Gamification psychotherapy healing system based on multi-agent architecture

Abstract

The invention belongs to the technical field of artificial intelligence, and discloses a multi-agent architecture-based gamification psychotherapy healing system, which comprises four agents, namely a trigger, a negative thought generation module, a guider module and a planner module, and is characterized in that the trigger is used for generating a dynamic virtual scene; the negative thought generation module is used for generating a negative thought matched with the psychological state of the patient; the guider module is used for generating a structured suggestion of the corresponding iteration process according to the dynamic virtual scene and the negative thought of the current iteration process and compressing the structured suggestion into a key abstract; and the planner module is used for generating a strategy instruction according to the key abstract of the current iteration process and the patient feedback information. According to the method, the dynamic scene is generated through multi-agent cooperation, so that a patient can more easily substitute own emotion in a virtual environment, the immersive experience is enhanced, and the common emotion is remarkably improved; feedback of a patient can be sensed in real time, scene content and role behaviors are dynamically adjusted, and a personalized chemotherapy healing path is achieved; and various psychological themes are supported, and the method can be expanded to different psychological health intervention scenes.

Description

Technical Field

[0001] The present invention belongs to the technical field of artificial intelligence, relates to a psychotherapy robot, and particularly relates to a gamified psychotherapy system based on a multi-agent architecture. Background Art

[0002] Mental illnesses are caused by physiological, psychological, and social factors that lead to brain disorders, disrupting the integrity of the human brain function and the unity between the individual and the external environment. When mental illnesses occur, psychotherapy by psychiatrists is usually required. However, the reality is that the number of psychiatrists is far from sufficient, and there are problems such as uneven professional qualities and high consultation fees in the psychiatrist team.

[0003] The method based on empathy training can help people solve psychological problems by themselves. For example, the empathy training method based on virtual reality (VR) helps patients achieve self-talk and emotional regulation through constructing virtual scenarios and role-playing (see Halim I, Stemmet L, Hach S, Porter R, Liang HN, Vaezipour A, Henry JD, Baghaei N. Individualized Virtual Reality for Increasing Self-Compassion: Evaluation Study. JMIR Ment Health. 2023 Oct 2;10:e47617. doi: 10.2196 / 47617. PMID: 37782537; PMCID: PMC10580135); its core mechanism is the "illusion of virtual body ownership", that is, the patient interacts with the "inner self" in a preset scenario through a virtual avatar (Avatar); the specific implementation steps are as follows: (1) Role reversal mechanism: Stage 1 (empathy output): The patient interacts with the "crying child" in the virtual scenario from the perspective of an adult avatar, providing emotional support through language and actions, such as Figure 1 (a) shown; Stage 2 (empathy reception): The patient switches to the perspective of the child avatar and receives comforting words from the "past self", such as Figure 1 (b) shown. This therapy creates a space for patients to reflect on themselves. By listening to their own expressions from different perspectives, patients can better understand the reasons for their emotions and then achieve emotional regulation.

[0004] However, traditional psychotherapy methods and existing chatbots have the following deficiencies:

[0005] (1) The virtual scenarios of VR empathy training (such as room layout and character behavior) rely on manual predefined settings, lacking the ability to generate dynamically, unable to adjust according to the patient's real-time emotions, and difficult to dynamically map the patient's actual life stressors (such as workplace conflicts, family contradictions, etc.). Lack of personalization leads to the disconnection between the treatment environment and the actual psychological needs.

[0006] (2) Interaction rigidity: The system adopts pre-programmed character behaviors and provides generalized feedback, unable to handle complex cognitive restructuring tasks according to the patient's emotional state, and lacking the capture of subtle changes in the patient's situation (such as the transition from anxiety to despair). This results in the mismatch between the feedback and the patient's real-time psychological state, making the interaction process mechanical and lacking emotional depth, thus weakening the treatment effect. Summary of the Invention

[0007] The purpose of the present invention aims to address the above-mentioned problems existing in the prior art, and provides a gamified psychological healing system based on a multi-agent architecture. Combining the generation and role-playing capabilities of LLMs, it can achieve a dynamic, adaptive, and highly immersive mental health intervention scenario to enhance the depth of emotional interaction and treatment effect.

[0008] To achieve the above object, the present invention provides a gamified psychological healing system based on a multi-agent architecture, which includes:

[0009] A receiving module, used to receive information sent by the patient, and the information sent by the patient is a request message or a feedback message;

[0010] A trigger, used to generate a dynamic virtual scene; at the initial iteration, the trigger generates the initial scene of the dynamic virtual scene according to the patient's request message; for the i-th iteration in the iterative process, the trigger generates the dynamic virtual scene of the corresponding iteration according to the patient's request message, the feedback message and the strategy instruction of the previous iteration process.

[0011] A negative thought generation module, used to generate negative thoughts that match the patient's psychological state; at the initial iteration, the negative thought generation module generates negative thoughts that match the patient's psychological state according to the dynamic virtual scene and the patient's request message; for the i-th iteration in the iterative process, the negative thought generation module generates negative thoughts that match the patient's psychological state according to the dynamic virtual scene of the current iteration, and the patient's feedback message and the strategy instruction of the previous iteration process.

[0012] A guide module, used to generate structured suggestions for the corresponding iteration process according to the dynamic virtual scene and negative thoughts of the current iteration process and compress them into a key summary;

[0013] A sending module, used to send the dynamic virtual scene, negative thoughts, and structured suggestions to the patient;

[0014] A strategist module for generating strategy instructions based on the key summary of the current iteration process and patient feedback information.

[0015] In an implementable manner, for the receiving module, the request information includes the subject involved and the content of the trouble.

[0016] In an implementable manner, at the initial iteration, the trigger generates an initial virtual scenario S0 through chain-of-thought prompting, which is expressed as:

[0017]

[0018] For the i-th iteration in the iteration process, the trigger generates a dynamic virtual scenario S i which is expressed as:

[0019]

[0020] where W is the content of the patient's trouble; T is the subject; π t is the scenario generation function of the trigger, which uses a large language model; C i-1 is the patient feedback information in the (i - 1)-th iteration process; P i-1 is the strategy instruction information in the (i - 1)-th iteration process.

[0021] In an implementable manner, the negative thought generation module determines the types of cognitive distortions and personality dimensions based on the information sent by the patient, and then generates negative thoughts matching the patient's mental state in combination with the dynamic virtual scenario; the types of cognitive distortions include emotional reasoning, overgeneralization, mental filtering, should statements, all-or-nothing thinking, mind reading, magnification and minimization, self-blame and blaming others, labeling, jumping to conclusions, etc.; the personality dimensions include openness, conscientiousness, neuroticism, extraversion, openness, agreeableness.

[0022] The patient's personality dimensions are determined by the content of the trouble in the patient's request information;

[0023] At the initial iteration, the types of the patient's cognitive distortions are determined by the content of the trouble in the patient's request information; at the i-th iteration, the types of the patient's cognitive distortions are determined by the patient's feedback information.

[0024] Specifically, at the initial iteration, the negative thought generation module generates negative thoughts D0 matching the patient's mental state, which is expressed as:

[0025]

[0026] where W is the content of the patient's trouble; S0 is the dynamic virtual scenario obtained from the initial iteration;

[0027] For the i-th iteration in the iterative process, the negative thought generation module generates negative thoughts D that match the patient's mental state i It is expressed as:

[0028]

[0029] where π d is the generation function of the negative thought generation module, and a large language model is used; S i is the dynamic virtual scene obtained in the i-th iteration process; C i-1 is the patient feedback information in the (i - 1)-th iteration process; P i-1 is the policy instruction information in the (i - 1)-th iteration process.

[0030] In an implementable manner, the guide module generates a structured suggestion G i for the corresponding iteration process and compresses it into a key summary M i It is expressed as:

[0031] (G i , M i ) = π g (S i , D i ), i ≥ 0;

[0032] where π g is the generation function of the guide module, and a large language model is used; S i is the dynamic virtual scene obtained in the i-th iteration process; D i is the negative thought generated in the i-th iteration process.

[0033] The key summary includes a summary of the dynamic virtual scene and a summary of the negative thought.

[0034] In an implementable manner, the strategist module generates a policy instruction expressed as:

[0035] P i = π s (M i , C i );

[0036] where π s is the generation function of the strategist module, and a large language model is used; M i is the key summary obtained in the i-th iteration process; C i is the patient feedback information generated in the i-th iteration process.

[0037] Therefore, the gamified psychological healing system based on a multi-agent architecture provided by the present invention has the following characteristics:

[0038] (1) Dynamic scenario generation: Abandoning predefined scripts, combining the patient's real-time emotional state with the historical feedback information, and using large language models (LLMs) to generate coherent and adaptive virtual scenarios, enabling the treatment environment to dynamically adjust according to the patient's psychological changes and enhancing the narrative immersion;

[0039] (2) Personalized interaction optimization: Simulating the complex cognitive restructuring process through the division of labor of multiple agents (trigger, negative thought generation module, guide module, and planner module), providing targeted feedback for the patient's specific types of cognitive distortions (such as "overgeneralization", "black-and-white thinking", etc.), and breaking through the limitations of traditional templated interactions;

[0040] (3) Efficient computing and real-time response: Compressing redundant dialogue information through the guide module, optimizing the processing efficiency of the LLM for long texts, and shortening the average system response time to the millisecond level to ensure smooth interaction;

[0041] (4) Treatment path adaptability: Dynamically adjusting the behavior of agents in real time through the planner module to ensure that the system can still stably converge to the preset treatment effect goal under non-ideal conditions (such as sudden changes in the patient's mood or deviation of the treatment goal).

[0042] Compared with the prior art, the gamified psychological healing system based on a multi-agent architecture provided by the present invention has the following beneficial effects:

[0043] (1) Enhanced immersive experience: By the collaborative generation of dynamic scenarios by multiple agents, it is easier for patients to substitute their own emotions in the virtual environment, enhancing the immersive experience and significantly improving empathy;

[0044] (2) Strong dynamic response ability: The system can perceive the patient's feedback in real time, dynamically adjust the scene content and character behavior, and achieve a personalized healing path;

[0045] (3) Effective cognitive restructuring guidance: Through round-by-round cognitive restructuring and multi-role conversations, it helps patients identify cognitive distortions and promotes self-reflection and psychological recovery;

[0046] (4) High interaction flexibility: The multi-agent architecture provides multi-perspective interactions, avoiding the rigidity of a single conversation style, and enhancing the fun of interaction and the patient's participation;

[0047] (5) Wide applicability: It supports multiple psychological themes (such as anxiety, depression, work stress, etc.) and can be extended to different mental health intervention scenarios. Description of the Drawings

[0048] Figure 1Scenario diagram of an empathy training method based on virtual reality (VR); among them, (a) corresponds to the schematic diagram of the patient comforting the "little girl" in stage 1, and (b) corresponds to the schematic diagram of the patient switching to the perspective of the "little girl" to receive comfort in stage 2;

[0049] Figure 2 Principle block diagram of a gamified psychotherapy system based on a multi-agent architecture provided by an embodiment of the present invention;

[0050] Figure 3 The first iteration process of learning a learning case using the gamified psychotherapy system based on a multi-agent architecture provided by an embodiment of the present invention;

[0051] Figure 4 Evaluation results of the ablation experiment content of the gamified psychotherapy system based on a multi-agent architecture provided by an embodiment of the present invention;

[0052] Figure 5 Comparison results of different psychotherapy paradigms. Detailed implementation manners

[0053] The technical solutions of the embodiments of the present invention will be clearly and completely described below with reference to the accompanying drawings. Obviously, the described embodiments are only a part of the embodiments of the present invention, rather than all of the embodiments. All other embodiments obtained by those of ordinary skill in the art based on the embodiments of the present invention without creative efforts belong to the scope protected by the present invention.

[0054] Embodiment 1

[0055] This embodiment provides a gamified psychotherapy system based on a multi-agent architecture, as Figure 2 shown. It realizes immersive psychotherapy through the dynamic cooperation of four types of agents. The four types of agents are trigger, devil, guide, and strategist, and all of these four types of agents adopt large language models (LLMs), specifically Gemini-2.0-flash.

[0056] In addition to the above four types of agents, the above gamified psychotherapy system further includes a receiving module and a sending module.

[0057] The receiving module is used to receive information sent by the patient, and the information sent by the patient is a request message or a feedback message. The request message includes the involved topic and the disturbing content. The topic mainly refers to the main types involved in mental illnesses, such as anxiety, depression, work stress, etc. The disturbing content refers to the specific events that cause trouble to the patient, such as working hard but getting poor grades.

[0058] A sending module, which is used to send the dynamic virtual scenarios generated by the trigger, the negative thoughts generated by the negative thought generation module, and the structured suggestions generated by the guide module to the patient. The patient will also feedback the feedback information given based on this information to the system through the receiving module.

[0059] The following will explain in detail the specific implementation methods of the trigger, the negative thought generation module (devil), the guide module, and the strategist module.

[0060] 1. Trigger

[0061] The trigger is used to generate dynamic virtual scenarios, including the initial scenario generated during the initial iteration and the dynamic scenario expansion during the iteration process.

[0062] During the initial iteration, the trigger generates the initial scenario of the dynamic virtual scenario based on the patient request information (including the theme T and the troubled content W). Specifically, the trigger generates the initial virtual scenario S0 through chain-of-thought prompting, which is expressed as:

[0063]

[0064] For the i-th iteration during the iteration process, the trigger generates the dynamic virtual scenario Si of the corresponding iteration based on the patient request information, the feedback information of the patient in the previous iteration process, and the strategy instructions given by the strategist module. i , which is expressed as:

[0065]

[0066] Among them, W is the patient's troubled content; T is the theme; π t is the scenario generation function of the trigger, which uses a large language model, specifically Gemini-2.0-flash; C i-1 is the feedback information of the patient in the previous iteration process; P i-1 is the strategy instruction information in the previous iteration process.

[0067] 2. Negative Thought Generation Module (devil)

[0068] The negative thought generation module is used to generate negative thoughts that match the patient's psychological state. Here, "matching the patient's psychological state" means matching the patient's cognitive distortion type and personality dimension. The cognitive distortion types include emotional reasoning, overgeneralization, mental filtering, should statements, all-or-nothing thinking, mind reading, making a mountain out of a molehill, self-blame and blaming others, labeling, jumping to conclusions, etc. The personality dimensions include openness, conscientiousness, neuroticism, extraversion, openness, agreeableness, etc.

[0069] Specifically, at the initial iteration, the negative thought generation module generates negative thought D0 that matches the patient's mental state based on the dynamic virtual scene and the patient's request information, which is expressed as:

[0070]

[0071] where W is the content of the patient's distress; S0 is the dynamic virtual scene obtained from the first iteration.

[0072] At the initial iteration, the negative thought generation module determines the cognitive distortion type and personality dimension based on the distress content in the patient's request information, and then generates negative thoughts in combination with the dynamic virtual scene.

[0073] For example, based on the patient's distress content, the cognitive distortion type is determined to be emotional reasoning, and in combination with the dynamic virtual scene generated by the trigger, the negative thought generated is "Feeling frustrated, so the fact must be bad"; for another example, based on the patient's distress content, the cognitive distortion type is determined to be making a mountain out of a molehill, and in combination with the dynamic virtual scene generated by the trigger, the negative thought generated is "I will never succeed".

[0074] Then, according to the personality dimension, the generated negative thought style is adjusted. For example, when the personality dimension is determined to be neurotic based on the distress content, the negative thought generated by the negative thought generation module is a more intense expression of negative emotions.

[0075] For the i-th iteration in the iterative process, the negative thought generation module generates negative thought D that matches the patient's mental state based on the dynamic virtual scene of the current iteration, and the patient feedback information and strategy instruction of the previous iteration process i , which is expressed as:

[0076]

[0077] where π d is the generation function of the negative thought generation module, which uses a large language model, specifically Gemini-2.0-flash; S i is the dynamic virtual scene obtained from the i-th iteration process; C i-1 is the patient feedback information of the i-1-th iteration process;

[0078] P i-1 is the strategy instruction information of the i-1-th iteration process.

[0079] The negative thought generation module first determines the cognitive distortion type based on the patient feedback information of the previous iteration process, and then generates negative thoughts in combination with the personality dimension (determined by the distress content at the initial iteration), the dynamic virtual scene of the current iteration process, and the strategy instruction information of the previous iteration process.

[0080] The output of the negative thought generation module includes cognitive distortion types, personality dimensions, and negative thoughts.

[0081] 3. Guide Module (guide)

[0082] The guide module is used to generate structured suggestions for the corresponding iteration process based on the dynamic virtual scenario and negative thoughts of the current iteration process, such as "record successful cases to refute overgeneralization" and "break down problems to reduce anxiety"; and generate a key summary of the dynamic virtual scenario and negative thoughts of the current iteration process to improve processing efficiency. Therefore, the key summary includes a summary of the dynamic virtual scenario and a summary of negative thoughts.

[0083] The guide module generates structured suggestions G for the corresponding iteration process i and compresses them into a key summary M i which is expressed as:

[0084] (G i , M i ) = π g (S i , D i ), i ≥ 0;

[0085] where π g is the generation function of the guide module, which uses a large language model, specifically Gemini-2.0-flash; S i is the dynamic virtual scenario obtained in the i-th iteration process; D i is the negative thought generated in the i-th iteration process.

[0086] 4. Strategist Module (strategist)

[0087] The strategist module is used to generate strategy instructions based on the key summary of the current iteration process and patient feedback information, and control the trigger scenario and the idea transformation of the negative thought generation module.

[0088] The strategy instructions generated by the strategist module are expressed as:

[0089] P i = π s (M i , C i );

[0090] where π s is the generation function of the strategist module, which uses a large language model, specifically Gemini-2.0-flash; M i is the key summary obtained in the i-th iteration process; C i is the patient feedback information generated in the i-th iteration process.

[0091] The strategy instructions generated by the strategist module include instructions for guiding the trigger to generate a dynamic virtual scene and instructions for guiding the negative thought generation module to generate negative thoughts.

[0092] The overall implementation steps of the above gamified psychotherapy system based on a multi-agent architecture include:

[0093] S1 Generate a dynamic virtual scene

[0094] The receiving module receives the patient's current distress content and selected healing theme and sends them to the trigger; the trigger generates an initial scene based on the patient's input. The initial scene vividly describes the patient's current environment and inner feelings, providing a background basis for subsequent conversations.

[0095] In subsequent iterations, the trigger generates the corresponding iterative dynamic virtual scene based on the patient's request information, the feedback information from the previous iteration process, and the strategy instructions.

[0096] S2 Simulate cognitive distortion

[0097] Based on the initial scene, the negative thought generation module, as the patient's "avatar", simulates the patient's negative self-talk and generates negative thought content with representative cognitive distortion types (such as making a mountain out of a molehill, taking responsibility for everything oneself, and blaming others for everything) and personality dimensions. This process uses psychological knowledge to accurately simulate the possible cognitive biases of the patient and truly restore the patient's inner negative dialogue.

[0098] S3 Guide cognitive restructuring

[0099] The guide module generates structured suggestions based on the dynamic virtual scene and negative thoughts and compresses them into a key summary, providing targeted psychological guidance. By raising doubts and alternative thinking in a gentle tone, it helps the patient identify the irrational components of negative thinking and provides cognitive restructuring strategies, such as self-comfort and evidence testing.

[0100] S4 Guide the patient's self-comfort dialogue

[0101] According to the suggestions of the guide module, guide the patient to give corresponding feedback information when facing negative thoughts in the corresponding dynamic scene, so as to carry out psychological comfort. This process encourages the patient to actively participate in the healing and enhances the patient's self-cognitive awareness and emotion regulation ability.

[0102] S5 Treatment strategy planning

[0103] The planner module generates strategy instructions based on the key summary given by the guide module and the feedback information given by the patient, comprehensively analyzes the patient's interaction history, and summarizes and plans the next round of scenes. The planner module is responsible for ensuring the coherence of the plot progression and gradually guiding the patient towards a positive emotional state.

[0104] The system repeats the above steps S1 - S5. Through multiple rounds of cyclic interaction, it continuously strengthens the patient's emotional regulation and cognitive restructuring abilities until the patient achieves inner reconciliation. The entire process fully considers the patient's emotional fluctuations and dynamically adjusts the interaction rhythm and guiding strategies.

[0105] Figure 3 Given the first iteration process of a learning case by the gamified psychotherapy system based on the multi - agent architecture provided in this embodiment, including steps S1 - S5. First, an initial scene is generated through a trigger; the cognitive distortion types (over - generalization, all - or - nothing) and personality dimensions (emotional reasoning) are determined by the negative thought generation module, and then negative thoughts are generated; then, a structured suggestion is generated by the guide module and compressed into a key summary; according to the suggestions of the guide module, the patient is guided to give corresponding feedback information when facing negative thoughts in the corresponding dynamic scene; the strategist module generates strategy instructions based on the patient's feedback information and the key summary. The system provided in this embodiment considers the patient's emotional fluctuations and dynamically adjusts the interaction rhythm and guiding strategies.

[0106] Table 1 shows the content evaluation results during the processing by the gamified psychotherapy system based on the multi - agent architecture provided in this embodiment under different themes. Among them, all the test theme cases come from the C2D2 dataset (see Bichen Wang, Pengfei Deng, Yanyan Zhao, and Bing 861 Qin. 2023. C2d2 dataset: A resource for the cog - 862 nitive distortion analysis and its impact on mental 863 health. In Findings of the Association for Compu - 864 tational Linguistics: EMNLP 2023, pages 10149–865 10160). It can be seen from Table 1 that the present invention has good evaluations under each theme.

[0107] Table 1 Content evaluation results under different themes

[0108]

[0109] Note: IM = Immersion, CO = Consistency, EN = Engagement, ER = Emotional comfort, SA = Satisfaction, IN = Interest.

[0110] This embodiment also gives the average effect of the content evaluation results of different themes when the key brief, guide module, and planner module are missing in the system (MIND) of this embodiment, as Figure 4As shown. From Figure 4 It can be seen that the absence of any one of the key brief, guide module, and planner module will lead to a significant reduction in the evaluation effect, indicating the irreplaceability of the system provided by the present invention.

[0111] This embodiment also gives the average effect of the evaluation results of the system (MIND) of this embodiment, traditional psychological counseling, traditional empathy training (see Halim I, Stemmet L, Hach S, Porter R, Liang HN, Vaezipour A, Henry JD, Baghaei N. Individualized Virtual Reality for Increasing Self-Compassion: Evaluation Study. JMIR Ment Health. 2023 Oct 2;10:e47617. doi:10.2196 / 47617. PMID:37782537; PMCID:PMC10580135) and chatbots on the content of different topics, such as Figure 5 As shown. From Figure 5 It can be seen that the system provided by the present invention has the highest evaluation result.

[0112] Those of ordinary skill in the art will realize that the embodiments here are to help readers understand the principles of the present invention, and it should be understood that the protection scope of the present invention is not limited to such specific statements and embodiments. Those of ordinary skill in the art can make various other specific deformations and combinations that do not deviate from the essence of the present invention based on these technical revelations disclosed in the present invention, and these deformations and combinations are still within the protection scope of the present invention.

Claims

1. A gamification psychological healing system based on a multi-agent architecture, characterized in that, Including: A receiving module for receiving information sent by a patient, where the information sent by the patient is a request message or a feedback message; A trigger for generating a dynamic virtual scene; at the initial iteration, the trigger generates an initial scene of the dynamic virtual scene according to the patient's request message; For the i-th iteration in the iteration process, the trigger generates a dynamic virtual scene for the corresponding iteration according to the patient's request message, the feedback message and the policy instruction of the previous iteration process; A negative thought generation module for generating negative thoughts matching the patient's mental state; at the initial iteration, the negative thought generation module generates negative thoughts matching the patient's mental state according to the dynamic virtual scene and the patient's request message; For the i-th iteration in the iteration process, the negative thought generation module generates negative thoughts matching the patient's mental state according to the dynamic virtual scene of the current iteration, and the patient's feedback message and the policy instruction of the previous iteration process; A guide module for generating a structured suggestion for the corresponding iteration process according to the dynamic virtual scene and the negative thoughts of the current iteration process and compressing it into a key summary; A sending module for sending the dynamic virtual scene, the negative thoughts, and the structured suggestion to the patient; A strategist module for generating a policy instruction according to the key summary of the current iteration process and the patient's feedback message.

2. The gamification psychological healing system based on a multi-agent architecture according to claim 1, characterized in that, For the receiving module, the request message includes the involved topic and the disturbing content.

3. The gamification psychotherapy system based on a multi-agent architecture according to claim 2, wherein At the initial iteration, the trigger generates an initial virtual scene S0 through chain-of-thought prompting, expressed as: For the i-th iteration in the iterative process, the trigger generates a dynamic virtual scene S for the corresponding iteration i Denoted as: Among them, W is the content that troubles the patient; T is the theme; π t is the scenario generation function of the trigger, which uses a large language model; c i-1 is the patient feedback information in the (i - 1)-th iteration process; p i-1 is the policy instruction information in the (i - 1)-th iteration process.

4. The gamification psychological healing system based on a multi-agent architecture according to claim 1, characterized in that, The negative thought generation module determines the cognitive distortion type and personality dimension according to the information sent by the patient, and then combines the dynamic virtual scene to generate negative thoughts matching the patient's mental state; the cognitive distortion types include emotional reasoning, overgeneralization, mental filtering, should statements, all-or-nothing, mind reading, magnification and minimization, self-blame and blaming others, labeling, and jumping to conclusions; the personality dimensions include openness, conscientiousness, neuroticism, extraversion, openness, and agreeableness.

5. The gamification psychological healing system based on a multi-agent architecture according to claim 4, characterized in that The patient's personality dimension is determined by the disturbing content in the patient's request message; At the initial iteration, the patient's cognitive distortion type is determined by the disturbing content in the patient's request message; at the i-th iteration, the patient's cognitive distortion type is determined by the feedback message of the patient in the (i - 1)-th iteration process.

6. The gamification psychological healing system based on a multi-agent architecture according to claim 1, 4 or 5, characterized in that, At the initial iteration, the negative thought generation module generates negative thoughts D0 matching the patient's mental state, expressed as: Where W is the patient's disturbing content; S0 is the dynamic virtual scene obtained in the initial iteration; For the i-th iteration in the iterative process, the negative thought generation module generates a negative thought D that matches the patient's mental state i Expressed as: Among them, π d is the generation function of the negative thought generation module, which uses a large language model; s i is the dynamic virtual scene obtained in the i-th iteration process; c i-1 is the patient feedback information in the (i - 1)-th iteration process; p i-1 is the policy instruction information in the (i - 1)-th iteration process.

7. The gamification psychological healing system based on a multi-agent architecture according to claim 1, characterized in that, The guide module generates structured suggestions g for the corresponding iterative process i and compresses them into a key summary m i which is expressed as: (g i ,M i ) = π g (S i ,D i ), i ≥ 0; Among them, π g is the generation function of the guide module, which uses a large language model; S i is the dynamic virtual scene obtained in the i-th iteration process; D i is the negative thought generated in the i-th iteration process.

8. The gamification psychotherapy system based on a multi-agent architecture according to claim 7, characterized in that, The key summary includes a summary of the dynamic virtual scene and a summary of the negative thoughts.

9. The gamification psychological healing system based on a multi-agent architecture according to claim 1, wherein The strategist module generates a policy instruction, expressed as: P i = π s (M i , C i ); Among them, π s is the generation function of the strategist module, which uses a large language model; M i is the key summary obtained in the i-th iteration process; C i is the patient feedback information generated in the i-th iteration process.

Images