Information processing apparatus, task solving method, and program

Abstract

To provide an information processing apparatus capable of appropriately solving a given task.SOLUTION: An information processing apparatus comprises: an access unit for accessing a memory that stores, for each of a plurality of tasks, a successful process in solving the task as an example; and a solving unit for retrieving example problems similar to a solution process of the task to be solved from the memory, training a large-scale language model, and solving the task to be solved using the trained large-scale language model.SELECTED DRAWING: Figure 3

Description

[Technical Field]

[0001] The present disclosure relates to an information processing device, a task solving method, and a program. [Background technology]

[0002] Large language models (LLMs) are attracting attention as a technology that can replace human behavior and decision-making. For example, LLMs are being used to solve tasks in situations such as automatically operating systems or automatically operating robots.

[0003] There is a technique called In-Context Learning, in which an LLM is given several examples and asked to solve a task. An example is, for example, a pair of an example task and an example solution. The examples are given to the LLM as examples of how to solve the task, and are included in the prompt. By learning from the several examples given, the LLM solves the task by taking into account the context corresponding to the examples. In-Context Learning is also sometimes called Few-shot Learning. [Prior art documents] [Patent documents]

[0004] [Patent Document 1] Japanese Patent Publication No. 2024-043563 Summary of the Invention [Problem to be solved by the invention]

[0005] Currently, it is not easy to prepare appropriate example problems that correspond to the task being solved. A configuration in which the user must constantly think up example problems can lead to unstable task-solving accuracy and place a heavy burden on the user. Furthermore, a configuration in which fixed example problems are provided to the LLM may not guarantee that the example problems are suitable for solving the current task. This can lead to the LLM considering contexts that do not match the task being solved, resulting in slow task resolution or even failure to solve the task.

[0006] Non-limiting examples of the present disclosure contribute to providing an information processing device, a task solving method, and a program that can appropriately solve a given task. [Means for solving the problem]

[0007] An information processing device according to one embodiment of the present disclosure includes an access unit that accesses a memory that stores successful processes for solving each of a plurality of tasks as examples, and a solving unit that retrieves example problems similar to the solution process of the task to be solved from the memory, trains a large-scale language model, and solves the task to be solved using the trained large-scale language model.

[0008] A task solving method according to one embodiment of the present disclosure accesses a memory that stores successful process steps for solving each of a plurality of tasks as example problems, retrieves example problems similar to the solution process of the task to be solved from the memory, trains the example problems into a large-scale language model, and solves the task to be solved using the trained large-scale language model.

[0009] A program according to one embodiment of the present disclosure causes a processor to access a memory that stores, for each of a plurality of tasks, a successful process for solving the task as an example, retrieve from the memory an example similar to the solution process of the task to be solved, train a large-scale language model, and solve the task to be solved using the trained large-scale language model.

[0010] These comprehensive or specific aspects may be realized as a system, an apparatus, a method, an integrated circuit, a computer program, or a recording medium, or may be realized as any combination of a system, an apparatus, a method, an integrated circuit, a computer program, and a recording medium. [Effects of the Invention]

[0011] According to an embodiment of the present disclosure, a given task can be solved appropriately.

[0012] Further advantages and benefits of an embodiment of the present disclosure will become apparent from the specification and drawings. Such advantages and / or benefits may be provided by some of the embodiments and features described in the specification and drawings, respectively, but not necessarily all of them may be provided to obtain one or more identical features. [Brief explanation of the drawings]

[0013] [Figure 1] FIG. 1 is a diagram showing an example of a block configuration of an information processing device 1 according to an embodiment. [Figure 2] FIG. 1 is a diagram illustrating an outline of the operation of the information processing device 1. [Figure 3] A diagram showing an example of the block configuration of Agent 2 [Figure 4] FIG. 10 is a diagram showing an example of information stored in memory 3. [Figure 5] Diagram illustrating the prompts for generating a full plan [Figure 6] Diagram illustrating the prompt for generating a retrieval key [Figure 7] Diagram illustrating the retrieval of past experiences related to an action [Figure 8] Diagram illustrating prompts for generating actions [Figure 9] 1 is a flowchart showing an example of the operation of the information processing device 1. [Figure 10] An example of text generation on the platform [Figure 11]FIG. 10 is a diagram illustrating an outline of the operation of the information processing device 1 according to a modified example. DETAILED DESCRIPTION OF THE INVENTION

[0014] Hereinafter, embodiments of the present disclosure will be described in detail with reference to the drawings as appropriate. However, more detailed explanation than necessary may be omitted. For example, detailed explanation of already well-known matters or redundant explanation of substantially identical configurations may be omitted. This is to avoid unnecessary redundancy in the following explanation and to facilitate understanding by those skilled in the art.

[0015] The accompanying drawings and the following description are provided to enable those skilled in the art to fully understand the present disclosure, and are not intended to limit the subject matter described in the claims.

[0016] <Block configuration of information processing device> Fig. 1 is a diagram showing an example of a block configuration of an information processing device 1 according to an embodiment. As shown in Fig. 1, the information processing device 1 includes an agent 2, a memory 3, and a communication unit 4. The agent 2 may be referred to as an LLM, an LLM agent, or a control unit. The information processing device 1 may be, for example, a personal computer or a server.

[0017] The memory 3 stores programs such as an OS (Operating System) program and application programs executed by the agent 2. The memory 3 also stores various data. For example, the memory 3 stores successful experiences (examples) in solving tasks. Here, the successful experiences include at least one pair of a task and an example of a solution to that task.

[0018] The successful experience may be a pair of a task that has been implemented or simulated in the past and a solution that actually succeeded in solving the task, i.e., a successful experience in solving a past task. The successful experience may also include a pair of an example task and an example solution that has not been implemented or simulated in the past but that the user has set as the correct solution. In this way, for example, for a task for which the user can easily imagine a solution, a successful experience can be created without implementing or simulating it, thereby expanding the variety of successful experiences stored in memory 3.

[0019] The memory 3 is configured by a storage device such as a hard disk drive (HDD), a solid state drive (SDD), a random access memory (RAM), a read only memory (ROM), and a flash memory.

[0020] The agent 2 is configured by a processor such as a CPU (Central Processing Unit) and / or a GPU (Graphics Processing Unit). The agent 2 executes a program stored in the memory 3, for example, to realize its functions.

[0021] Agent 2 retrieves (obtains) successful experiences that correspond to the current situation from memory 3. Agent 2 learns the retrieved successful experiences through in-context learning (ICL) and solves the current task.

[0022] The communication unit 4 communicates with other devices via a network such as the Internet.

[0023] <Outline of operation of information processing device> Fig. 2 is a diagram for explaining an outline of the operation of the information processing device 1. In Fig. 2, the agent 2 shown in Fig. 1 is omitted. In Fig. 2, the memory 3 shown in Fig. 1 is shown.

[0024] The "ReAct" diagram shown in Figure 2 is a diagram explaining the behavior of a conventional agent. The "RAP (Retrieval-Augmented Planning)" diagram shown in Figure 2 is a diagram explaining the behavior of an agent 2 according to the present disclosure. Figure 2 also explains the behavior of a conventional agent on ALFWorld.

[0025] ALFWorld is a platform (simulator) for agents to learn in a virtual 3D environment. The agents solve assigned tasks while interacting within the ALFWorld environment (see Current Task in Figure 2). The inventors conducted evaluation experiments of this embodiment using ALFWorld, and therefore, the following description of this embodiment will be based on simulations in the ALFWorld environment. Since ALFWorld is a simulator that reproduces characteristics similar to those of a real environment, this embodiment will operate effectively even when applied to a real environment.

[0026] As shown in "ReAct" in Figure 2, conventional agents learn from fixed examples (Manual Examples) using ICL and lead to the solution of the current task (Current Task).

[0027] As shown in "RAP" in FIG. 2, the agent 2 of the present disclosure searches (acquires) an experience corresponding to the current situation (e.g., task, Act:think, etc.) from a memory 3 that stores past successful experiences (hereinafter, sometimes simply referred to as experiences or examples). The agent 2 learns the searched experience using the ICL and leads the current task to a solution. The current situation can be considered as the situation or process (solution process) when leading the task to a solution.

[0028] For example, Agent 2 retrieves (obtains) from Memory 3 an experience A2b similar to Situation A2a shown in Figure 2. Agent 2 learns the retrieved experience A2b, i.e., an example problem corresponding to the current situation A2a, using the ICL, and leads to the solution of the current task as shown by arrow A2c.

[0029] For example, Agent 2 retrieves an experience A2e similar to situation A2d shown in Figure 2 from Memory 3. Agent 2 learns the retrieved experience A2e, i.e., an example problem corresponding to the current situation A2d, using the ICL, and leads the current task to a solution, as shown by arrow A2f.

[0030] In this way, the agent 2 searches (calls) the successful experience corresponding to the current situation from the memory 3. The agent 2 learns the successful experience searched from the memory 3 using the ICL and leads to the solution of the current task. This operation enables the information processing device 1 to appropriately solve the given task.

[0031] As will be described below, the number of successful experiences retrieved from memory 3 may be two or more. For example, agent 2 may retrieve two or more successful experiences similar to situation A2a from memory 3. Agent 2 may learn the two or more retrieved successful experiences using the ICL and use them to solve the current task.

[0032] <Agent block configuration> FIG. 3 is a diagram showing an example of the block configuration of an agent 2. As shown in FIG. 3, the agent 2 has a reasoner 11, a retriever 12, and an executor 13. The reasoner 11 and the executor each have a language model (LM). The LM is, for example, a large-scale language model LLM. In addition to the block of the agent 2, FIG. 3 also shows a memory 3 and a current task A3a on a platform such as ALFworld.

[0033] The first and second lines in the current task A3a are the task (issue) given to the agent 2 (information processing device 1). The third, fourth, and sixth lines are the output of the agent 2. The fifth line is the output (answer) of the platform (environment).

[0034] (1) Memory Successful experiences (as examples) are stored in the memory 3. The experiences include, for example, task-related information such as tasks and overall plans, and trajectory information such as action plans, actions, and observations (Obs).

[0035] The successful experiences stored in memory 3 are, for example, past successful experiences, and are logs of when agent 2 solved a task. That is, the memory 3 stores logs of when agent 2 solved a task in the past. The successful experiences stored in memory 3 can be considered as successful processes.

[0036] The user may manually generate a log of tasks that the agent 2 has never solved and store it in the memory 3 as a successful experience. However, since there may be constraints, etc., that are difficult for the user to imagine when solving a task, a log of tasks that the agent 2 has actually solved may be more appropriate as a successful experience than a log set by the user based on imagination. Therefore, if accuracy is important, it is advisable to store only tasks that the agent 2 has solved in the past as successful experiences. On the other hand, if the number of tasks that the agent 2 has solved in the past is limited, even if the content of the task with the correct answer is easily known from the user's experience, the log of that task may not be stored in the memory 3. Since storing a larger number of successful experiences in the memory 3 allows for a wider variety of successful experiences to be referenced, in such cases the user may manually generate a log and store it in the memory 3 as an additional successful experience. Below, an example of using past successful experiences as successful experiences will be described. However, since the process when using the successful experiences set by the user and the process when using the past successful experiences are similar, a detailed description will be omitted.

[0037] FIG. 4 is a diagram showing an example of a log stored in memory 3. Past successful experiences (logs) are stored in memory 3 in the format shown in FIG. 4. For example, one successful experience has a format such as task, category, plan, and trajectories, and is stored in memory 3. FIG. 4 shows an example in which a successful experience of placing a hot tomato in a garbage can is stored, but various successful experiences are stored in memory 3.

[0038] Here, task indicates the task that was solved in the successful experience. category indicates the classification of the task. plan indicates the planned series of operations performed to solve the task. trajectories indicates the planned actions or behaviors performed to realize the plan. Note that successful experiences do not necessarily need to include all of this information. In the example described below, the information contained in task, plan, and trajectories is used, but the information in category is not used, so category may be omitted.

[0039] 4 may correspond to an overall plan. Here, the overall plan refers to a plan of a series of operations required to solve a task. The overall plan stored in memory 3 may be called a success plan.

[0040] The "think" in Figure 4 may correspond to an action plan, where the action plan refers to a plan of next steps to be taken to carry out a part of the overall plan.

[0041] The statements with ">" without think shown in Figure 4 (for example, lines 7 and 9 in Figure 4) may correspond to actions. Here, an action refers to an action actually performed on the simulator. Here, the current situation, such as the placement of objects, does not change when an action plan is generated, but the current situation changes when the action is performed.

[0042] The statements without ">" in Figure 4 (for example, statements on lines 6 and 8 in Figure 4) may correspond to Obs. Obs indicates the results of the platform's evaluation of an action plan or the observation results of the environment changed by the action. Obs may be considered the output of the platform.

[0043] (2) Reasonable The reasoner 11 generates an overall plan, an action plan, and a retrieval key.

[0044] (2-1) Overall plan The reasoner 11 makes an overall plan for solving the task from the task information (generates an overall plan). For example, the reasoner 11 generates the overall plan from "task" included in the current task A3a in FIG. 3. The task information may be considered as information including the task and the conditions related to the task (see, for example, the first to third lines in FIG. 10). In this embodiment, the overall plan is generated by giving instructions to the large-scale language model by a prompt including the task information. The overall plan generated by the instructions of the prompt (large-scale language model) may be called a solution plan.

[0045] FIG. 5 is a diagram illustrating a prompt for generating an overall plan. Note that the prompt shown in FIG. 5 is different from the current task A3a shown in FIG. 3. The prompt may be considered as the ICL and the executor 13. Note that the task shown underlined A5a in FIG. 5 may correspond to, for example, "Task: put a hot tomato on desk" shown in the current task A3a in FIG. 3.

[0046] The reasoner 11 retrieves similar tasks similar to the task shown underlined A5a in Fig. 5 and the overall plan for the similar tasks from the memory 3 via the retriever 12. Hereinafter, the process of retrieving information similar to specified information from the memory 3 using the retriever 12 is also referred to as "retrieving." The reasoner 11 provides (inputs) the retrieved similar tasks and the overall plan for the similar tasks to the prompt.

[0047] For example, the reasoner 11 retrieves from the memory 3 the similar task (similar task similar to the task underlined A5a) shown in underlined A5b in Figure 5 and the overall plan for the similar task shown in underlined A5b, and provides them to the prompt. That is, the reasoner 11 retrieves from the memory 3 an example problem (text underlined A5b and A5c) from the memory 3 to solve the task shown in underlined A5a, and provides them to the prompt.

[0048] In the above example, one example similar to the text shown in underline A5a is retrieved from memory 3, but this is not limiting. The reasoner 11 may retrieve the top n example problems similar to the text shown in underline A5a from memory 3. The reasoner 11 may provide the retrieved top n example problems in a prompt. By providing an appropriate number of example problems in the prompt, the task shown in underline A5a can be properly solved.

[0049] Furthermore, the retrieval of similar tasks and overall plans for similar tasks is actually performed by the retriever 12. For example, the retriever 12 is equipped with an Embedding Model that performs a process of vectorizing text, and vectorizes the text (tokens) written in "task" using a technique such as embedding, and calculates the similarity with the text of "task" (vectorized text) stored in the memory 3. The similarity is evaluated, for example, using the cosine similarity between the vectorized texts. The retriever 12 retrieves highly similar texts (similar tasks) and overall plans for similar tasks from the memory 3.

[0050] Here, "high similarity" means, for example, that the similarity is equal to or greater than a predetermined threshold. However, the criteria for determining whether the similarity is high are not limited to this. For example, the similarity may be determined to be high if the similarity is within a predetermined rank from the top. Alternatively, the similarity may be determined to be high if the similarity is equal to or greater than a predetermined threshold and is within a predetermined rank from the top. Note that "high similarity" is also expressed as "similar." The reasoner 11 provides the retrieved similar tasks and the entire plant to a prompt.

[0051] (2-2) Action Plan The action plan indicates a plan for the next action to be taken (see "Act: think: ..." in FIG. 3). The reasoner 11 generates the action plan for the executor 13 to generate an action, which will be described later.

[0052] For example, the reasoner 11 generates an action plan based on the answers (Obs) from the task or the platform. The generated action plan is given to the executor 13 (ICL) and becomes the basis for the executor 13 to generate an action.

[0053] (2-3) Retrieval Bulky The retrieval key is a keyword that corresponds to the current situation. The reasoner 11 provides a prompt with a fixed example (Manual Example) and uses a large-scale language model to generate a (relevant) retrieval key based on the content of the action plan.

[0054] 6 is a diagram illustrating a prompt for generating a retrieval key. The reasoner 11 provides the prompt with an example shown in area A6b in FIG. 6 to generate a retrieval key based on the contents of the action plan (think) shown in underlined area A6a in FIG. 6.

[0055] The reasoner 11 generates an appropriate retrieval key for the action plan shown in underlined A6a by learning the example shown in area A6b in Fig. 6. For example, in the action plan shown in underlined A6a, the reasoner 11 generates a retrieval key such as "watch." The reasoner 11 outputs the generated retrieval key to the current task (see, for example, box A7a in Fig. 7 or the seventh line in Fig. 10).

[0056] The number of example problems (fixed example problems) for obtaining a retrieval key is not limited to one. A plurality of example problems for obtaining a retrieval key may be given.

[0057] Although the example described here involves generating a retrieval key in a large-scale language model by providing a prompt containing a predetermined example, a retrieval key may also be generated using a simpler method, such as extracting frequently occurring words in an action plan or summarizing an action plan in a large-scale language model without providing an example as a prompt.

[0058] (3) Retriever The retriever 12 retrieves the similar tasks and the overall plans for the similar tasks from the memory 3, as described above in "(2-1) Overall Plan."

[0059] Furthermore, the retriever 12 retrieves experiences (past experiences) related to the action from the memory 3 based on the task information and the retrieval key.

[0060] Figure 7 is a diagram illustrating the retrieval of past experiences related to an action. Box A7a in Figure 7 indicates the current task. Box A7b indicates one experience out of multiple experiences stored in memory 3. Box A7c indicates a prompt.

[0061] The retriever 12 calculates the similarity between the text of "Task" in the current task and the text of "Task" stored in the memory 3, as indicated by an arrow A7d in FIG.

[0062] The retriever 12 calculates the similarity between the text of the "Overall Plan" in the current task and the text of the "Overall Plan" stored in the memory 3, as indicated by an arrow A7e in FIG.

[0063] The retriever 12 calculates the similarity between the "Retrieval Key" in the current task and the text of the "trajectories" stored in the memory 3, as indicated by an arrow A7f in FIG.

[0064] The similarity may be calculated using the method described above in "(2-1) Overall Plan."

[0065] The retriever 12 retrieves the most similar experience from the memory 3, for example, based on the sum of the degrees of similarity described above. The experience retrieved by the retriever 12 is provided to the prompt (executor 13), as shown by arrow A7g in FIG.

[0066] In the above example, one experience is retrieved from memory 3, but multiple highly similar experiences may be retrieved from memory 3. For example, the top k most similar experiences may be retrieved from memory 3. The retrieved top k experiences may be provided as examples to a prompt. By providing an appropriate number of examples to the prompt, an appropriate action can be generated.

[0067] (4) Executor The executor 13 generates the next action based on the task information and past experience. If an action plan has been generated, the executor 13 also takes the action plan into consideration when generating the next action.

[0068] FIG. 8 is a diagram illustrating a prompt for generating an action. The executor 13 inputs task information into the prompt, as shown in areas A8a and A8b in FIG. 8. The executor 13 inputs past experiences retrieved by the retriever 12 into the prompt, as shown in area A8c in FIG. 8. For example, the executor 13 may input n past experiences into the prompt. This generates an action (not shown).

[0069] The past experience shown in area A8c in FIG. 8 may correspond to the "Top experience from memory" indicated by arrow A7g in FIG.

[0070] The functional blocks of agent 2 shown in Figure 3 are an example. The function in agent 2 that learns a given example problem and solves a task may be called a solving unit. The function in agent 2 that accesses memory 3 may be called an accessing unit.

[0071] <Operation flow> 9 is a flowchart showing an example of the operation of the information processing device 1. It is assumed that task information (see the first to third lines in FIG. 10) is given to the information processing device 1.

[0072] The information processing device 1 generates an overall plan (S1). For example, as described in Fig. 5, the information processing device 1 retrieves similar tasks that are similar to a given task and overall plans for the similar tasks from the memory 3. The information processing device 1 provides the retrieved similar tasks and their overall plans as examples to the prompt, and generates the overall plan.

[0073] The information processing device 1 retrieves (extracts) examples (past experiences) similar to the task and the overall plan from the memory 3 (S2).

[0074] The information processing device 1 generates an action or an action plan based on the example problem extracted in S2 or the example problem (past experience) extracted in S7 (described later) (S3). For example, the information processing device 1 provides the example problem extracted in S2 or the example problem extracted in S7 to a prompt as described in FIG. 8, and generates an action or an action plan.

[0075] The information processing device 1 determines whether or not an action has been generated (S4).

[0076] When an action is generated (Yes in S4), the information processing device 1 inputs the action into the environment (current task) and runs a framework (simulator) such as ALFWorld (S5), which changes the current state of the simulator.

[0077] On the other hand, if an action is not generated (No in S4), in other words, if an action plan is generated, the information processing device 1 generates a retrieval key (S6). For example, as described in FIG. 6, the information processing device 1 provides a predetermined example problem to the prompt and generates a retrieval key related to the current action plan. In other words, the information processing device 1 generates a retrieval key according to the current situation.

[0078] After generating the retrieval key in S6, the information processing device 1 extracts example problems similar to the task, overall plan, and retrieval key from the memory 3 (S7). For example, as described in Fig. 7, the information processing device 1 calculates the similarity between each of the task, overall plan, and retrieval key, and retrieves one or more example problems that are most similar based on the sum of the calculated similarities.

[0079] After running the simulator in S5 or extracting example questions in S7, the information processing device 1 determines whether the task is completed or whether the generation of text in the current task has reached a predetermined number of steps (for example, 50 steps) (S8). Note that the number of steps is not limited to 50. As will be described later, this number of steps is a condition for determining that the task has been unsuccessful, and therefore may be changed as appropriate depending on how strictly the failure is to be evaluated.

[0080] If the task is not completed and 50 steps have not been reached (No in S8), the information processing device 1 shifts the process to S3.

[0081] On the other hand, if the task is completed or reaches 50 steps, the information processing device 1 ends the processing of the flowchart in Fig. 9. Note that if the task reaches 50 steps without being completed, the information processing device 1 may output an error indicating that the task has failed to be resolved.

[0082] <Text generation example> Fig. 10 is a diagram showing an example of text generation on a platform. When the information processing device 1 is operated on the ALFWorld platform, the text shown in Fig. 10 is generated as an example. Note that the parenthesized sentences in Fig. 10 are sentences explaining the example of text generation, and are not text generated on the platform.

[0083] The first to third lines in FIG. 10 are task information. The fourth line is generated, for example, by the processing of S1 in FIG. 9. The fifth line is generated, for example, by the processing of S3. The seventh line is generated, for example, by the processing of S4 (No) and S6. The eighth line is generated, for example, by the processing of S7, S8 (No), and S3. The tenth and twelfth lines are generated, for example, by the processing of S4 (Yes), S8 (No), and S3. The fourteenth line is generated, for example, by the processing of S4 (No) and S6. The fifteenth line is generated, for example, by the processing of S7, S8 (No), and S3. The seventeenth line is generated, for example, by the processing of S3. The fourth and subsequent lines may be considered as a solution process or decision-making process for resolving the task information.

[0084] <Summary of the embodiment> The information processing device 1 learns from given example problems to solve a task. The information processing device 1 acquires and learns example problems corresponding to the task solving process from a memory 3 that stores successful processes in solving tasks in the past as examples. Through this operation, the information processing device 1 can call and learn example problems (successful experiences) corresponding to the current situation, and can solve the given task appropriately.

[0085] <Modification> Although the operation of the information processing device 1 in a text environment has been described above, the information processing device 1 can also be applied to a multimodal environment (text + image). In a modified example, the text described in the above embodiment is replaced with text and image (picture).

[0086] Fig. 11 is a diagram illustrating an outline of the operation of the information processing device 1 in the modified example. Fig. 11 shows the memory 3 shown in Fig. 1. Also, Fig. 11 omits the illustration of the agent 2 shown in Fig. 1. Fig. 11 explains the operation of the agent 2 of the present disclosure on Franka Kitchen.

[0087] Franka Kitchen is a platform (simulator) for agents to learn how to operate a robot with nine degrees of freedom in a kitchen environment equipped with household items.

[0088] Agent 2 retrieves experiences (text and images) that correspond to the current situation from Memory 3, which stores text and images of past successful experiences. Agent 2 learns the retrieved experiences using a Vision Language Model (VLM) and leads to the solution of the current task.

[0089] For example, Agent 2 retrieves from its memory an experience A11b similar to the situation A11a shown in Fig. 11. Agent 2 learns the retrieved experience A11b, i.e., an example problem corresponding to the current situation A11a, using a visual language model, and leads the task to a solution as shown in box A11c.

[0090] Here, the visual language model may include a first model that converts natural language text and images into a single embedding space, and a second model that accepts input of natural language text (prompts) and images and generates text.

[0091] By using the first model, it is possible to compare the text and images corresponding to the current situation with the text and images included in past successful experiences in the same embedding space, and to calculate a similarity that comprehensively considers the text and images. This makes it possible to select successful experiences to be learned by comprehensively considering the text and images. Furthermore, by using the second model, it is possible to generate an action or action plan from an image and a task description. Therefore, it is possible to train a second model using successful experiences selected using the first model, and to generate an action or action plan from an image and a task that corresponds to the current situation using this second model. Note that CLIP and the like are known as first models, and CPT-4V, LLaVA and the like are known as second models, so a detailed description of each model is omitted.

[0092] In addition, in the visual language model, either the first model or the second model may be omitted.

[0093] When the first model is omitted, for example, an experience A11b similar to the situation A11a may be searched for based on the similarity between the text portion included in the situation A11a and the text portion of the successful experience stored in the memory 3. In this case, the procedure for searching for the successful experience is the same as the operation of the retriever of the information processing device 1 in the text environment.

[0094] In addition, if the second model is omitted, the text portion of the text and image included in the selected successful experience can be learned by the ICL. In this case, the specific operation is the same as that of the information processing device 1 in a text environment, and therefore a detailed description thereof will be omitted.

[0095] Furthermore, the second model may be used to calculate the similarity taking into consideration the text and the image comprehensively. For example, the image included in the situation A11a and the image of the successful experience stored in memory 3 may be converted into text in a natural language using the second model, and the experience A11b may be searched for based on the similarity between the texts. In this case, if the situation A11a or the successful experience includes text, this text may also be used to search for the experience A11b.

[0096] In this way, the agent 2 searches the memory 3 for successful experiences that correspond to the current situation. The agent 2 learns the searched successful experiences using a visual language model and leads them to the solution of the current task. This operation enables the information processing device 1 to appropriately solve the given task.

[0097] (Other variations) In the above-described embodiment, successful experiences used as examples are selected based on the similarity between tasks when generating an overall plan, and based on the similarity between the retrieval key corresponding to the current situation and the trajectories of the successful experience when generating an action or action plan. However, successful experiences used as examples may also be selected based on the similarity evaluated for other combinations. For example, successful experiences may be selected based on the similarity between the retrieval key corresponding to the current situation and the task of the successful experience, or based on the similarity between the task to be solved and the trajectories of the successful experience. Furthermore, if images are also stored as successful experiences as in a modified example, successful experiences used as examples may be selected based on the similarity between text information such as the task and retrieval key and the image included in the successful experience. In this case, the image included in the successful experience may be converted into text or a vector corresponding to the text using a visual language model, and then the similarity may be evaluated. Furthermore, in the above-described embodiment, successful experiences may be selected using category information included in the successful experience. In this case, a category corresponding to the task to be solved or the current situation may be estimated based on either the task, overall plan, or retrieval key, and the similarity may be compared with the category included in the successful experience. In this case, a large-scale language model may be used to suggest categories based on the task, overall plan, or retrieval key, and these categories may be used as categories for the task to be solved or the current situation. Alternatively, the similarity between the task, overall plan, or retrieval key and the categories included in the successful experience may be directly evaluated.

[0098] Alternatively, the similarity may be evaluated as the sum of the similarities between each of the tasks, overall plans, and retrieval keys and the tasks, overall plans, and trajectories of the successful experience. In this case, for example, the similarity between tasks, the similarity between overall plans, and the similarity between the retrieval keys and trajectories may be calculated as cosine similarities, and the sum of these similarities may be treated as the similarity between the task to be solved or its solution process and the task or its successful process stored as a successful experience.

[0099] Furthermore, the text or image may be recorded or used in a format that can be used by the information processing device 1. In other words, as long as the information indicates text or an image, it may be data in a format that humans cannot recognize as text or an image. In this case, the information processing device 1 may use the text or image in a format that humans cannot recognize without converting the text or image into a format that humans can recognize. Note that both formats that can be recognized by humans and formats that cannot be recognized by humans are collectively referred to as information indicating text or information indicating an image.

[0100] Although the embodiments have been described above with reference to the drawings, the present disclosure is not limited to such examples. It is clear that a person skilled in the art can conceive of various modifications or alterations within the scope of the claims. It is understood that such modifications or alterations also fall within the technical scope of the present disclosure. Furthermore, the components in the embodiments may be combined in any manner without departing from the spirit of the present disclosure.

[0101] In the above-described embodiments, the notation "... part" used for each component may be replaced with other notations such as "... circuitry," "... assembly," "... device," "... unit," or "... module."

[0102] The present disclosure can be realized by software, hardware, or software linked to hardware. Each functional block used in the description of the above embodiments may be partially or entirely realized as an LSI, which is an integrated circuit, and each process described in the above embodiments may be partially or entirely controlled by a single LSI or a combination of LSIs. The LSI may be composed of individual chips, or may be composed of a single chip that includes some or all of the functional blocks. The LSI may have data input and output. Depending on the degree of integration, the LSI may be called an IC, system LSI, super LSI, or ultra LSI.

[0103] The integrated circuit method is not limited to LSI, but may be realized by a dedicated circuit, a general-purpose processor, or a dedicated processor. Also, a field programmable gate array (FPGA) that can be programmed after LSI manufacturing, or a reconfigurable processor that can reconfigure the connections and settings of circuit cells within the LSI, may be used. The present disclosure may be realized as digital processing or analog processing.

[0104] Furthermore, if an integrated circuit technology that can replace LSI emerges due to advances in semiconductor technology or other derivative technologies, it is natural that such technology may be used to integrate functional blocks. The application of biotechnology, etc. is also a possibility. [Industrial Applicability]

[0105] The present disclosure is useful for solving tasks in situations such as automatically operating a system or automatically operating a robot. [Explanation of symbols]

[0106] 1. Information processing equipment 2 Agent 3. Memory 11 Reasonable 12 Retriever 13 Executor

Claims

1. an access unit that accesses a memory that stores, for each of a plurality of tasks, a successful process in solving the task as an example; a solving unit that retrieves example problems similar to a solution process of a task to be solved from the memory, trains a large-scale language model, and solves the task to be solved using the trained large-scale language model; An information processing device having the above.

2. the solving unit causes the large-scale language model to learn the example problem by including the acquired example problem in a prompt that instructs the large-scale language model to solve the task to be solved. The information processing device according to claim 1 .

3. the solving unit selects the successful process similar to the solved process as an example problem to be trained by the large-scale language model; The information processing device according to claim 2 .

4. The solving unit generates information indicating keywords related to the solving process, and retrieves from the memory, as examples, successful processes including information indicating text similar to the information indicating the keywords. The information processing device according to claim 3 .

5. the solving unit retrieves a plurality of example problems from the memory in order of most similarity to the solving process; The information processing device according to claim 3 .

6. The memory further stores a success plan for solving each of the plurality of tasks; The resolution unit a large-scale language model is trained on a successful plan for a task similar to the task to be solved, and a solution plan for the task to be solved is created using the trained large-scale language model; retrieving from the memory an example problem similar to a solution process of the task to be solved, and training the example problem in a large-scale language model; and generating the solution process using the trained large-scale language model. The information processing device according to claim 1 .

7. The memory stores a successful process as an example, the successful process including at least information indicating a text; the solving unit learns examples of information representing text and solves the task; The information processing device according to claim 1 .

8. The memory stores information representing text and information representing images as examples of successful processes; the solving unit learns examples of information representing text and information representing images and solves the task; The information processing device according to claim 7 .

9. accessing a memory storing, for each of a plurality of tasks, an example of a successful process in solving the task; retrieving from the memory example problems similar to the solution process of the task to be solved, and training the large-scale language model to learn the problem; and solving the task to be solved using the trained large-scale language model. Task solving methods.

10. The processor accessing a memory storing, for each of a plurality of tasks, an example of a successful process in solving the task; retrieving from the memory example problems similar to the solution process of the task to be solved, and training the large-scale language model to learn the problem; and solving the task to be solved using the trained large-scale language model. A program that executes a process.

Images