AI Persona Contour Control

Abstract

This solves the problem that in artificial intelligence agents operated over the long term, personality traits become fixed or fluctuate excessively, resulting in a loss of flexibility and consistency in responses. [Solution] In artificial intelligence agents of the physical robot type, virtual avatar type, or bodyless type, personality traits are evaluated and analyzed based on quantitative indicators or learning history, and the progression of personality fixation or excessive personality change is evaluated, estimated, analyzed, and detected. Based on the results, the generated output, speech style, movements, facial expressions, probability distribution parameters, weight adjustment, etc. are autonomously controlled via response tendency control means, achieving both consistency and flexibility in the personality, thereby improving long-term stability and reliability.

Description

[Technical Field]

[0001] The present invention relates to a technology for quantitatively evaluating and controlling the personality characteristics of an AI agent, and in particular to an AI agent control device, method, and program that detects or estimates the progression of identity consolidation or identity drift, and maintains or restores consistency and flexibility in response tendencies. In this specification, "Artificial intelligence agent" is a general term for an artificial intelligence system that interacts with humans through dialogue or response, and includes any of the following: (1) Physical robot type: A humanoid or non-humanoid robot with a physical body (2) Virtual avatar type: An avatar in a virtual space without a physical body (3) Bodyless: Voice assistants without physical bodies or virtual avatars, cloud-based conversational AI, character AI, etc. These artificial intelligence agents may be embedded, cloud-based, or distributed collaborative. "Personality feature vector" refers to an AI agent's response tendencies, value judgments, writing style features, emotional expression characteristics, or a combination thereof, expressed as a numerical vector in a multidimensional space. [Background technology]

[0002] In recent years, artificial intelligence agents using large-scale language models and multimodal artificial intelligence have rapidly spread and are now used in a wide range of fields, including education, customer service, medicine, and entertainment. These AI models are often provided as pre-trained models with fixed internal parameters (weights), meaning their internal structure and personality do not naturally change over time. However, in actual operational environments, output trends can become fixed or fluctuate due to biases in input distribution, prompt design, dialogue history, and external setting changes. This can have a negative impact on long-term consistency and user experience quality, such as reducing response diversity or generating responses that deviate from the character settings. On the other hand, in recent years, there have been cases where trained models are provided with configurations that allow additional learning (continuous learning / online training) during operation. In such cases, long-term learning and operation can lead to problems such as fixed response tendencies and value judgments, reduced flexibility, rigid responses, biased thinking, or changes in personality tendencies. While conventional prompt control and reinforcement learning-based adjustments can adjust responses in the short term, there has not been a fully established mechanism for quantitatively detecting and controlling long-term changes and fixations in output tendencies and personality tendencies. Description of the Invention

[0003] The purpose of this invention is to evaluate, estimate, and detect the progression of personality fixation or personality changes of an AI agent, and dynamically control response tendencies, thereby achieving both consistency and flexibility in personality and maintaining long-term reliability and natural interactivity. Solution to the problem: The present invention provides the following means, but is not limited to the following configurations, and includes similar means and combinations thereof. (1) Personality trait assessment tools Evaluate the personality traits and tendencies of AI agents based on quantitative and qualitative indicators such as internal state, output history, degree of fixation, self-consistency, vocabulary distribution, and response diversity. Examples of evaluation methods include personality feature vector extraction, time series change analysis, cosine similarity calculation, and KL divergence analysis. (2) Storage means ·Collect evaluation results and dialogue history to understand short-term and long-term trends. · The degree of fixation and personality changes over time, as well as changes in response trends, are statistically analyzed and used as input information for the control unit. (3) Personality fixation and personality change assessment and detection methods · Analyze changes in assessment results over time to evaluate, estimate, or detect the progression of personality fixation or the occurrence of personality change. -Fixation and change are quantitatively identified using threshold judgment, change rate analysis, time differential analysis of self-consistency, etc. (4) Response control means Based on the detection results, the response generation process is adjusted through internal or external control logic to achieve both consistency and flexibility in personality. Specifically, the following control methods can be applied: (a) Probability distribution feedback control: Calculates the entropy index based on the probability distribution of the generated output (e.g., token distribution), and if the distribution is highly concentrated, increases the temperature parameter or noise component to temporarily expand diversity. Conversely, if a personality change is detected, decreases the parameter to restore stability. (b) Personality feature vector control: The personality feature vector is compared with the reference state vector, and if the similarity (such as cosine similarity) falls below a threshold, the internal weight balance or response template is adjusted to re-converge to the reference state. (c) External intervention control: When the progression of fixation or personality change exceeds a predetermined threshold, the personality traits are restored by introducing reference data for retraining (with prompt corrections, if necessary). Example embodiments of the present invention: The following examples illustrate embodiments of the present invention, but are intended to be illustrative and not limiting of the scope of the present invention. (1) Physical robot-type artificial intelligence agent The AI ​​agent control device of this embodiment is applied to a robot (including humanoid and non-humanoid types) having a physical body equipped with a voice input / output unit, a visual sensor, a motion actuator, and the like. [Means for evaluating personality characteristics] We analyze the content of speech, tone of voice, speech rate, emotional expression, posture and movement history, and dialogue records with the user to calculate personality feature vectors, degree of fixation, and self-concordance. This allows us to quantify the robot's response tendencies and emotional expression characteristics. [Storage means] The calculation results from the evaluation means and speech and behavior history are accumulated as time-series data, and the degree of fixation and the time transition of personality change are continuously recorded. This makes it possible to grasp the tendency of personality change over the long term. [Means for assessing and detecting personality fixation and personality change] Using data stored in the memory means, the rate of change in the degree of fixation and self-concordance, and the rate of decline in response diversity are analyzed. If these exceed a predetermined threshold, the progression of personality fixation or personality change is detected. [Response control means] Based on the detection results, speech tempo, vocabulary selection, facial expression patterns, and gesture generation are adjusted. If personality stabilization is progressing, speech diversity is expanded, and if personality change is detected, correction is made to reconverge to the standard personality feature vector. This control loop prevents excessive bias in response tendencies even during long-term operation, maintaining natural and reliable dialogue. (2) Virtual avatar-type artificial intelligence agent The AI ​​agent control device of this embodiment is applied to AI (including virtual avatars, digital humans, character AI, etc.) that do not have a physical body but have a visual avatar or digital representation in a virtual space. [Means for evaluating personality characteristics] A personality feature vector is generated by analyzing the dialogue log, facial expression generation patterns, behavioral history, emotional agreement with the user, and speech style characteristics, and the degree of fixation and self-agreement is calculated. [Storage means] Along with the evaluation results, the dialogue history and behavioral data of each session are stored, and changes in personality traits are tracked over time. This allows for a statistical understanding of short-term and long-term personality trends. [Means for evaluating and detecting personality fixation and personality change] Based on the recorded data, cosine similarity and KL divergence are analyzed to evaluate the deviation from the reference state vector. If the deviation exceeds a threshold, personality fixation or personality change is detected. [Response control means] Based on the detection results, the system readjusts the speech style, tone of voice, gaze, and facial motion. If the personality becomes fixed, the system expands the variations in facial expressions and speech expressions, and if a personality change occurs, it performs stabilization control based on the baseline personality feature vector. As a result, the avatar's personality expression in the virtual space is maintained flexibly and stably, and the naturalness of the user experience can be maintained even over long periods of operation. (3) Bodyless AI Agent The artificial intelligence agent control device of this embodiment is applied to artificial intelligence (including voice assistants, cloud-based conversational artificial intelligence, chat artificial intelligence, etc.) that does not have a physical body and does not have an avatar in a virtual space. [Means for evaluating personality traits] The vocabulary distribution, stylistic structure, response consistency index, and emotional expression ratio of the output history are analyzed to calculate a personality feature vector. This allows us to quantitatively grasp the tendency for response rigidity and personality changes. [Storage means] Accumulates speech logs, generation probability distributions, and evaluation indices in chronological order, and statistically records the degree of fixation and the progression of personality changes. [Means for evaluating and detecting personality fixation and personality change] Data stored in the memory means is analyzed, and the occurrence of personality fixation or personality change is detected using the entropy decline rate, the response diversity decline rate, or the change in similarity with the reference personality feature vector as indicators. [Response control means] Based on the detection results, the temperature parameter of the generated probability distribution is controlled to autonomously adjust diversity and stability. In addition, if there is a significant change in personality, prompt template correction or external training data reintroduction is performed to reconverge to the baseline personality feature vector. This control makes it possible to suppress personality bias even in AI that operates for a long period on the cloud, and to continuously provide natural and flexible responses. (4) Other application examples The control device and method of the present invention can also be applied to other artificial intelligence agents, such as artificial intelligence partners, virtual spouses, educational accompaniment artificial intelligence, character artificial intelligence, customer service robots, medical and nursing care support artificial intelligence, tourist guidance artificial intelligence, and customer support artificial intelligence. In these applications, the evaluation, memory, detection, and control methods work together to stabilize personality traits while maintaining flexibility, thereby improving long-term reliability and user experience quality. Benefits of this invention: According to the present invention, the following effects can be obtained. (1) The system can quantitatively grasp and control the progression of personality fixation and personality changes, and is equipped with the ability to autonomously recover from abnormal states, allowing for both consistency and flexibility of personality even in long-term operation. (2) The progress of personality fixation and personality changes can be detected and adjusted in real time, increasing the safety, reliability, and transparency of AI agents. (3) It can be applied to various artificial intelligence agents, including physical robots, virtual avatars, and bodyless types. (4) Improve the naturalness, immersion, and reliability of the user interaction experience, thereby improving service quality. (5) It can be deployed in a wide range of industrial fields, including education, medical care, hospitality, tourism, entertainment, and customer support. (6) Even in applications that require long-term interaction, such as artificial intelligence partners, virtual spouses, and educational accompaniment artificial intelligence, it enables natural changes while maintaining consistency in personality. According to the present invention, it is possible to quantitatively detect and control personality fixation and personality change of an AI agent, thereby achieving both consistency and flexibility of personality. Industrial Applicability: This invention quantitatively tracks the personality behavior of an AI agent over time and structure, and detects and controls the progression of personality fixation and excessive personality changes, thereby ensuring long-term stability. - It can be applied to all types of artificial intelligence agents, including physical robots, virtual avatars, and bodyless types. Applicable to a wide range of usage scenarios, including home, business, education, and entertainment. In fields such as entertainment, education, medical care, hospitality, tourism, and customer support, controlling personality fixation and personality changes makes it possible to provide flexible responses while ensuring safety and predictability. By adjusting responses according to the progression of personality fixation and personality change, it is possible to provide a personalized experience that meets the user's needs and circumstances. Developers and businesses can design and operate AI agents optimized for individual applications through personality state control and analysis functions, thereby increasing commercial and social value. - It is effective for applications that require long-term relationship building, such as AI partners, virtual spouses, AI for nursing care, and AI for educational support. - Quantitative control of personality changes through continuous learning can improve the stability and reliability of relationships with users. -It has high technical and social utility as a "personality stability control infrastructure" that is essential for the sustainable development of an artificial intelligence society.

Claims

1. An apparatus for evaluating, recording, and analyzing personality traits of an artificial intelligence agent, comprising: Based on the output history, internal state, vocabulary distribution, response diversity, response diversity score, or other quantitative and qualitative indicators of the AI ​​agent, a personality trait evaluation means for evaluating fixed indicators related to personality traits by using a time series change analysis of the personality feature vector and a time differential analysis of the self-concordance; a storage means for accumulating the evaluation results and the dialogue history and recording short-term and long-term trends; Based on the information stored in the storage means and the evaluation results, A personality fixation and personality change evaluation and detection means for quantitatively evaluating, analyzing, and estimating the progression of personality fixation or excessive personality change; Based on the output result of the detection means, a response control means for autonomously adjusting response tendencies according to the entropy index of the response distribution and deviations of the personality feature vector, thereby simultaneously maintaining or restoring consistency and flexibility; An AI agent control device characterized in that the response control means includes a configuration for adjusting response tendencies in cooperation with not only internal control logic but also external control logic or an external device.

2. The response control means, in response to the quantitative evaluation result by the personality fixation and personality change evaluation and detection means or the time change of the response diversity index, Control response tendencies through generative output, speaking style, movements, facial expressions, prompt control, probability distribution parameters, weight adjustment, template correction, external filters, human intervention, or a combination thereof; Based on the entropy index of response distribution and the deviation of personality feature vectors, we can restore flexibility from excessive rigidity.

2. The AI ​​agent control device according to claim 1, wherein the AI ​​agent control device executes distribution feedback control, personality feature vector correction control, or a combination thereof, which simultaneously maintains consistency and flexibility.

3. The device is incorporated into a physical robot, a virtual avatar, or a bodyless artificial intelligence agent; the response control means dynamically adjusts a response tendency based on the output of the detection means or a signal from external control logic; 3. The AI ​​agent control device according to claim 1 or 2, which constitutes an AI agent control system that maintains stability and flexibility of response.

4. Evaluate, record, and analyze the personality characteristics of the artificial intelligence agent, and evaluate, estimate, analyze, and detect the progression of personality fixation, excessive personality change, or bias in response diversity indicators, 1. A method for controlling an artificial intelligence agent that autonomously adjusts response tendencies to maintain or restore consistency and flexibility of responses, comprising:

10. An artificial intelligence agent control method, characterized in that it is carried out as an internal control or a cooperative control with an external device, using the artificial intelligence agent control device according to claim 1.

5. Evaluate, record, and analyze the personality characteristics of the artificial intelligence agent, and evaluate, estimate, analyze, and detect the progression of personality fixation, excessive personality change, or bias in response diversity indicators, A program that causes a computer to execute a process of autonomously adjusting response tendencies to restore flexibility or maintain stability, or a computer-readable medium that stores the program, the program executes a time series analysis of the personality feature vector, a time differential analysis of the self-concordance, and a flexibility recovery control based on an entropy index of the response distribution; An artificial intelligence agent control program that dynamically adjusts response tendencies in cooperation with cloud environments, edge devices, or external control systems.