forced Pause: As described in Example 3 above, when the system's state awareness component detects that the user exhibits characteristics of a dependent cognitive state, a forced pause is triggered, and the system switches to a forced independent thinking mode. Unlike the previous two modes, this pause mode is dynamically driven by the cognitive state perception results, rather than based on preset content types or dialogue nodes. The timing of the pause depends on the changes in the user's cognitive state perceived by the system in real time.

[0106] How to implement boundary declarations: Boundary declaration is a technical mechanism by which the system explicitly informs the user of the limitations of its AI capabilities after outputting analytical content. The content and form of the boundary declaration can be dynamically adapted based on the current dialogue topic and the nature of the output content, rather than using a fixed template. The following are some typical forms of boundary declarations (all are examples; in actual implementation, they can be adapted and expanded according to the specific dialogue topic): Form 1, General Logical Boundary Declaration: Applicable to scenarios where the system output is based on general knowledge and logical deduction. Example: The above analysis is based on deduction from general logic and publicly available information in the relevant field. It provides a general and referential analytical framework and cannot cover all special variables and constraints that may exist in your specific environment. Please make your own independent final judgment and decision based on your actual situation.

[0107] Form Two, Data Timeliness Boundary Declaration: Applicable to scenarios where the content output by the system may be affected by data time factors. Example: The industry data and market information cited in the above analysis reflect the publicly available information as of a specific point in time. The market environment and industry landscape are constantly changing. If you are concerned about the latest developments and trends, it is recommended that you further consult recent industry research reports and authoritative data sources to obtain the latest information for comprehensive judgment.

[0108] Form Three, Declaration of Professional Boundaries: Applicable when the system outputs content involving fields requiring specific professional qualifications. Example: The above analysis is a general overview and framework construction from a general knowledge perspective and does not constitute professional field-specific advice (e.g., legal opinions, medical diagnoses, financial advisory advice, etc.). If you require legally binding professional judgment, it is recommended that you consult a professional holding the corresponding professional qualifications.

[0109] Form Four, Application Boundary Declaration: Applicable to scenarios where the system provides users with multiple alternative solutions or strategic frameworks. Example: The alternative directions and analytical frameworks provided above are general summaries based on the information you described. The actual applicability and effectiveness of different solutions will depend on the specific conditions of your environment, resource constraints, and priority ranking. The above solutions are not standard answers, but rather thinking tools for your reference and comparison during the decision-making process.

[0110] How to implement confirmatory questions: Confirmatory questions are one of the key pieces of information the system outputs to the user after a forced pause. The core function of confirmatory questions is to require the user to explicitly confirm their understanding of the system or their own choices, technically ensuring that the user has engaged in independent thought rather than passively accepting information. Confirmatory questions can include at least one of the following types (listed as examples): Type 1, Understanding Confirmation: This type requires the user to confirm whether the system's understanding of their question is accurate. Example: This is my understanding of the question you raised. Is this understanding accurate? Is there anything that needs to be added or corrected? Type Two, Confirmation-Based: This type requires users to make a clear choice among multiple analytical directions or frameworks provided by the system. Example: Among the analytical directions mentioned above, which one would you prefer to begin exploring in depth? Type 3, Self-Assessment and Confirmation: This type requires users to express their independent preliminary judgment based on the analytical framework provided by the system. Example: Based on the above analytical framework, which option do you currently prefer? Please briefly explain the main reasons for your preference.

[0111] Type 4, Action Confirmation Type: This type requires users to specify the concrete action they will take next. Example: Based on your analysis and judgment, what is the first specific action you plan to take? Control of forced pause duration: Forced pause, as implemented by the system, involves the system proactively ceasing to output new content to the user after providing guided or analytical questions, boundary statements, and / or confirmatory questions, and entering a state of waiting for the user's response. The actual pause duration depends on the time required for the user to complete independent thinking and respond; the system does not impose a preset limit on this duration, allowing users anywhere from a few seconds to several tens of minutes to reflect before replying. If the user has not sent a new message within the preset time threshold (e.g., 30 minutes), the system may proactively send a gentle reminder message, such as, "Are you still thinking about the question? If you need any help or a different approach, please let me know." This reminder message is designed not to impose time pressure, but merely to confirm the continuity of the conversation.

[0112] After detecting that the core problem has been fully decomposed and analyzed, this invention generates a cognitive progress report. The cognitive progress report is a summary document of the system's progress in the current guided interactive session. It includes at least some of the following elements: (1) Problem Review Elements: Briefly restate the original problem or core demand raised by the user at the beginning of the conversation in one or two sentences, and show the initial cognitive state as a benchmark for subsequent progress. For example: The problem you raised at the beginning of this conversation is: I want to improve team efficiency but I don't know where to start.

[0113] (2) Summary elements of the process: Present the cognitive progress achieved through multiple rounds of guided interaction in this session in a structured manner. Specific forms include: which specific sub-problems were broken down from the original broad problem; the current analysis progress of each sub-problem (preliminary analysis completed / key factors identified / solution direction formed, etc.).

[0114] (3) Elements of the executable task list: Transform the results of the breakdown and analysis into several specific, executable task descriptions. Each task should include: a prompt of the executing entity (the user or a specific role in their team), a description of the core actions of the task, and verifiable completion criteria. For example: Task 1: Compile the weekly work task list of each member of the current team, calculate the workload of each member, and the completion criteria: produce a table containing all members and the corresponding number of tasks.

[0115] (4) Next Step Prompt Elements: Using an open-ended question format, the decision-making power for subsequent actions is explicitly returned to the user. The user can decide which task to prioritize, which direction to analyze in depth, or whether the current analysis path needs adjustment. For example: The above are the interim results of this analysis. Next, which task would you like to start with? The cognitive progress report's structure follows a four-stage progressive logic: review, summary, outlook, and handover. Its technical effect is twofold: it allows users to intuitively see the complete process of their cognition evolving from an initial vague and broad state to a clear and actionable state through structured guidance from the system, thereby enhancing users' confidence in their independent thinking and decision-making abilities; simultaneously, through the design of next-step action prompts, after completing a stage of assistance, the system explicitly returns the choice and decision-making power of subsequent actions to the user, ensuring that the system always plays the role of a facilitator rather than a substitute.

[0116] like Figure 2 As shown, this embodiment describes a cognitive empowerment system based on AI-guided interaction. This system is used to execute the cognitive empowerment method based on AI-guided interaction described in any of the above method embodiments. The composition, function, and connection relationship of each module of the system are described in detail below.

[0117] The receiving module is used to receive natural language input from the user. Physically, the receiving module can be deployed on the server side and run as a network service (such as an HTTP / WebSocket interface service). This service continuously listens for user message requests from various front-end applications (including but not limited to: mobile terminal applications, web applications, desktop client programs, and bot programs connected to instant messaging platforms).

[0118] The receiving module performs preprocessing operations on the received raw messages, including at least: text encoding standardization (converting all inputs to the standard encoding format used internally by the system); filtering and removing invalid characters or formatting marks; and message content integrity verification (ensuring that the received message content is complete and not truncated or damaged). After preprocessing, the receiving module sends the formatted natural language question text to the input of the intent parsing and state awareness module.

[0119] In addition, the receiving module is also responsible for establishing and managing user sessions, assigning a globally unique session identifier to each independent user interaction session. This session identifier is passed between modules of the system with each round of user messages and system responses to support context association and state tracking during multi-round interactions.

[0120] The intent parsing and state awareness module connects its input to the output of the receiving module, receiving preprocessed user natural language question text. This module contains two functional sub-components: an intent parsing sub-component and a state awareness sub-component.

[0121] The function of the intent parsing subcomponent is to perform semantic analysis on the received user input text to identify the core problem (i.e., core need) that the user wishes to solve, as well as relevant key entity information. Intent parsing can be implemented based on a pre-trained semantic understanding model, such as an encoder-decoder model based on the Transformer architecture, which performs intent classification and multi-label entity recognition tasks after semantically encoding the input text. The output of the intent parsing subcomponent is a structured data object, which at least contains a core need description field and a key entity list field.

[0122] The state-aware subcomponent performs multi-dimensional linguistic feature analysis on user input text to determine the user's current cognitive state. This subcomponent maintains a set of feature extractors, each responsible for extracting a specific dimension of linguistic features, including but not limited to: text length features, sentiment polarity features, uncertainty word density features, sentence structure complexity features, decision-making subject externalization features, and risk assessment word density features. The extracted multi-dimensional feature vectors are input into the state classifier, which outputs the cognitive state type determination.

[0123] The state classifier can be implemented using any combination of one or more of the methods described above: a rule-based system based on multi-dimensional feature weighted scoring, a machine learning classification model trained on labeled training data, or a judgment system based on few-shot reasoning using a large language model. The output of the intent parsing and state awareness module simultaneously outputs the intent parsing result and the state awareness result, and sends them to the input of the guidance strategy dynamic generation module.

[0124] The guidance strategy dynamic generation module connects its input to the output of the intent parsing and state awareness module, receiving intent parsing results (user's core needs and key entity information) and cognitive state awareness results (current cognitive state type). Based on the received cognitive state type, this module dynamically generates guided interactive content. The core characteristic of the generated guided interactive content is that it is not a direct answer to the user's question or a definitive conclusion, but rather structured interactive content containing guided questions to guide the user to think independently.

[0125] This module internally supports at least two strategy generation methods: (a) retrieving strategy templates based on a preset strategy template library, using the cognitive state type as the query key, and filling the variable slots in the templates with key information from the intent parsing results to generate instantiated guided interactive content; (b) assembling the current dialogue history, cognitive state perception results, and strategy generation constraints into reasoning prompts based on real-time reasoning from a large language model, and submitting them to the large language model for reasoning to generate personalized guided interactive content. This module can automatically select or combine the two methods based on the matching confidence level.

[0126] The output of the guidance strategy dynamic generation module is connected to the input of the multi-turn dialogue management module. The guided interaction content (including guided questions) generated by this module is sent to the multi-turn dialogue management module for further processing.

[0127] The multi-turn dialogue management module connects its input end to the output end of the guidance strategy dynamic generation module, receiving the guided interaction content generated by that module. The multi-turn dialogue management module is responsible for maintaining the complete multi-turn interaction session context with the user and continuously tracking changes in the user's cognitive state throughout the entire session.

[0128] Internally, this module maintains a session state storage structure (e.g., a key-value store with the session identifier as the key), which stores the following information for the current session: user input text records for each round of interaction, system output text records, and the cognitive state determination results obtained from the intent parsing and state awareness module for each round. After each round of user responses, this module transmits the user input text for that round to the intent parsing and state awareness module, triggering a new round of intent parsing and state awareness analysis.

[0129] This module compares the updated cognitive state obtained from the current analysis with the cognitive state recorded in the previous round. If they differ (i.e., the cognitive state has changed), the module performs the following actions: It sends a notification message to the guidance strategy dynamic generation module, informing it that the cognitive state has changed from type X to type Y, along with the user input text for this round and the updated intent parsing results. Based on this notification, the guidance strategy dynamic generation module re-executes strategy matching and content generation based on the changed cognitive state, thereby achieving dynamic adjustment of the guidance strategy.

[0130] The output of the multi-turn dialogue management module is connected to the input of the forced pause and sovereignty relinquishment module. Guided interaction content, determined after policy adjustment (or policy maintenance), is sent to the forced pause and sovereignty relinquishment module.

[0131] The forced pause and sovereignty relinquishment module connects its input to the output of the multi-turn dialogue management module, receiving guided interactive content to be output to the user. This module determines the type of this content. If the content is analytical (i.e., includes outputs with information analysis and cognitive assistance properties such as problem breakdown and analysis, multi-solution comparison, causal relationship deduction, and information framework organization), then the following forced pause and sovereignty relinquishment operation sequence is triggered before the analytical content is output to the user terminal: Operation 1: Attach boundary declarations at the end of the guided interactive content. Based on the topic type and content nature of the current dialogue, dynamically select the appropriate boundary declaration form from the boundary declaration type set (such as general logical boundaries, data timeliness boundaries, professional domain boundaries, solution applicability boundaries, etc.) to clearly inform the user of the analysis scope and capability boundaries of the system's current output.

[0132] Operation 2: After the boundary statement, attach a confirmatory question. Based on the current decision-making stage of the dialogue, dynamically select an appropriate question form from the set of confirmatory question types (such as understanding confirmation, choice confirmation, self-judgment confirmation, action confirmation, etc.) and require the user to make a clear confirmation.

[0133] Operation 3: Perform a forced pause. After outputting the complete content mentioned above (analytical content + boundary declaration + confirmatory question), the system will proactively stop further output to the current user session and enter a state of waiting for user response. In this state, the system will not proactively push new content until it receives the user's next round of input messages before resuming output.

[0134] In addition, this module also receives dependency state detection notifications from the multi-turn dialogue management module. When it receives a notification that the user's cognitive state has shifted to dependency, the module triggers a forced independent thinking mode: in this mode, the module intercepts direct answers or alternative judgments that the guidance strategy dynamic generation module may output, and requires the guidance strategy dynamic generation module to regenerate guiding questions that conform to the constraints of refusing to provide direct answers and prompting the user to think independently.

[0135] The data flow and connections between the modules are as follows: The output of the receiving module connects to the input of the intent parsing and state awareness module, transmitting the preprocessed user natural language question text. The output of the intent parsing and state awareness module connects to the input of the guidance strategy dynamic generation module, transmitting the intent parsing results and cognitive state awareness results. The output of the guidance strategy dynamic generation module connects to the input of the multi-turn dialogue management module, transmitting the generated guided interaction content. The output of the multi-turn dialogue management module connects to the input of the forced pause and sovereignty relinquishment module, transmitting the final guided interaction content after strategy management. The output of the forced pause and sovereignty relinquishment module returns the final guided interaction content, with added boundary declarations and / or confirmatory questions, to the user terminal for display via the application layer interface.

[0136] Meanwhile, the system contains two feedback loops: the first feedback loop involves the multi-turn dialogue management module sending each user's new response back to the intent parsing and state awareness module, forming a continuous perception loop of user input, state awareness, strategy generation, user input, and state update; the second feedback loop involves the forced pause and sovereignty relinquishment module sending a regeneration request to the guidance strategy dynamic generation module when it detects a dependency state, requiring the generation of new content that conforms to the constraints of the forced independent thinking mode, forming a cognitive sovereignty protection loop of dependency detection, forced pause, strategy regeneration, and new content output.

[0137] The defense layer is a crucial component of the system architecture of this invention. During the normal execution of guided interactions by the system, without affecting the system's service functions and user experience for normal users, the defense layer incorporates technical mechanisms to prevent unauthorized third parties from interfering with the system's model fitting or knowledge distillation. The specific implementation of the defense layer is described in detail below.

[0138] In the processing stage after the guidance strategy dynamic generation module generates the guided interactive content and before the forced pause and sovereignty handover module performs final output, the defense layer performs semantic pattern embedding processing on the natural language representation of the guided interactive content. This processing, while maintaining the semantic correctness of the guided interactive content and the integrity of the guidance function, performs at least one of the following operations on the text representation: (1) Sentence structure transformation according to pre-defined rules: While keeping the semantic content of the text unchanged, the grammatical structure of the sentence is transformed according to a set of pre-defined transformation rules. The transformation rules may include: switching between active and passive voice (e.g., we suggest you analyze from the following three directions to the following three analysis directions for your reference); switching between positive and negative expressions of causal sentences (e.g., "A caused B" is transformed into "B originates from A" or "A is a major reason for B"); sentence splitting and merging (e.g., splitting a long complex sentence into two short simple sentences, or merging two semantically related short sentences into a complex sentence).

[0139] The set of transformation rules can be unique to each system deployment instance; that is, system instances running under different customers or in different deployment environments will use slightly different sets of transformation rules. The technical effect of this design is that third parties attempting to train a fitting model by collecting large amounts of dialogue data from different system instances will encounter training samples with high inconsistency and unpredictability at the sentence structure level. This structured noise at the sample level will significantly increase the difficulty for the fitting model to learn consistent strategies from the training data, thereby reducing the quality and usability of the fitted model.

[0140] (2) Random Replacement of Functional Terms: In guided interactive content, non-critical functional terms and modifiers that do not affect the core semantics are randomly replaced according to a preset synonym or near-synonym mapping table. For example, when expressing a request, "please" can be randomly replaced with "please," "invite," "suggest," "hope," etc.; when expressing a conditional relationship, "if" can be randomly replaced with "if," "suppose," "if," etc.; when expressing a summary, "in general" can be randomly replaced with "generally speaking," "in summary," "overall," etc. The randomness of the replacement means that even for the same user and the same cognitive state, there are random differences in the literal expression of the system's output text. This difference increases the difficulty of reverse engineering analysis and technical solution restoration based on the consistency of text output.

[0141] (3) Semantic watermark embedding: In the output guided interactive text, a semantic pattern with specific statistical characteristics is embedded in a way that is imperceptible to or will not be noticed by normal end users. One feasible implementation is as follows: In the guided interactive content, the selection of optional words is carried out according to a preset binary encoding rule. For example, when the system needs to choose between synonym A and synonym B, choosing A represents embedding binary bit 0, and choosing B represents embedding binary bit 1, thereby embedding a binary watermark encoding sequence that can be recognized by the system owner in the output text sequence of the entire interactive session.

[0142] When a system owner suspects that a third-party AI model was obtained through unauthorized knowledge distillation of their system, they can verify this by sending a specially designed set of test inputs to the third-party model, analyzing the word selection patterns in its output text, and detecting whether these patterns statistically contain statistical features that match the watermark encoding sequence of their system. If a significant statistical correlation is detected, these statistical features can serve as one of the technical bases for tracing the source of the distillation.

[0143] The defense layer also includes a sub-layer for preventing bulk data collection. This sub-layer is deployed at the system's application layer interface to monitor and analyze request traffic from external networks in real time.

[0144] The anti-bulk collection sublayer maintains a request behavior analyzer. This analyzer continuously collects the following behavioral characteristics for requests from the same source (such as the same IP address range, the same user agent identifier, the same device fingerprint, etc.): the distribution of request time intervals (whether they exhibit high regularity or automated characteristics), the frequency density of requests (whether they exceed the upper limit of the interaction frequency of normal human users), the diversity of request content (whether the semantic diversity of natural language questions in the request matches the natural dialogue pattern of normal users), and the completeness of session behavior (whether there are a large number of short sessions or single-turn interactions, while lacking the characteristics of multi-turn deep interactions of normal users).

[0145] When a request behavior pattern from a certain source is detected to deviate significantly from the preset baseline of normal user behavior (e.g., the request intervals are highly regular and the frequency far exceeds the upper limit of human typing speed; the request content exhibits traversal or enumeration characteristics rather than natural dialogue characteristics), the anti-batch collection sublayer automatically performs at least one of the following protective operations: (a) injecting random time delays into the response returned to the source to reduce its collection efficiency; (b) performing the above semantic pattern embedding operation with higher probability and stronger intensity in the returned guided interactive content to increase the semantic noise density in the collected data; (c) temporarily limiting the upper limit of the request frequency of the source.

[0146] The defense layer is designed with strict functional isolation principles to ensure that its various protective operations do not affect the user experience of normal end users and the core boot functions of the system. (1) Sentence structure transformation and random replacement of functional words only change the form of the text, without changing the semantic content or the correctness of the guiding logic. The text before and after the transformation remains equivalent in two dimensions: whether the user can correctly understand the system's guiding intent and whether the user can think effectively based on the guidance.

[0147] (2) Semantic watermarking embedding only introduces statistical biases at the word selection level that are imperceptible to normal users. The magnitude of this bias is controlled within the natural fluctuation range of word selection in normal human natural language expression, and will not cause a decrease in the user's perception of text quality.

[0148] (3) The triggering conditions of the anti-mass collection mechanism are set with a reasonable sensitivity threshold. The setting of this threshold ensures that only requests that significantly deviate from the normal user behavior pattern will trigger the protection operation. The natural interactive behavior of normal human users (including faster typing speed, more frequent questions, etc.) are all within the safe range of the threshold and will not be affected by any perceptible delay or functional limitation.

[0149] The above-described design of the defense layer ensures that the system of this invention provides high-quality guided interactive services to normal users while possessing technical protection against unauthorized fitting and distillation behaviors. It should be noted that the specific implementation details and parameter settings of the defense layer can be customized according to the security requirements of the actual deployment scenario; the above description is merely illustrative.

[0150] It should be clarified that the specific scenarios (team efficiency improvement, new product pricing, career development decisions, technical solution evaluation, health management, etc.), specific parameter values ​​(such as threshold settings, state judgment indicators, etc.), specific state types (confused, anxious, dependent, conservative), and specific strategy forms (structured decomposition, simplified framework, forced independent thinking, progressive framework, etc.) described in the above embodiments are only used to exemplify the technical solutions of the present invention, and are not intended to limit the scope of protection of the present invention. After reading the above embodiments, those skilled in the art can make at least one of the following modifications or combinations without departing from the core inventive concept of the present invention, and these modifications and combinations all fall within the scope of protection of the present invention: (1) Arbitrarily combine the different cognitive state perception methods described in the above embodiments (rule-based multidimensional feature scoring, machine learning-based classification model, and small-sample reasoning based on large language model) to form a hybrid state perception system. For example, use the rule-based system for rapid initial screening, and then call the reasoning system based on large language model for fine judgment in boundary cases with insufficient confidence.

[0151] (2) In addition to the four cognitive states mentioned above, new cognitive state types may be added according to the needs of specific application scenarios (e.g., but not limited to: confused type, where the user is confused by the information provided by the system and needs a different interpretation; confident type, where the user has formed a relatively clear judgment and only needs the system to provide confirmatory feedback; avoidant type, where the user avoids certain sensitive topics or difficult decisions and needs gradual guidance; curious type, where the user is in an active exploration state and needs open-ended information expansion rather than structured convergent guidance). Define corresponding language feature judgment criteria and corresponding guidance strategies for each newly added cognitive state type.

[0152] (3) Apply the technical solution of the present invention to other fields and scenarios that are not explicitly listed in the above embodiments but also have the characteristic of "users having cognitive assistance needs", including but not limited to: student learning tutoring and knowledge construction assistance, writing ideation and logic building assistance, personal investment and financial management decision assistance, information sorting assistance in medical and health consultation, element decomposition assistance in legal problem analysis, multi-dimensional comparison assistance in engineering design scheme evaluation, career planning and skill development path assistance, and feasibility analysis assistance for entrepreneurial projects.

[0153] (4) Adjust the module division method of the system, for example, split the intent parsing and state awareness module into two physically independent microservice modules, or merge the guidance strategy dynamic generation module and the multi-turn dialogue management module and deploy them on the same computing node. Regardless of how the modules are divided, merged or reorganized, as long as the overall operation flow of the system executes the steps of receiving user input, intent parsing and state awareness, dynamically generating guided interactive content, continuously sensing state changes and dynamically adjusting strategies in multi-turn interactions, and performing forced pause and outputting boundary declarations and / or confirmatory questions after outputting analytical content, it falls within the protection scope of this invention.

[0154] (5) The semantic pattern embedding method, watermark encoding rules, anti-batch collection trigger threshold and protection operation type of the defense layer can be customized and combined according to the security requirements, performance constraints and operation and maintenance strategies of the specific deployment environment.

[0155] The cognitive empowerment method and system based on AI-guided interaction constructed in this invention have the following technical effects compared with existing passive response AI interaction technologies: By proactively generating guided questions instead of directly providing answers, the system liberates users from the constraints of needing precise questioning skills. Through a closed-loop process of intent parsing, state awareness, and dynamic guidance, the system proactively helps users clarify ambiguous questions, break down complex goals, and promote structured thinking. This enables users with different knowledge backgrounds and technical abilities to communicate with the system in natural language and efficiently analyze and solve problems under the system's structured cognitive guidance.

[0156] By continuously and multidimensionally analyzing the linguistic features of user input text, the system achieves real-time dynamic perception of the user's cognitive state and adjusts guidance strategies accordingly. This technology ensures that the AI's interaction mode is not a static, unchanging pattern, but a dynamically adaptable mode that precisely matches the user's current cognitive state. For example, when the user's cognitive state changes from confusion to dependence, the system can immediately detect this change and switch to a forced independent thinking mode; when the user changes from conservative to initial orientation, the system can correspondingly switch from a progressive framework to a targeted in-depth exploration strategy. This state-aware, dynamic strategy matching mechanism provides optimal cognitive assistance while effectively preventing AI from misleading or providing ineffective guidance.

[0157] Through a unique forced pause and boundary declaration mechanism, the system proactively stops outputting analytical content and delivers confirmatory questions to the user, thus forcibly returning the final analysis, judgment, and decision-making process to the user at the technical architecture level. This mechanism ensures that humans maintain an irreplaceable dominant position throughout the entire decision-making loop, fundamentally addressing the potential for users to become dependent on AI by directly obtaining answers from it. Simultaneously, by proactively declaring the analytical boundaries of the system's capabilities, it guides users to maintain critical thinking, establishing a safer, more trustworthy, and healthier human-machine collaborative relationship.

[0158] The above description represents the preferred embodiments of the present invention. It should be noted that those skilled in the art can make various improvements and modifications without departing from the principles of the present invention, and these improvements and modifications should also be considered within the scope of protection of the present invention.

Claims

1. A cognitive empowerment method based on AI-guided interaction, characterized in that, include: The problem of receiving natural language input from users; The natural language problem is analyzed to determine the user's core needs, and the user's current cognitive state is perceived by analyzing the language features of the input. Based on the perceived cognitive state, guided interactive content is dynamically generated, which includes guided questions to guide the user's thinking. During multiple rounds of interaction with users, continuously perceive changes in the user's cognitive state and dynamically adjust the guidance strategy based on the changed cognitive state; After outputting analytical content, a forced pause is executed, and boundary statements and / or confirmatory questions are presented to the user to return the final decision-making power to the user and ensure that the user completes independent thinking.

2. The method according to claim 1, characterized in that, The cognitive state includes at least two or more of the following: confused, anxious, dependent, and conservative.

3. The method according to claim 2, characterized in that, When the perceived cognitive state is dependent, the dynamic adjustment guidance strategy includes: automatically switching to a forced independent thinking mode, refusing to provide direct answers in the forced independent thinking mode, and generating guiding questions that prompt the user to think independently.

4. The method according to claim 1, characterized in that, The guided interactive content includes providing structured options, frameworks, or follow-up questions to help users organize and break down the issues.

5. The method according to claim 1, characterized in that, The forced pause also includes: actively stopping the current interactive output after outputting guided questions or analysis content, and waiting for the user's response; the confirmatory questions are used to require the user to explicitly confirm their understanding or selection.

6. The method according to claim 1, characterized in that, The dynamically generated guided interactive content generates corresponding guidance strategies in real time based on the cognitive state.

7. The method according to claim 1, characterized in that, Also includes: After breaking down the core issues, a cognitive progress report is generated, summarizing the breakdown results and actionable tasks, and returning the initiative to the user.

8. The method according to claim 1, characterized in that, The boundary declaration includes declaring the scope of the AI ​​system's analysis based on general logic and prompting users to make final decisions based on their own actual circumstances.

9. The method according to claim 1, characterized in that, The continuous sensing of changes in the user's cognitive state includes: Based on the user's responses in multiple rounds of interaction, language features are re-analyzed to update cognitive state.

10. A cognitive empowerment system based on AI-guided interaction, characterized in that, The system is used to perform the method as described in any one of claims 1 to 9, comprising: The receiving module is used to receive natural language questions input by the user; The intent parsing and state awareness module is used to perform intent parsing on the natural language question, determine the user's core needs, and perceive the user's current cognitive state by analyzing the language features of the input. The guidance strategy dynamic generation module is used to dynamically generate guided interactive content based on the perceived cognitive state. The guided interactive content includes guided questions to guide the user to think. The multi-turn dialogue management module is used to continuously sense changes in the user's cognitive state during multiple rounds of interaction and dynamically adjust the guidance strategy based on the changed cognitive state. The Forced Pause and Sovereignty Relinquishment module is used to perform a forced pause after outputting analytical content, outputting boundary statements and / or confirmatory questions to the user, so as to return the final decision-making power to the user.

Images