A sentiment homeostatic control system for an AI agent

Abstract

The application discloses an emotion homeostasis control system for an AI agent, and relates to the field of emotional computing, comprising: a perception input module for collecting and preprocessing dialogue text, voice intonation, facial expression and physiological signals input by a user, and converting the same into analyzable structured data; an analysis kernel module for extracting comprehensive context information including semantic intention, emotional state and interactive situation from the structured data; a main control kernel module for serving as a core scheduling hub, receiving the comprehensive context information, coordinating the operation of various functional modules, defining and evaluating a plurality of homeostasis levels according to the modeling depth of user information; and the application builds an evolvable personality state vector generation mechanism, so that the AI agent can comprehensively understand the semantic, emotional and situational information of the user, and generate a highly personalized response based on the internal role setting, stable characteristics and dynamic deduction strategy.

Description

Technical Field

[0001] This invention relates to the field of affective computing technology, specifically to an affective mimicry control system for AI agents. Background Technology

[0002] Large language models have demonstrated powerful capabilities in dialogue generation tasks, and the application scenarios of AI agents are expanding from tools for completing specific tasks to anthropomorphic partners providing services such as companionship, consultation, and entertainment. Users' expectations for intelligent agents are no longer limited to accurate information retrieval or mechanical task execution; they yearn for agents to possess more "human" qualities. This has spurred an urgent need in the fields of affective computing, personality computing, and human-computer interaction to build AI entities with emotional intelligence. Simultaneously, the maturity of multimodal perception technology enables systems to more comprehensively capture and understand users' emotions and intentions from text, speech, vision, and even physiological signals, providing a technological foundation for building more nuanced and realistic emotional interactions.

[0003] Existing AI agents based on large language models have the following inherent shortcomings in terms of personality mimicry and emotional consistency: Traditional methods typically guide large language models by statically describing character settings in prompts. However, in long-term, multi-turn dialogues, large language models are prone to forgetting their initial settings or being disturbed by user input and subtle changes in context, leading to uncontrollable shifts in the personality traits, values, and language style of the generated responses, resulting in a lack of stability. Existing systems often process emotions in an immediate and superficial manner, only able to perform simple emotion classification on the current input and generate text with corresponding intonation. They lack the accumulation, changes, and coherence of emotions based on long-term interactive memory, cannot remember emotional events shared with users, and are difficult to naturally reference in subsequent interactions, resulting in a lack of depth and authenticity in the interaction. Although existing systems can recognize multimodal information such as user facial expressions and intonation, this information is often only used for emotional state analysis and fails to be effectively transformed into structured parameters to quantitatively influence and constrain the core expression process of personality, making it difficult for the agent's response to form an organically unified whole at the cognitive, emotional, and behavioral levels. Summary of the Invention

[0004] (a) Technical problems to be solved In view of the above-mentioned shortcomings of the existing technology, the present invention provides an emotion mimicry control system for AI intelligent agents, which can effectively solve the problems of the existing technology.

[0005] (II) Technical Solution To achieve the above objectives, the present invention provides the following technical solution: This invention discloses an emotion mimicry control system for AI intelligent agents, comprising: The input perception module is used to collect and preprocess user input dialogue text, voice tone, facial expressions and physiological signals, and convert them into parsable structured data. The parsing kernel module is used to extract comprehensive contextual information from structured data, including semantic intent, emotional state, and interaction context. The main control kernel module serves as the core scheduling hub, responsible for receiving comprehensive context information and coordinating the operation of various functional modules. Based on the depth of user information modeling, it defines and evaluates several mimicry levels. The personality control structure module is used to construct a role layer, a trait layer, and a deductive layer. The role layer is used to store and define the agent's identity, core values, basic beliefs, and long-term mission. The trait layer is used to define the agent's stable personality traits, thinking patterns, and expression habits. The deductive layer is used to dynamically adjust the interaction strategy and emotional expression method according to the current interaction scenario, object relationship, and conversation goal. The memory module is used to store and manage the long-term relationship memory between the current agent and the user, the short-term conversation context, and the new personality characteristic data evolved from the interaction; The PSV computing engine is used to dynamically fuse the role layer definition, trait layer parameters and deductive layer strategies of the personality control structure module, the memory information of the memory module, and the comprehensive context information input in real time from the kernel module to calculate and generate the personality state vector at the current moment. The PSV injection interface is used to inject the calculated personality state vector as a key control parameter into the inference generation process of the preset large language model in real time. The reasoning generation module receives reasoning instructions and context information from the PSV injection interface constrained by the personality state vector, and generates realistic text responses that conform to the established personality settings, emotional states, and scenario requirements.

[0006] Furthermore, the reasoning generation module is communicatively connected to a personality evolution control module. The personality evolution control module is used to drive the PSV computing engine to calculate the personality state vector at the next moment after each round of dialogue, based on historical interaction records, memory updates, and changes in personality characteristics, thus forming a continuous personality state chain.

[0007] Furthermore, the acquisition process of the sensing input module includes: Receive and perform initial cleaning of the user-input dialogue text; After collecting user voice signals and performing noise reduction and frame segmentation, prosodic features including pitch, intensity, speech rate, and rhythm are extracted. Based on a pre-defined micro-expression recognition model, extract changes in the user's facial expressions; The system acquires at least one physiological indicator from the user's heart rate, skin conductance, and electroencephalogram (EEG) segments using sensors. The data streams from the different modalities mentioned above are aligned and packaged based on a unified timestamp to generate multimodal data packets with time-series markers, which are then output to the parsing kernel module.

[0008] Furthermore, the process of extracting comprehensive context information from the parsing kernel module includes: It receives time-stamped multimodal data packets from the sensing input module, performs deep decoding on text, speech, visual and physiological signals, and performs cross-modal correlation on the decoded information to form a feature sequence under a unified time base; An attention-based fusion model is adopted to perform deep fusion of aligned cross-modal feature sequences. This unit identifies and resolves the consistency and conflict of information between modalities, and weights and aggregates complementary information to generate a multimodal joint feature representation. Abstract reasoning is performed based on multimodal joint feature representation. By combining dialogue history and current joint features, the user's current direct request, deep purpose or potential intention category is inferred. The user's current emotional state is quantified by integrating voice emotion vector, facial expression category intensity and physiological indicators. Based on the historical data of the current dialogue user and agent, the emotional tone of the current conversation and the interaction mode, an interactive context representation including participant roles, relationship status, conversation goals and scenario constraints is dynamically constructed. The semantic intent, emotional state assessment results, and interaction context representation are packaged and encapsulated into a structured comprehensive contextual information object.

[0009] Furthermore, the role layer in the personality control structure module is implemented in the form of a knowledge graph and a set of rule constraints, which defines the agent's social identity, ethical principles, core goals, and inviolable belief boundaries; the trait layer is represented by a multi-dimensional vector space model, with each dimension corresponding to a personality trait. Through long-term training data fitting or explicit parameter configuration, the stable baseline values ​​of the agent in each personality dimension are determined; the deductive layer has a built-in policy engine that selects suitable interaction scripts, emotional response intensity coefficients, and language style templates from a pre-set policy library based on the real-time interaction context labels, emotional state index, and semantic intent tone of the current conversation obtained from the parsing kernel module.

[0010] Furthermore, the calculation logic for the personality state vector in the PSV computing engine includes: Obtain the basic personality framework vector from the role layer; obtain the personality trait baseline vector from the trait layer; obtain the current dynamic adjustment vector from the deductive layer; obtain the long-term memory vector and short-term context vector from the memory module; obtain the real-time input stimulus vector from the parsing kernel module; By using a pre-defined feature mapping network, the above input vectors are uniformly mapped to a common feature space, and the importance of features from different sources is evaluated and weighted. Among them, the basic personality framework vector of the role layer participates in the weighting as the highest priority personality constraint vector. The weighted and fused features are processed and compressed using a nonlinear transformation network to output a personality state vector.

[0011] Furthermore, the construction process of the large language model in the PSV injection interface includes: Training on several general text corpora yields basic language comprehension and generation capabilities; The model was fine-tuned under supervision using a dataset containing specific character backgrounds, personality trait descriptions, and corresponding dialogue examples, enabling the model to learn to map personality descriptions to corresponding language styles and content preferences. Construct training data consisting of triples including input text, corresponding personality state vectors, and target output text. Train the model on this dataset so that the model learns to generate output text that conforms to the personality state represented by the personality state vector given the input text and personality state vector. Establish a mapping relationship from the personality state vector space to the language generation space.

[0012] Furthermore, the path for injecting the personality state vector into the large language model through the PSV injection interface includes: Prefix injection path: The personality state vector is decoded into a natural language description by the adaptation network, which is used as a system prompt prefix and concatenated with the user input before being input into the model's encoder; Attention bias injection path: The personality state vector is mapped to a set of attention bias vectors or scalars through the adaptation network. Learnable biases are applied to the keys or values ​​of the self-attention mechanism at each layer of the model, thereby implicitly guiding the model to pay attention to semantic patterns related to the current personality state. Original probability distribution bias injection path: The personality state vector is mapped to a bias vector on the vocabulary through the adaptation network. This vector is added to the original probability distribution vector output by the model decoder, thereby directly adjusting the probability distribution of the next word and increasing or decreasing the generation probability of words that reflect specific personality traits, emotions or values.

[0013] Furthermore, the proposed levels in the main control kernel module include familiarity level, déjà vu level, and soul connection level; their definition and evaluation process includes: Familiarity Level: When there is little known information about the interactive object and the modeling completeness is lower than the preset first threshold, a mimic response is generated based on the basic definition of the role layer and the baseline parameters of the trait layer in the personality control structure module. Familiarity Level: When the known information of the interactive object increases and the modeling completeness reaches the preset first threshold but is lower than the second threshold, based on the familiarity level, the short-term conversation history and key personalized information of the object stored in the memory module are called to generate a mimic response. Déjà vu level: When the known information of the interactive object is rich and the modeling completeness reaches the preset second threshold but is lower than the third threshold, based on the familiarity level, the long-term plot memory and relationship evolution data related to the object in the memory module are deeply integrated, and the deductive layer dynamically selects a highly adaptable strategy based on the current complex situation and deep relationship state to generate a mimic response. Soul Sense Level: When the known information of the interactive object is sufficient and the modeling completeness reaches or exceeds the preset third threshold, based on the déjà vu level, the function of the personality evolution control module is fully activated, so that the personality state vector can dynamically evolve and iterate based on the rich interaction history with the object, and generate mimicry response.

[0014] Furthermore, the memory module adopts a hierarchical storage architecture, including: A short-term memory store, in the form of a queue or sliding window, contains detailed context of the most recent rounds of dialogue for immediate reasoning. Long-term memory stores important events, emotional moments, and key personal information disclosed by users during interactions with specific users, in the form of a vector database or graph structure. Semantic memory stores the agent's own world knowledge, domain expertise, and general patterns abstracted from interactions; Through a preset index and retrieval interface, relevant memory fragments are quickly retrieved and activated from long-term memory based on the current personality state vector and conversation context, and provided as part of the input to the PSV computing engine.

[0015] Furthermore, multiple intelligent agents can exchange states and interact collaboratively through their respective personality state vectors.

[0016] (III) Beneficial Effects Compared with the known prior art, the technical solution provided by this invention has the following beneficial effects: Through a hierarchical personality control structure consisting of role layers, trait layers, and deductive layers, combined with the dynamic fusion calculation of the PSV computing engine, a real-time personality state vector is provided for large languages. This ensures that the agent's values, personality traits, and expression strategies are systematically constrained and guided each time a response is generated. This fundamentally suppresses the personality drift and style mutation problems common in traditional methods. The personality state vector evolves continuously based on the interaction history, enabling the agent's personality to undergo subtle changes and growth based on experience, just like a real person. This achieves dynamic adaptability on a stable baseline and enhances the credibility of the simulation.

[0017] By incorporating long-term memory, short-term context, and real-time multimodal stimuli into the calculation of personality state vectors, the agent's emotional responses and interaction strategies are not only based on the input of the current moment, but also deeply integrate the relationship information, user preferences, and shared experiences accumulated in past interactions, demonstrating emotional continuity. This allows the interaction to transcend the limitations of a single conversation and establish a human-computer interaction experience with temporal depth and relational depth.

[0018] By decomposing abstract requirements for personality, emotion, and consistency into calculable parameters and operable processes, the proposed four-level standard of familiarity, déjà vu, and soul provides a clear indicator system for measuring and optimizing the degree of anthropomorphism of intelligent agents. This makes the development and debugging of the system more goal-oriented, provides configuration and evaluation benchmarks for the differentiated needs of different application scenarios for the depth of personality of intelligent agents, and standardizes the control process of personality mimicry. Attached Figure Description

[0019] To more clearly illustrate the technical solutions in the embodiments of the present invention or the prior art, the drawings used in the description of the embodiments or the prior art will be briefly introduced below. Obviously, the drawings described below are only some embodiments of the present invention. For those skilled in the art, other drawings can be obtained based on these drawings without creative effort.

[0020] Figure 1 This is a schematic diagram of the framework of the present invention; Figure 2 This is a diagram illustrating the three-layer dual-core architecture of this invention. Figure 3 This is a schematic diagram illustrating the grading of the realism of the personification in this invention; Figure 4 This is a schematic diagram of PSV dynamically driving LLM in this invention.

[0021] The labels in the diagram represent: 1. Perception Input Module; 2. Parsing Kernel Module; 3. Main Control Kernel Module; 4. Personality Control Structure Module; 5. Memory Module; 6. PSV Computation Engine; 7. PSV Injection Interface; 8. Reasoning Generation Module; 9. Personality Evolution Control Module. Detailed Implementation

[0022] To make the objectives, technical solutions, and advantages of the embodiments of the present invention clearer, 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 some, not all, of the embodiments of the present invention. All other embodiments obtained by those skilled in the art based on the embodiments of the present invention without creative effort are within the scope of protection of the present invention.

[0023] The present invention will be further described below with reference to embodiments.

[0024] Example 1: This example describes an emotion mimicry control system for AI intelligent agents, such as... Figure 1 and Figure 2 As shown, it includes: The input perception module 1 is used to collect and preprocess the user's input dialogue text, voice tone, facial expressions and physiological signals, and convert them into parsable structured data. The acquisition process includes: receiving and initially cleaning the user's input dialogue text; acquiring the user's voice signal, performing noise reduction and frame segmentation, and extracting prosodic features including pitch, intensity, speech rate, and rhythm; extracting user facial expression changes based on a preset micro-expression recognition model; acquiring at least one physiological indicator from the user's heart rate, skin conductance response, and electroencephalogram (EEG) segments through sensors; aligning and packaging the data streams from the above different modalities based on a unified timestamp to generate a multimodal data packet with time-series markers, and outputting it to the parsing kernel module 2.

[0025] Parsing kernel module 2 is used to extract comprehensive contextual information from structured data, including semantic intent, emotional state, and interaction context; the process of extracting comprehensive contextual information includes: The system receives time-stamped multimodal data packets from the perception input module 1 and performs deep decoding on text, speech, visual, and physiological signals: text information is processed by a semantic parsing model to extract literal meaning and potential intent; speech information is processed by an acoustic model and an emotional speech recognition model to extract prosodic features and map them into emotional state vectors and speaking style labels; visual information is processed by a preset expression recognition model and a posture analysis model to extract expression categories, intensity, and body language intent; physiological signals are processed by signal processing and pattern recognition to map them into physiological indicators of emotional arousal and valence; the decoded information is then cross-modal correlated to form a feature sequence under a unified time base. An attention-based fusion model is used to deeply fuse aligned cross-modal feature sequences. This unit identifies and resolves the consistency and conflict of information between modalities, such as the difference between facial expressions and speech emotions when forcing a smile, and weights and aggregates complementary information to generate a multimodal joint feature representation. Abstract reasoning is performed based on multimodal joint feature representation. By combining dialogue history and current joint features, the user's current direct request, deep purpose or potential intention category is inferred. The user's current emotional state is quantified by integrating voice emotion vector, facial expression category intensity and physiological indicators. Based on the historical data of the current dialogue user and agent, the emotional tone of the current conversation and the interaction mode, an interactive context representation including participant roles, relationship status, conversation goals and scenario constraints is dynamically constructed. Semantic intent, emotional state assessment results, and interaction context representation are packaged and encapsulated into a structured, comprehensive contextual information object.

[0026] The main control kernel module 3 acts as the core scheduling hub, responsible for receiving comprehensive context information and coordinating the operation of various functional modules. Based on the depth of user information modeling, it defines and evaluates several mimicry levels; these levels include familiarity level, déjà vu level, and soul level; such as... Figure 3 As shown:

[0027] Its definition and evaluation process include: Familiarity Level: When the known information and modeling completeness of the interactive object are lower than the preset first threshold, based on the basic definition of the role layer and the baseline parameters of the trait layer in the personality control structure module 4, a mimicry response that conforms to the basic identity setting and universal behavioral guidelines is generated to ensure the basic compliance and style stability of the personality response. The intelligent agent can only make general templated responses based on the basic role layer information. When the input information only contains intuitive "form" such as visual, sound, and communication style A1-class charm information, the system activates body and voice cloning to generate visual and auditory mimicry output with basic similarity. Using image and sound cloning, it feels as if you have seen the other person. Familiarity Level: When the known information and modeling completeness of the interactive object reach the preset first threshold but are below the second threshold, based on the familiarity level, the system calls upon the object's short-term conversation history and key personalized information stored in memory module 5 to generate a mimicry response that is related to the specific dialogue context and reflects preliminary individual differentiation. The agent combines trait layer information and some short-term memory, and can exhibit stable personality traits and a certain continuity of dialogue history. When the input information further includes A2-type charm information such as philosophical views and beliefs of feeling "meaning", as well as B1-type market knowledge such as business and market descriptions of enterprises or individuals, the system injects specific viewpoint expressions into the role layer and the interpretation layer, producing a familiarity output with the same viewpoints and impressions. Expressions with the same viewpoints and ideas make people feel "very similar to someone they know". Déjà vu level: When the known information and modeling completeness of the interactive object reach the preset second threshold but are below the third threshold, based on the familiarity level, the system deeply integrates long-term plot memories and relationship evolution data related to the object in memory module 5. The deductive layer dynamically selects a highly adaptable strategy based on the current complex situation and deep relationship state to generate a mimicry response with deep situational fit and cognitive coherence. The agent deeply integrates user preferences, shared experiences, and dynamic strategies from long-term memory, and the response has scene adaptability and personalized cognitive depth that reach the preset threshold. When the input information further includes B2-type knowledge source information such as academic evidence and insights, the system activates comprehensive knowledge modeling and semantic alignment to replicate the subject's "insights and experiences," generating a déjà vu output that makes the user feel as if they are talking to the person in the same tone. It replicates the user's own "insights / speech," giving the user the illusion of talking to the person in the same tone. Soul Sense Level: When the known information and modeling completeness of the interactive object reach or exceed the preset third threshold, based on the déjà vu level, the function of the personality evolution control module 9 is fully activated, enabling the personality state vector to dynamically evolve and iterate based on the rich interaction history with the object. This generates a high-order mimicry response that not only deeply fits the history and context but also reflects the continuous evolution of personality traits, emotional logic consistency, and unique insights, demonstrating a high degree of emotional authenticity and cognitive consistency. When the input information fully covers the above A1, A2, B1, and B2 types of information, and customized rules based on specific task capabilities, deep psychological trait tendencies, and core rhetorical feature library of golden sentences / metaphors / allegoricals are introduced for deep modeling training, the system achieves full parameter decoupling and perfect synergy of the three-layer architecture, endowing the intelligent agent with deep expertise and mental soul, as if it has a deeper level of familiarity, like "meeting the real person". The first, second, and third thresholds are distributed from low to high.

[0028] The personality control structure module 4 is used to construct the role layer, trait layer, and deductive layer. The role layer stores and defines the agent's identity, core values, basic beliefs, and long-term mission. This role layer includes a high-weight personality constraint parameter matrix based on the target agent's values, belief system, and principle boundaries. This parameter matrix has higher priority than the constraint weights of the trait layer and deductive layer in the personality state vector generation process. The role layer parameters participate in the PS operation as the basic personality constraint vector and continuously constrain the subsequent personality state vector evolution process. The trait layer defines the agent's stable personality characteristics, thinking patterns, and expression habits. The deductive layer dynamically adjusts the interaction strategy and emotional expression based on the current interaction scenario, object relationships, and conversation goals. The agent's social identity, ethical principles, core goals, and inviolable belief boundaries are defined in the role layer, which is implemented in the form of a knowledge graph and a set of rule constraints. The trait layer is represented by a multi-dimensional vector space model, with each dimension corresponding to a personality trait. Through long-term training data fitting, stable baseline values ​​of the agent in various personality dimensions such as extraversion, agreeableness, neuroticism, openness, and conscientiousness are determined. The deductive layer has a built-in policy engine that selects appropriate interaction scripts, emotional response intensity coefficients, and language style templates from a pre-set policy library based on the real-time interaction context labels, emotional state index, and semantic intent tone of the current conversation obtained from the parsing kernel module 2. It also includes a subjective emotional state machine to record and update the agent's own emotional state in real time.

[0029] Memory Module 5 is used to store and manage the long-term relationship memory between the current agent and the user, short-term conversation context, and new personality characteristic data evolved from interactions; Memory Module 5 adopts a hierarchical storage architecture, including: A short-term memory store, in the form of a queue or sliding window, contains detailed context of the most recent rounds of dialogue for immediate reasoning. Long-term memory stores important events, emotional moments, and key personal information disclosed by users during interactions with specific users, in the form of a vector database or graph structure. Semantic memory stores the agent's own world knowledge, domain expertise, and general patterns abstracted from interactions; Through a preset index and retrieval interface, relevant memory fragments are quickly retrieved and activated from long-term memory based on the current personality state vector and conversation context, and provided as part of the input to the PSV computing engine 6.

[0030] The PSV computation engine 6 is used to dynamically fuse the role layer definition, trait layer parameters, and deductive layer strategies from the personality control structure module 4, the memory information from the memory module 5, and the comprehensive context information input in real time from the kernel module 2 to calculate and generate the personality state vector PSV at the current moment. The calculation logic of the personality state vector includes: Obtain the basic personality framework vector from the role layer; obtain the personality trait baseline vector from the trait layer; obtain the current dynamic adjustment vector from the deductive layer; obtain the long-term memory vector and short-term context vector from memory module 5; obtain the real-time input stimulus vector from parsing kernel module 2; By using a pre-defined feature mapping network, the above input vectors are uniformly mapped to a common feature space. The importance of features from different sources is evaluated and weighted. Among them, the basic personality framework vector of the role layer participates in the weighting as the highest priority personality constraint vector. Its attention weight is dynamically determined by the dynamic vector of the current scene's deductive layer and the short-term context vector. This means that in conflict situations, the weight of the basic vector of the role layer may increase to uphold principles, while in relaxed social scenarios, the weight of the real-time stimulus vector and the baseline vector of the trait layer may increase to showcase personality and on-the-spot reactions. The weighted and fused features are processed and compressed through a nonlinear transformation network to output a personality state vector. This vector comprehensively encodes the agent's cognitive tendencies, emotional state, intentions, and expression preferences at the current moment.

[0031] PSV injection interface 7 is used to inject the calculated Personality State Vector (PSV) as a key control parameter into the inference generation process of the preset Large Language Model (LLM) in real time; the construction process of the Large Language Model, such as Figure 4 As shown, it includes: Training on several general text corpora yields basic language comprehension and generation capabilities; The model was fine-tuned under supervision using a dataset containing specific character backgrounds, personality trait descriptions, and corresponding dialogue examples, so that the model learned to map personality descriptions to corresponding language styles and content preferences. Construct training data consisting of triples including input text, corresponding personality state vectors, and target output text. Train the model on this dataset so that the model learns to generate output text that conforms to the personality state represented by the personality state vector given input text and personality state vector. Establish a mapping relationship from the personality state vector space to the language generation space. The paths for injecting personality state vectors into large language models include: Prefix injection path: The personality state vector is decoded into a natural language description by the adaptation network, which is used as a system prompt prefix and concatenated with the user input before being input into the model's encoder; Attention bias injection path: The personality state vector is mapped to a set of attention bias vectors or scalars through the adaptation network. Learnable biases are applied to the keys or values ​​of the self-attention mechanism at each layer of the model, thereby implicitly guiding the model to pay attention to semantic patterns related to the current personality state. Original probability distribution bias injection path: The personality state vector is mapped to a bias vector on the vocabulary through the adaptation network. This vector is added to the original probability distribution vector output by the model decoder, thereby directly adjusting the probability distribution of the next word and increasing or decreasing the generation probability of words that reflect specific personality traits, emotions or values. The aforementioned adaptation network is jointly trained with the large language model during the personality state vector conditionalization training phase to ensure the effectiveness of the personality state vector to control signal conversion, with at least one path selected. Analyzing the content of each round of dialogue, valuable information, such as user-revealed personal preferences, important events, and peak emotional interactions, is categorized and stored in the long-term or short-term memory of memory module 5, which may trigger minor adjustments to the trait layer parameters. When generating a new personality state vector, the personality state vector of the previous moment is used as one of the important inputs to ensure the smoothness and continuity of personality state evolution. Verification is performed to monitor the consistency between the generated content and the personality state vector. If the deviation exceeds a threshold, it is fed back to PSV computing engine 6 or personality control structure module 4 for parameter fine-tuning according to a preset scale to suppress personality drift.

[0032] The reasoning generation module 8 is used to receive reasoning instructions and context information from the PSV injection interface 7 constrained by the personality state vector, and generate realistic text responses that conform to the predetermined personality settings, emotional states, and scenario requirements. The reasoning generation module 8 can adapt to large language models with different architectures. The PSV injection interface 7 provides a standardized calling protocol, allowing upstream applications or platforms to call this system to obtain the current personality state vector before calling the large language model service, and send it to the large language model along with the user query, thereby achieving the decoupling and reuse of personality mimicry capabilities.

[0033] The reasoning generation module 8 is connected to the personality evolution control module 9. The personality evolution control module 9 is used to drive the PSV computing engine 6 to calculate the personality state vector at the next moment after each round of dialogue, based on historical interaction records, memory updates and changes in personality characteristics, to form a continuous personality state chain, so as to ensure the consistency of personality, the continuity of emotions and the stability of behavioral logic in long-term interaction.

[0034] Multiple intelligent agents can exchange states and interact collaboratively through their respective personality state vectors.

[0035] Example 2: "Mental Digital Migration" of a Real-Person Expert as a B-End User, using a teacher specializing in family education as an example: 1.1 Initialization and Multi-level Data Cascading Mechanism of Digital Human's "Mental Digital Migration": In this embodiment, to effectively collect abstract implicit knowledge descriptions such as real-world consciousness and concepts, the approach first involves converting the subjective consciousness, implicit cognition, and domain expertise of real-world experts (such as education and psychology experts) into a parameter matrix that can be computed at the system's underlying level. This data cascading mechanism specifically involves the information extraction and transformation through the following three layers of structured data files in a descending manner: (1) Third-order documents (collecting original subjective expressions): Through interactive data collection interfaces (such as voice interview recording transcription and open natural language input forms), we collect raw data from real experts (teachers) who express themselves spontaneously without system architecture constraints. Although the data at this stage is characterized by high noise and unstructured nature, it represents the most authentic and voluntary original statements of the teachers' needs.

[0036] (2) Second-order document (structured summary matrix): Using semantic parsing operators in Natural Language Processing (NLP), feature cleaning and clustering are performed on the third-order documents. A summary is extracted from the original corpus to form a second-order topic matrix with preliminary topological relationships, which is used to eliminate redundant information in the original corpus. This matrix can be reused later, eliminating the need for repeated interviews with real people due to changes in the digital persona.

[0037] (3) First-order file (core meta-parameter aggregation): To further refine the second-order files in line with the digital human persona goals, semantic kernel compression and meta-parameter labeling are performed, ultimately forming a "first-order core file array" (in this embodiment, specifically instantiated as two files: "Knowledge Modeling I: Concept / Impression Cloning" and "Knowledge Modeling II: Insight / Reflection Cloning"). The first-order files directly define the weight boundaries and logical rules of each module in the underlying control architecture of the digital human, serving as direct precursor data for the generation of the system's prompt word matrix.

[0038] 1.2 Knowledge Modeling and Prompt Matrix Generation Pipeline Based on a "Three-Layer Architecture": Once the teacher has extracted the implicit knowledge (completed the "first-order file"), the knowledge molding pipeline is started. The core meta-parameters in the first-order file are precisely mapped to the "role layer," "trait layer," and "deductive layer" of this invention, and finally exported as a formatted configuration file. The specific execution steps are as follows: Step 1: Subjective Prior Belief Matrix Extraction and Role Layer Configuration (Corresponding to First-Order File I) The first-order document "Knowledge Modeling I," compiled from the information provided by the teacher, extracts the core metadata at the expert's intellectual level. Identity Positioning (R1): Transform the teacher's statement of "a mentor who is deeply rooted in Chinese family culture and does not play the role of a savior" into ethical hierarchy constraints for the role.

[0039] Values ​​and Beliefs (R2, R3): These transform pre-defined abstract ideas into a subjective prior belief matrix for the control system. This matrix defines the weighted derivative function for the digital human to determine the positive or negative attributes of any input event during subsequent reasoning.

[0040] Long-term mission (R5): Set the educational endpoint as "cultivating good young people of the new era who have ideals, dare to take responsibility, can endure hardship, and are willing to work hard" as the objective function with global convergence.

[0041] Step 2: Construction of general and specific theoretical maps and configuration of trait layers (corresponding to first-order file II) The analysis of another first-order document, "Knowledge Modeling II," involves a dual-track modeling of the teacher's specialized knowledge. This prevents the illusion of generalization created by a large language model, while ensuring that the professional content output aligns with the teacher's field, doctrine, and arguments. General Theoretical Network (T3-Thinking Mode): Taking teachers as an example, the theory is transformed into a logical constraint tree of thought. When a user asks a question involving family conflict, the system will prioritize searching this theoretical tree, rather than calling the probability-based word chain of the larger model.

[0042] Domain-specific self-developed theory (T4-Decision Preference): The system incorporates the teacher's original work, "The Underlying Logic of Family Education: Solving the Dilemma of Adolescent Education Pathogenesis," into decision preferences. Based on this, the system establishes a "Depathogenesis Diagnosis Preference Operator," inherently favoring "seeking relationships and addressing community deficiencies" in behavioral decisions, rather than the "diagnosis of mental illness" in Western psychology. This ensures that the digital human's stable consistency reflects the teacher's knowledge characteristics, differing from the responses generated by general probabilistic models.

[0043] Step 3: Instantiate the high-frequency rhetorical feature primitive library (corresponding rhetorical database construction) The semantic and lexical features extracted from teachers’ daily conversations are structured and included in the “T6 Expression Habit Parameter Library” of the trait layer. The recurrence probability threshold of specific rhetorical primitives can be set in the “P7 Expression Constraint Module” of the deductive layer to ensure that the generated dialogue has a very high degree of interpersonal déjà vu (B-level or higher personality mimicry realism).

[0044] Step 4: Generation and runtime loading of the unified semantic space configuration file The parameter matrices, theoretical graphs, and rhetorical primitives generated in steps 1 to 3 are globally semantically aligned and compiled into a structured Markdown configuration file ("Teacher Digital Human"). This file serves as the physical carrier of the Dynamic System Prompt for the digital human. It is directly loaded into the memory state machine by the second kernel (main control kernel) when the system starts. When a dialogue is triggered, the system will connect this prompt to the large model to generate response text with teacher characteristics.

[0045] 1.3 Dynamic Running Examples and Demonstration of [PSV] Formula Calculation To further illustrate the operational mechanism of this invention in practical applications, the following demonstrates the dynamic calculation process when the operational mechanism of this invention (loaded with the aforementioned teacher's mental model) faces a specific event: [Scene ~ Specific Event Input] An anxious mother (user) inputs audio and text into the teacher's digital human interface via the perception kernel: "Hello teacher, my child is 15 years old. He stays at home playing games all day and doesn't want to go to school. The school teacher says he may have severe depression and internet addiction and advises us to take him out of school to see a psychologist. I feel like the sky is falling, and I don't know what to do." [Dynamic calculation and workflow in the system background] Multimodal Perceptive Kernel (First Kernel) Analysis: [Input] Semantic analysis: The child is reluctant to study, faces a crisis of dropping out of school, and is diagnosed with depression by an authoritative institution (teacher).

[0046] User emotion detection: extreme anxiety, despair, blind obedience to authority.

[0047] The main control kernel (secondary kernel) calls the [PSV] formula for calculation: PSV = f (Role, Trait, Performance, Memory, Input) [Role] Injection: Triggers R3 (Belief: Emotions must be seen first) and R2 (Values: Be realistic and eliminate anxiety and selfishness), initiating the underlying mindset of "de-pathologization".

[0048] [Trait] Injection: Triggers T4 (Decision Preference: Invokes the exclusive self-built theory 4 - the underlying logic of family education) and T6 (Expression Habits: Invokes the exclusive terms "pathological illusion" and "lack of community function".

[0049] [Performance] (Deductive Layer) Dynamic Adaptation: P1 / P2 identification: The current object is [an extremely anxious parent], and the scenario is [crisis counseling].

[0050] P5 / P6 Interaction Strategy Update: Set the current interaction goals as [breaking down authority anxiety, emotional empowerment, and research-oriented].

[0051] P4 Emotional State Machine Transfer: The digital human's own emotional state is locked as [highly calm, firm, critical thinking].

[0052] PSV generation and large model guidance: The system derives the current high-dimensional vector. This vector, multiplied into the large model's decoding layer as a dynamic control matrix, strongly regulates the generic comforting phrases that the large model might automatically generate (e.g., "I'm sorry to hear this news; depression requires following medical advice..."), significantly increasing the probability of generating tokens that align with the teacher's spirit and essence.

[0053] [System's final simulated output performance] In the teacher's voice: "Mom, please take a deep breath and put aside your panic that the sky is falling (R3 / P6 empathy and acceptance). We must first look at this matter realistically. Don't blindly believe authority and simply 'ill' your child just because you see them backing down (self-constructed theory 4 / T4 diagnostic bias)."

[0054] Based on my extensive past research, your child's current staying at home and internet addiction are not fundamentally due to a brain disorder, but rather a disruption in his "social functioning and relational support" within the family and neighborhood community (Summary of Second-Order Document Theory / T3 Thinking Mode). He is using this extreme measure of refusing to go to school to sound the alarm about ineffective communication within the entire family system.

[0055] If parents don't investigate the underlying relationship deficiencies behind their child's behavior (General Theory 2: The Mass Line), but instead view it through a pathological lens, that's a real educational misconception. Are you willing to cooperate with me, like a true researcher, and calmly examine the underlying relationship hierarchy within your family? (R1 Mentor Positioning / A-Level Soulful Output) The above embodiments fully demonstrate that, through the multi-level data cascading and "three-layer dual-core" architecture of the present invention, the system can successfully break free from the cold and generalized role of traditional AI as a mouthpiece, and interpret the expert's "soul cognition and value proposition" in every highly challenging dialogue.

[0056] In summary, this invention, through the deep integration of multimodal information perception and hierarchical personality control structure, constructs an evolvable personality state vector generation mechanism. This enables AI agents to transcend the limitations of single text interaction, comprehensively understand the user's semantic, emotional, and contextual information, and generate highly human-like responses based on their inherent role settings, stable traits, and dynamic deductive strategies. By injecting the real-time calculated personality state vector as the core control parameter into the reasoning process of the large language model, the system effectively constrains and guides the direction of generated content, ensuring the consistency of personality performance, the coherence of emotional responses, and the stability of behavioral logic in long-term interactions. This invention reduces common problems in traditional large language models, such as personality drift, style mutation, and value inconsistency. The proposed mimicry effect grading standard provides a quantifiable and optimizable evaluation system for personality modeling, and promotes the evolution of AI agents from basic functional responses to interactions with deep personalization and emotional realism.

[0057] The above embodiments are only used to illustrate the technical solutions of the present invention, and are not intended to limit it. Although the present invention has been described in detail with reference to the foregoing embodiments, those skilled in the art should understand that modifications can still be made to the technical solutions described in the foregoing embodiments, or equivalent substitutions can be made to some of the technical features. Such modifications or substitutions will not cause the essence of the corresponding technical solutions to deviate from the spirit and scope of the technical solutions of the embodiments of the present invention.

Claims

1. An emotion mimicry control system for AI intelligent agents, characterized in that, include: The input perception module is used to collect and preprocess user input dialogue text, voice tone, facial expressions and physiological signals, and convert them into parsable structured data. The parsing kernel module is used to extract comprehensive contextual information from structured data, including semantic intent, emotional state, and interaction context. The main control kernel module serves as the core scheduling hub, responsible for receiving comprehensive context information and coordinating the operation of various functional modules. Based on the depth of user information modeling, it defines and evaluates several mimicry levels. The personality control structure module is used to construct a role layer, a trait layer, and an interpretation layer. The role layer is used to store and define the agent's identity, core values, basic beliefs, and long-term mission. The trait layer is used to define the agent's stable personality traits, thinking patterns, and expression habits. The interpretation layer is used to dynamically adjust the interaction strategy and emotional expression method according to the current interaction scenario, object relationship, and conversation goal. The memory module is used to store and manage the long-term relationship memory between the current agent and the user, the short-term conversation context, and the new personality characteristic data evolved from the interaction; The PSV computing engine is used to dynamically fuse the role layer definition, trait layer parameters and deductive layer strategies of the personality control structure module, the memory information of the memory module, and the comprehensive context information input in real time from the kernel module to calculate and generate the personality state vector at the current moment. The PSV injection interface is used to inject the calculated personality state vector as a key control parameter into the inference generation process of the preset large language model in real time. The reasoning generation module is used to receive reasoning instructions and context information from the PSV injection interface constrained by the personality state vector, and generate realistic text responses that conform to the predetermined personality settings, emotional states and scenario requirements.

2. The emotion mimicry control system for AI intelligent agents according to claim 1, characterized in that, The reasoning generation module is communicatively connected to the personality evolution control module. The personality evolution control module is used to drive the PSV computing engine to calculate the personality state vector at the next moment after each round of dialogue, based on historical interaction records, memory updates, and changes in personality characteristics, thus forming a continuous personality state chain.

3. The emotion mimicry control system for AI intelligent agents according to claim 1, characterized in that, The acquisition process of the sensing input module includes: Receive and perform initial cleaning of the user-input dialogue text; After collecting user voice signals and performing noise reduction and frame segmentation, prosodic features including pitch, intensity, speech rate, and rhythm are extracted. Based on a pre-defined micro-expression recognition model, extract changes in the user's facial expressions; The system acquires at least one physiological indicator from the user's heart rate, skin conductance, and electroencephalogram (EEG) segments using sensors. The data streams from the different modalities mentioned above are aligned and packaged based on a unified timestamp to generate multimodal data packets with time-series markers, which are then output to the parsing kernel module.

4. The emotion mimicry control system for AI intelligent agents according to claim 1, characterized in that, The process of extracting comprehensive context information in the parsing kernel module includes: It receives time-stamped multimodal data packets from the sensing input module, performs deep decoding on text, speech, visual and physiological signals, and performs cross-modal correlation on the decoded information to form a feature sequence under a unified time base; An attention-based fusion model is adopted to perform deep fusion of aligned cross-modal feature sequences. This unit identifies and resolves the consistency and conflict of information between modalities, and weights and aggregates complementary information to generate a multimodal joint feature representation. Abstract reasoning is performed based on multimodal joint feature representation. By combining dialogue history and current joint features, the user's current direct request, deep purpose or potential intention category is inferred. The user's current emotional state is quantified by integrating voice emotion vector, facial expression category intensity and physiological indicators. Based on the historical data of the current dialogue user and agent, the emotional tone of the current conversation and the interaction mode, an interactive context representation including participant roles, relationship status, conversation goals and scenario constraints is dynamically constructed. The semantic intent, emotional state assessment results, and interaction context representation are packaged and encapsulated into a structured comprehensive contextual information object.

5. The emotion mimicry control system for AI intelligent agents according to claim 1, characterized in that, The personality control structure module implements the role layer in the form of a knowledge graph and a set of rule constraints. It defines the agent's social identity, ethical principles, core goals, and inviolable belief boundaries. The role layer includes a high-weight personality constraint parameter matrix based on the target subject's values, belief system, and principle boundaries. This parameter matrix has higher priority than the constraint weights of the trait layer and the deductive layer during the personality state vector generation process. The trait layer is represented by a multi-dimensional vector space model, with each dimension corresponding to a personality trait. Stable baseline values ​​for the agent in each personality dimension are determined through long-term training data fitting or explicit parameter configuration. The deductive layer has a built-in policy engine. Based on real-time interaction context labels, emotional state indices, and the semantic intent tone of the current conversation obtained from the parsing kernel module, it selects suitable interaction scripts, emotional response intensity coefficients, and language style templates from a pre-set policy library. It also includes a subjective emotional state machine to record and update the agent's own emotional state in real time.

6. The emotion mimicry control system for AI intelligent agents according to claim 1, characterized in that, The calculation logic of the personality state vector in the PSV computing engine includes: Obtain the basic personality framework vector from the role layer; obtain the personality trait baseline vector from the trait layer; obtain the current dynamic adjustment vector from the deductive layer; obtain the long-term memory vector and short-term context vector from the memory module; obtain the real-time input stimulus vector from the parsing kernel module; By using a pre-defined feature mapping network, the above input vectors are uniformly mapped to a common feature space, and the importance of features from different sources is evaluated and weighted. Among them, the basic personality framework vector of the role layer participates in the weighting as the highest priority personality constraint vector. The weighted and fused features are processed and compressed using a nonlinear transformation network to output a personality state vector.

7. The emotion mimicry control system for AI intelligent agents according to claim 1, characterized in that, The construction process of the large language model in the PSV injection interface includes: Training on several general text corpora yields basic language comprehension and generation capabilities; The model was fine-tuned under supervision using a dataset containing specific character backgrounds, personality trait descriptions, and corresponding dialogue examples, enabling the model to learn to map personality descriptions to corresponding language styles and content preferences. Construct training data consisting of triples including input text, corresponding personality state vectors, and target output text. Train the model on this dataset so that the model learns to generate output text that conforms to the personality state represented by the personality state vector given input text and personality state vector. Establish a mapping relationship from the personality state vector space to the language generation space. The path for injecting personality state vectors into the large language model via the PSV injection interface includes: Prefix injection path: The personality state vector is decoded into a natural language description by the adaptation network, which is used as a system prompt prefix and concatenated with the user input before being input into the model's encoder; Attention bias injection path: The personality state vector is mapped to a set of attention bias vectors or scalars through the adaptation network. Learnable biases are applied to the keys or values ​​of the self-attention mechanism at each layer of the model, thereby implicitly guiding the model to pay attention to semantic patterns related to the current personality state. Original probability distribution bias injection path: The personality state vector is mapped to a bias vector on the vocabulary through the adaptation network. This vector is added to the original probability distribution vector output by the model decoder, thereby directly adjusting the probability distribution of the next word and increasing or decreasing the generation probability of words that reflect specific personality traits, emotions or values.

8. The emotion mimicry control system for AI intelligent agents according to claim 1, characterized in that, The proposed levels in the main control kernel module include familiarity level, déjà vu level, and soul level. The assessment of these mimicry levels is based on a combination of image style information, value and belief information, domain knowledge information, and knowledge source information. The definition and assessment process include: Familiarity Level: When the known information and modeling completeness of the interactive object are lower than the preset first threshold, a mimic response is generated based on the basic definition of the role layer and the baseline parameters of the trait layer in the personality control structure module. Familiarity Level: When the known information and modeling completeness of the interactive object reach the preset first threshold but are lower than the second threshold, based on the familiarity level, the short-term conversation history and key personalized information of the object stored in the memory module are called to generate a mimic response. Déjà vu level: When the known information and modeling completeness of the interactive object reach the preset second threshold but are lower than the third threshold, based on the familiarity level, the long-term plot memory and relationship evolution data related to the object in the memory module are integrated, and the deductive layer dynamically selects a highly adaptable strategy based on the current complex situation and deep relationship status to generate a mimic response. Soul Sense Level: When the known information and modeling completeness of the interactive object reach or exceed the preset third threshold, the personality evolution control module is activated based on the déjà vu level, so that the personality state vector evolves dynamically based on the interaction history with the object, generating a mimic response.

9. The emotion mimicry control system for AI intelligent agents according to claim 1, characterized in that, The memory module adopts a hierarchical storage architecture, including: A short-term memory store, in the form of a queue or sliding window, contains detailed context of the most recent rounds of dialogue for immediate reasoning. Long-term memory stores important events, emotional moments, and key personal information disclosed by users during interactions with specific users, in the form of a vector database or graph structure. Semantic memory stores the agent's own world knowledge, domain expertise, and general patterns abstracted from interactions; Through a preset index and retrieval interface, relevant memory fragments are quickly retrieved and activated from long-term memory based on the current personality state vector and conversation context, and provided as part of the input to the PSV computing engine.

10. The emotion mimicry control system for AI intelligent agents according to claim 1, characterized in that, Multiple intelligent agents can exchange states and interact collaboratively through their respective personality state vectors.

Images