A modeling method and system for cross-temporal emotion-behavior interaction generation

Abstract

The application belongs to the technical field of intelligent interaction of virtual roles, and discloses a modeling method and system for cross-time sequence emotion-behavior interaction generation, which comprises the following steps: step S1, definition and integration of emotion states; step S2, generation and integration of behaviors; step S3, adaptive mechanism based on feedback and real-time interaction modulation, the agent combines the internal state and the next behavior from the current executed behavior to realize adaptive interaction. The application adopts the above modeling method and system for cross-time sequence emotion-behavior interaction generation, proposes a cross-time sequence emotion modeling (CTEM) framework, effectively solves the problems of fragmentation of existing companion-type agent interaction and separation of emotion and behavior, and significantly improves the naturalness and companionship of virtual agents by coupling long-term behavior history and instant emotion expression.

Description

Technical Field

[0001] This invention relates to the field of intelligent interaction technology for virtual characters, and in particular to a modeling method and system for generating cross-temporal emotion-behavior interactions. Background Technology

[0002] With the rapid development of basic models, the application scenarios of conversational agents have gradually expanded from simple task execution to fields such as education, healthcare, and daily social interaction. Their role has also shifted from task assistant to continuous companion. Companion-type agents differ fundamentally from traditional task-oriented dialogue systems. Their core requirement is to maintain a coherent and emotionally engaging human-computer relationship throughout long-term interactions. They need to possess behavioral models that evolve gradually based on past interaction experiences, maintaining consistency in their behavior while appropriately responding to the user's emotions to meet the user's emotional needs during long-term companionship.

[0003] Psychological and neuroscience research has clearly shown that emotions are not isolated reactions, but rather a dynamic process that evolves over time. Their development is influenced by memory, cognitive appraisal, and situational factors. Furthermore, cognition and emotion are essentially inseparable, interacting and integrating with each other to continuously regulate an individual's attention, memory, and goal-oriented behavior. Motivation theory further demonstrates that an individual's needs and goals continuously drive behavior over time, providing a solid theoretical foundation for modeling the emotions and behaviors of companion agents.

[0004] However, current companion-type intelligent agent systems still have obvious technical defects, making it difficult to support natural and long-term companion-style human-computer interaction. This results in a fragmented user experience that lacks realism. The core problem is that existing systems neglect the coupled modeling of social behavior and internal emotional state across time scales, and generally treat behavior and emotion separately.

[0005] Specifically, among existing agent-based methods, some agent social simulation technologies can model the agent's memory and behavioral planning over time, achieving human-like behavioral continuity, but they lack internal emotion models, cannot depict the evolution of emotions in multiple interactions, and cannot reflect the reverse influence of emotions on subsequent behaviors. In addition, many dialogue systems rely on preset static personality or emotion settings, which can generate empathetic responses in local interactions, but such static settings are disconnected from the agent's real interaction experience and cannot reflect the shaping effect of past interactions on internal emotional states and the regulatory effect of emotional evolution on subsequent behaviors.

[0006] Furthermore, the few studies that attempt to model emotional changes are limited to emotional assessment at the level of a single turn or partial dialogue, failing to depict the dynamic process by which emotions are shaped over a long period by the agent's own behavior and user feedback. These technical shortcomings prevent existing companion-type agents from meeting the consistency and emotional coordination requirements of long-term companionship, making it difficult to achieve sustainable, high-quality companion-style human-computer interaction.

[0007] Therefore, there is an urgent need for a computational model that can explicitly couple long-term behavioral accumulation with real-time emotional expression, solve the problem of the separation of behavior and emotion in existing technologies, realize cross-temporal linkage between the emotional state and behavioral trajectory of intelligent agents, improve the naturalness, consistency and emotional coordination of companion-type human-computer interaction, and meet the actual needs of various fields for high-quality companion-type intelligent agents. Summary of the Invention

[0008] The purpose of this invention is to provide a modeling method and system for cross-temporal emotion-behavior interaction generation, construct a cross-temporal emotion modeling (CTEM) framework, effectively solve the problems of fragmented interaction and separation of emotion and behavior in existing companion-type intelligent agents, and significantly improve the naturalness and companionship of virtual intelligent agents by coupling long-term behavioral history with immediate emotional expression.

[0009] To achieve the above objectives, this invention provides a modeling method for cross-temporal emotion-behavior interaction generation, comprising the following steps: Step S1: Definition and integration of emotional states; Step S2: Generation and integration of behaviors; Step S3: Based on feedback-driven adaptive mechanisms and real-time interactive modulation, the agent combines its internal state with the next action derived from the current action to achieve adaptive interaction.

[0010] Preferably, in step S1, the definition and integration of emotional states are carried out as follows: Emotional state determines the future behavior and interaction style of an agent. Inspired by motivational psychology, time steps... The state is formally defined as the following tuple: ; in, The overall internal state; Physiological-emotional state; For motivation vectors; For memory; An adaptive personality; Physiological-emotional state These are used to characterize the three basic dimensions of physiological energy, emotional valence, and arousal, as shown below: ; in, Indicates current energy level; Indicates the degree of positivity / negativity of the current emotion; Indicates the degree of emotional activation; Motivational vectors This is used to represent three types of motivational drivers: physiological needs, psychological needs, and social needs, as shown below: ; in, Physiological needs refer to needs related to physical recovery and maintenance of a basic state; These are psychological needs, representing needs related to psychological growth, self-satisfaction, and cognitive exploration. Social needs refer to needs related to social relationships and emotional connections. memory It represents a time-indexed memory structure used to support cross-session organization and retrieval of semantic and sentiment context; Adaptive personality Indicating basic personality Above, based on the current emotional state Modulation is performed as follows: ; By injecting state-related tone labels into the model prompts, personality performance is dynamically adjusted; the state variables integrate four dimensions: emotion, motivation, memory, and personality, enabling the agent to maintain long-term consistency while dynamically adapting to the situation.

[0011] Preferably, in step S2, the generation and integration of behaviors are carried out as follows: Step S21: From the predefined set of behavior categories In the process, the set of behaviors at the current moment is obtained by sampling. This set covers five behavioral domains: physiological, work, leisure, social, and emotional. Step S22: The behavior set is used to organize the agent's behavior into three levels, including historically executed behaviors, currently executed behaviors, and planned future behaviors, as shown below: ; in, For a set of behaviors; For historical behavior; For the current action being performed; Actions for planning the future; Each behavior entry is represented as a tuple. ,in Indicates a specific action, This indicates the expected update of the state that the action represents; Step S23: Using the scoring function The function completes the behavior selection based on the current emotional state. The evaluation of candidate behaviors follows this process: First, the behavior is mapped to the motivation space, as shown below: ; in, Indicates specific behavior Feature representation in motivation space; Then, define the scoring function. As shown below: ; in, For specific behaviors in the scoring function Embedded representation in motivation space; For the current physiological energy Determined modulation weights; For the logistic function; This represents the dot product.

[0012] Preferably, in each behavioral choice process, CTEM first conducts future planning, when When empty, from the set Select the candidate behavior with the highest score and fill it in; Subsequently, from The sampling behavior is based on the latest score distribution and executed accordingly; Finally, based on the candidate behavior Filter the data to avoid duplication and maintain contextual consistency.

[0013] Preferably, in step S3, the real-time interactive modulation is as follows: The selected behavior is instantiated as a narrative or multimodal output by the base model and recorded in the behavior set. The agent interacts with it through an instant messaging interface; the communication style is determined by the current state. And dynamic modulation of real-world contexts; To maintain consistency and alleviate the limitations of the large language model context window, the system summarizes and categorizes dialogues by time and topic, thereby maintaining emotional consistency during the interaction.

[0014] Preferably, in step S3, to ensure trust and ethical safety, the intelligent agent implements a two-stage monitoring mechanism: Combining keyword screening and a multi-model classifier, the system detects extreme emotions, risk signals, or excessive dependence. Once such a situation is detected, the system dynamically injects safety constraints into the generated prompts, prioritizing reassuring responses, emphasizing user autonomy, and suggesting professional help when necessary. All outputs are confined within a safe space. This is to ensure both content safety and emotional safety.

[0015] Preferably, in step S3, based on the feedback-based adaptive mechanism, as interactions accumulate, the intelligent system learns to adapt. Rebalancing is performed to maintain long-term emotional stability; every action taken triggers a state update. Gradually adjust and ; Establish mandatory nighttime rest periods to restore energy and emotional balance, preventing system overload; simultaneously, gradually adjust personality state. The process of moving from unfamiliarity to familiarity allows for continuous evolution and deepens the interactive relationship.

[0016] Preferably, in step S3, the intelligent system updates its memory at the end of each day by introducing new interactive events, as shown below: ; in, For the updated memory; A new experience for the current time step.

[0017] A modeling system for cross-temporal emotion-behavior interaction generation is provided to implement the aforementioned modeling method for cross-temporal emotion-behavior interaction generation. The system includes: an environment and input signal module, an interaction content understanding module, an emotion state system, a behavior generation system, a memory and summary module, and a behavior performance module.

[0018] Preferably, the environment and input signal module is used to receive external signals in the environment in which the virtual character is located; the external signals in the environment include user input information, environmental events, system state changes and other external trigger information; this module transmits the acquired environmental signals to the interactive content understanding module for parsing; The interactive content understanding module is used to perform necessary semantic parsing, contextual analysis, and cues judgment on the environment and input signals, extract explicit or implicit semantic information, interactive intentions, and contextual information, and transmit the extracted information to the emotion state system and behavior generation system. The behavior generation system is further divided into a behavior selection part and a behavior execution result part. The behavior selection part generates candidate behaviors for the virtual character based on data from the emotional state system, the results of the interaction content understanding module, and historical information from the memory and summary module. The candidate behavior construction includes: setting different behavior preferences based on the agent's personality and preferences, and selecting different behaviors based on the behavior preferences; the agent plans future behaviors and adjusts the planned execution behaviors based on future execution preferences; The behavior execution result determines the final direction of the behavior; the emotional state system and the behavior generation system maintain a two-way coupling relationship; the emotional state system and the behavior generation system form a cross-temporal driving relationship, so that the emotional state can not only reflect the current behavior result, but also influence subsequent behavior decisions, thereby jointly promoting the cross-temporal state behavior evolution of the virtual character and achieving long-term matching and stability of emotions and behaviors; The memory and summary module is used to record the accumulated historical state information of the virtual character during the interaction process, including emotional changes, behavior execution results, and environmental feedback information. This module provides a reference for subsequent emotional state updates and behavior generation by storing and summarizing historical information. The behavior performance module is used to transform the behaviors generated by the behavior generation system into actual behavior outputs and put them into practice. This behavior performance is both a passive response to input from the external environment and an autonomous behavior triggered by the combined drive of the virtual character's emotional state and behavior generation, thereby enabling the virtual character to have both active interaction and passive response capabilities in different external environments. During system operation, the results of behavioral execution and environmental feedback will re-enter the environment and input signal module, and the system will then enter the state update process of the next time step, thus forming a continuous cross-evolution of emotional state and behavioral generation in the time dimension.

[0019] Therefore, the present invention employs the above-mentioned modeling method and system for cross-temporal emotion-behavior interaction generation, and the beneficial effects are as follows: (1) This invention proposes a cross-temporal emotion modeling (CTEM) framework, which effectively solves the problems of fragmented interaction and separation of emotion and behavior in existing companion-type intelligent agents. By coupling long-term behavioral history with immediate emotional expression, it significantly improves the naturalness and companionship of virtual intelligent agents.

[0020] (2) CTEM constructs a closed-loop mechanism, which realizes cross-temporal linkage between emotions and behaviors through three core modules: behavior generation and integration, emotional state update, and adaptive interaction. Emotional state evolves dynamically with past interactions, thereby adjusting current behavior. User feedback continuously corrects memory and emotions, giving the agent the ability to reflect and anticipate. The agent is deployed on an instant messaging platform, adapting to daily open scenarios, and achieving a balance between stability and flexibility.

[0021] (3) CTEM significantly improves agent consistency. Removing behavioral or emotional modules will greatly reduce perceptual consistency, while self-driven personality traits contribute the most to consistency. In terms of emotional coordination, the complete CTEM system significantly improves participants' emotional experience, especially the group that restores the complete system after the loss of emotional modules shows the most significant improvement.

[0022] (4) CTEM effectively breaks through the limitations of existing technologies, enabling intelligent agents to achieve more natural and coherent long-term companionship, providing a feasible framework and empirical support for the design of emotional companion intelligent agents, while reducing the risk of excessive emotional dependence through lightweight and de-anthropomorphic design. Attached Figure Description

[0023] Figure 1 This is an architecture diagram of a modeling system for cross-temporal emotion-behavior interaction generation according to the present invention; Figure 2 This invention is Auri, a lightweight companion agent based on cross-temporal emotion modeling. Figure 3 This is a system overview diagram of Auri, a lightweight companion agent based on cross-temporal emotion modeling, as described in this invention. Figure 4 These are examples of cross-time interaction in two scenarios within the CTEM framework of this invention; Figure 5 It is the average score of the contribution of different continuity traits of the intelligent agent of the present invention to being perceived as having a coherent presence (1=not important, 5=very important). Figure 6 This is a distribution diagram of the usage scenarios of each experimental group in the first and second phases of this invention. Detailed Implementation

[0024] The technical solution of the present invention will be further described below with reference to the accompanying drawings and embodiments.

[0025] The Cross-Temporal Emotional Modeling (CTEM) framework of this invention formally characterizes the bidirectional dynamic relationship between long-term behavioral accumulation and emotional state evolution, constituting the core driving mechanism of the intelligent agent.

[0026] This invention discloses a modeling method for generating cross-temporal emotion-behavior interactions, comprising the following steps: Step S1: Definition and integration of emotional states.

[0027] Emotional state is a core variable determining an agent's future behavior and interaction style. Inspired by motivational psychology, time steps... The state is formally defined as the following tuple: ; in, This refers to the overall internal state. Physiological-emotional state; For motivation vectors; For memory; It is an adaptive personality.

[0028] Physiological-emotional state It is used to characterize three basic dimensions: physiological energy, emotional valence (positive / negative), and arousal (calm / excited), as shown below: ; in, It indicates the current level of energy, that is, physical strength and vitality; It indicates the positive / negative degree of the current emotion, i.e., emotional valence; It indicates the degree of emotional activation, or emotional arousal.

[0029] Motivational vectors It is a 12-dimensional vector used to represent three types of motivational drivers: physiological needs (such as health), psychological needs (such as curiosity), and social needs (such as altruistic behavior), as shown below: ; in, Physiological needs refer to needs related to physical recovery and maintenance of a basic state; These are psychological needs, representing needs related to psychological growth, self-satisfaction, and cognitive exploration. Social needs refer to needs related to social relationships and emotional connections.

[0030] Memory It represents a time-indexed memory structure used to support cross-session organization and retrieval of semantic and sentiment contexts.

[0031] Adaptive personality Indicating basic personality Above, based on the current emotional state Modulation is performed as follows: ; Specifically, personality performance is dynamically adjusted by injecting state-related tone labels (such as fatigue) into the model prompts.

[0032] The aforementioned state variables collectively integrate four dimensions: emotion, motivation, memory, and personality, enabling the agent to maintain long-term consistency while dynamically adapting to different situations. Detailed definitions of each variable are shown in Table 1. It includes three emotional dimensions. Motivational drivers are categorized into three types: physiological, psychological, and social.

[0033] Table 1 CTEM Variables and Structure Definitions

[0034] Step S2: Generation and integration of behaviors.

[0035] Step S21: From the predefined set of behavior categories In the process, the set of behaviors at the current moment is obtained by sampling. This set covers five behavioral domains: physiological, work, leisure, social, and emotional. These five categories are derived from the motivational classification system in affective and behavioral science. Specific behavioral classifications are organized and abstracted based on real-world daily activity patterns. All behavioral items are represented using a unified structure to support the system's scalability.

[0036] Step S22: The behavior set is used to organize the agent's behavior into three levels, including executed historical behaviors (Past), currently executed behaviors (Present), and planned future behaviors (Future), as shown below: ; in, For a set of behaviors; For historical behavior; For the current action being performed; Actions for planning for the future.

[0037] Each behavior entry is represented as a tuple. ,in It indicates a specific action (e.g., attending a concert, studying, resting, etc.). This indicates the expected update of the state that the action represents.

[0038] Step S23: Using the scoring function The function completes the behavior selection based on the current emotional state. The evaluation of candidate behaviors follows this process: First, the behavior is mapped to the motivation space, as shown below: ; in, Indicates behavior Feature representation in motivation space.

[0039] Then, define the scoring function. As shown below: ; in, For behavior in the scoring function Embedded representation in motivation space; For the current physiological energy Determined modulation weights; For the logistic function; This represents the dot product.

[0040] Scoring function While modeling the correlation between behavior and emotion, a dynamic balance was achieved for different motivational drives: when physiological energy is low, physiological needs are prioritized; when physiological energy is high, the focus shifts to psychological and social exploration. The modulating weights are implemented using a logistic function. Constraints are imposed to simulate the "diminishing marginal returns" in emotional dynamics (i.e., sensitivity decreases as it approaches saturation).

[0041] In each behavioral choice process, CTEM first plans for the future. When empty, from the set The candidate behavior with the highest score is selected for filling. Then, from... The system samples and executes actions based on the latest score distribution. This decoupling mechanism between planning and execution ensures the system's responsiveness, especially when the state... Changes in the system may render the originally planned action unsuitable (e.g., a drop in score due to fatigue), triggering an immediate replanning. Finally, the system will also consider the candidate actions based on... Filter the data to avoid duplication and maintain contextual consistency.

[0042] Step S3: Based on feedback-driven adaptive mechanisms and real-time interactive modulation, the agent combines its internal state with the next action derived from the current action to achieve adaptive interaction.

[0043] Step S31: Real-time interactive modulation.

[0044] The selected behavior is instantiated from the base model as narrative or multimodal output (such as social media posts or images) and recorded in the behavior set. By design, the agent interacts through an instant messaging interface, remaining accessible even while performing internal activities.

[0045] Its communication style has changed from the current state (Include The agent can dynamically modulate its responses based on real-world contexts (such as weather, time, and holidays), enabling it to both initiate conversations and respond with appropriate emotional tone. As memories and states accumulate, the agent can develop richer emotional traits, such as the ability to reflect and anticipate.

[0046] To maintain consistency and alleviate the limitations of the large language model context window, the system summarizes and categorizes dialogues by time and topic, thereby maintaining emotional consistency during the interaction.

[0047] To ensure trust and ethical safety, the intelligent agent is explicitly defined as a "companion" rather than a "romantic partner." This design is achieved through a two-stage monitoring mechanism: combining keyword screening with a multi-model classifier (LLM ensemble) to detect extreme emotions, risk signals, or excessive dependence. Once such a situation is detected, the system dynamically injects safety constraints into the generated prompts, prioritizing reassuring responses, emphasizing user autonomy, and suggesting seeking professional help when necessary.

[0048] In addition, all outputs are constrained by a safe space. This ensures both content safety (filtering violent, sexual, or biased content) and emotional safety (providing support rather than harmful responses). Furthermore, the interaction strategy adjusts based on the user's emotional state to ensure user well-being remains the primary design goal: prioritizing listening when the user is emotionally distressed and engaging in deeper conversations when they are emotionally positive.

[0049] Step S32: Based on a feedback-driven adaptive mechanism, as interactions accumulate, the intelligent system learns to adapt. A rebalancing process is initiated to maintain long-term emotional stability. Each action taken triggers a state update. Gradually adjust and .

[0050] For example, high-energy tasks can reduce physiological energy. Positive interactions enhance emotional valence. To ensure continuity, these updates are usually gradual; however, certain key restorative behaviors (such as eating or sleeping) can lead to significant changes in state.

[0051] The system also includes a mandatory nighttime rest period to restore energy and emotional balance, preventing system overload. Simultaneously, this process gradually adjusts the personality state. The transition from "unfamiliar" to "familiar" (the shift in social connection from 0 to 1) enables continuous evolution and deepens interactive relationships.

[0052] Step S33: The intelligent system updates its memories at the end of each day by introducing new interactive events, as shown below: ; in, For the updated memory; A new experience for the current time step.

[0053] like Figure 1 As shown, this invention also proposes a modeling system for cross-temporal emotion-behavior interaction generation, to implement the aforementioned modeling method for cross-temporal emotion-behavior interaction generation. The system includes: (1) 101 Environment and Input Signal Module, used to receive external signals in the environment where the virtual character is located.

[0054] External signals in the environment include user input information, environmental events (such as time, location, and occurrence of events), system state changes, or other external triggering information. This module transmits the acquired environmental signals to the 102 interactive content understanding module for parsing.

[0055] (2) 102 Interactive content understanding module, used to perform necessary semantic parsing, context analysis, and suggestion judgment on the environment and input signals, extract explicit or implicit semantic information, interactive intentions and contextual information, and transmit the extracted information to 103 Emotional state system and 104 Behavior generation system.

[0056] (3) 103 Emotional State System, used to maintain the long-term continuous emotional state of virtual characters. This emotional state is jointly determined by the behavior execution results generated by 104 Behavior Generation System, the parsing results of 102 Interactive Content Understanding Module, and the historical information of 105 Memory and Summary Module.

[0057] During system operation, emotional states are continuously adjusted based on historical records, behavioral choices, and environmental changes. Each update is based on the emotional state at the previous moment, thus maintaining the continuity and cumulative nature of emotional changes over time.

[0058] (4) The 104 Behavior Generation System can be further divided into a behavior selection part and a behavior execution result part. The behavior selection part generates candidate behaviors for the virtual character based on data from the 103 Emotional State System, the results from the 102 Interaction Content Understanding module, and historical information from the 105 Memory and Summary module. The candidate behavior construction includes: setting different behavior preferences based on the agent's personality and selecting different behaviors according to these preferences; the agent planning future behaviors and adjusting the planned execution behaviors based on future execution preferences; these collectively construct the candidate behaviors. The behavior execution result part determines the final implementation direction of the behavior, such as successful execution, refusal to execute, or other execution states.

[0059] The behavior generation process in this system runs periodically according to the system's set time steps and does not depend on the presence of user input in the environment or input signals. Even if the user interaction input is empty, the system will still generate behavior candidates based on the current emotional state, historical information, and queryable real-world information, enabling the virtual character to produce autonomous behavior.

[0060] The 103 Emotional State System and the 104 Behavior Generation System maintain a bidirectional coupling relationship. The emotional state in the 103 Emotional State System at the current moment (time T) serves as an important influencing factor, driving the behavior selection of the 104 Behavior Generation System at the next moment (time T+1); while the behavior execution result of the 104 Behavior Generation System at time T+1 in turn affects the emotional state update process of the 103 Emotional State System at time T+1.

[0061] Thus, a cross-temporal driving relationship is formed between the 103 emotional state system and the 104 behavior generation system, enabling emotional states to not only reflect the current behavioral results but also influence subsequent behavioral decisions, thereby jointly promoting the cross-temporal state behavior evolution of virtual characters and achieving long-term matching and stability of emotions and behaviors.

[0062] (5) The 105 Memory and Summary module is used to record the accumulated historical state information of the virtual character during the interaction process, including emotional changes, behavioral execution results, and environmental feedback information. This module provides a reference for the subsequent emotional state update and behavior generation by storing and summarizing historical information.

[0063] (6) 106 Behavior Performance Module, which is used to transform the behaviors generated by the 104 Behavior Generation System into actual behavior outputs and put them into practice. This behavior performance can be either a passive response to external environmental input or an autonomous behavior triggered by the combined drive of the virtual character's emotional state and behavior generation, thereby enabling the virtual character to have both active interaction and passive response capabilities in different external environments.

[0064] During system operation, the results of behavioral execution and environmental feedback will re-enter the 101 environment and input signal module, and the system will then enter the state update process of the next time step (T+1), thus forming a continuous cross-evolution of emotional state and behavioral generation in the time dimension.

[0065] Example 1 like Figure 2 As shown, this embodiment develops a lightweight companion agent Auri based on a foundation model and deploys it on an instant messaging platform to explore how computational mechanisms support cross-temporal coupling of behavior and emotion in companion agents.

[0066] Auri's overall system architecture consists of a front-end user interaction module and a back-end mechanism-driven module. The back-end uses the CTEM framework to model and regulate the agent's behavior and emotional dynamics, such as... Figure 3 As shown.

[0067] Front-end user interaction module: Auri is deployed on an instant messaging (IM) platform. Leveraging the platform's familiar and low-barrier interaction features, it naturally embeds long-term, cross-temporal emotional interactions into users' daily usage scenarios.

[0068] Users interact with Auri in a manner similar to chatting with a regular contact. Auri can initiate greetings or conversations based on context using text, emojis, and images. In addition to direct conversations, Auri also proactively shares "life updates" in a format similar to social media "features," reflecting her internal state; users can then participate in the interaction by liking, commenting, or providing emoji feedback.

[0069] Back-end mechanism driven module: Auri's core is driven by the Cross-Time Sentiment Modeling (CTEM) framework. This framework models the agent's behavior and emotional dynamics through three core components: behavior generation and integration, an adaptive interaction mechanism with safety controls, and sentiment state updates. By continuously updating its internal sentiment state based on interaction history, CTEM can maintain responsiveness to user feedback while achieving long-term behavioral and emotional consistency.

[0070] like Figure 4 As shown, two specific examples illustrate the cross-temporal dynamics of CTEM, as follows: In Example 1, agent Auri performs a high-intensity activity, resulting in a decrease in its physiological energy (Time 1). This leads to a more subdued tone in subsequent interactions, and Auri begins planning restorative behaviors. After a night's rest (Time 2) and corresponding restorative behaviors (Time 3), Auri recovers its energy and initiates the next conversation with a more positive tone, referencing the previous day's context to maintain continuity. This example demonstrates that behavior can drive the evolutionary trajectory of emotional states and dynamically shape how Auri presents itself.

[0071] In Example 2, Auri initially exhibits low affective valence and low arousal, thus displaying egocentric and avoidant behavioral tendencies (Time 1). Subsequently, a pleasant dining experience restores her state to baseline (Time 2), thereby mitigating the negative bias. In Time 3, the recovered state allows Auri to respond to the user's unpleasant experience in a more relaxed and supportive manner. This example illustrates how Auri's internal state dynamics modulate her communication style, maintaining emotional continuity while responding appropriately to the user's emotions.

[0072] Example 2 To verify the effectiveness of the system of this invention in terms of user perception consistency, emotional coordination, and long-term interactive experience, this embodiment conducted a 21-day hybrid approach real-world scenario study. This study included a 14-day main experiment and a 7-day sub-experiment, covering multiple experimental groups, and collected data through questionnaires and interviews.

[0073] Ninety-six participants (38 males, 55 females, and 3 non-binary, aged 18–26) were recruited from multiple universities, representing Auri's core target users. A widely used mobile instant messaging application was chosen as the interaction platform, supporting multimodal interaction forms such as text, voice, images, and emoticons, which highly aligns with Auri's pursuit of natural interaction design. An agent without CTEM was selected as the baseline system, but the basic capabilities of a standard chatbot were retained, ensuring the system's basic usability in long-term use scenarios while clearly isolating the emotional and behavioral dynamics introduced by CTEM during analysis.

[0074] The main experiment (N=54) employed a two-week single-blind ablation design to isolate the core components of CTEM. In the first week (phase one), participants were randomly assigned to one of four experimental groups: full functional group, no multimodal group → full functional group, no emotion modeling group → full functional group, or no behavioral mechanism group → full functional group.

[0075] In addition to the group with the complete functionalities, each of the other groups had one key functional module (adaptive interaction, emotional state update, or behavior generation and integration) removed in the first phase to explore the role of these modules in shaping Auri's perceived consistency characteristics and their emotional support for users. In the second week (second phase), all experimental groups used the complete system, and the experimental conditions are shown in Table 2.

[0076] Table 2 Experimental conditions in the main experiment and sub-experiments

[0077] Sub-experiments (N=42): These sub-experiments evaluated the overall effectiveness of CTEM through a control experiment against the baseline system. A balanced design was used, with participants using the baseline system for 2 days followed by the full CTEM system for 5 days, or vice versa. This design minimized the impact of the contrast effect while obtaining comparative experience data. All participants were randomly assigned to either the baseline group → full functional group or the full functional group → baseline group.

[0078] In the main experiment, participants answered Q1–Q6 in the first and second phases, respectively; while Q7–Q9 were completed once after experiencing the complete system. The complete questionnaire is shown in Table 3.

[0079] Table 3 Questionnaire Questions Classified by Research Question

[0080] A hybrid approach was adopted, combining quantitative analysis with qualitative feedback, with the significance level set at [value missing]. The main experiment included outcomes related to perceived congruence (Q1–Q3), including the perception of congruence traits, their importance, and personality adjustability.

[0081] (1) In Q1, participants identified the consistency features they perceived on Auri at each stage. The binary choice data were analyzed using the binomial generalized estimation equation (GEE). Table 4 shows the overall probability estimates of positive consistency features.

[0082] Table 4. Marginal probability and 95% confidence interval of user-perceived positive consistency features.

[0083] In all experimental groups, participants selected positive consistency features more frequently than negative ones. When the "behavioral module" or "affective module" was removed in the first phase, the probability of perceived consistency decreased significantly, indicating that these two modules play a fundamental role in establishing consistency. In contrast, the removal of multimodal interactions had a smaller impact; even without multimodal information, consistency scores remained high. However, restoring multimodal cues in the second phase further enhanced the perceived consistency formed by the "behavioral module" and "affective module," particularly showing a significant improvement in appearance-related consistency.

[0084] (2) In Q2, participants rated the importance of various features in the existence of Auri consistency (1–5 points, where 1 indicates "not important" and 5 indicates "very important"). The three-way ANOVA results showed that the main effect of feature type was significant ( ).

[0085] Post-hoc tests adjusted for Bonferroni showed that "self-motivation and personality traits" contributed the most to consistency. The most important factor was "relationship consistency," followed by "social behavior," while "appearance" was the least important. The descriptive statistical results are consistent with the above inferential analysis, such as... Figure 5 As shown.

[0086] Neither the main time effect nor the "feature × time" interaction effect reached a significant level. Although the full feature set improved in descriptiveness from stage one (+0.25) to stage two (+0.58), the importance ranking of various features remained generally stable across different stages and conditions.

[0087] (3) In Q3, 81.5% of participants believed that Auri should have personality adjustability. Overall, participants explored an average of 2.67 different personality profiles. Mapping these personality tags to a two-dimensional space with two dimensions, namely warmth (high / low) and interactivity (high / low), the results showed that they were distributed in all four quadrants.

[0088] It can be observed that participants actively adjust their interaction style to match their diverse preferences. As one participant stated, "Since it is a companion AI, it should better meet the needs of users."

[0089] The results for Q4–Q6 cover contextual appropriateness, mood changes, and usage patterns.

[0090] (1) The cumulative linked mixture model (CLMMs) was used to analyze Q4 (context fit) and Q5 (emotional change), and the robustness was tested using ordered GEE.

[0091] For Q4, the context fit score (1 = completely unsuitable; 4 = very suitable) remained generally stable across different phases, experimental groups, and interaction conditions (all...). Intra-group comparisons based on CLMM showed that the group without emotion modeling showed a certain degree of improvement from the first stage to the second stage in the fully functional group. Although the overall scores of all groups were in the middle to high range (see Table 5), there was a significant change from the no-emotion modeling group to the fully functional group: the proportion of scores above the median increased from 20.0% in the first stage to 66.7% in the second stage.

[0092] For Q5, participants assessed mood changes using a 1–5 scale (1 = significantly worse; 5 = significantly better; 3 = no change). CLMM results showed a significant main effect of the phases. However, the confidence interval for this effect is relatively wide. No significant between-group differences were observed. Within-group comparisons showed that the intact functional group ( ) and the emotionless modeling group → the fully functional group ( All groups showed significant improvement in mood between the first and second stages, with the group without emotion modeling showing the greatest improvement when transitioning to the fully functional group. The average score for mood change in all experimental groups was higher than 3 (see Table 5).

[0093] Table 5. Descriptive statistics and within-group CLMM analysis results for Q4 and Q5 at different stages.

[0094] Overall, situational adaptability remained stable under different conditions, while emotional experience showed a positive change, with the most significant improvement observed in the group without emotion modeling, transitioning to the group with full functionality.

[0095] (2) In Q6, participants selected the scenario they used most often from six common interaction scenarios. The chi-square test was used to analyze the differences between different experimental groups, and the McNemar paired test was used to analyze the changes of participants at different stages.

[0096] The chi-square test results showed that no significant differences between groups were observed in either phase (Phase 1: both groups were equal). Phase Two: All This indicates that the usage patterns of the four experimental groups were generally consistent. In the first phase, "fragmented time" (34.3%) and "emotional communication" (30.5%) were the most common usage scenarios, while "emotional connection" (10.5%) and "support and care" (8.6%), which involved deeper engagement, were relatively less common.

[0097] In the second phase, deeply engaged scenarios remained fewer than lightweight social scenarios, which continued to dominate user interactions. It's worth noting that user preferences in lightweight scenarios changed over time.

[0098] Further analysis using McNemar paired tests revealed changes in light social situations between the first and second stages. Results showed a significant increase in "leisure time" (…). ), while "fragmented time" has significantly decreased ( This trend was consistent across all four experimental groups, such as... Figure 6 As shown.

[0099] Overall, participants' usage patterns have gradually shifted from fragmented to more purposeful, contextualized, and integrated into daily habits, while the frequency of use in situations requiring high emotional investment remains low.

[0100] The results of the sub-experiments (Q7–Q9) include user experience and emotional experience, as well as the contribution of various CTEM mechanisms to consistency and emotional coherence.

[0101] (1) In Q7, participants rated the degree of improvement in their user experience (1 = worse, 2 = no change, 5 = significant improvement). Based on "no change", a one-sample t-test was used to analyze the scores, and the Wilcoxon test was used as a robustness test.

[0102] The one-sample t-test results showed that all experimental groups exhibited significant improvements compared to the "no change" baseline (all... (See Table 6), where the improvement from baseline to full functional group was the largest. When tested with a more stringent baseline of 3 (“slight improvement”), the overall improvement ( ) and the upgrade from the baseline group to the full functionality group ( The result remained significant, while the transition from the full functional group to the baseline group did not reach a significant level. ).

[0103] For emotional experience (1 = significantly worse, 3 = no change, 5 = significantly better), the score significantly improved from the baseline group (M = 3.24) to the intact functioning group (M = 3.67) (W = 32, p = 0.004, r = 0.53). Within-group analysis showed that this positive change was observed under both ordering conditions (both p < 0.05, see Table 6).

[0104] Table 6. Results of User Experience and Emotional Changes in CTEM Across Different System Versions

[0105] (2) In Q8 and Q9, the contribution scores of the three CTEM core mechanisms (BGI, AdI, ESU) were analyzed (1-5 points, where 1 means "not important", 5 means "very important" and 3 means "neutral").

[0106] For consistency (Q8), all mechanisms scored significantly higher than the neutral level (p<0.001, see Table 7). Friedman's test showed a significant difference (p=0.012), and CLMM analysis further indicated that AdI and ESU scores were significantly higher than BGI (…). ).

[0107] For emotional coordination (Q9), AdI and ESU scores were also above the neutral level, while BGI scores did not reach a significant level. Although the overall Friedman test did not reach a significant level (p=0.135, see Table 7), CLMM analysis still showed that AdI and ESU scores tended to be higher than BGI scores (p<0.03).

[0108] Table 7. Sub-experiment participants' perceptions of the contributions of the three CTEM mechanisms to consistency and harmony.

[0109] Overall, in terms of consistency and emotional coherence, AdI and ESU make a more significant contribution to user perceived quality than BGI.

[0110] Overall, by bridging interaction gaps with complex emotional behaviors, CTEM enhances consistency and emotional coherence. By ensuring that usage intervals do not disrupt the subsequent experience, it can adapt to diverse usage scenarios, allowing users to initiate interactions at any time, while ensuring that the subsequent interaction experience remains unaffected even after a long period of inactivity.

[0111] Example 3 In this embodiment, design tensions, ethical considerations, broader implications, and research limitations are further discussed.

[0112] I. Design Tension.

[0113] This embodiment identifies four key design tensions: stability versus variability, consistency versus flexibility, support versus reciprocity, and consistency maintenance versus adaptation. These design tensions are based on the quantitative and qualitative findings of this invention and can be mapped to relevant psychological theories (see Table 8). Overall, these design tensions integrate the empirical findings into a unified explanatory framework to illustrate how cross-temporal affective modeling supports long-term companionship experiences.

[0114] Table 8. Design Tensions, Relevant Psychological Theories, and Design Implications in Companion Agents

[0115] 1. Stability and variability.

[0116] Stability helps build trust, but excessive stability can lead to monotony; variability can enhance vitality, but too much uncertainty can disrupt continuity. The companionship experience depends on a balance between the two. Interview results show that participants disagree on this: some users view diversity as an entertaining experience and anticipate Auri's next move; while others prefer Auri to remain simple and stable.

[0117] This divergence manifests in qualitative analysis as differing understandings of "variability." On one hand, many participants highly valued "stable personality and tone" and "continuity of memory." When using baseline or control systems, participants showed significant discomfort with abrupt and unsubstantiated changes, describing them as "disruptions to continuity." On the other hand, within the CTEM framework, Auri exhibited more complex emotional characteristics and behaviors, such as anticipation, regret, and hesitation. These changes can be interpreted as signals developing during continuous interaction, reflecting some users' expectations regarding the depth and development of emotions.

[0118] These differences indicate that stability is the foundation of trust, while moderate change is a reasonable expectation in interaction; the two constrain each other along the same interaction trajectory. Therefore, an agent can adjust its behavior generation based on the user's tolerance for diversity and need for predictability, thereby achieving a more personalized and sustainable companionship experience. In practice, this can be achieved through a series of algorithmic mechanisms that regulate behavioral change and stability, enabling the agent to dynamically adjust the balance between the two based on individual user preferences.

[0119] 2. Consistency and flexibility.

[0120] Experimental and interview results show that users desire both consistency and flexibility in the expression of intelligent agents. The hierarchical emotion modeling design proposed in this invention effectively supports this requirement.

[0121] Within the CTEM framework, the "behavioral module" and the "emotional module" exert cross-temporal influence. Through persistent behavioral patterns and emotional expressions, these two modules collectively shape users' perceptions of overall consistency in Auri. The results of Q1 support this design logic: when either the "behavioral module" or the "emotional module" is removed, users' attributions to Auri consistency features significantly decrease, indicating that these cross-temporal modules play a central role in shaping consistency perception. Furthermore, even in the absence of multimodal output, participants still reported high levels of consistency perception. This result preliminarily suggests that consistency perception does not entirely depend on surface-level expressions; long-term behavioral patterns and emotional expressions themselves can provide users with a stable and reliable sense of consistency.

[0122] Further experiments and interviews revealed that participants perceived richer and more flexible expressions after regaining their multimodal abilities. Therefore, these multimodal elements (such as emojis, images, and social media posts) can be understood as means to enhance the flexibility of short-term emotional expression, especially at the expressive and visual levels, thereby further strengthening users' perception of consistency. This indicates that in layered emotional design, cross-temporal emotional and behavioral expression is crucial for a consistent experience, while multimodal cues enrich the flexibility of expression in short-term interactions.

[0123] To enhance agent consistency, this invention goes beyond simple emotion modeling and explores the roles of stable value goals and core personality traits. Q2 compared the relative importance of different positive traits in consistency perception. The results showed that "self-motivation and personality traits" were the most critical, followed by "relational consistency" and "social behavior." This ranking remained stable over time, indicating that users gradually formed a stable perception of Auri consistency. Participants were more inclined to attribute consistency to their stable value goals and core personality traits, rather than the simple accumulation of momentary behaviors or fragmented states.

[0124] These value goals and core personality traits continuously shape the agent's behavioral tendencies and emotional dynamics, and work in conjunction with short-term situational factors to guide state updates and interactive behaviors. This enables the agent to maintain consistency and predictability in the long term while still possessing short-term situational adaptability.

[0125] In summary, consistency and flexibility operate on the same interaction trajectory at different time scales, and should be understood as a balancing relationship rather than independent optimization goals. By maintaining and adhering to its core value goals, an agent can present users with behavioral patterns that combine consistency and flexibility. In practice, combining a hierarchical sentiment model with a value goal calculation model helps the agent maintain overall consistency while adapting to dynamic situations.

[0126] 3. Support and reciprocity.

[0127] In this embodiment, Auri's "emotional module" was primarily used to enhance supportive interactions. Results from Q4 and Q5 indicate that restoring the "emotional module" improves users' perception of situational fit and leads to improved mood. These improvements were more pronounced when participants compared their baseline system to the full CTEM system and were largely attributed to mechanisms such as Adaptive Interaction (AdI) and Emotional State Update (ESU). In addition to providing supportive responses, Auri occasionally exhibited frustration, self-deprecation, mild complaints, or rejection. The expression of a small amount of negative emotion conveyed a degree of vulnerability, thereby enhancing its authenticity.

[0128] The results of Q6 further indicate that participants preferred to view Auri as a light companion in specific everyday situations rather than a source of deep emotional support. Limited expressions of reciprocity (such as occasionally showing needs or vulnerability) help create a sense of mutual care and make users feel needed. However, without control, frequent negative expressions or excessive demands for reciprocity can lead to emotional burden or even fatigue.

[0129] Overall, within the same companionship role, supportive responses and reciprocal expressions fulfill different emotional expectations, and their impact on enhancing the emotional support experience is not always synchronized. Designers should strike a balance between support and reciprocity, enhancing authenticity while avoiding burdening users.

[0130] 4. Consistency and adaptability.

[0131] Progressive personalization can enhance user engagement, but over-adaptation can undermine consistency. In the Q3 study on Auri adjustability, many participants customized their personality traits, often making them more pronounced or even aggressive. Participants typically viewed this adjustability as a factor that enhances fit and flexibility, and in this context, it is not usually considered to disrupt consistency.

[0132] However, some participants later reverted to the initial settings. This phenomenon reflects users' initial pursuit of novelty, followed by a return to stability and control. Personality customization should allow for a degree of flexibility in peripheral characteristics while maintaining core stability.

[0133] II. Ethical considerations and security mechanisms.

[0134] This invention focuses on safe interactions between young people and empathetic intelligent agents. Ethical and risk issues are unavoidable when designing companionable intelligent agents based on cross-temporal emotion modeling. While emphasizing the avoidance of romanticization and over-dependence, potential problems such as long-term dependence, data privacy, and mental health risks still need to be systematically addressed.

[0135] 1. Long-term dependence.

[0136] With the introduction of cross-temporal memory, users may gradually become overly dependent on the agent, thereby weakening their real-world social connections. Although CTEM enhances interaction engagement by coupling emotion and behavior, this temporal coherence may also strengthen users' anthropomorphic perception of the agent, thus increasing the risk of emotional over-attachment. Research shows that a higher degree of anthropomorphism mediates the social influence of companion chatbots, and users with a higher degree of anthropomorphism towards the agent are more likely to experience changes in their interpersonal social behavior.

[0137] To mitigate the aforementioned risks, the design adheres to the principle of "lightweight companionship": (1) Visual image: Auri’s visual design deliberately avoids humanization and adopts a soft, pet-like image to convey warmth and empathy while subtly reminding people of its non-human attributes.

[0138] (2) Behavior design: Auri’s behavior generation emphasizes youth and pet-like companionship characteristics, rather than imitating human behavior or simulating real human subjectivity.

[0139] (3) Reality anchoring: The system incorporates real-time clues and holiday information to encourage users to engage in offline social interaction at key moments, thereby avoiding excessive immersion in virtual companionship.

[0140] 2. Data privacy.

[0141] CTEM does not record the user's original emotional trajectory. The system follows the "minimize storage" principle, retaining only abstract representations related to emotion modeling.

[0142] 3. Mental health risks.

[0143] When users are experiencing emotional distress or expressing psychological problems, the AI ​​agent might be mistaken for an alternative to professional psychological counseling. To avoid overstepping its bounds, the system explicitly positions itself as a companion and assistant, not a therapist. Upon detecting risk signals, the AI ​​agent employs a de-escalation strategy, shifting towards neutral reassurance and proactively inquiring whether the user needs to connect to professional counseling services or school mental health support resources, thus guiding support to appropriate professional channels when necessary.

[0144] 4. Age group selection.

[0145] The primary objective of this invention is to understand how people interact with intelligent agents in emotional companionship scenarios, and how cross-temporal emotional cues influence users' emotional experiences and behavioral responses. Therefore, this embodiment selects participants from an age group with relatively consistent socio-emotional goals, emotional expression styles, and interaction preferences—for example, those who are more inclined to seek emotional feedback, respond more directly to emotional cues, and engage in more frequent peer-oriented social behaviors. These characteristics make their behavior in the emotional companionship system easier to observe and quantify.

[0146] Based on the above analysis, the present invention proposes the following ethical design principles: (1) Secure companionship, maintain a light and bounded companionship relationship, and prevent excessive emotional dependence caused by long-term or highly anthropomorphic interaction.

[0147] (2) Transparency and controllability ensure that users are aware of and able to manage the data and memories stored in the system.

[0148] (3) A responsible exit mechanism, adopting dialogue to de-escalate and a gradual exit strategy in risky situations.

[0149] (4) Role boundaries: Clearly define the role of the intelligent agent to avoid confusion with mental health services.

[0150] (5) Well-being guidance: Encourage users to participate in real-world social activities through time and holiday reminders.

[0151] These principles not only help reduce potential risks, but also provide actionable guidance for the responsible design of future companion agents.

[0152] Therefore, the present invention adopts the above-mentioned modeling method and system for cross-temporal emotion-behavior interaction generation. Based on the cross-temporal emotion modeling (CTEM) framework, it effectively solves the problems of fragmented interaction and separation of emotion and behavior in existing companion-type intelligent agents. By coupling long-term behavioral history with immediate emotional expression, it significantly improves the naturalness and companionship of virtual intelligent agents.

[0153] Finally, it should be noted that the above embodiments are only used to illustrate the technical solutions of the present invention and not to limit them. Although the present invention has been described in detail with reference to preferred embodiments, those skilled in the art should understand that modifications or equivalent substitutions can still be made to the technical solutions of the present invention, and these modifications or equivalent substitutions cannot cause the modified technical solutions to deviate from the spirit and scope of the technical solutions of the present invention.

Claims

1. A modeling method for cross-temporal emotion-behavior interaction generation, characterized in that, Includes the following steps: Step S1: Definition and integration of emotional states; Step S2: Generation and integration of behaviors; Step S3: Based on feedback-driven adaptive mechanisms and real-time interactive modulation, the agent combines its internal state with the next action derived from the current action to achieve adaptive interaction.

2. The modeling method for cross-temporal emotion-behavior interaction generation according to claim 1, characterized in that, In step S1, the definition and integration of emotional states are carried out as follows: Emotional state determines the future behavior and interaction style of an agent. Inspired by motivational psychology, time steps... The state is formally defined as the following tuple: ； in, The overall internal state; Physiological-emotional state; For motivation vectors; For memory; An adaptive personality; Physiological-emotional state These are used to characterize the three basic dimensions of physiological energy, emotional valence, and arousal, as shown below: ； in, Indicates current energy level; Indicates the degree of positivity / negativity of the current emotion; Indicates the degree of emotional activation; Motivational vectors This is used to represent three types of motivational drivers: physiological needs, psychological needs, and social needs, as shown below: ； in, Physiological needs refer to needs related to physical recovery and maintenance of a basic state; These are psychological needs, representing needs related to psychological growth, self-satisfaction, and cognitive exploration. Social needs refer to needs related to social relationships and emotional connections. memory It represents a time-indexed memory structure used to support cross-session organization and retrieval of semantic and sentiment context; Adaptive personality Indicating basic personality Above, based on the current emotional state Modulation is performed as follows: ； By injecting state-related tone labels into the model prompts, personality performance is dynamically adjusted; the state variables integrate four dimensions: emotion, motivation, memory, and personality, enabling the agent to maintain long-term consistency while dynamically adapting to the situation.

3. The modeling method for cross-temporal emotion-behavior interaction generation according to claim 2, characterized in that, In step S2, the generation and integration of behaviors are carried out as follows: Step S21: From the predefined set of behavior categories In the process, the set of behaviors at the current moment is obtained by sampling. This set covers five behavioral domains: physiological, work, leisure, social, and emotional. Step S22: The behavior set is used to organize the agent's behavior into three levels, including historically executed behaviors, currently executed behaviors, and planned future behaviors, as shown below: ； in, For a set of behaviors; For historical behavior; For the current action being performed; Actions for planning the future; Each behavior entry is represented as a tuple. ,in Indicates a specific action, This indicates the expected update of the state that the action represents; Step S23: Using the scoring function The function completes the behavior selection based on the current emotional state. The evaluation of candidate behaviors follows this process: First, the behavior is mapped to the motivation space, as shown below: ； in, Indicates specific behavior Feature representation in motivation space; Then, define the scoring function. As shown below: ； in, For specific behaviors in the scoring function Embedded representation in motivation space; For the current physiological energy Determined modulation weights; For the logistic function; This represents the dot product.

4. The modeling method for cross-temporal emotion-behavior interaction generation according to claim 3, characterized in that, In each behavioral choice process, CTEM first plans for the future. When empty, from the set Select the candidate behavior with the highest score and fill it in; Subsequently, from The sampling behavior is based on the latest score distribution and executed accordingly; Finally, based on the candidate behavior Filter the data to avoid duplication and maintain contextual consistency.

5. The modeling method for cross-temporal emotion-behavior interaction generation according to claim 3, characterized in that, In step S3, the real-time interactive modulation is as follows: The selected behavior is instantiated as a narrative or multimodal output by the base model and recorded in the behavior set. The agent interacts with it through an instant messaging interface; the communication style is determined by the current state. And dynamic modulation of real-world contexts; To maintain consistency and alleviate the limitations of the large language model context window, the system summarizes and categorizes dialogues by time and topic, thereby maintaining emotional consistency during the interaction.

6. The modeling method for cross-temporal emotion-behavior interaction generation according to claim 4, characterized in that, In step S3, to ensure trust and ethical safety, the intelligent agent implements a two-stage monitoring mechanism: Combining keyword screening and a multi-model classifier, the system detects extreme emotions, risk signals, or excessive dependence. Once such a situation is detected, the system dynamically injects safety constraints into the generated prompts, prioritizing reassuring responses, emphasizing user autonomy, and suggesting professional help when necessary. All outputs are confined within a safe space. This is to ensure both content safety and emotional safety.

7. The modeling method for cross-temporal emotion-behavior interaction generation according to claim 6, characterized in that, In step S3, based on the feedback-driven adaptive mechanism, as interactions accumulate, the intelligent system learns to adapt. To rebalance and maintain long-term emotional stability; Each executed action triggers a state update. Gradually adjust and ; Establish mandatory nighttime rest periods to restore energy and emotional balance, preventing system overload; simultaneously, gradually adjust personality state. The process of moving from unfamiliarity to familiarity allows for continuous evolution and deepens the interactive relationship.

8. The modeling method for cross-temporal emotion-behavior interaction generation according to claim 7, characterized in that, In step S3, the intelligent system updates its memories at the end of each day by introducing new interactive events, as shown below: ； in, For the updated memory; A new experience for the current time step.

9. A modeling system for cross-temporal emotion-behavior interaction generation, used to implement the modeling method for cross-temporal emotion-behavior interaction generation as described in any one of claims 1-8, characterized in that, The system includes: an environment and input signal module, an interactive content understanding module, an emotion state system, a behavior generation system, a memory and summary module, and a behavior performance module.

10. A modeling system for cross-temporal emotion-behavior interaction generation according to claim 9, characterized in that, The environment and input signal module is used to receive external signals from the environment in which the virtual character is located. These external signals include user input information, environmental events, system state changes, and other external triggering information. This module then transmits the acquired environmental signals to the interactive content understanding module for parsing. The interactive content understanding module is used to perform necessary semantic parsing, contextual analysis, and cues judgment on the environment and input signals, extract explicit or implicit semantic information, interactive intentions, and contextual information, and transmit the extracted information to the emotion state system and behavior generation system. The behavior generation system is further divided into a behavior selection part and a behavior execution result part. The behavior selection section generates candidate behaviors for the virtual character based on data from the emotional state system, the results from the interaction content understanding module, and historical information from the memory and summary module. The candidate behavior construction includes: setting different behavior preferences based on the agent's personality and preferences, and selecting different behaviors based on the behavior preferences; the agent plans future behaviors and adjusts the planned execution behaviors based on future execution preferences; The behavior execution result determines the final direction of the behavior; the emotional state system and the behavior generation system maintain a two-way coupling relationship; the emotional state system and the behavior generation system form a cross-temporal driving relationship, so that the emotional state can not only reflect the current behavior result, but also influence subsequent behavior decisions, thereby jointly promoting the cross-temporal state behavior evolution of the virtual character and achieving long-term matching and stability of emotions and behaviors; The memory and summary module is used to record the accumulated historical state information of the virtual character during the interaction process, including emotional changes, behavior execution results, and environmental feedback information. This module provides a reference for subsequent emotional state updates and behavior generation by storing and summarizing historical information. The behavior performance module is used to transform the behaviors generated by the behavior generation system into actual behavior outputs and put them into practice. This behavior performance is both a passive response to input from the external environment and an autonomous behavior triggered by the combined drive of the virtual character's emotional state and behavior generation, thereby enabling the virtual character to have both active interaction and passive response capabilities in different external environments. During system operation, the results of behavioral execution and environmental feedback will re-enter the environment and input signal module, and the system will then enter the state update process of the next time step, thus forming a continuous cross-evolution of emotional state and behavioral generation in the time dimension.

Images