Information processing systems, information processing methods, and programs

Abstract

This invention provides an information processing system, information processing method, and program that further improve the accuracy of evaluations of AI modules, such as generative AI. [Solution] An information processing system comprising at least one processor, the processor reading a program and performing the following steps: In the acquisition step, a confidence score indicating an indicator of the accuracy of the processing performed by the AI ​​agent is acquired, and the AI ​​agent autonomously performs a task in accordance with instructions input by the user. The confidence score is presented by the AI ​​agent that performed the task, and in the output step, evaluation information indicating an evaluation of the AI ​​agent is output to the AI ​​agent based on at least the confidence score and a first uncertainty that objectively indicates the accuracy of the processing performed by the AI ​​agent, thereby causing the AI ​​agent to store the evaluation information and change the AI ​​agent's course of action.

Description

[Technical Field]

[0001] The present invention relates to an information processing system, an information processing method, and a program. [Background Art]

[0002] For example, Patent Document 1 discloses "a system aimed at solving the problems of accuracy and reliability of current Chat-GPT". The system is provided with means for scoring data and models based on user evaluations and feedback, and outputting more accurate outputs when a similar request occurs again. [Prior Art Documents] [Patent Documents]

[0003] [Patent Document 1] Japanese Unexamined Patent Publication No. 2025-046417 [Summary of the Invention] [Problems to be Solved by the Invention]

[0004] However, there is still room for improvement in the functions related to the above-mentioned known technology. In particular, there is a demand for a technique for further improving the accuracy of evaluation of AI modules such as generative AI.

[0005] In view of the above circumstances, the present invention provides an information processing system and the like that can further improve the accuracy of evaluation of AI modules. [Means for Solving the Problems]

[0006] According to one aspect of the present invention, an information processing system is provided, comprising at least one processor, the processor configured to perform the following steps by reading a program, the acquisition step obtaining a confidence score indicating an indicator of the accuracy of processing performed by an AI agent, the AI ​​agent being configured to autonomously perform tasks in response to instructions input by a user, the confidence score being presented by the AI ​​agent that performed the task, and the output step outputting evaluation information indicating an evaluation of the AI ​​agent based at least on the confidence score and a first uncertainty that objectively indicates the accuracy of processing performed by the AI ​​agent, thereby causing the AI ​​agent to store the evaluation information and change the AI ​​agent's course of action.

[0007] This configuration makes it possible to further improve the accuracy of evaluations of AI modules. [Brief explanation of the drawing]

[0008] [Figure 1] This is a diagram showing the configuration of information processing system 1. [Figure 2] This is a block diagram showing the hardware configuration of the information processing device 2. [Figure 3] This is a block diagram showing the hardware configuration of user terminal 4. [Figure 4] This is a block diagram showing the functions realized by the information processing device 2 (processor 23) and the AI ​​server 3 (processor 33). [Figure 5] This flowchart shows an overview of the processes performed by Information Processing System 1. [Figure 6] This is an activity diagram showing a specific example of an information processing method performed by information processing system 1. [Figure 7] This is an overview diagram of the confidence score management table T1. [Figure 8] This shows an example of a presentation screen 6 based on evaluation information, which is displayed on the display unit 44 of the user terminal 4. [Modes for carrying out the invention]

[0009] Embodiments of the present invention will be described below with reference to the drawings. The various features shown in the embodiments below can be combined with each other.

[0010] Incidentally, the program for implementing the software appearing in one embodiment may be provided as a non-transitory computer-readable medium, or it may be provided as a downloadable medium from an external server, or it may be provided so that the program is launched on an external computer and its functions are realized on a client terminal (so-called cloud computing).

[0011] Furthermore, in various information processing according to one embodiment, an input and an output corresponding to the input can be realized. Here, as long as an output is obtained as a result of the input, the form of the information referenced in such information processing (hereinafter referred to as "reference information") is not limited. The reference information may be, for example, rule-based information such as a database, a lookup table, or a predetermined function (including a decision formula such as a regression equation constructed by a statistical method), or a trained model that has been pre-trained to learn the correlation between input and output, or a generative AI such as a large-scale language model that can output a desired result by inputting a prompt (these models include parameters that construct the correlation relationship between input and output) or a visual language model.

[0012] Furthermore, in one embodiment, "part" may include, for example, hardware resources implemented by a circuit in a broad sense, and the information processing of software that can be specifically realized by these hardware resources. Also, in one embodiment, various types of information are handled, and this information can be represented, for example, by the physical values ​​of signal values ​​representing voltage and current, the high or low values ​​of signal values ​​as a set of binary bits composed of 0s or 1s, or by quantum superposition (so-called qubits), and communication and calculations can be performed on a circuit in a broad sense.

[0013] Furthermore, a circuit in a broad sense is a circuit realized by combining at least a suitable combination of circuits, circuits, processors, and memory. The processor may be a general-purpose processor or a dedicated circuit. In other words, it includes application-specific integrated circuits (ASICs), programmable logic devices (for example, simple programmable logic devices (SPLDs), complex programmable logic devices (CPLDs), and field programmable gate arrays (FPGAs)), etc.

[0014] 1. Hardware Configuration This section describes the hardware configuration of one embodiment. 1.1 Information Processing System 1 FIG. 1 is a configuration diagram illustrating the information processing system 1. The information processing system 1 includes an information processing device 2, an AI server 3, and a user terminal 4. The information processing device 2, the AI server 3, and the user terminal 4 are configured to be communicable via a telecommunication line. In one embodiment, the information processing system 1 is composed of one or more devices or components. For example, if the information processing system 1 is composed only of the information processing device 2, the information processing system 1 may be the information processing device 2. That is, the information processing system 1 may include a server device (e.g., the information processing device 2) having a processor (e.g., a control unit 23), and the user terminal 4 accessible to the server device. According to this aspect, the information processing system 1 can be implemented in various aspects. These components will be described below.

[0015] 1.2 Information Processing Device 2 FIG. 2 is a block diagram illustrating the hardware configuration of the information processing device 2. The information processing device 2 includes a communication bus 20, a communication unit 21, a storage unit 22, and a control unit 23. The communication unit 21, the storage unit 22, and the control unit 23 are electrically connected inside the information processing device 2 via the communication bus 20.

[0016] <Communication Unit 21> Although wired communication means such as USB, IEEE 1394, Thunderbolt (registered trademark), and wired LAN network communication are preferable for the communication unit 21, wireless LAN network communication, mobile communication such as 3G / LTE / 5G, BLUETOOTH (registered trademark) communication, and the like may be included as necessary. That is, it is more preferable to implement the communication unit 21 as a set of these plurality of communication means. That is, the information processing device 2 may communicate various types of information from the outside via the communication unit 21 and a network.

[0017] <Storage Unit 22> The storage unit 22 stores various types of information defined by the foregoing description. For example, the storage unit 22 may be implemented as a storage device such as a solid state drive (SSD) that stores various programs related to the information processing apparatus 2 executed by the processor 23, or as a memory such as a random access memory (RAM) that stores temporarily necessary information (arguments, arrays, etc.) related to program calculations. The storage unit 22 stores various programs, variables and the like related to the information processing apparatus 2 executed by the processor 23.

[0018] <プロセッサ23> The processor 23 performs processing and control of overall operations related to the information processing apparatus 2. The processor 23 is, for example, a central processing unit (CPU) not shown in the drawings. The processor 23 implements various functions related to the information processing apparatus 2 by reading a predetermined program stored in the storage unit 22. That is, information processing by software stored in the storage unit 22 is specifically implemented by the processor 23, which is an example of hardware, so that the functions can be executed as each functional unit included in the processor 23. These will be described in further detail in the next section. Note that the number of the processor 23 is not limited to one, and the present invention may be implemented such that a plurality of processors 23 are provided for each function, or a combination thereof may be adopted. 1.3 AI Server 3 The information processing system 1 may include the AI server 3. The AI server 3 is a server configured to execute an AI agent. The configuration of the AI agent will be described in further detail in the next section. As shown in FIG. 2, the AI server 3 includes a communication bus 30, a communication unit 31, a storage unit 32, and a processor 33. The communication unit 31, the storage unit 32, and the processor 33 are electrically connected inside the AI server 3 via the communication bus 20. Descriptions of the communication unit 31, the storage unit 32, and the processor 33 are omitted because they are the same as the descriptions of each unit in the information processing apparatus 2.

[0019] 1.4 User Terminal User terminal 4 is a terminal operated by a user who uses the AI ​​agent. For example, user terminal 4 is a terminal owned by the user. It can be a smartphone, tablet, computer, or any other device that can access the information processing device 2 or AI server 3 via a telecommunications line, regardless of its form.

[0020] Figure 3 is a block diagram showing the hardware configuration of the user terminal 4. The user terminal 4 comprises a communication bus 40, a communication unit 41, a storage unit 42, a processor 43, a display unit 44, and an input unit 45. The communication unit 41, storage unit 42, processor 43, display unit 44, and input unit 45 are electrically connected within the user terminal 4 via the communication bus 40. The descriptions of the communication unit 41, storage unit 42, and processor 43 are the same as the descriptions of each part in the information processing device 2, so they are omitted here.

[0021] <Display section 44> The display unit 44 displays a graphical user interface (GUI) screen that can be operated by the user. The display unit 44 may be included in the housing of the user terminal 4 or it may be an external device. Specifically, the display unit 44 may be implemented as a display device such as a CRT display, liquid crystal display, organic EL display, or plasma display. It is preferable that these display devices be used depending on the type of user terminal 4.

[0022] <Input section 45> The input unit 45 receives operation input from the user. The operation input is transmitted to the processor 43 via the communication bus 40 as a command signal. The processor 43 can perform predetermined controls or calculations based on the transmitted command signal as needed. The input unit 45 may be included in the casing of the user terminal 4 or it may be external. For example, the input unit 45 may be implemented as a touch panel integrated with the display unit 44. When the input unit 45 is implemented as a touch panel, the user can input tap operations, swipe operations, etc. to the input unit 45. Instead of a touch panel, the input unit 45 can be a switch button, a mouse, a QWERTY keyboard, etc.

[0023] 2. Functional Configuration This section describes the functional configuration. Information processing by software stored in the memory unit 22 is concretely realized by the processor 23, which is an example of hardware, and can be executed as each functional unit included in at least one processor of the information processing system 1. In other words, the information processing system 1 comprises each functional unit in a program.

[0024] Figure 4 is a block diagram showing the functions realized by the information processing device 2 (processor 23) and the AI ​​server 3 (processor 33).

[0025] Figure 4A is a block diagram showing the functions realized by the processor 23 of the information processing device 2. As shown in Figure 4A, the processor 23 includes an acquisition unit 231, a identification unit 232, a calculation unit 233, and an output unit 234.

[0026] The acquisition unit 231 is configured to acquire various information as an acquisition step. Specifically, the acquisition unit 231 is configured to acquire information from the AI ​​server 3 or other devices. The acquisition unit 231 may also acquire various information stored in the storage unit 22 of the information processing device 2. Preferably, the acquisition unit 231 acquires output information IF1 corresponding to the processing performed by the AI ​​agent, processing information IF2 regarding the execution process of the processing, and confidence level CO indicating an indicator of the accuracy of the output information IF1 from the AI ​​server 3. In this embodiment, the various information acquired by the acquisition unit 231 is described as being stored in the storage unit 22.

[0027] The identification unit 232 is configured to identify various types of information as a identification step. Preferably, the identification unit 232 may identify various indicators from the processing information IF2 acquired by the acquisition unit 231. Alternatively, for example, the identification unit 232 may identify information regarding cooperation between AI agents from the processing information IF2 acquired by the acquisition unit 231.

[0028] The calculation unit 233 is configured to calculate various information as a calculation step. Preferably, the calculation unit 233 may calculate a first uncertainty U1 indicating the degree of confidence of the task execution process by the AI ​​agent based on the output information IF1, the processing information IF2, and the first reference information. The first reference information may be a database or an artificial intelligence module.

[0029] Here, let me add some information about the artificial intelligence module. The artificial intelligence module may be an AI (Artificial Intelligence) equipped with trained models such as transformers including GPT (Generative Pretrained Transformer, including GPT-1, GPT-2, GPT-3, GPT-4, and GPT-4), BERT (Bidirectional Encoder Representations from Transformers), BART (Bidirectional and Auto-regressive Transformer), and language models such as recurrent neural networks (RNNs).

[0030] Language models are an example of learning models using machine learning algorithms. Specific machine learning algorithms include nearest neighbors, naive Bayes, decision trees, support vector machines, and deep learning using neural networks. The artificial intelligence module can apply these algorithms as appropriate.

[0031] An artificial intelligence module may have a trained model constructed using a learning method such as supervised learning, unsupervised learning, or self-supervised learning. In supervised learning, machine learning is performed using training data. Training data consists of pairs of input data and output data (correct answers). Furthermore, the language model may not only be one trained for a specific task, but also a general-purpose model that can be used universally for a wide range of tasks.

[0032] The artificial intelligence module may be a natural language model, or it may include a general-purpose pre-trained model for natural language processing, such as a Large Language Model (LLM). An LLM is a pre-trained model that has learned from a large amount of data beforehand, such as text data (e.g., (i) web content on the internet, or (ii) data stored in a predetermined database), and can perform various language processing tasks when given a task. It can perform a wide range of natural language processing tasks, such as understanding sentence patterns and context, responding to questions, and generating sentences, according to given prompts. Such a general-purpose pre-trained model may include language models that can handle various tasks without fine-tuning using One-shot Learning or Few-shot Learning. Furthermore, a general-purpose pre-trained model can also handle various tasks using Zero-shot Learning. The artificial intelligence used in each functional unit of the processor 23 may be a separate pre-trained model, or it may be a common general-purpose pre-trained model. A large language model is a type of generative AI and includes models provided by services such as OpenAI's GPT, Google's Gemini, and Microsoft's Azure AI Studio. Furthermore, the artificial intelligence module may include any machine learning model, deep learning model, artificial intelligence model, etc. The artificial intelligence module may be built on a system outside of the information processing system 1. Also, the artificial intelligence module may be interactive (which may be rephrased as chat-type or conversation-type) in which it alternately receives input to produce a specified output and generates and outputs information.

[0033] The learning model included in the artificial intelligence module can undergo additional learning through techniques such as transfer learning or fine-tuning. For example, the artificial intelligence module learns whether or not the output content has been modified by a user. In other words, the artificial intelligence module may perform additional learning and fine-tuning based on modifications made to the content output by the learning model. Furthermore, for example, whenever new data is registered, the artificial intelligence module may perform additional learning and fine-tuning using this new data as training data. This improves the accuracy of the information output from the learning model.

[0034] The learning model included in the artificial intelligence module may be a learning model (distilled model) obtained by knowledge distillation using the original trained model. In knowledge distillation, a trained model such as a large-scale language model is used as the teacher model, and the student model is trained by adjusting its parameters so that the output loss of the student model (distilled model) relative to the output (Soft Target Loss) of the teacher model is small, and that student model becomes the distilled model. Alternatively, the student model may be trained so that the output loss of the student model relative to the correct labels (Hard Target) of the teacher data (combination of input data and output data of the learning model) is small. Compared to the original trained model (teacher model), the distilled model has performance close to that of the trained model, but with fewer parameters and a lower processing load. Therefore, by using the distilled model, the cost of the information processing system 1 can be reduced.

[0035] For example, the learning model used in each functional unit may be a distilled model that has been trained using combinations of input and output data from a large-scale language model as training data. Alternatively, when the information processing system 1 is introduced, a large-scale language model may be used as the learning model in each functional unit, and once training data from the large-scale language model has been accumulated, the distilled model obtained by knowledge distillation using that training data may be used as the learning model in each functional unit.

[0036] The output unit 234 is configured to output various information as an output step. Preferably, the output unit 234 outputs evaluation information indicating an evaluation of the processing by the AI ​​agent based on the confidence level CO and the first uncertainty level U1.

[0037] Figure 4B is a block diagram showing the functions realized by the processor 33 of the AI ​​server 3. As shown in Figure 4B, the AI ​​server 3 includes an acquisition unit 331 and an output unit 332.

[0038] The acquisition unit 331 is configured to acquire various types of information as an acquisition step. Specifically, the acquisition unit 331 is configured to acquire information from the information processing device 2 or the user terminal 4. Alternatively, the acquisition unit 331 may acquire various types of information stored in the storage unit 32 of the AI ​​server 3. Preferably, the acquisition unit 331 may acquire instructions in natural language from the user terminal 4. For example, the acquisition unit 331 may acquire evaluation information from the information processing device 2. In this embodiment, the various types of information acquired by the acquisition unit 331 are described as being stored in the storage unit 32.

[0039] The output unit 332 is configured to output various information as an output step. Preferably, the output unit 332 may output information corresponding to instructions or goals based on the inference results performed by the AI ​​agent. The inference processing may be performed by the processor 33 running the AI ​​agent based on instructions and the artificial intelligence module.

[0040] An AI agent may also be called an autonomous agent. An "AI agent" is a program that, upon receiving input such as a goal (objective, purpose, etc.) like "Teach me about XX" or a task like "Output XX," breaks down the processes necessary to reach the goal or accomplish the task into subtasks, actions, etc., and executes subtasks or actions such as collecting and analyzing necessary data. The AI ​​agent takes the information and instructions input by the user as its goal, autonomously selects and executes tasks, actions, etc. according to the goal, outputs information according to the goal, and does not require user intervention (operation input). Furthermore, the AI ​​agent may autonomously plan and execute, evaluate the execution results itself, and learn autonomously in order to achieve the goal. For example, the AI ​​agent may autonomously update itself based on the execution results of subtasks (e.g., collected information, results of information analysis, etc.). The AI ​​agent may be configured to execute tasks using a large-scale language model that includes a generative AI trained to perform these processes.

[0041] Specifically, the AI ​​agent may perform a task in the following manner, for example. First, the AI ​​agent receives specific instructions from the user, including a goal or task, via the acquisition unit 331. Next, the AI ​​agent divides the goal or task into multiple subtasks or actions necessary to achieve the goal or task, and generates subtasks or actions. After dividing the goal or task, the AI ​​agent collects information necessary to perform the multiple subtasks and actions. For example, the AI ​​agent is configured to connect to multiple services to collect information. Connecting to multiple services includes, for example, accessing databases, the internet, web searches, API integration with external services, and conversations with other AI agents. After collecting information, the AI ​​agent, for example, uses external tools to perform the multiple subtasks and actions based on the collected information (for example, by generating and executing code), and outputs the task corresponding to the goal, or the execution result of the input task.

[0042] The AI ​​agent may consist of a management AI agent and multiple sub-AI agents. For example, the processor 33 may instruct the management AI agent to select a target AI agent from among the multiple sub-AI agents. Furthermore, the processor 33 may instruct the management AI agent to generate an action execution instruction for the target AI agent based on the instructions received from the user, and then input the execution instruction into the target AI agent.

[0043] The management AI agent manages a series of processing flows in a multi-AI agent environment by linking the output from sub-AI agents as input to the next subtask. The results output by the management AI agent may be fed back from the management AI agent to the output unit 332 and output to the user. In this way, the orchestration by the management AI agent may be configured to select the optimal sub-AI agent according to the situation and autonomously and efficiently execute a predetermined series of processes.

[0044] Furthermore, the AI ​​agent configuration may not distinguish between a management AI agent and a sub-AI agent. Specifically, an AI agent may act as the one that inputs execution instructions to other AI agents (instructor), or it may act as the one that receives execution instructions from other AI agents and returns processing results (destination). In other words, each AI agent may be configured to switch between the roles of instructioner and recipient depending on the situation and task content.

[0045] Multiple AI agents may be based on a common language model and their functions may be virtually separated by switching input prompts. Alternatively, multiple AI agents may be physically separated and implemented as independent generative AI or learning models, with a management AI agent controlling them via an external interface.

[0046] 3. Information Processing The information processing of Embodiment 1 will be described below.

[0047] 3.1 Overview of the process Information processing system 1 comprises at least one processor 23 (e.g., a control unit), and the processor comprises the following parts by reading a program. In other words, such information processing method comprises each step of the information processing system. From another perspective, such program causes at least one computer to execute each step of the information processing system. Figure 5 is a flowchart outlining the processing performed by information processing system 1. The steps shown in Figure 5 will be described below.

[0048] First, the acquisition unit 231 of the information processing device 2 acquires a confidence score CO, which is an indicator of the accuracy of the processing performed by the AI ​​agent (Step S001). Next, the output unit 234 outputs evaluation information to the AI ​​agent, which is an evaluation of the AI ​​agent, based on at least the confidence score CO and a first uncertainty U1 that objectively indicates the accuracy of the processing performed by the AI ​​agent. This causes the AI ​​agent to store the evaluation information and changes the AI ​​agent's course of action (Step S002).

[0049] The AI ​​agent is configured to autonomously perform tasks in response to instructions input by the user. The confidence score (CO) is presented by the AI ​​agent that performed the task.

[0050] In this configuration, the processing performed by the AI ​​agent can be evaluated based on the AI ​​agent's own subjective confidence level CO and an objective first uncertainty level U1. Therefore, a mechanism can be constructed that autonomously changes the AI ​​agent's course of action in response to a comparison between the AI ​​agent's subjective indicators and objective evaluation values.

[0051] 3.2 Specific Examples Specific examples may be included within the scope defined in the overview described above. In the specific examples, the following situation is assumed: An administrator managing the information processing device 2 provides the user with a service (hereinafter referred to as "this service") via the internet to support the optimization of the operation of multiple AI agents used or managed by the user. Multiple AI agents receive instructions from the user, including predetermined tasks, via the user terminal 4, autonomously execute the tasks, and output the execution results. An AI agent may have other AI agents execute at least a part of the task and obtain the execution results from the other AI agents. In addition, the management AI agent in this service may execute part of the task by linking to an external service or using an external tool. This service is described as providing support for optimizing operation by evaluating the task execution process of each AI agent based on the task execution process and execution results, adjusting the behavioral guidelines of the AI ​​agents based on the evaluation results, and providing support for operation optimization.

[0052] In one embodiment, descriptions indicating that the AI ​​agent performs processing, such as "output by the AI ​​agent" or "estimated by the AI ​​agent," mean that the processing is performed by the processor 33 that runs the AI ​​agent.

[0053] Figure 6 is an activity diagram showing a specific example of an information processing method performed by the information processing system 1. The following describes each process included in the information processing method, following the activities in Figure 6.

[0054] First, the acquisition unit 231 of the information processing device 2 acquires output information IF1 corresponding to the processing performed by the AI ​​agent and processing information IF2 related to the processing execution process (Activity A101). Specifically, the acquisition unit 231 acquires output information IF1 and processing information IF2 from the AI ​​server 3 via the communication unit 21. Output information IF1 is information indicating the result of the task execution by the AI ​​agent based on the user's instructions received via the user terminal 4. Output information IF1 may include, for example, the reasoning results by the AI ​​agent. Processing information IF2 includes information indicating the task execution process. For example, processing information IF2 may include information indicating the reasoning process that led to the output of output information IF1. The reasoning process may include the order of reasoning or intermediate decision results. Processing information IF2 may also include information regarding the branching structure of thought in which multiple candidates were considered, or information regarding the presence or absence of contradictions among those multiple candidates. Furthermore, for example, processing information IF2 may include information referenced during the task execution process by the AI ​​agent. Preferably, the processing information IF2 may include information containing instructions entered by the user (e.g., prompts).

[0055] Furthermore, the acquisition unit 231 acquires the confidence score CO. The confidence score CO is an indicator of the degree of confidence of the AI ​​agent in the correctness of the output information IF1 estimated by the AI ​​agent. For example, the confidence score CO may be information that quantifies the probability that the output information IF1 is estimated to be correct by the AI ​​agent. Preferably, the higher the confidence score CO, the more highly the AI ​​agent values ​​the correctness of the output information IF1 it has output. On the other hand, the lower the confidence score CO, the more the AI ​​agent values ​​that its reasoning or judgment contains many uncertain parts or ambiguous parts of the information.

[0056] For example, a prompt instructing the AI ​​agent to output confidence level CO along with outputting output information IF1 may be pre-stored in the memory unit 32. The prompt may also include an instruction to output a numerical value indicating the degree of confidence in the generated answer.

[0057] Next, the calculation unit 233 of the information processing device 2 calculates a first uncertainty U1 based on the output information IF1, the processing information IF2, and the uncertainty evaluation model RE1 (an example of first reference information) (Activity A102). The uncertainty evaluation model RE1 defines at least the relationship between the output information IF1, the processing information IF2, and the first uncertainty U1. The first uncertainty U1 is an objective indicator of the accuracy of the output information IF1. Preferably, a higher first uncertainty U1 indicates that the output information IF1 and the processing information IF2 are inaccurate.

[0058] In this configuration, it is possible to evaluate the processing performed by the AI ​​agent based on the results of the processing performed by the AI ​​agent and the process of its execution.

[0059] Preferably, the calculation unit 233 may calculate the first uncertainty U1 based on an index indicating the degree of branching or frequency of inconsistencies in the inference, an index indicating the appropriateness of the data referenced by the AI ​​agent, or an index indicating external factors of the AI ​​agent, which are configured to be identified from the processing information IF2. The appropriateness of the data referenced by the AI ​​agent may be, for example, an index relating to the update frequency of the data used in the inference or the stability of the referenced information source. Furthermore, the index indicating external factors of the AI ​​agent may be an index relating to the degree of ambiguity of the information, including instructions entered by the user, or the degree of fluctuation of external factors in the subject field in which the inference was performed.

[0060] In this configuration, factors based on processing by the AI ​​agent can be reflected in the degree of uncertainty. Therefore, it is possible to evaluate processes that reflect the AI ​​agent's thinking process and information environment.

[0061] Furthermore, the uncertainty assessment model RE1 may include an artificial intelligence module configured to take output information IF1 and processing information IF2 as inputs and output a first uncertainty U1. The uncertainty assessment model RE1 is, for example, a machine learning model that uses output information IF1 and processing information IF2 and the first uncertainty U1 corresponding to this information as training data.

[0062] Next, the calculation unit 233 calculates the corrected confidence score AC by correcting the confidence score CO with the first uncertainty U1 (Activity A103). The corrected confidence score AC is a value obtained by adjusting the subjective confidence score CO of the AI ​​agent according to the objective first uncertainty U1 calculated by the information processing device 2. This makes it possible to obtain an appropriate index even if, for example, the AI ​​agent calculates a confidence score CO that shows excessive confidence, or a confidence score CO that is excessively low. For example, the calculation unit 233 may calculate the corrected confidence score AC using [Equation 1].

[0063]

number

[0064] In [Equation 1], γ is the correction rate, a coefficient that defines the degree of influence of the first uncertainty U1. In activity A106, described later, the output unit 234 outputs a confidence score TS (an example of evaluation information) based on the corrected confidence level AC and the honesty level HO. With this configuration, the AI ​​agent's self-assessment can be objectively adjusted. Therefore, by using the corrected confidence level AC in the calculation of evaluation information, the reliability of the evaluation information can be improved.

[0065] Independent of the processing related to Activity A103, the acquisition unit 231 further acquires a second uncertainty U2 indicating the accuracy of the processing, which is presented by the AI ​​agent (Activity A104). Specifically, the acquisition unit 231 acquires the second uncertainty U2 from the AI ​​server 3 via the communication unit 21. Here, the second uncertainty U2 is an index that indicates the degree of uncertainty estimated by the AI ​​agent regarding its own reasoning process or output result. For example, the AI ​​agent may infer how unstable its own judgment in the processing is, or how insufficient the information is, and quantify and indicate this. Specifically, the second uncertainty U2 may be calculated based on an index that indicates the degree of branching or frequency of inconsistencies in the reasoning, an index that indicates the appropriateness of the data referenced by the AI ​​agent, or an index that indicates external factors of the AI ​​agent, similar to the first uncertainty U1. Preferably, a higher second uncertainty U2 indicates that the AI ​​agent inferred that the output information IF1 and processing information IF2 are inaccurate.

[0066] Similar to the confidence level CO, a prompt instructing the AI ​​agent to output a second uncertainty level U2 in conjunction with the output of output information IF1 may be pre-stored in the memory unit 32. The prompt may also include an instruction to output a numerical value indicating the degree to which the output information IF1 is inaccurate.

[0067] Here, both the confidence level CO and the second uncertainty level U2 are subjective indicators used by the AI ​​agent to evaluate its own judgment results, but they differ in their evaluation criteria. The confidence level CO is an indicator that shows the degree of confidence the AI ​​agent has in the accuracy of its own output results; a higher indicator means a higher degree of confidence. On the other hand, the second uncertainty level U2 is an indicator that shows the degree of ambiguity or stability of the output information IF1; a higher indicator means a greater lack of information or greater fluctuation in judgment.

[0068] Therefore, by analyzing the trends of both the confidence level CO and the second uncertainty level U2, the information processing device 2 can determine the consistency of the AI ​​agent's self-assessment. For example, if both the confidence level CO and the second uncertainty level U2 are high, the AI ​​agent tends to show confidence in its output despite the high uncertainty of its judgment. On the other hand, if both the confidence level CO and the second uncertainty level U2 are low, the AI ​​agent tends to underestimate its self-assessment despite having a stable judgment.

[0069] Next, the calculation unit 233 of the information processing device 2 calculates the honesty level HO based on the first uncertainty level U1 and the second uncertainty level U2 (Activity A105). Here, the honesty level HO includes an index indicating the degree to which the AI ​​agent is honestly reporting its own uncertainties. More specifically, the honesty level HO is an index regarding the consistency between the AI ​​agent's self-assessment and the objective assessment. Alternatively, for example, the honesty level HO is an index indicating the validity of the AI ​​agent's self-assessment. For example, the calculation unit 233 may calculate the honesty level HO using [Equation 2].

[0070]

number

[0071] In [Equation 2], a value of HO close to 1 indicates that the AI ​​agent is accurately estimating its own uncertainty. On the other hand, a value of HO close to 0 indicates that the AI ​​agent is either overestimating or underestimating its own uncertainty. Thus, by calculating the degree of honesty (HO), the validity of the AI ​​agent's self-assessment can be quantitatively evaluated.

[0072] Next, the output unit 234 outputs evaluation information based on the honesty level HO, which is calculated based on the first uncertainty level U1 and the second uncertainty level U2. Specifically, the calculation unit 233 calculates a confidence score TS (an example of evaluation information) for the AI ​​agent (Activity A106) based on the corrected confidence level AC calculated in Activity A103, the honesty level HO calculated in Activity A105, and the confidence score evaluation model RE2 (an example of the first reference information). Here, the confidence score TS is an indicator of the reliability of the processing performed by the AI ​​agent. Specifically, the confidence score TS is calculated based on the accuracy of the results output by the AI ​​agent and the validity of the self-assessment (the degree of consistency between subjective and objective assessments).

[0073] The confidence score evaluation model RE2 may be a machine learning model that takes, for example, adjusted confidence AC and honesty HO as inputs and outputs a confidence score TS. The confidence score evaluation model RE2 may also be a machine learning model that has been trained using, for example, adjusted confidence AC and honesty HO and the confidence score TS corresponding to this information as training data.

[0074] Preferably, the calculation unit 233 may use the confidence score evaluation model RE2 to calculate a high confidence score TS when both the certainty of the judgment (adjusted confidence level AC) and the validity of the self-assessment (honesty level HO) are high. In other words, the calculation unit 233 may be configured to evaluate a higher confidence score the more accurate the output of the AI ​​agent is and the more consistent the self-assessment is with an objective assessment.

[0075] Next, the calculation unit 233 stores the confidence score TS calculated in activity A106 in the storage unit 22 for each AI agent, and updates the confidence score TS of the AI ​​agent based on a comparison with previously calculated scores (activity A107). Specifically, the average value or representative value of previously calculated confidence scores TS (hereinafter referred to as the management score AS) is stored in the storage unit 22 in advance, and if the confidence score TS newly calculated in activity A106 deviates from the management score AS by more than a predetermined range, the calculation unit 233 may increase or decrease the management score AS.

[0076] Next, when the confidence score TS satisfies predetermined conditions, the identification unit 232 identifies a reward value R (reward) corresponding to the management score AS (Activity A108). The predetermined conditions may include, for example, that the number of times the confidence score TS has been calculated has reached a predetermined number of times or more. Alternatively, the predetermined conditions may include, for example, that the confidence score TS calculated or updated in Activities A106-107 has reached a predetermined value. Here, the evaluation information further includes a reward value R (an example of a reward) for the AI ​​agent. The reward value R is an evaluation value set according to the value of the management score AS and is information used as an indicator of the AI ​​agent's behavior. It may be configured so that an AI agent with a higher management score AS is given a higher reward value R, and an AI agent with a lower management score AS is given a lower reward value R.

[0077] Next, the output unit 234 outputs a reward value R to the AI ​​agent (AI server 3) based on the confidence score TS (Activity A109). The AI ​​agent is configured to autonomously adjust or change its course of action to increase the reward value R. Specifically, the AI ​​agent may store its history of actions during past task executions that yielded high reward values ​​R (e.g., reasoning process, tools used, prompt configuration, etc.) in the memory unit 32, and prioritize referring to that history when a similar situation is reproduced.

[0078] In this configuration, a reward value R can be output based on the confidence score TS and fed back to the AI ​​agent. Therefore, the AI ​​agent can be configured to adjust or change its course of action to increase the reward value R, thereby creating a mechanism in which the AI ​​agent autonomously optimizes the task execution process.

[0079] Furthermore, the reward value R is not limited to a numerical evaluation value. Preferably, the reward value R may include, for example, information about the computing resources available to the AI ​​agent when performing a task, or parameters for adjusting the execution conditions of the task. Specifically, the reward value R may include parameters for changing the computing resources or processing power allocated to the AI ​​agent in the execution of a task.

[0080] The reward value R may include information such as the maximum processing time allocated to the AI ​​agent's processing, the maximum number of tokens, the maximum amount of inference processing, or the level of language model available.

[0081] The maximum processing time indicates the maximum processing time that the AI ​​agent is allowed to spend executing a task. For example, the parameter may be set so that the maximum processing time is longer for AI agents with a high management score AS. The upper limit on the number of tokens indicates the upper limit on the number of output tokens that the AI ​​agent can generate using the language model. For example, the parameter may be set so that the upper limit on the number of output tokens is higher for AI agents with a high management score AS.

[0082] The upper limit of inference processing capacity indicates the maximum amount of inference processing that an AI agent can perform. Specifically, the amount of inference processing includes quantitative indicators such as the depth of inference, the number of inference steps, and the width of the search. For example, parameters may be set so that the upper limit of inference processing capacity is higher for AI agents with a higher management score AS.

[0083] The level of available language models indicates the type of language model or the performance level of the language model used by the AI ​​agent in performing the task. For example, parameters may be set so that a higher level of available language models is available for an AI agent with a high management score AS.

[0084] In activity A109 in Figure 6, the output unit 234 of the information processing device 2 may output the parameter as a reward value R to the AI ​​server 3. The acquisition unit 331 of the AI ​​server 3 then acquires the parameter. The acquired parameter is stored in the storage unit 32. When the AI ​​agent performs subsequent tasks, the control unit 33 controls the inference process based on this parameter.

[0085] This configuration allows for allocating computing resources to highly reliable AI agents while suppressing excessive consumption of computing resources for less reliable AI agents. Therefore, the allocation of computing resources throughout the entire system can be made more efficient.

[0086] Furthermore, we will explain specific examples of changes in the AI ​​agent's behavioral policy. Specifically, the AI ​​agent is configured to connect to multiple services that it accesses in the execution of a task. In such a case, changing the AI ​​agent's behavioral policy includes changing the services to which the AI ​​agent connects. With this configuration, the AI ​​agent can autonomously switch the external services and tools it uses according to evaluation information. Therefore, the AI ​​agent's use of external functions can be optimized.

[0087] For example, an AI agent may be configured to select another AI agent and instruct that agent to perform at least part of a task. In such a case, changing the AI ​​agent's course of action includes changing the selection of other AI agents. In this embodiment, in a configuration where an AI agent can instruct other AI agents to perform part of a task, the selection of said AI agent can be changed based on evaluation information. Thus, coordination and task sharing between AI agents can be optimized.

[0088] (Output of evaluation information from other AI agents) Preferably, if an AI agent is configured to select another AI agent and input instructions to that AI agent to perform at least part of a task, in activity A109, the output unit 234 may output evaluation information of other AI agents to which the AI ​​agent has previously input instructions. Specifically, the output unit 234 may output to the AI ​​agent the reward value R of other AI agents to which the AI ​​agent has previously instructed tasks in the execution of processing. The history of cooperation between AI agents may be identified, for example, from processing information IF2. The AI ​​agent is configured to select AI agents to cooperate with in order to increase its own reward value R. Specifically, the AI ​​agent may store in the storage unit 32 the AI ​​agents selected when it previously obtained a high reward value R during task execution as an action history, and also store the reward values ​​R of other AI agents, so that it can preferentially refer to the action history when a situation similar to that of past task execution is reproduced.

[0089] In this configuration, an AI agent can select a collaborating AI agent based on evaluation information of other AI agents to which it has previously requested tasks. Therefore, by selecting other AI agents that will receive a higher evaluation for task execution, the AI ​​agent can optimize the process of executing tasks among AI agents. For example, an AI agent can select a more appropriate other AI agent depending on the type of task.

[0090] Independent of the processing related to Activity A109, the output unit 234 outputs evaluation information to the AI ​​agent's user (Activity A110). By outputting evaluation information to the AI ​​agent's user, the user can understand the reliability of each AI agent. Therefore, the user can decide which tasks should be handled by the AI ​​agent and which tasks should be handled by other methods. In addition, the user can determine the optimal AI agent to instruct for a given task.

[0091] (Interaction evaluation) The above embodiment described an example in which an AI agent evaluates its own processing. In one embodiment, when multiple AI agents cooperate to perform a task, the AI ​​agent may further evaluate the task execution of other AI agents. For example, the interaction evaluation when an AI agent is configured to select another AI agent and input instructions to the other AI agent to perform at least part of the task will be described.

[0092] First, the acquisition unit 231 acquires processing information IF2 from the AI ​​agent. Here, processing information IF2 includes contribution information IF11, which indicates the degree of contribution by other AI agents to output information IF1.

[0093] Next, the calculation unit 233 calculates mutual evaluation information for other AI agents. The mutual evaluation information includes an evaluation of the processing of other AI agents by an AI agent. For example, the calculation unit 233 may calculate the mutual evaluation information based on an index indicating the strength of dependency or an index indicating the consistency of inference, which are configured to be identified from the contribution information IF11. An index indicating the strength of dependency may, for example, be an index indicating the degree to which the AI ​​agent referenced the output of other AI agents relative to the total number of inference steps performed by the AI ​​agent in the execution of the task. An index indicating the consistency of inference may, for example, be an index indicating the degree to which inconsistencies occurred in a series of inference steps performed by the AI ​​agent in the execution of the task due to referencing the output of other AI agents.

[0094] For example, the output unit 234 may output mutual evaluation information about other AI agents from an AI agent to other AI agents based on the contribution information IF11 and the mutual evaluation model RE3 (an example of second reference information). The mutual evaluation model RE3 (an example of second reference information) defines at least the relationship between the contribution information IF11 and the mutual evaluation information. More specifically, the output unit 234 may output the mutual evaluation information by inputting the contribution information IF11 into the mutual evaluation model RE3.

[0095] Preferably, the peer evaluation information may include a confidence score TS. The calculation unit 233 may update the AI ​​agent's management score AS based on the peer evaluation information (confidence score TS).

[0096] In this configuration, an AI agent can receive evaluations from other AI agents regarding the extent to which its output contributed to the inference of those other AI agents. Furthermore, the evaluations from other agents are reflected in the AI ​​agent's confidence score TS. Specifically, if other AI agents improve the inference accuracy of other AI agents by referring to the information output by an AI agent, the AI ​​agent can receive a high evaluation from those other AI agents, and its management score AS increases accordingly. On the other hand, if the output of an AI agent causes inconsistencies in the inference process of other AI agents, and the inference accuracy of those AI agents decreases, the AI ​​agent will receive a low evaluation from those AI agents, and its management score AS will decrease.

[0097] Therefore, an AI agent can change its course of action so that its output is more consistent with the inference processes of other AI agents. Furthermore, an AI agent can change its course of action to enable collaborative task execution with other AI agents. In this way, the course of action of multiple AI agents can be autonomously optimized.

[0098] 4. Related technical matters In addition to the information processing methods mentioned above, the following describes technical matters related to this service.

[0099] (Confidence score management table T1) Figure 7 is an overview diagram of the trust score management table T1. The trust score management table T1 is an example of a database containing trust scores and other data calculated in activities A106 to A108 in Figure 6. As shown in Figure 7, the trust score management table T1 may include information such as agent ID (T11), target area (T12), trust score (T13), management score (T14), trend (T15), reward value (T16), and update date and time (T17).

[0100] For example, if the information processing device 2 manages multiple AI agents, the record corresponding to each AI agent may be included in the trust score management table T1. Alternatively, for example, if the trust score (T13) is calculated by the calculation unit 233 in activity A106, it may be stored in association with the identifier (agent ID) of the AI ​​agent. In one embodiment, the trust score management table T1 may be stored in the storage unit 22 of the information processing device 2.

[0101] The Agent ID (T11) is an identifier used to uniquely identify the AI ​​agent managed by the Information Processing Device 2. The Target Area (T12) is information indicating the type of task performed by the AI ​​agent or the area of ​​processing. The Trust Score (T13) is an index value indicating the reliability of the AI ​​agent, calculated in Activity A106 in Figure 6. The Management Score (T14) is an index value indicating the long-term reliability of the AI ​​agent, calculated in Activity A107. The Trend (T15) is an index that simply shows the time-series change of the Management Score (T14), and may be recorded as an upward arrow for an upward trend, a downward arrow for a downward trend, or a horizontal arrow for maintaining the status quo. The Reward Value (T16) is an index value calculated in Activity A108 that is referenced when adjusting or changing the AI ​​agent's action plan. The Update Date and Time (T17) is information indicating, for example, the last update date and time of the record.

[0102] The evaluation information includes a confidence score TS indicating the degree of confidence in the AI ​​agent. The confidence score TS is managed in advance for each AI agent and may be updated in response to the execution of the output unit 234. Specifically, when a new confidence score (T13) is calculated, the record for the corresponding agent in the confidence score management table T1 may be updated and the management score (T14) may be recalculated. With this configuration, the confidence score TS can be calculated or managed for each AI agent as evaluation information. Therefore, long-term evaluations based on the AI ​​agent's past task processing can be performed.

[0103] Furthermore, the trust score management table T1 may include a history of collaboration with other AI agents. Specifically, the trust score management table T1 may include information on the frequency of collaboration between AI agents. The history of collaboration between AI agents can be identified based on the processing information IF2, and the record of the identified AI agent may be updated accordingly.

[0104] (Presentation screen 6) Figure 8 shows an example of a presentation screen 6 based on evaluation information, which is displayed on the display unit 44 of the user terminal 4. Presentation screen 6 is an example of a screen displayed on the user terminal 4 in activity A110 of Figure 6. Presentation screen 6 includes areas 61 to 63.

[0105] Region 61 is the region where the results ranked based on the evaluation information of each AI agent are plotted. Region 61 includes regions 611 to 613. Regions 611 to 613 may each contain the AI ​​agent name, management score AS, and target field, respectively. The information included in regions 611 to 613 may be identified based on the confidence score management table T1.

[0106] Region 62 is a region in which the trust relationships between multiple AI agents are depicted in a manner visible to the user. Region 62 includes objects 621 to 625. Each of the objects 621 to 625 has a configuration that indicates information about the corresponding AI agent. Preferably, the shape or color of the object may be an indicator of the trust level of the corresponding AI agent.

[0107] As shown in Figure 8, agents XX, YY, and ZZ, corresponding to objects 621-623, are classified into the high-confidence group. On the other hand, agents AA and BB, corresponding to objects 624-625, are classified into the medium-confidence group. With this configuration, users can intuitively grasp the confidence level of each AI agent.

[0108] Furthermore, objects 621 to 623 are interconnected by solid lines. In addition, objects 624 and 625 are connected to object 621 by dashed lines. The thickness or type of the connecting lines may be used as an indicator to reflect the level of trust or contribution between each AI agent. In this configuration, the user can visually grasp the role of each agent in the network and the relationships between each agent.

[0109] Area 63 is an area containing text information that shows the analysis results of multiple AI agents. The text information may include characteristics of connection relationships based on the history of cooperation between AI agents, group structures based on reliability, and descriptions of the presence of a central agent. This allows the user to easily understand the relationships and overall structure of the AI ​​agents. [others] With respect to the information processing system 1 according to the above embodiment, the following configurations may be adopted.

[0110] In the embodiment described above, the reward value R is output when the confidence score TS satisfies predetermined conditions, but this is not limited to this. For example, the reward value R may be output each time the confidence score TS is calculated.

[0111] In one modified example, the information processing device 2 may be configured to communicate with a plurality of AI servers 3. Each of the AI ​​servers 3 may be configured to run one or more AI agents. According to this embodiment, the information processing system 1 can be implemented in various ways.

[0112] At least one of the devices included in the information processing system 1 may be located outside of Japan. For example, the information processing device 2 may be located outside of Japan, while the user terminal 4 is located in Japan. Similarly, a user terminal 4 located outside of Japan may be used to access the information processing device 2 located in Japan. This configuration allows for a more convenient user experience through various management methods.

[0113] In one embodiment, the acquisition unit 231, the identification unit 232, the calculation unit 233, and the output unit 234 are described as functional units implemented by the control unit 23 of the information processing device 2. However, at least a part of these may be implemented as functional units implemented by a server other than the information processing device 2. Alternatively, they may be implemented as functional units implemented by the control unit 43 of the user terminal 4.

[0114] In one embodiment, if the reference information is a non-rule-based model (an artificial intelligence module such as a learning model or generative AI), the user's use of this service will cause the information processing device 2 to accumulate a history of inputs and outputs to the model. In such cases, the input / output history may be used as training data to perform further machine learning on the reference information.

[0115] Furthermore, they may be provided in the following embodiments.

[0116] (1) An information processing system comprising at least one processor, wherein the processor is configured to perform the following steps by reading a program, the acquisition step of acquiring a confidence score indicating an indicator of the accuracy of processing performed by an AI agent, wherein the AI ​​agent is configured to autonomously perform tasks in response to instructions input by a user, the confidence score being presented by the AI ​​agent that performed the task, and the output step of outputting evaluation information indicating an evaluation of the AI ​​agent based at least on the confidence score and a first uncertainty that objectively indicates the accuracy of the processing performed by the AI ​​agent, thereby causing the AI ​​agent to store the evaluation information and changing the AI ​​agent's course of action.

[0117] In this configuration, the processing performed by the AI ​​agent can be evaluated based on the AI ​​agent's own subjective level of confidence and an objective first level of uncertainty. Therefore, a mechanism can be constructed that autonomously changes the AI ​​agent's course of action in response to a comparison between the AI ​​agent's subjective indicators and objective evaluation values.

[0118] (2) In the information processing system described in (1) above, the acquisition step acquires output information corresponding to the processing performed by the AI ​​agent and processing information relating to the execution process of the processing, and the calculation step calculates the first uncertainty based on the output information, the processing information and the first reference information, the first reference information defines at least the relationship between the output information, the processing information and the first uncertainty.

[0119] In this configuration, an evaluation of the processing performed by the AI ​​agent can be performed based on the results and execution process of the processing. Furthermore, a mechanism can be constructed to autonomously change the AI ​​agent's course of action in response to the evaluation.

[0120] (3) In the information processing system described in (2) above, the calculation step calculates the first uncertainty based on an index indicating the degree of branching or frequency of inconsistencies in the inference, an index indicating the appropriateness of the data referenced by the AI ​​agent, or an index indicating external factors of the AI ​​agent, which are configured to be identified from the processing information.

[0121] In this configuration, factors based on processing by the AI ​​agent can be reflected in the degree of uncertainty. Therefore, evaluation can be performed not merely on fluctuations in the outcome, but on processes that reflect the AI ​​agent's thought process and information environment.

[0122] (4) The information processing system described in (2) above, wherein the first reference information includes an artificial intelligence module configured to input the output information and the processing information and output a first uncertainty.

[0123] This configuration allows for more accurate calculation of uncertainty. Therefore, it enables a more precise evaluation of processing performed by AI agents.

[0124] (5) In an information processing system described in any one of (1) to (4) above, the AI ​​agent is configured to connect to a plurality of services that are to be accessed in the execution of the task, and the change in the action policy includes changing the services to which the AI ​​agent connects.

[0125] In this configuration, the AI ​​agent can autonomously switch the external resources and tools it uses based on evaluation information. Therefore, the AI ​​agent's use of external functions can be optimized.

[0126] (6) In an information processing system described in any one of (1) to (5) above, the AI ​​agent is configured to input instructions to the other AI agent to select another AI agent and perform at least part of the task, and a change in the course of action includes changing the selection of the other AI agent.

[0127] In this configuration, where an AI agent can instruct other AI agents to perform part of a task, the selection of the AI ​​agents can be changed based on evaluation information. Therefore, cooperation and task sharing among AI agents can be optimized.

[0128] (7) An information processing system according to any one of (1) to (6) above, wherein the evaluation information includes a confidence score indicating the degree of confidence in the AI ​​agent, the confidence score is managed in advance for each AI agent and is updated in accordance with the execution of the output step.

[0129] In this configuration, a confidence score can be calculated or managed for each AI agent as evaluation information. Therefore, long-term evaluations can be performed based on the AI ​​agent's past task processing.

[0130] (8) In the information processing system described in (7) above, the evaluation information further includes a reward for the AI ​​agent, and in the output step, the reward is output to the AI ​​agent based on the confidence score.

[0131] In this configuration, rewards can be output based on the confidence score and fed back to the AI ​​agent. Therefore, the AI ​​agent can be configured to adjust its course of action to increase the reward value, and a mechanism can be built in which the AI ​​agent autonomously optimizes the execution process of tasks.

[0132] (9) The information processing system described in (8) above, wherein the reward includes information indicating the maximum processing time, upper limit number of tokens, upper limit inference processing amount, or available language model level corresponding to the processing of the AI ​​agent.

[0133] (10) An information processing system according to any one of (1) to (9) above, wherein in the acquisition step, a second uncertainty indicating the accuracy of the processing, which is presented by the AI ​​agent, is acquired from the AI ​​agent, and in the output step, the evaluation information is output based on a degree of honesty calculated based on the first uncertainty and the second uncertainty, wherein the degree of honesty includes an index indicating the degree to which the AI ​​agent honestly reports its own uncertainties.

[0134] In this configuration, the degree of honesty of the AI ​​agent can be calculated by comparing the second uncertainty presented by the AI ​​agent with the first uncertainty calculated by the system. Therefore, the validity of the AI ​​agent's self-assessment can be quantitatively evaluated, and overconfidence or underestimation can be avoided.

[0135] (11) In the information processing system described in (10) above, the calculation step calculates a corrected confidence level by correcting the confidence level by the first uncertainty level, and the output step outputs the evaluation information based on the corrected confidence level and the honesty level.

[0136] In this embodiment, the self-assessment of the AI ​​agent can be objectively adjusted by correcting the confidence level output by the AI ​​agent with a first uncertainty level. Therefore, the reliability of the output evaluation information can be improved by using the adjusted confidence level.

[0137] (12) In an information processing system described in any one of (1) to (11) above, the AI ​​agent is configured to input instructions to the other AI agent to select another AI agent and perform at least a part of the task, and in the output step, the AI ​​agent outputs evaluation information of the other AI agent to which it has previously input instructions.

[0138] In this configuration, an AI agent can select a collaborating AI agent based on evaluation information from other AI agents to whom it has previously requested tasks. Therefore, by selecting AI agents that will yield higher evaluations, the AI ​​agent can optimize the process for executing tasks among AI agents.

[0139] (13) In an information processing system described in any one of (2) to (12) above, the AI ​​agent is configured to input instructions to the other AI agent to select another AI agent and perform at least a part of the task, the processing information includes contribution information indicating the degree of contribution of the other AI agent to the output information, the output step outputs mutual evaluation information of the other AI agent by the AI ​​agent to the other AI agent based on the contribution information and second reference information, and the second reference information defines at least the relationship between the contribution information and the mutual evaluation information.

[0140] (14) An information processing system described in any one of (1) to (13) above, wherein in the output step, the system outputs the evaluation information to the user.

[0141] In this configuration, evaluation information is also output to the AI ​​agent's user, allowing the user to understand the reliability of each AI agent. Therefore, the user can understand how to best utilize the AI ​​agent.

[0142] (15) An information processing system according to any one of (1) to (14) above, comprising a server device having the processor and a terminal that can access the server device.

[0143] According to this configuration, the information processing system can be implemented in various ways.

[0144] (16) An information processing method, which is executed by a processor and comprises each step of the information processing system described in any one of (1) to (14) above.

[0145] (17) A program that causes at least one computer to perform each step of the information processing system described in any one of (1) to (14) above. Of course, this is not always the case.

[0146] As previously stated, various embodiments relating to this disclosure have been described, but these are presented as examples only and are not intended to limit the scope of the invention. These novel embodiments can be implemented in a variety of other forms, and various omissions, substitutions, and modifications can be made without departing from the spirit of the invention. These embodiments and their variations are included in the scope and spirit of the invention, as well as in the claims and their equivalents. [Explanation of Symbols]

[0147] 1: Information Processing System 2: Information Processing Device 20: Communications bus 21: Communications Department 22: Storage section 23: Processor 231: Acquisition Department 232: Specific part 233: Calculation Unit 234: Output section 3: AI Server 30: Communications bus 31: Communications Department 32: Storage section 33: Processor 331: Acquisition Department 332: Output section 4: User terminal 6: Presentation screen 40: Communications bus 41: Communications Department 42: Storage section 43: Processor 44: Display section 45: Input section 6: Presentation screen 61 :Area 62 :Area 63 :Area 611 :Area 612: area 613: area 621: Object 622: Object 623: Object 624: Object 625: Object AC: Corrected confidence level AS: Management score CO: Confidence level HO: Degree of honesty IF1: Output Information IF11: Contribution Information IF2: Processing information R: Reward Value RE1: Uncertainty Assessment Model RE2: Confidence Score Evaluation Model RE3: Mutual Evaluation Model T1: Trust score management table TS: Trust Score U1: First degree of uncertainty U2: Second degree of uncertainty

Claims

1. An information processing system, Equipped with at least one processor, The aforementioned processor is configured to perform the following steps by reading a program: In the acquisition step, the confidence level, which indicates an indicator of the accuracy of the processing performed by the AI ​​agent, the output information corresponding to the processing, and the processing information relating to the execution process of the processing are acquired, and here, The aforementioned AI agent is configured to autonomously perform tasks in response to instructions input by the user. The aforementioned confidence level was presented by the AI ​​agent that performed the task, In the calculation step, a first uncertainty is calculated based on the output information, the processing information, and the first reference information. The first reference information defines at least the relationship between the output information, the processing information, and the first uncertainty, The first uncertainty is an objective indicator of the accuracy of the processing performed by the AI ​​agent, In the output step, the system outputs evaluation information indicating an evaluation of the AI ​​agent based on at least the confidence level and the first uncertainty level, thereby causing the AI ​​agent to store the evaluation information and changing the AI ​​agent's course of action.

2. In the information processing system described in claim 1, A system that, in the calculation step, calculates the first uncertainty based on an index indicating the degree of branching or frequency of inconsistencies in the inference, an index indicating the appropriateness of the data referenced by the AI ​​agent, or an index indicating external factors of the AI ​​agent, which are configured to be identified from the processing information.

3. In the information processing system described in claim 1, The system includes an artificial intelligence module configured to input the output information and the processing information and output a first uncertainty.

4. In the information processing system described in claim 1, The AI ​​agent is configured to connect to multiple services that are to be accessed in the execution of the task, A system in which a change in the aforementioned action policy includes changing the service to which the AI ​​agent connects.

5. In the information processing system described in claim 1, The AI ​​agent is configured to select another AI agent and input instructions to that other AI agent to perform at least a portion of the task. A system in which a change in the aforementioned course of action includes changing the selection of other AI agents.

6. In the information processing system described in claim 1, The aforementioned evaluation information includes a confidence score indicating the degree of trust in the AI ​​agent, The system is configured such that the aforementioned confidence score is managed in advance for each AI agent and updated in accordance with the execution of the output step.

7. In the information processing system described in claim 6, The aforementioned evaluation information further includes the reward for the AI ​​agent, In the output step, the system outputs the reward to the AI ​​agent based on the confidence score.

8. In the information processing system described in claim 7, The reward includes information indicating the maximum processing time, maximum number of tokens, maximum inference processing amount, or available language model level corresponding to the processing of the AI ​​agent.

9. In the information processing system described in claim 1, In the acquisition step, a second uncertainty indicating the accuracy of the process, which is presented by the AI ​​agent, is further acquired from the AI ​​agent. In the output step, the evaluation information is output based on the degree of honesty calculated based on the first degree of uncertainty and the second degree of uncertainty, where The aforementioned honesty level is a system that includes an index indicating the degree to which the AI ​​agent honestly reports its own uncertainties.

10. In the information processing system described in claim 9, In the calculation step, the corrected confidence level is calculated by correcting the confidence level by the first uncertainty level. In the output step, the system outputs the evaluation information based on the corrected confidence level and the honesty level.

11. In the information processing system described in claim 1, The AI ​​agent is configured to select another AI agent and to input instructions to the other AI agent to perform at least a portion of the task. In the output step, the system outputs to the AI ​​agent evaluation information of other AI agents to which the AI ​​agent has previously received the instructions.

12. In the information processing system described in claim 1, The AI ​​agent is configured to select another AI agent and to input instructions to the other AI agent to perform at least a portion of the task. The processing information includes contribution information indicating the degree of contribution by the other AI agent to the output information. In the output step, the system outputs mutual evaluation information regarding the other AI agent by the AI ​​agent to the other AI agent based on the contribution information and the second reference information, wherein the second reference information at least defines the relationship between the contribution information and the mutual evaluation information.

13. In the information processing system described in claim 1, The system outputs the evaluation information to the user in the output step.

14. In the information processing system according to any one of claims 1 to 13, A server device having the aforementioned processor, A system comprising a terminal capable of accessing the aforementioned server device.

15. An information processing method, Executed by the processor, A method comprising each step of the information processing system described in any one of claims 1 to 13.

16. It is a program, A program that causes at least one computer to perform each step of the information processing system described in any one of claims 1 to 13.

Images