Information processing device, information processing method, and information processing program
The information processing device addresses the challenge of reflecting human intent in complex tasks by assigning, redesigning, and updating tasks using a generation and update unit, ensuring accurate and efficient alignment with human instructions through reinforcement learning.
Patent Information
- Application Number
- PCT/JP2026/017424
- Authority / Receiving Office
- WO · WO
- Patent Type
- Applications
- Current Assignee / Owner
- Priority Date
- 2025-05-08
- Filing Date
- 2026-05-01
- Publication Date
- 2026-11-12
AI Technical Summary
Conventional technologies using Large Language Models (LLMs) struggle to accurately reflect human intent in complex tasks, particularly when multiple agents are involved, leading to deviations from intended outcomes due to the complexity and dynamic nature of human instructions.
An information processing device that includes a generation unit to assign tasks to multiple agents, a design unit to redesign composite tasks based on agent approaches, and an update unit to incorporate human instructions, utilizing reinforcement learning to optimize the LLM's performance and ensure alignment with human intent.
The solution enables accurate reflection of human intent in complex tasks, reducing the time and effort required for instruction, while maintaining creativity and consistency across multiple processes and agents, thereby improving the accuracy of outputs.
Smart Images

Figure JP2026017424_12112026_PF_FP_ABST
Abstract
Description
Information Processing Apparatus, Information Processing Method, and Information Processing Program
[0001] The present invention relates to an information processing apparatus, an information processing method, and an information processing program.
[0002] Conventionally, there is known a technique for executing tasks by coordinating agents equipped with Large Language Models (LLMs).
[0003] For example, a technique is known for constructing a team in which a plurality of specialized agents adaptively cooperate by defining the role of each agent and extracting knowledge related to the role from the LLM (see, for example, Non-Patent Document 1, Non-Patent Document 2, and Non-Patent Document 3).
[0004] In addition, a technique is known for dividing a predefined complex task into individual tasks for each agent, so that individual solutions generated by the agents are shared within the team to address the overall composite task (see, for example, Non-Patent Document 1 and Non-Patent Document 3).
[0005] Guangyao Chen, Siwei Dong, Yu Shu, Ge Zhang, Jaward Sesay, Boerje Karlsson, Jie Fu, and Yemin Shi. 2024a. Autoagents: A framework for automatic agent generation. In Proc. I, JCAI'24, pages 22-30. Weize Chen, Yusheng Su, Jingwei Zuo, Cheng Yang, Chenfei Yuan, Chi-Min Chan, Heyang Yu, Yaxi, Lu, Yi-Hsin Hung, Chen Qian, Yujia Qin, Xin Cong, Ruobing Xie, Zhiyuan Liu, Maosong Sun, and Jie Zhou. 2024c. Agentverse: Facilitating multi-agent collaboration and exploring emergent behaviors. In Proc. ICLR'24. Zhenhailong Wang, Shaoguang Mao, Wenshan Wu, Tao Ge, Furu Wei, and Heng Ji. 2023. Unleashing cognitive synergy in large language models: A task-solving agent through multi-persona self-collaboration. CoRR, abs / 2307.05300.
[0006] However, conventional technologies have a problem in that they make it difficult to reflect human intent in solving complex tasks.
[0007] For example, tasks such as creating service proposals or novels tend to improve in accuracy when humans provide detailed instructions to the LLM (Language-Limited Modeling). On the other hand, repeatedly writing detailed instructions from humans within a multi-agent framework is time-consuming.
[0008] Furthermore, in complex tasks, it is not easy to reflect human intentions that change during task progress in the prompts. As a result, the generated output may deviate from human intentions. Also, when solving a task with multiple processes using multiple agents, the large number and diversity of processes and agents make it difficult to accurately reflect human instructions in the output.
[0009] Therefore, the present invention aims to reflect human intent in the resolution of complex tasks.
[0010] To solve the problem, the information processing device of the present invention includes a generation unit that assigns each of a plurality of tasks obtained based on a given task to each of a plurality of agents, a design unit that designs a composite task based on the task-solving approach presented by each of the plurality of agents, and an update unit that updates the composite task designed by the design unit using instructions given by a human.
[0011] According to the present invention, human intentions can be reflected in the resolution of complex tasks.
[0012] Figure 1 is a diagram showing an example configuration of the information processing device of the first embodiment. Figure 2 is a flowchart showing the processing flow of the information processing device. Figure 3 is a diagram illustrating the procedure for task resolution. Figure 4 is a diagram illustrating the procedure for task resolution. Figure 5 is a diagram illustrating reinforcement learning of the control model. Figure 6 is a diagram illustrating reinforcement learning of the transmission model. Figure 7 is a diagram illustrating an overview of Example 1. Figure 8 is a diagram showing an example configuration of the information processing device of Example 1. Figure 9 is a flowchart showing the processing flow of the information processing device of Example 1. Figure 10 is a diagram showing an example output of Example 1. Figure 11 is a diagram illustrating the processing of multiple services. Figure 12 is a diagram showing an example configuration of a computer that executes an information processing program.
[0013] The embodiments for carrying out the present invention will be described below with reference to the drawings. The present invention is not limited to these embodiments.
[0014] <Configuration of the First Embodiment> The configuration of the information processing device will be explained using Figure 1. Figure 1 is a diagram showing an example of the configuration of the information processing device according to the first embodiment.
[0015] As shown in Figure 1, the information processing device 10 includes a communication unit 11, an input unit 12, an output unit 13, a storage unit 14, and a control unit 15.
[0016] The communication unit 11 is a module for data communication with other devices. The communication unit 11 is, for example, a NIC (Network Interface Card). The input unit 12 is an interface connected to input devices such as a mouse and a keyboard. The output unit 13 is an interface connected to output devices such as a speaker and a display.
[0017] The memory unit 14 stores data, programs, etc., that are referenced when the control unit 15 performs various processes. The memory unit 14 is implemented by semiconductor memory elements such as RAM (Random Access Memory) and flash memory, or by storage devices such as hard disks and optical discs. The memory unit 14 stores model information 141.
[0018] Model information 141 consists of parameters of the language model. For example, if the language model is a neural network, then model information 141 consists of parameters such as weights and biases. The language model in the first embodiment is an LLM (Language Language Model). Furthermore, the LLM described below is constructed based on model information 141.
[0019] The control unit 15 is responsible for controlling the entire information processing device 10. The functions of the control unit 15 are realized, for example, by the CPU (Central Processing Unit) executing a program stored in the memory unit 14. The control unit 15 includes a generation unit 151, a design unit 152, an update unit 153, and a learning unit 154.
[0020] The agent solves the task. Specifically, the agent can be described as a system (or program) that uses LLM to obtain a solution to the task. Each operation of the agent described below can be realized by the information processing device 10 prompting the LLM or by the information processing device 10 obtaining the output of the LLM. The information given to the agent and the information output by the agent may be data in natural language text format. Furthermore, techniques for having an agent perform processes that mimic human activities such as meetings and conversations are known (see, for example, Non-Patent Documents 1, 2, and 3).
[0021] The generation unit 151 performs the process of generating output (products, solutions) for a composite task. For example, the composite task is a text describing information about a product. For example, the output is a marketing plan for the product.
[0022] The design unit 152 redesigns the composite task. The update unit 153 updates the composite task. The learning unit 154 trains the model. For example, the learning unit 154 fine-tunes the LLM.
[0023] Here, we will explain the processing of each part included in the control unit 15 using Figure 2. Figure 2 is a flowchart showing the processing flow of the information processing device. The information processing device 10 can perform each process by giving prompts to the LLM (including the LLM installed in the agent). Furthermore, in the process where each part gives prompts to operate the agent, the agent may be described as the main entity performing the processing.
[0024] A human presents a team of agents, comprised of multiple agents, with a complex task to be solved. For example, the complex task at this stage may simply be a combination of several different tasks, as is done in the Trivia Creative Writing task described in Non-Patent Documents 1 and 3. Furthermore, the complex task may not be one that the human has been able to fully refine.
[0025] As shown in Figure 2, first, the generation unit 151 divides the complex task (step S101 in Figure 2). The generation unit 151 then assigns each of the multiple tasks obtained by the division to an agent (step S102 in Figure 2, (1) in Figure 3). Figure 3 is a diagram illustrating the task resolution procedure.
[0026] For example, the generation unit 151 automatically divides a complex, large task presented by a human into multiple agents configured as a team using natural language processing, structuring it in a way that maintains orthogonality and coordination. Note that when the generation unit 151 assigns a complex task to an agent, there may be overlaps in the tasks sent to each agent when assigning two or more tasks that make up the complex task. In such cases, the generation unit 151 does not necessarily have to divide each task that makes up the complex task individually; it may divide the task in a way that includes two or more tasks. When the generation unit 151 assigns an agent a task that includes two or more tasks, it may assign the agent by providing the agent with task assignment information indicating which task the receiving agent will be responsible for.
[0027] As shown in Figure 3(1), the generation unit 151 assigns each of the multiple tasks (tasks 1 to 4) obtained by dividing the composite task to each agent.
[0028] The multiple tasks after division are orthogonal to each other, and their concretization is driven by discussions among the agents.
[0029] Next, each agent generates knowledge of the assigned task (step S103 in Figure 2, (2) in Figure 3). Each agent generates knowledge to solve the assigned task. Each agent's initial knowledge may be empty. Each agent adds the generated knowledge to itself.
[0030] Next, each agent presents an approach to solving the task (step S104 in Figure 2). This allows agents to understand the tasks and knowledge possessed by other agents. At this time, each agent conducts a team meeting as appropriate ((3) in Figure 3). In the team meeting, each agent shares their respective objectives and corresponding tasks within the team and consolidates their opinions. In other words, the team meeting involves the integration of each agent's approach. For example, by viewing, comparing, and understanding the task-solving approaches presented by each agent, agents can complement each other's knowledge gaps and improve the consistency of their task designs.
[0031] Here, the design unit 152 redesigns the composite task (step S105 in Figure 2, (4) in Figure 3). For example, the meta-agent combines the approaches presented by each agent and redesigns the composite task. At this time, the design unit 152 issues a redesign prompt. This redesign prompt is named the "DESIGN prompt". The design unit 152 generates a redesign prompt (DESIGN prompt) and redesigns the task at the stage of integrating the task-solving approaches of multiple agents. Alternatively, the redesign may be performed by the leader agent randomly selected from among the multiple agents.
[0032] After receiving the instruction, the design unit 152 designs a composite task that reflects the instruction. For example, the design unit 152 has in advance stored multiple possible solutions for the task and instruction given by a human, and reflects the candidate for the optimal solution in the generated output while preserving the differences in expression and intent of each.
[0033] The generation unit 151 accepts input of instruction statements for the redesigned composite task (step S106 in Figure 2, (5) in Figure 4). Tasks 1, 2, 3, and 4 are assumed to have been changed to Task 1', Task 2', Task 3', and Task 4' through redesign. Figure 4 is a diagram illustrating the task resolution procedure.
[0034] For example, a human reviews the redesigned composite task in step S105, and if they are not satisfied, they instruct the information processing device 10 to swap, add, delete, or modify the task. Specifically, the human reviews the tasks assigned to each agent and instructs the swapping, adding, deleting, or modifying of the tasks. When instructing a change to an agent's task, the human provides clear instructions on whether to add an item to the content of the task already assigned to that agent, or to delete the assigned task itself. Hereinafter, the text describing the above instructions given by the human will be called an "instruction statement."
[0035] Examples of instructions include: "Swap Task 1' with Task 5'," "Add item X to Task 1'," and "Delete Task 1' and assign Task 7."
[0036] The generation unit 151 determines whether the complex task has been resolved (step S107 in Figure 2). The generation unit 151 determines that the complex task has been resolved if the instruction indicates that the human was satisfied with the complex task.
[0037] If the composite task is not resolved (step S107; No), the update unit 153 links the instruction statement to the composite task and updates the composite task (step S108). After that, the generation unit 151 returns to step S101 and repeats the process.
[0038] If the complex task is resolved (step S107; Yes), the generation unit 151 proceeds to the next process (step S109). For example, the next process is the production process. In the production process, the generation unit 151 generates products based on the complex task using the agent's knowledge. The products are documents such as proposals and novels. The products may also be described as creative descriptions, creative documents, etc.
[0039] For example, the update unit 153 updates the corresponding agent task based on any operation of swapping, adding, deleting, or modifying tasks provided by a human after the redesign of the complex task. As a result, the generation unit 151 can repeat the task division, knowledge generation, and integration processes to realize a Human-in-the-Loop process that gradually constructs an output (e.g., a creative description) that reflects human intent.
[0040] The learning unit 154 aims to optimize an objective function that measures how accurately the instructive statements are reflected in the creative descriptions, which are the products. For example, the learning unit 154 uses an index called Rouge (see Reference 1) to evaluate the accuracy of the products. Rouge is an index that measures the extent to which the text generated as a product contains the correct reference response.
[0041] Reference 1: Lin, C.-Y.: ROUGE: A Package for Automatic Evaluation of Summaries, in Text Summarization Branches Out, pp. 74{81 (2004)
[0042] The learning unit 154 fine-tunes the LLM based on the Rouge generated by the LLM on the composite task. Specifically, the learning unit 154 uses the Rouge-L score as a direct reward and optimizes the model using reinforcement learning (RL). In this method, there are two types of reward scores: a control score and a transfer score.
[0043] Here, the LLM is considered by dividing it into a control model and a communication model. The control model performs division of the composite task (step S101 in Figure 2) and assignment of tasks to agents (step S102 in Figure 2). The communication model performs redesign of the composite task (step S105 in Figure 2). In the second and subsequent loops, in the redesign of the composite task, redesign that takes the instruction text into consideration is performed. Therefore, according to the redesign of the composite task, the instruction text is reflected in the composite task. Note that it is assumed that the execution subject of step S101, step S102, and step S105 in Figure 2 is a meta-agent.
[0044] Figure 5 is a diagram for explaining reinforcement learning of the control model. As shown in Figure 5, there may be a plurality of processes for producing a product. In each process, the processing of Figure 2 is performed. For example, the process of generating a proposal document is performed by agent A, agent B, agent C, and agent D. For example, the process of generating a specification document is performed by agent E, agent F, agent G, and agent H.
[0045] The learning unit 154 performs learning of the control model by optimizing a control score that indicates how much the entire sentence as a product, which is not included in the instruction text, reflects the content of the correct solution (corresponding to the aforementioned correct reference response) corresponding to the task assigned to the agent. The correct solution may be learning data that has the role of a product.
[0046] Note that the meta-agent may be common among a plurality of processes. That is, the information processing apparatus 10 includes a meta-agent that detects human requests in each process and reflects them in downstream processes through prompt generation and instruction transmission in multi-agent task execution spanning a plurality of processes. For tasks with complex processes, the meta-agent structures the examination results, progress status, and discussion history of each process as incidental knowledge, and distributes and utilizes the incidental knowledge to downstream processes, thereby improving the consistency and quality of the entire task.
[0047] In addition, the meta-agent structures the review history (auxiliary knowledge) of each agent and shares it between processes, and optimizes instruction transmission through reinforcement learning so that the content of human instructions is accurately reflected in the deliverables.
[0048] FIG. 6 is a diagram for explaining reinforcement learning of a transmission model. As shown in FIG. 6, there may be a plurality of processes for producing a product. In each process, the processing of FIG. 2 is performed. For example, within the process, a deliverable that is a composite task before reflecting the instruction text, and a review result that is a composite task after reflecting the instruction text are generated.
[0049] The learning unit 154 performs learning of the transmission model by optimizing a transmission score indicating how much the entire sentence as the product reflects the content of the correct solution (corresponding to the aforementioned correct reference response). The correct solution may be learning data having the role of an instruction text.
[0050] There may be a plurality of the aforementioned correct solutions for each instruction text or task. This is to enable joint design of complex tasks as an agent team including humans while having multiple potential solutions for each individual LLM task, and to enable exploration of diverse solutions that can demonstrate creativity. In other words, these solutions represent different interpretations and representations of the composite task, each of which can be considered a correct answer.
[0051] In the design of the reward score, the importance of these different types of correct answers is considered, the score for the correct answer of the instruction text is set high, then the score for the correct answer of the task as a background element is set, and learning adjustment is performed.
[0052] Note that since the Rouge-L score is calculated in the range from 0 to 1, the absolute value of the score does not fluctuate significantly even if the length of the generated sentence is different. This characteristic can reduce the risk that the convergence of learning is hindered due to large fluctuations in the reward score according to the length of the generated sentence. As a result, the evaluation of the quality of the generated sentence is stabilized, and efficient learning is enabled.
[0053] Furthermore, the learning unit 154 utilizes LoRA (Low-Rank Adaptation) to limit the model's learning target to low-rank matrices, thereby conserving computational resources while achieving efficient model optimization.
[0054] Specifically, the learning unit 154 designs the Rouge-L score as the reward function and updates the model's policy using Proximal Policy Optimization (PPO). Furthermore, the learning unit 154 introduces a KL divergence penalty to adjust the distribution of the pre-trained model so that it does not fluctuate significantly, thereby ensuring learning stability.
[0055] This method allows LLM to maximize the incorporation of diverse information from multiple reference texts into the overall text while maintaining consistency and balance across the entire task. This enables efficient resolution of complex tasks while accurately reflecting human intent.
[0056] The information processing device 10 can simplify human-provided instructions while simultaneously building a system that allows the model to interpret instructions with high accuracy through automatic optimization of complex task design. This approach reduces the time and effort required to create instructions, enabling the information processing device 10 to create an environment that does not hinder the inherent creativity of humans in creative tasks.
[0057] The information processing device 10 utilizes machine learning to appropriately interpret ambiguous instructions and intentions that change during the process. Through prompt optimization, the information processing device 10 builds a mechanism that reflects implicit human knowledge and background information, thereby improving the accuracy with which the generated output matches human intentions. This improves the success rate of creative description writing tasks.
[0058] The information processing device 10 incorporates human instructions into the task design through automatic optimization of complex task design, and then distributes them to agents, thereby ensuring that tasks proceed consistently across multiple processes and agents. This ensures that instructions from each process are appropriately passed on to the next process, and that human intentions are accurately reflected in the final product. As a result, complex interactions between processes and agents are coordinated.
[0059] In this way, the learning unit 154 generates prompts for a complex task using concise instructions given by a human, evaluates the degree to which the instructions are reflected in the creative description generated based on the generated prompts using a Rouge-L score, and performs reinforcement learning using the score as a reward, thereby achieving both simplification of instructions and high-precision reflection.
[0060] Furthermore, the learning unit 154 uses two types of Rouge-L scores as a reward function: a degree of reflection of content corresponding to human instructions (transmission score) and a degree of reflection of background knowledge based on the task (control score). By adjusting the weights of each score, it simultaneously optimizes the accuracy and consistency of the generated output.
[0061] Furthermore, in the execution of a multi-agent task spanning multiple processes, the learning unit 154 can accumulate a history of discussions, decisions, and instructions that occurred in each process, along with evaluation indicators, and reinforce the communication model based on this history.
[0062] (Other learning) Furthermore, the learning unit 154 can perform the following optimizations. When a human gives a revised instruction, the learning unit 154 learns the patterns and characteristics of expressions contained in that instruction and optimizes the prompt content in future task redesigns to bring it closer to the human intention.
[0063] Furthermore, the learning unit 154 maintains multiple potential correct answer candidates for each task and selects and outputs the optimal solution based on its respective Rouge score, thereby achieving both creativity and reflection of intent.
[0064] Furthermore, the learning unit 154 can improve the task performance and knowledge generation accuracy of each agent using a non-factoid QA dataset. This makes it possible to improve output accuracy while retaining diverse expressions and creative answer options.
[0065] Furthermore, the learning unit 154 models a series of loops in each process, including prompt generation, instruction transmission, task redesign, and evaluation of the generation results, and can achieve both instruction accuracy and generation quality while adjusting to avoid burdening human involvement.
[0066] The following describes embodiments based on the first embodiment. These embodiments can be realized by adding functions to the first embodiment, or by adding some functions. Furthermore, these embodiments can be realized in the first embodiment by providing the agent with prompts appropriate to the purpose.
[0067] <Example 1> Figure 7 is a diagram illustrating the overview of Example 1. As shown in Figure 7, the meta-agent 250a selects multiple agents from the agent pool 242a and creates a team of the selected agents. The meta-agent 250a controls the team and obtains output regarding tasks (initial values for composite tasks) based on instructions given by humans. The meta-agent 250a performs this process for multiple phases such as "team task design," "planning design," and "specification design."
[0068] Figure 8 shows an example configuration of the information processing device of Embodiment 1. The information processing device 20 in Figure 8 may be implemented based on the information processing device 10 of the first embodiment.
[0069] As shown in Figure 8, the information processing device 20 has a communication unit 21, an input unit 22, an output unit 23, a storage unit 24, and a control unit 25. The communication unit 21, input unit 22, output unit 23, storage unit 24, and control unit 25 have the same functions as the communication unit 11, input unit 12, output unit 13, storage unit 14, and control unit 15 of the information processing device 10.
[0070] The memory unit 24 stores model information 241. It also stores the agent pool 242 and the knowledge database 243. Model information 241 is the same information as model information 141.
[0071] The agent pool 242 contains information about each of the multiple agents. For example, the agent pool 242 contains the knowledge, roles, personas, etc., of each of the multiple agents. The knowledge database 243 stores the history of the discussions by the agents (for example, a summary of what discussions took place or what conclusions were reached) and the knowledge acquired during the discussions. In the initial stages before the discussions take place, the knowledge database 243 may contain information that serves as the basis for the agents to generate knowledge.
[0072] The control unit 25 includes an agent selection unit 251, a discussion control unit 252, an intention reflection unit 253, a learning unit 254, and an output control unit 255.
[0073] The agent selection unit 251 selects one or more agents from among the agents included in the agent pool 242. The agent selection unit 251 mainly performs the same processing as the generation unit 151.
[0074] The discussion control unit 252 performs processing to obtain output (products, solutions) for the complex task. For example, the discussion control unit 252 causes the agent to propose an approach. Also, for example, the discussion control unit 252 designs the task. The discussion control unit 252 mainly performs the same processing as the generation unit 151 and the design unit 152.
[0075] The intent reflection unit 253 performs processing to reflect the user's intent in the task. The intent reflection unit 253 mainly performs the same processing as the update unit 153.
[0076] The learning unit 254 performs the same processing as the learning unit 154. The output control unit 255 outputs information via the output unit 23, for example, by displaying it on a screen.
[0077] The processing flow of Example 1 will be explained using Figure 9. Figure 9 is a flowchart showing the processing flow of the information processing device of Example 1. Note that the main entity performing the processing in each step of Figure 9, except for step S204, can be described as a meta-agent. Step S204 is performed by each agent of the team.
[0078] As shown in Figure 9, first, the agent selection unit 251 receives text and images as instructions (step S201). Thus, in this embodiment 1, in addition to text, images can be input.
[0079] For example, the agent selection unit 251 accepts input such as images of the interior, exterior, and store information of a commercial building, along with the text, "Please consider a marketing strategy for store A." For example, the images may be of the specific exterior of a commercial building that the user wants to use as a reference.
[0080] Next, the agent selection unit 251 selects one or more agents that have knowledge of the instruction from the agent pool 242 (step S202). The number of agents selected by the agent selection unit 251 may be variable. The agent selection unit 251 also determines the subtasks that the selected agents should fulfill based on the instruction (step S203). For example, the agent selection unit 251 considers the instruction as a composite task and assigns each of the multiple subtasks obtained by dividing the composite task to each of the multiple agents. The subtasks are the objectives of each agent and will be used in the design of the composite task in a manner that satisfies these objectives as much as possible in the discussion described later. For example, the agent selection unit 251 uses a control model to divide the composite task and assign it to each agent.
[0081] In the agent pool 242, each agent is pre-assigned specialized knowledge. The agent selection unit 251 selects agents by considering whether they possess knowledge useful for discussion and whether orthogonality between the agents' knowledge (whether the knowledge held by the agents does not overlap).
[0082] For example, consider a case where the instructions are the text "Think of a tea shop that sells tea" and an image related to a tea shop. In this case, the agent selection unit 251 selects Agent A (Tea Sommelier), Agent B (Cafe Management Consultant), Agent C (Interior Designer), and Agent D (Marketing Specialist) from the agent pool 242.
[0083] The words in parentheses, such as "Tea Sommelier," "Cafe Management Consultant," "Interior Designer," and "Marketing Specialist," are keywords representing the knowledge of each agent. Agent Pool 242 may also store the specific content of each agent's knowledge, not just keywords.
[0084] For example, Agent Selection Department 251 might assign Agent A (Tea Sommelier) the subtask of "Thinking about types of tea, origins, brewing methods, menu development, and quality control." Similarly, Agent Selection Department 251 might assign Agent B (Cafe Management Consultant) the subtask of "Thinking about store operations, revenue models, customer acquisition strategies, and business plans." Furthermore, Agent Selection Department 251 might assign Agent C (Interior Designer) the subtask of "Creating the cafe's atmosphere, traffic flow design, and spatial design." Finally, Agent Selection Department 251 might assign Agent D (Marketing Manager) the subtask of "Approaching the target audience, utilizing social media, and pre- and post-opening promotions." Note that the subtasks assigned by Agent Selection Department 251 here are initial values and may be changed in later processing.
[0085] Next, under the control of the discussion control unit 252, each agent presents an approach to solving the task (step S204). Based on the presented approaches, the discussion control unit 252 designs a composite task (step S205). The discussion control unit 252 also calculates the degree to which the composite task encompasses the content of the subtasks (step S206).
[0086] If the degree of inclusion does not exceed the threshold (step S207; No), the argument control unit 252 updates the subtask (step S208) and returns to step S204. If the degree of inclusion exceeds the threshold (step S207; Yes), the argument control unit 252 proceeds to step S209. The degree of inclusion may be the Rouge-L score described in Reference 1. Here, if the condition that the degree of inclusion exceeds the threshold is not met, the argument control unit 252 returns to step S205 and designs the composite task. The conditions are not limited to those listed here; for example, the composite task may contain a specific keyword.
[0087] The discussion control unit 252 stores the content of discussions between agents, approaches, or the history of consideration of complex tasks in the knowledge database 243.
[0088] For example, the discussion control unit 252 designs a complex task: "to develop stores where users can explore their favorite tea recipes and brewing methods, and purchase tea at reasonable prices, and to communicate this intuitively to users."
[0089] Furthermore, the agent updates its knowledge during the discussion. In doing so, the agent can update its knowledge based on images input as instructions. The agent also stores the discussion content and the history of the deliberations in the knowledge database 243. This allows for consistency between services and between phases by utilizing the knowledge database 243. For example, even if different agents participate in the discussion due to a phase change, the knowledge database 243 ensures the consistency of the discussion.
[0090] In Agent Pool 242, Agent C (Interior Designer) is assumed to have originally possessed the following knowledge: • Principles of minimalist interior design • Selection and application of natural materials • Lighting design for hospitality spaces • Furniture placement for optimal traffic flow • Creation of community-focused gathering spaces
[0091] Let's assume the image included in the instructions is a photograph of the interior of a Victorian-style cafe. At this point, Agent C updates their knowledge based on the image as follows: The keyword representing Agent C's knowledge changes from "cafe interior designer" to "Victorian style specialist." • Victorian-era interior design and architecture • Selection and placement of period-appropriate furniture • Historical lighting design (chandeliers, wall sconces) • Victorian-era color theory and textile patterns • Antique appraisal and authentication • Traditional tea salon etiquette • 19th-century social customs and cultural background • Woodwork and decorative craftsmanship • Atmosphere design of historical spaces
[0092] The discussion control unit 252 can provide the agent with the image itself. In this case, the agent can obtain a lot of information from the image and engage in discussions that are closer to the user's image.
[0093] Furthermore, the discussion control unit 252 can provide the agent with text obtained by transforming the image features. In this case, the agent can conserve resources for referencing the image (e.g., LLM consumption tokens).
[0094] Here, the user can provide instructions to modify the composite task and the updated subtasks. These instructions may be provided as text. However, instructions are not limited to text; they may also be provided by other means such as voice or images. In the following explanation, the term "instruction text" may be simply replaced with "instruction."
[0095] The output control unit 255 can display the composite task and subtasks on the screen. The output control unit 255 can display the composite task and subtasks in a chat-style UI and accept user instructions as chat messages.
[0096] If no correction instructions are given (step S209; No), the argument control unit 252 generates an output based on the composite task and subtasks (step S213). For example, the argument control unit 252 generates an output (product) by inputting the composite task and task into the LLM.
[0097] If there are correction instructions (step S209; Yes), the intent reflection unit 253 accepts input of correction instructions for the composite task and subtasks (step S210). Then, the intent reflection unit 253 uses a communication model to convert the instructions into a prompt for intent reflection (step S211).
[0098] The discussion control unit 252 passes the prompt obtained by the intent reflection unit 253 to the agent (step S212). The discussion control unit 252 then prompts the agent to propose further approaches. At this time, the agent selection unit 251 assigns each of the multiple subtasks obtained by dividing the pre-designed composite task to each of the multiple agents.
[0099] For example, suppose the instruction for the revision was "Make it a child-friendly space." After the prompt obtained by the intent reflection unit 253 based on the instruction is passed on, the discussion control unit 252 updates the composite task in step S205 and the subtask in step S208 as follows. Specifically, the discussion control unit 252 updates the composite task from "Consider developing a store where customers can explore their favorite tea recipes and brewing methods, and purchase tea at a reasonable price range, and communicate this intuitively to users" to "A comprehensive project for the launch and success of a child-friendly tea cafe that balances tea expertise with family safety." Furthermore, the discussion control unit 252 updates Agent C's (interior designer) subtask from "Think about creating the cafe's atmosphere, designing traffic flow, and creating the space" to "Creating a space that emphasizes safety, comfort, and friendliness."
[0100] When an image is input, the discussion control unit 252 can obtain an approach to solving the assigned task by having multiple agents conduct a discussion based on the given composite task, knowledge, and image. The image may be input as the first instruction in step S201, or it may be input along with the instruction text at the time of the modification instruction in step S209. If the image is input in step S201, the content of the image will be reflected in the discussion with the minimum number of references. If the image is input at the time of the modification instruction in step S209, the content of the image will be reflected in the discussion more specifically, improving the consistency of the discussion.
[0101] The learning unit 254 performs training on the control model and the transmission model. For example, the learning unit 254 trains the control model using a control score that represents the extent to which a predetermined background task is reflected in each agent's subtask. Alternatively, the learning unit 254 can train the transmission model by measuring the degree to which instructions are reflected in a composite task using Rouge-L and rewarding the agent for this.
[0102] Furthermore, the learning unit 254 can learn the communication model using human evaluation as a reward. For example, based on the instruction text, the learning unit 254 can reward the level of human satisfaction with the communication model, the small burden on the human for making correction instructions, the low frequency of correction instructions, the small number of correction instructions required, or the short time it took for the human's intention to be reflected in the complex task.
[0103] Furthermore, the learning unit 254 can fine-tune the control model and the transmission model using techniques such as LoRA (Low-Rank Adaptation) and PPO (Proximal Policy Optimization). Through learning, the transmission model is expected to be able to output prompts that are simplified and reflect instructions with high accuracy.
[0104] The processing by the agent selection unit 251, the discussion control unit 252, the intent reflection unit 253, and the learning unit 254 can be realized by the generation unit 151, the design unit 152, and the update unit 153 shown in Figure 1. The generation unit 151 can perform the processing of the agent selection unit 251 and the processing of the discussion control unit 252, excluding the design of complex tasks. The design unit 152 can perform the design of complex tasks among the processing of the discussion control unit 252. The update unit 153 can perform the processing of the intent reflection unit 253.
[0105] Here, we consider the case where Example 1 is applied to a service consisting of multiple phases. A service is a collection of multiple phases. In this case, the service includes four phases: team task design, proposal creation, specification creation, and implementation. Note that a phase may also be referred to as a process or step.
[0106] The information processing device 20 performs the processing described in Figure 9 for each phase in a predetermined order. For example, the information processing device 20 processes in the order of team task design, planning, specification review, and implementation. The information processing device 20 generates output for each phase as shown in step S213. The information processing device 20 may also re-select agents in each phase, or it may omit agent selection and have the same team of agents as in the previous phase conduct the discussion.
[0107] (1) Team Task Design Phase The team task, which is the output of the team task design phase, may be the composite task and subtask itself. The information processing device 20 passes the composite task and subtask output in the team task design phase to the next phase.
[0108] (2) Planning Phase In the planning phase, in step S201, the composite task and subtasks are accepted as instructions. The composite task and subtasks may reflect the user's intentions through modification instructions. Furthermore, in subsequent phases, the text and images entered in the team design phase may or may not be included in the instructions.
[0109] The discussion control unit 252 has the agent team discuss specific project details based on the team task and output a project proposal and a marketing strategy. In step S205, a composite task for outputting the project proposal is designed. In addition, the discussion control unit 252 may not only reflect any modification instructions for the composite task, but may also present the project proposal generated based on the composite task to the user, receive instructions on the proposal, and reflect them in the composite task.
[0110] The following paragraph shows an example of a generated project proposal.
[0111] 1. Project Name: ChocoTea Haven 2. Main Service Experience: ChocoTea Haven offers a unique tea experience centered around chocolate-flavored tea and re-steeping technology. 3. Target Persona: The target audience is tea lovers aged 25-45, health-conscious urban dwellers, and people seeking unique flavor experiences. 4. Executive Summary: ChocoTea Haven enhances the customer experience by focusing on chocolate-flavored tea and re-steeping technology. 5. Service Space Layout: Separation of tasting and sales areas ensures smooth traffic flow. Centralized re-steeping experience station improves visibility and accessibility. Movable furniture is introduced to allow for flexible space changes during events. A kids' corner and safe materials are used to accommodate families with children. 6. Service Definition: Includes tea tasting, re-steeping education, community events, and product sales. Tastings and workshops will also be held to increase customer engagement and loyalty. 7. Monetization Strategy: Revenue sources include product sales, subscriptions, participation fees for tasting events and workshops, and sponsorships from tea brands. Seasonal products and themed events attract a diverse customer base and improve profitability.
[0112] Suppose the user requests a revision to the above proposal, stating that they would like it to be able to host medium-sized events. For example, the intent reflection unit 253, based on the revision request, adds the following phrase to the proposal item "5. Service Space Layout": "Expand the community event space to accommodate workshops and social gatherings with up to approximately 20 participants."
[0113] Furthermore, the intent reflection unit 253 may display options for revisions to each item in the proposal and accept selections. The discussion control unit 252 recreates the proposal according to the options selected by the user. For example, the intent reflection unit 253 may display options such as "Reconsider '5. Service Space Layout'" or "Make '6. Service Definition' a little more specific."
[0114] (3) Specification Review Phase In the specification review phase, the project proposal is accepted as an instruction in step S201. The project proposal may reflect the user's intentions through revision instructions.
[0115] The information processing device 20 has the agent team discuss specific specifications based on the project proposal and output a specification document. In step S205, a composite task for outputting the specification document is designed. The information processing device 20 can also reflect modification instructions for the composite task into the composite task. Furthermore, the information processing device 20 may present the specification document generated based on the composite task to the user, receive instructions regarding the specification document, and reflect them in the composite task.
[0116] The following paragraph shows an example of the generated specification.
[0117] UI specifications, screen transition flow, input / output information, order processing structure and operation design, processing content, backend / API specifications, product database specifications, user authentication, real-time communication, synchronous session.
[0118] Suppose the user requests a modification to the above specification document, stating that they want to "show what products are available." For example, the information processing device 20, based on the modification request, adds the phrase "Display product information" under "Input / Output Information" in the specification document.
[0119] (4) Implementation Phase In the implementation phase, the specification document is accepted as an instruction in step S201. The specification document may reflect the user's intentions through modification instructions.
[0120] The information processing device 20 has the agent team discuss specific implementation details based on the specifications and output data for visualization. In step S205, a composite task for data output is designed. In addition, the information processing device 20 may not only reflect modification instructions for the composite task in the composite task, but may also present the data generated based on the composite task to the user, receive instructions on the data, and reflect them in the composite task.
[0121] The discussion control unit 252 can prompt the user to specify a format compatible with a particular file format (such as ".blender" or ".obj") for the 3D model, and cause the LLM to generate the 3D model data. For example, the information processing device 20 displays a model of a store implemented based on the specifications in 3D space.
[0122] Furthermore, any user-instructed modifications made in later phases are retroactively reflected in previous phases by the information processing device 20. For example, modifications made in later phases are retroactively reflected to ensure consistency between the project proposal, specifications, and implementation (3D model).
[0123] Furthermore, especially during the planning, specification review, and implementation phases, the information processing device 20 may impose constraints on the agent. For example, the information processing device 20 may set constraints such as the format of the project proposal and specification document, and modeling software for creating 3D models.
[0124] The information processing device 20 executes these four phases of processing, so that, for example, in response to a human instruction such as "Think of a marketing strategy for store A," it proposes a marketing strategy for store A that includes planning, interior and exterior design, product list, and operational plan, and can further visualize the marketing strategy in a 3D model to make it easier to intuitively understand and modify.
[0125] Figure 10 shows an example of the output from Example 1. As shown in Figure 10, a 3D model 31 and a 3D model 32 that reproduce the interior of store A based on the specifications can be presented.
[0126] Furthermore, if store A is one of several stores located within a large commercial facility, the information processing device 20 can process services related to stores other than store A, as well as intermediate spaces (such as plazas). In this case, as shown in Figure 11, the information processing device 20 can devise an integrated marketing strategy in which each project is independent while the stores cooperate with each other.
[0127] Figure 11 is a diagram illustrating the processing of multiple services. As shown in Figure 11, the information processing device 20 can use the output generated in the phases related to store A and store B (team tasks, proposals, specifications, implementation) to generate the output in the phase related to the intermediate space.
[0128] As described above, the information processing device 20 can store the output generated in each phase, the content of discussions between agents, and the history of consideration of approaches or complex tasks in the knowledge database 243. Agents can refer to the stored information even when discussing other services or other phases.
[0129] <Example of prompts in Example 1> The following are specific examples of prompts entered into the LLM or agent in Example 1.
[0130] The prompt P1 shown in the following paragraph is a prompt for the agent selection unit 251 (meta-agent) to select an agent and divide and assign a complex task. For example, prompt P1 is used in steps S202 to S204 in Figure 9.
[0131] (プロンプトP1) ``` You are the leader of a team of experts. Your goal is to solve a complex and interdisciplinary problem by selecting specialists from different fields, assigning them complementary and orthogonal tasks, and deriving a precise and in-depth solution through discussions. Please address the following challenge. {Challenge Description} Instructions for Task Division To solve this challenge, construct 4 orthogonal tasks that meet the following requirements: Define each task from a different perspective of the problem. Ensure that each task complements the others. For each task, specify the following: Subject: The primary theme of the task. Detail: Detailed objectives or actions for the task. Use bullet points for multiple items. Design tasks that can be discussed and integrated by the entire team.Output Format Define the tasks using the following format: Task 1: [Task Theme] Subject: [Describe the main theme of the task] Detail: [Provide detailed objectives or actions for the task] Task 2: [Task Theme] Subject: [Describe the main theme of the task] Detail: [Provide detailed objectives or actions for the task] Task 3: [Task Theme] Subject: [Describe the main theme of the task] Detail: [Provide detailed objectives or actions for the task] Task 4: [Task Theme] Subject: [Describe the main theme of the task] Detail: [Provide detailed objectives or actions for the task] Notes Show mutual complementarity between tasks. Consider how each task facilitates discussions and enables eventual integration.Additional Rule The output must be in plain text only - no Markdown symbols, no bullets, no formatting characters (#, *, -, >, _, ```` etc.), and no indentation. Only line breaks may be used to show structure. ````.
[0132] Prompt P1 stipulates that the relationships between the multiple tasks (subtasks) obtained must be "complementary and orthogonal."
[0133] The agent selection unit 251 (meta-agent) selects agents and provides another example of a prompt (prompt P2) for dividing and assigning complex tasks, as shown in the following paragraph.
[0134] (プロンプトP2) ``` You are the leader of a team of experts. Your mission is to solve the following challenge as the ultimate objective of this task. To achieve this, you must select specialists from different fields, assign them complementary and orthogonal tasks, and derive a precise and in-depth solution through structured discussion. Primary Objective: The utmost priority is to analyze and solve the challenge described below by dividing it into the minimum number of essential and independent tasks.{Challenge Description} Referenced passage: {passage} Instructions for Task Division To solve this challenge, analyze the referenced passage above and divide the problem into only the number of orthogonal tasks that are truly necessary to solve it completely, following these guidelines: The reason for dividing the challenge into subtasks is to enable collaborative problem-solving with synergy, similar to tackling complex tasks like DROP, and to allow thorough cross-checking from multiple perspectives to ensure completeness and accuracy. DROP(A Reading Comprehension Benchmark Requiring Discrete Reasoning Over Paragraphs) is a groundbreaking benchmark dataset proposed in the field of machine reading comprehension.This paper points out the "fragility" of conventional reading comprehension tasks, where AI achieves human-level performance but tends to rely on simple patterns, and introduces a dataset that requires more advanced discrete reasoning (such as addition, counting, sorting, etc.) applied to paragraph content. If the challenge is simple, produce only one task. If it is moderately complex, produce two or three. Do not show Task 4 unless it is genuinely required to cover an entirely different dimension of analysis. Do not include skipped or empty task slots. For each task, specify the following: Subject: The primary theme or focus of the task. Detail: Detailed objectives or actions for the task. Use bullet points for multiple items. The tasks should be designed so that they can be discussed and integrated by the entire team.Output Format Define the tasks using the following format: Task [number]: [Task Theme] Subject: [Describe the main theme of the task] Detail: [Provide detailed objectives or actions for the task] Notes Generate only as many tasks (from 1 to 4) as are necessary to appropriately divide the challenge based on its complexity and scope. Show mutual complementarity between tasks. Consider how each task facilitates discussions and enables eventual integration. Additional Rule The output must be in plain text only - no Markdown symbols, no bullets, no formatting characters (#, *, -, >, _, ``` etc.), and no indentation. Only line breaks may be used to show structure. ```.
[0135] Prompt P2 shows a dynamic method for determining the number of tasks (subtasks) that will be obtained.
[0136] The prompt P3 shown in the following paragraph is a prompt for generating the role name and expertise of the selected agents. In step S204, the discussion control unit 252 (meta-agent) uses prompt P3 to generate the role and expertise for each agent's discussion.
[0137] (プロンプトP3) ``` {name} is the member of the team {teamName}, and this team currently has a mission '{objective}'. Now, {name} recognizes the existence of the following aim: task: {task} subject: {subject} details: {details} In this context, to cope with it, what expertise does this person have, and what social role could he / she be considered to have? **IMPORTANT INSTRUCTIONS:** - Output EXACTLY 1 role name - Output EXACTLY 4 expertise items - Use this EXACT format with no additional text: Additional Rule The output must be in plain text only - no Markdown symbols, no bullets, no formatting characters (#, *, -, >, _, ``` etc.), and no indentation. Only line breaks may be used to show structure. role: [ONE ROLE NAME HERE] expertise: - [Expertise 1] - [Expertise 2] - [Expertise 3] - [Expertise 4] Respond ONLY with the format above. No explanations, no JSON, no other text. ``` ```
[0138] The prompt P4 shown in the following paragraph is a prompt issued by the discussion control unit 252 to prompt the agent to engage in discussion. For example, prompt P4 is used in step S204 of Figure 9.
[0139] (プロンプトP4) {{> profile}} {{> last-behavior}} {{> team-profile}} Here, your team has a mission: {{{objective}}}. To realize the above, your team is on the meeting for brainstorming. The goal of this meeting is "devising 'aim and approach' to realize the team's mission". {{> meeting-history}} {{> last-meeting-result}} Now, all participants have expressed their own opinions. Task: As the leader of the meeting, you need to integrate the opinions of the participants and come up with new "aim and approach" to realize the team's mission. You can find the opinions of the participants in the meeting history. At this time, you need to choose the best method for integrating opinions from the following three options, firstly. 1. Select a "core task" as a central concept of team's "aim and approach" from the participants' opinions and integrate other participants' opinion in a way that reinforces it. {{> integrate-strategy-primary}} 2.Extract the core value, that is, the "desired outcome" from the participants' opinions, find common points among them, and create a "core task" for the team that is different from each participant's "aim and approach". {{> integrate-strategy-newappr}} 3. Propose a new method for integrating opinions. If it is difficult to create attractive opinions using methods 1 or 2, you can propose a new method for integrating opinions through free description. However, this is a last resort, and it is usually desirable to choose method 1 or 2. Therefore, if you propose a new method for integrating opinions, you must also provide the reason for it. Please respond in the following JSON format. ``` { "thought": "Step-by-step thought of the method select", "option": number (number of the method you selected. 1, 2 or 3), "content": "content of the method (need to describe only if you select option 3.otherwise, it is OK to be a blank string)"} ``` Specify the number of the option you judged to be the best method for integrating opinions in the "option" property.
[0140] The prompt P5 shown in the following paragraph is a prompt that specifies the format of the project proposal. Prompt P5 may also be entered into the agent selection unit 251 as an instruction during the planning phase.
[0141] (Prompt P5) 1) A descriptive and appealing name for the product / service. 2) The target persona. 3) Executive summary. The experience that this product / service will provide. 4) Specific offerings, including the opinions of each attendee. If there are specific specifications such as product items, include them as well. 5) Revenue Model. How to monetize this product / service in the market. 6) Other extra remarks if exists.
[0142] The prompt P6 shown in the following paragraph is a prompt for the implementation of the 3D model. Prompt P6 may also be input to the agent selection unit 251 as an instruction during the implementation phase. Note that the content following "**Global Grid Specification (preamble, unnumbered)**" may be the content of the specification document output during the specification review phase.
[0143] (プロンプトP6) **Instruction**: The following document must include **all 8 sections (#1-#8)**. Partial outputs (e.g., only 5 sections) are invalid. This is the **proposal document** to be generated by agents. The multi-agent framework is applied. At this stage, the task is to draft a **preliminary spatial design proposal** for the **Service Space Layout and Service Definitions for a Tea Shop**, as a conceptual design for an architectural competition. To ensure operationalization within ACT and interoperability across teams, the following **Global Grid Specification** is introduced. --- **Global Grid Specification (preamble, unnumbered)** * **GridID**: G1 * **Cell size**: **0.25 m × 0.25 m** (uniform in X / Y) * **Extent**: **X = 0-12 m (48 cols) / Y = 0-15 m (60 rows)** * **Origin**: (0,0) = bottom-left corner (right-handed system) * **Note**: In Blender, `bpy.ops.mesh.primitive_plane_add(size=1)` generates a plane of **1 m × 1 m**. To align with this specification (`CellSize=0.25 m`), **Blender 1 = 16 grid cells** must be applied as a conversion. > All objects must fit within this range and reference their occupied cells via `GridRef`. --- **Agent Work Instructions (within ACT)** * For each segment in **Section 5 (Service Space Layout)**, always provide a **detailed description (Description)** including: * **Texture** (PBR-oriented material qualities) * **Design** (designer's intent / narrative identity) * **Flow** (ergonomic aspects: corridor width, circulation, lingering) * When making proposals or adjustments, always apply **§4.1 ergonomic and plotting rules**. * If objects span multiple segments, declare an **`AnchorSegment`** and list **`overhang_cells`** (avoid or validate circulation-critical cells). * Outputs must comply with the **Global Grid Specification** and **§4.2 GridRef notation rules**. * Each agent should adapt its level of detail to match its initial task, role, expertise, and the team's collective goal.* Agents must collaborate organically across perspectives to generate a **multi-view, integrative, creative outcome**. * **IMPORTANT**: The grid provides extensive space (Y=0-15m, 60 rows). **Utilize available empty areas** beyond the example layout to accommodate team member suggestions. * When team members propose additional spaces or modifications, **specify exact GridRef coordinates** and **discuss spatial placement** with other agents to avoid conflicts. --- # Project Name **“Citrus Bliss Tea Shop - A Harmonious Tea Experience”** --- ## 1. Service Experience Specifications The Citrus Bliss Tea Shop is designed to provide a seamless experience from entry to checkout. * **Checkout Area**: Staff will assist with payments and membership inquiries. * **Product Shelves**: Equipped with interactive digital displays that provide detailed information on flavor profiles, caffeine content, and health benefits.QR codes will allow customers to access preparation methods and detailed product descriptions. * **Tasting Area**: Customers can engage with aroma and tea simulations through interactive user interfaces. * **Resting Area**: Comfortable seating with ambient background music and subtle scent adjustments to enhance relaxation. Additional sensory elements, such as curated soundscapes and aroma diffusion, will create an immersive experience. Family-friendly seating and interactive features for children will encourage participation and community engagement. --- ## 2. Target Persona | Attribute | Details | | --------- | ------------------------------------------------------ | | Age | 25-45 | | Interests | Tea, interior design, relaxation | | Lifestyle | Seeks quality time; shares experiences on social media | --- ## 3.Executive Summary Visitors will flow naturally from the entrance, browse shelves, interact with staff at checkout, and relax at designated tables. The digital model visualizes **layout, circulation, experience points,** and **purchase flow**, ensuring an engaging and educational environment. The design includes visually striking displays and interactive stations for sharing brewing experiences, enhancing customer engagement and promoting social media sharing. --- ## 4. Grid Segmentation ### 4.1 Ergonomics & Plotting Rules * **Circulation clearances**: Primary aisles must be at least 1.5 m wide to accommodate two-way movement. * **Queues**: Width should be 0.9-1.2 m, with additional slack at turns. * **Reach envelope**: Effective shelf pick zone is between 0.6-1.6 m above the floor. * **Overhang & protrusions**: Head clearance must be >= 2.1 m, with protrusions into aisles limited to <= 0.2 m. ### 4.2 GridRef Notation Rules * Use row / column indices and range / set expressions.* All references assume **1 cell = 0.25 m** (Blender: 4 cells = 1 m). --- ## 5. Service Space Layout | SegmentID | SegmentName | ObjectType | ObjectName | Purpose | Interaction Spec | GridRef | Description | |---|---|---|---|---|---|---|---| | S1 | Entrance | Structure | Welcome Sign | Emphasize first impression | Lights up for 3s on pass-through | G1\:R4C4-R8C24 | **Texture**: Brass hairline; terrazzo floor. **Design**: Minimal Japan-modern; floating sign. **Flow**: Main aisle >= 1.5 m; photo dwell 0.8 × 1.0 m.| | S2 | Cashier Counter | Furniture / Staff | Register & Staff | Core of checkout and service | Conversation, membership, checkout | G1\:R12C36-R20C48 | **Texture**: Oak veneer; brass kick line. **Design**: “Quiet hospitality”; task lighting. **Flow**: Queue 0.9-1.2 m; exit S2\_out → S5 / S3. | | S3 | Product Shelf Area | Fixture / Product | Tea Shelves (3 types) | Enhance visibility | Show info on approach; selectable | G1\:R12C8-R24C32 | **Texture**: Matte melamine; oak edge band. **Design**: Low-chroma header band; monospaced price labels. **Flow**: Loop S1 → S3 → S2; aisle >= 1.2 m. | | S4 | Tasting Table | Furniture / Display | Tasting Table | Promote product experience | Click to show aroma / info UI | G1\:{R28C16,R28C20,R32C16,R32C20} | **Texture**: Ceramic matte top; powder-coated legs.**Design**: Round table; light ambient scent. **Flow**: One-way loop; perimeter clearance >= 1.0 m. | | S5 | Resting Tables | Furniture | 2 Round Tables | Encourage post-purchase stay | Slight BGM / scent increase on seating | G1\:R28C36-R36C48 | **Texture**: Fabric seats; wood-look herringbone floor. **Design**: Low-glare indirect light; diffuse ambient sound. **Flow**: Seat spacing >= 0.9 m; aisle >= 1.2 m. | --- ## 6. Service Definitions ### 6.1 Product List | ProductID | ProductName | Category | Price | SegmentID | Notes | | --------- | ----------------- | -------- | --------- | --------- | ------------------------ | | T001 | Citrus Bliss Tea | Tea | 1,800 JPY | S3 | Citrus × white tea blend | | T002 | Floral Meditation | Tea | 2,000 JPY | S3 | Lavender × black tea | | T003 | Original Cup Set | Goods | 2,500 JPY | S3 | Japanese-modern design | ### 6.2 Shopping Cart Schema | Field Name | Data Type | Description | | ---------- | --------- | -------------------------------- | | CartID | String | Unique cart ID | | CustomerID | String | Customer ID | | ProductID | String | Linked to Product List | | Quantity | Integer | Number of items | | AddedBy | String | “Customer” or “Staff” | | Timestamp | DateTime | Time of add / remove | | Status | String | “InCart”, “Removed”, “Purchased” | --- ## 7. Monetization * In-store sales * Pre-orders for limited items * Gift sets * Paid events (tea workshops) --- ## 8. Glossary * **Grid / Cell / CellSize / Origin / Extent**: See Global Grid Specification. * **Row (R) / Column (C)**: Vertical / horizontal indices in GridRef. * **GridRef**: Notation of occupied cells (range or set).* **AnchorSegment / overhang\_cells**: Base segment and explicit cross-segment declaration. * **PBR**: Physically Based Rendering parameters (albedo, roughness, metallic, etc.).
[0144] <System Configuration, etc.> Furthermore, the components of each part shown in the diagram are functional concepts and do not necessarily need to be physically configured as shown. In other words, the specific forms of distribution and integration of each device are not limited to those shown in the diagram, and all or part of them can be functionally or physically distributed and integrated in any unit according to various loads and usage conditions. In addition, all or any part of the processing functions performed by each device can be realized by a CPU and the program executed on that CPU, or by hardware using wired logic.
[0145] Furthermore, among the processes described in the embodiments described above, all or part of the processes described as being performed automatically can be performed manually, or all or part of the processes described as being performed manually can be performed automatically by known methods. In addition, the processing procedures, control procedures, specific names, and information including various data and parameters shown in the above document and drawings can be arbitrarily changed unless otherwise specified.
[0146] [Program] The information processing device 10 described above can be implemented by installing a program (information processing program) as packaged software or online software on a desired computer. For example, by having the computer run the above program, the computer can function as the information processing device 10. The term "computer" here includes mobile communication terminals such as smartphones, mobile phones and PHS (Personal Handyphone System), as well as terminals such as PDA (Personal Digital Assistant).
[0147] Figure 12 shows an example configuration of a computer that executes an information processing program. Computer 1000 has, for example, memory 1010 and CPU 1020. Computer 1000 also has a hard disk drive interface 1030, a disk drive interface 1040, a serial port interface 1050, a video adapter 1060, and a network interface 1070. These components are connected by a bus 1080.
[0148] Memory 1010 includes ROM (Read Only Memory) 1011 and RAM (Random Access Memory) 1012. ROM 1011 stores, for example, a boot program such as BIOS (Basic Input Output System). The hard disk drive interface 1030 is connected to the hard disk drive 1090. The disk drive interface 1040 is connected to the disk drive 1100. For example, a removable storage medium such as a magnetic disk or optical disk is inserted into the disk drive 1100. The serial port interface 1050 is connected to, for example, a mouse 1110 and a keyboard 1120. The video adapter 1060 is connected to, for example, a display 1130.
[0149] The hard disk drive 1090 stores, for example, the OS 1091, application programs 1092, program modules 1093, and program data 1094. That is, the programs that define each process executed by the information processing device 10 are implemented as program modules 1093 in which executable code for a computer is written. The program modules 1093 are stored, for example, in the hard disk drive 1090. For example, a program module 1093 for executing processes similar to the functional configuration of the information processing device 10 is stored in the hard disk drive 1090. Note that the hard disk drive 1090 may be replaced by an SSD (Solid State Drive).
[0150] Furthermore, the data used in the processing of the above-described embodiment is stored as program data 1094 in, for example, memory 1010 or hard disk drive 1090. The CPU 1020 then reads the program module 1093 and program data 1094 stored in memory 1010 or hard disk drive 1090 into RAM 1012 as needed and executes them.
[0151] Furthermore, the program module 1093 and program data 1094 are not limited to being stored in the hard disk drive 1090; for example, they may be stored in a removable storage medium and read by the CPU 1020 via a disk drive 1100 or the like. Alternatively, the program module 1093 and program data 1094 may be stored in another computer connected via a network (LAN (Local Area Network), WAN (Wide Area Network), etc.). The program module 1093 and program data 1094 may then be read by the CPU 1020 from the other computer via a network interface 1070.
[0152] The functions of the elements disclosed herein may be implemented using circuits or processing circuitry that include general-purpose processors, special-purpose processors, integrated circuits, ASICs (Application Specific Integrated Circuits), FPGAs (Field Programmable Gate Arrays), conventional circuits, and / or combinations thereof that are programmed using one or more programs stored in one or more memories, or otherwise configured to perform the disclosed functions. A processor is considered processing circuitry or circuitry because it includes transistors and other circuits. A processor may be a programmed processor that executes programs stored in memory. In this disclosure, circuits, units, or means are hardware that performs the enumerated functions, or hardware programmed to perform the enumerated functions. Hardware may be any hardware disclosed herein that is programmed or configured to perform the enumerated functions.
[0153] There is a memory for storing a computer program that includes computer instructions. These computer instructions provide logic and routines that enable hardware (e.g., processing circuitry or circuitry) to perform the methods disclosed herein. This computer program can be implemented in commonly known forms, such as computer-readable storage media, computer program products, memory devices, recording media such as CD-ROMs and DVDs, and / or memory of FPGAs and ASICs.
[0154] 10, 20 Information Processing Unit 11, 21 Communication Unit 12, 22 Input Unit 13, 23 Output Unit 14, 24 Storage Unit 15, 25 Control Unit 141, 241 Model Information 151 Generation Unit 152 Design Unit 153 Update Unit 154 Learning Unit 242 Agent Pool 243 Knowledge Database 251 Agent Selection Unit 252 Discussion Control Unit 253 Intent Reflection Unit 254 Learning Unit 255 Output Control Unit
Claims
1. An information processing device comprising: a generation unit that assigns each of several tasks obtained based on a given task to each of several agents; a design unit that designs a composite task based on the task-solving approaches presented by each of the several agents; and an update unit that updates the composite task designed by the design unit using instructions given by a human.
2. The information processing apparatus according to claim 1, wherein the generation unit assigns each of the plurality of tasks obtained by dividing the composite task updated by the update unit to each of the plurality of agents, and causes each of the plurality of agents to generate knowledge of the assigned task.
3. The information processing apparatus according to claim 1 or 2, wherein the generation unit generates a document using the knowledge of the plurality of agents based on the composite task designed by the design unit.
4. The information processing apparatus according to claim 3, further comprising a learning unit that trains a model for performing the processing of the generating unit by optimizing a control score that indicates how well the document reflects information that is not included in the instructions and corresponds to the task assigned to the agent.
5. The information processing apparatus according to claim 3, further comprising a learning unit that learns a model for executing the processing of the design unit by optimizing a transmission score that indicates how well the document reflects the content of the instructions.
6. The information processing apparatus according to claim 1, wherein the generation unit causes each agent to generate knowledge of the assigned task, and the design unit causes the plurality of agents to engage in discussion based on the given task, the knowledge, and the image, thereby obtaining an approach to solving the assigned task.
7. The information processing apparatus according to claim 6, wherein the design unit causes the plurality of agents to perform a given task, the image input along with the given task, and a discussion based on the knowledge.
8. The information processing apparatus according to claim 6, wherein the generation unit assigns each of the multiple tasks obtained based on the designed composite task to each of the multiple agents, causes each of the multiple agents to generate knowledge of the assigned task, and the design unit causes the multiple agents to conduct a discussion based on the image input along with the instructions and the knowledge.
9. The information processing apparatus according to claim 1, wherein the design unit generates 3D model data based on the composite task updated by the update unit.
10. The information processing device according to claim 1, wherein the design unit updates the multiple tasks and causes the multiple agents to propose further approaches to solving the tasks if the degree to which the multiple tasks are covered by the designed composite task does not satisfy a predetermined condition.
11. An information processing method in which a computer performs the following processes: assigning each of the multiple tasks obtained based on a given task to each of the multiple agents; designing a composite task based on the task-solving approaches presented by each of the multiple agents; and updating the designed composite task using instructions given by a human.
12. An information processing program that assigns each of several tasks obtained based on a given task to each of several agents, designs a composite task based on the task-solving approaches presented by each of the several agents, and causes a computer to perform the process of updating the designed composite task using instructions given by a human.
