World model-based task processing method and device, equipment and medium

Abstract

The application relates to the technical field of artificial intelligence, and discloses a task processing method and device based on a world model, equipment and a medium. The method comprises the following steps: acquiring task input information; the task input information comprises image information, voice information, text information, structured information and / or action information; based on the task input information, a target reasoning mode and a plurality of target task networks are determined; the task input information is input into the target task network to obtain prediction information; based on the target reasoning mode, the prediction information is reasoned to obtain reasoning information; and based on the reasoning information, a task request corresponding to the task input information is responded. The embodiments of the application can ensure the versatility and diversity of the capabilities of the world model to adapt to various task scenarios.

Description

Technical Field

[0001] This application relates to the field of artificial intelligence technology, and in particular to a task processing method, apparatus, device, and medium based on a world model. Background Technology

[0002] A world model is an artificial intelligence system that predicts future events in an environment by constructing an internal representation of that environment. In related technologies, perception technologies such as computer vision and speech recognition can acquire data from the environment. Then, techniques such as machine learning and deep learning are used to model, learn from, and train this data, thereby forming a world model capable of recognizing and understanding the environment. Using this world model, prediction and simulation of future events in the environment can be achieved.

[0003] However, existing world models are trained on specific tasks in a single application scenario (e.g., robot control) using data from specific modalities (e.g., control commands). The trained world model can only handle the specific task based on the data of that specific model and cannot be used for other application scenarios, thus limiting its application. Summary of the Invention

[0004] The purpose of this application is to provide a task processing method, apparatus, device, and medium based on a world model, which can ensure the universality of the world model and the diversity of its capabilities to adapt to various task scenarios.

[0005] This application provides a task processing method based on a world model, including: Obtain task input information; the task input information includes image information, voice information, text information, structured information and / or action information; Based on the task input information, a target reasoning pattern and several target task networks are determined; the target reasoning pattern is a constraint process with the task input information as a condition, and the target task network is a task network in a pre-trained world model that meets the matching conditions. The task input information is input into the target task network to obtain prediction information; Based on the target reasoning pattern, the predicted information is reasoned to obtain reasoning information; Based on the reasoning information, respond to the task request corresponding to the task input information.

[0006] In some embodiments, the method for determining the target reasoning pattern includes: The task input information is processed to extract request instructions, resulting in request instruction information. Calculate the matching degree between the request instruction information and multiple inference patterns to obtain the pattern matching degree; Based on the pattern matching degree, candidate reasoning patterns are determined; The candidate reasoning patterns are configured with constraint parameters to obtain the target reasoning pattern.

[0007] In some embodiments, a method for determining the target task network includes: Context extraction is performed on the task input information to obtain context information; The task input information and the context information are combined to obtain fused information; Calculate the matching degree between the fused information and the task network to obtain the network matching degree; The target task network is determined based on the network matching degree.

[0008] In some embodiments, fusing the task input information and the context information includes: The task input information is determined as a cross-attention query vector; The context information is transformed to obtain the cross-attention key vector and the cross-attention value vector; The cross-attention query vector, the cross-attention key vector, and the cross-attention value vector are subjected to cross-attention processing to obtain the fused information.

[0009] In some embodiments, reasoning about the predicted information based on the target reasoning pattern includes: Based on the constraint parameters of the target reasoning mode, the prediction information is denoised and diffused to reconstruct the corresponding task requirement information and obtain the reasoning information.

[0010] In some embodiments, the denoising and diffusion of the prediction information based on the constraint parameters of the target inference pattern includes: Gaussian noise is added to the predicted information to obtain initial noisy information; Based on the diffusion constraint parameters of the target reasoning mode, the initial noisy information is forward diffused to obtain noisy state information; Based on the denoising constraint parameters of the target reasoning mode, the noisy state information is denoised in reverse to obtain the reasoning information.

[0011] In some embodiments, before determining the target inference pattern and several target task networks based on the task input information, the method further includes: The task input information is preprocessed to obtain preprocessed task input information; The preprocessed task input information is encoded to obtain encoded task input information.

[0012] This application also provides a task processing device based on a world model, including: The first module is used to acquire task input information; the task input information includes image information, voice information, text information, structured information and / or action information; The second module is used to determine a target inference pattern and several target task networks based on the task input information; the target inference pattern is a constraint process with the task input information as a condition, and the target task network is a task network in a pre-trained world model that meets the matching conditions. The third module is used to input the task input information into the target task network to obtain prediction information; The fourth module is used to reason about the predicted information based on the target reasoning pattern to obtain reasoning information; The fifth module is used to respond to the task request corresponding to the task input information based on the reasoning information.

[0013] This application also provides an electronic device, which includes a memory and a processor. The memory stores a computer program, and the processor executes the computer program to implement the above-described task processing method based on a world model.

[0014] This application also provides a computer-readable storage medium storing a computer program that, when executed by a processor, implements the above-described task processing method based on a world model.

[0015] The beneficial effects of this application are as follows: A world model integrating multiple task networks is pre-defined. By acquiring multimodal task input information and dynamically determining the target inference mode and target task network, the target task network generates corresponding prediction information based on the task input information. Then, based on the target inference mode, the prediction information is inferred to obtain inferred information. Finally, the task request corresponding to the task input information is responded to based on the inferred information. Therefore, by pre-defining a world model integrating multiple task networks and dynamically determining the target inference mode and target task network based on multimodal task input information, the appropriate task network in the world model can be invoked and the appropriate inference mode determined for the task input information. This enables targeted prediction and inference of multimodal task input information, flexibly responding to different task requirements, ensuring the universality and diversity of the world model's capabilities to adapt to various task scenarios. Attached Figure Description

[0016] Figure 1 This is an application environment diagram of the task processing method based on the world model provided in the embodiments of this application.

[0017] Figure 2 This is a flowchart of a task processing method based on a world model provided in an embodiment of this application.

[0018] Figure 3 This is a flowchart of a specific method for determining a target reasoning pattern provided in an embodiment of this application.

[0019] Figure 4 This is a flowchart of a specific method for determining a target task network provided in an embodiment of this application.

[0020] Figure 5 This is a schematic diagram of the structure of the task processing device based on the world model provided in the embodiments of this application.

[0021] Figure 6 This is a schematic diagram of the hardware structure of the electronic device provided in the embodiments of this application. Detailed Implementation

[0022] To make the objectives, technical solutions, and advantages of this application clearer, the following detailed description is provided in conjunction with the accompanying drawings and embodiments. It should be understood that the specific embodiments described herein are merely illustrative and not intended to limit the scope of this application.

[0023] It should be noted that although functional modules are divided in the device schematic diagram and a logical order is shown in the flowchart, in some cases, the steps shown may be performed in a different order than the module division in the device or the order in the flowchart. The terms "first," "second," etc., in the specification, claims, and drawings are used to distinguish similar objects and are not used to describe a specific order or sequence.

[0024] Unless otherwise defined, all technical and scientific terms used herein have the same meaning as commonly understood by one of ordinary skill in the art to which this application belongs. The terminology used herein is for the purpose of describing embodiments of this application only and is not intended to limit this application. Furthermore, the information, data, and signals involved in the embodiments of this application are all authorized by relevant parties or have been fully authorized by all parties, and the collection, use, and processing of related data comply with the relevant laws, regulations, and standards of the relevant countries and regions.

[0025] In traditional world model applications, the model training process is only focused on specific tasks within a single application scenario and uses data from that specific modality. This results in world models that can only handle data and tasks within that specific modality, failing to adapt to multimodal inputs and cross-scenario application requirements. The essence of this problem lies in the modality-specificity and task-specificity of the world model, which makes it lack generalization ability when faced with diverse task inputs, thus affecting the overall adaptability and scalability of the system. For example, in a smart home control system, when a user simultaneously provides the voice command "raise the air conditioner temperature" and an image pointing to the air conditioner, the existing world model, supporting only single-modal processing, cannot effectively integrate voice and image information, causing the system to fail to accurately recognize the user's intent and execute the corresponding operation. Furthermore, in this scenario, the task input information contains multi-source heterogeneous data, but the existing model can only parse single-modal features, leading to misclassification or neglect of task requests, resulting in system response failure.

[0026] If these problems are not addressed, the world model will struggle to cope with complex multimodal interaction scenarios in real-world applications, leading to increased task failure rates, decreased system reliability, and limitations on the deployment and application of artificial intelligence technology across a wider range of fields. Therefore, the continued existence of this problem will severely restrict the practical value of the world model in dynamic environments, making the system unable to meet the adaptability requirements of diverse application scenarios.

[0027] Based on this, embodiments of this application provide a task processing method, apparatus, device, and medium based on a world model. A world model integrating multiple task networks is preset, and by acquiring multimodal task input information and dynamically determining the target reasoning mode and target task network, it can flexibly respond to different task requirements.

[0028] Figure 1 This diagram illustrates the application environment of the world-model-based task processing method provided in this embodiment. (See also...) Figure 1This method is applied to a task processing system based on a world model. The system includes a terminal 110 and a server 120. The terminal 110 and server 120 are connected via a network. The terminal 110 can be at least one of a mobile phone, tablet, laptop, vehicle-mounted terminal, or robot control module. The server 120 can be a standalone server or a server cluster consisting of several servers. The terminal 110 sends task input information to the server 120. The server 120 acquires the task input information, determines a target inference pattern and several target task networks based on the task input information, inputs the task input information into the target task networks to obtain prediction information, infers from the prediction information based on the target inference pattern to obtain inference information, and responds to the task request corresponding to the task input information based on the inference information. The task input information includes image-type information, voice-type information, text-type information, structured information, and / or action-type information. The target inference pattern is a constraint process conditioned on the task input information, and the target task network is a task network in a pre-trained world model that meets the matching conditions.

[0029] It should be understood that Figure 1 The application scenarios shown are merely examples. In practical applications, the task processing method based on the world model provided in this application embodiment can also be applied to other scenarios. For example, it can be directly applied to terminal 110, which is used to obtain task input information, determine a target inference mode and several target task networks based on the task input information, input the task input information into the target task network to obtain prediction information, infer the prediction information based on the target inference mode to obtain inference information, and respond to the task request corresponding to the task input information based on the inference information.

[0030] See Figure 2 In one embodiment, a task processing method based on a world model is provided. The execution subject of the method can be a terminal or a server, including but not limited to steps S201 to S205.

[0031] Step S201: Obtain task input information.

[0032] Task input information includes image-based information, voice-based information, text-based information, structured information, and / or action-based information. In essence, task input information refers to the various types of data received to initiate task processing. This information can include multiple modalities such as image-based information, voice-based information, text-based information, structured information, and / or action-based information to adapt to the task requirements of different application scenarios.

[0033] Obtaining task input information can be achieved by directly reading locally stored files to acquire image or text information; alternatively, the executing entity can connect to an external sensor interface to receive raw data streams from a camera or microphone to acquire image or voice information; or, the executing entity can receive structured data packets sent from other systems via a network communication interface.

[0034] Step S202: Based on the task input information, determine the target reasoning pattern and several target task networks.

[0035] The goal-oriented reasoning pattern is a constrained process based on task input information. In essence, it defines the specific rules and logic for reasoning based on task input information to ensure that the reasoning results meet the specific requirements and constraints of the task request.

[0036] The target task network is a task network within a pre-trained world model that meets matching conditions. A world model can be understood as an artificial intelligence system that predicts future events in its environment by constructing an internal representation of that environment. This model can model, learn, and train on environmental data, thereby forming a cognition and understanding of the environment to achieve prediction and simulation of future events. The target task network refers to a task network within a pre-trained world model that meets matching conditions, specifically designed to handle specific types of tasks or data modalities, and capable of providing corresponding predictive capabilities based on task input information.

[0037] Determining the target inference pattern and several target task networks based on task input information can be achieved by maintaining a pre-defined rule base and directly searching for and specifying one or more pre-defined inference patterns and task networks based on the type of task input information or the keywords it contains. Alternatively, the executing entity can provide a user interface that allows operators to manually select or configure appropriate inference patterns and task networks based on the received task input information.

[0038] Step S203: Input the task input information into the target task network to obtain prediction information.

[0039] Predictive information refers to the output information obtained after the task input information is processed by the target task network. This information is a preliminary understanding or simulation result of the task input under a specific task network.

[0040] Inputting task input information into the target task network can be achieved by directly using the original task input information as input data. Alternatively, the task input information can undergo simple format conversion or data type adaptation to meet the input requirements of the target task network before being input into it.

[0041] Step S204: Based on the target reasoning pattern, reason about the predicted information to obtain reasoning information.

[0042] Inference information refers to the final result obtained after the prediction information has been constrained and processed by the target inference model. This information has been refined and filtered, and is directly used to respond to task requests.

[0043] Reasoning about the predicted information based on the target reasoning pattern can be achieved by filtering, classifying, or aggregating the predicted information according to predefined logical judgment rules within the target reasoning pattern. Alternatively, it can be done by utilizing transformation functions included in the target reasoning pattern to adjust the predicted information according to specific format or content requirements, thereby generating reasoning information.

[0044] Step S205: Based on the reasoning information, respond to the task request corresponding to the task input information.

[0045] A task request refers to the specific task instructions or requirements corresponding to the task input information.

[0046] Responding to the task request corresponding to the task input information based on inference information can be achieved by directly displaying the inference information as the final result to the user. Alternatively, it can trigger preset automated operations based on the inference information, such as sending instructions to other systems, updating database records, or controlling external devices.

[0047] The following example will provide a more detailed explanation of the above technical solution: Suppose user A wants to identify a specific object in an image and perform a simple operation based on the identification result. Existing world models may only be able to handle single modalities (such as images) and have fixed tasks (such as identification only), and cannot flexibly cope with user A's complex needs (image recognition and subsequent operations).

[0048] First, the executing entity obtains the task input information. User A uploads an image containing multiple objects and verbally instructs, "Identify the chair in the image." At this point, the executing entity receives both image-related information (image data) and voice-related information (verbal instruction).

[0049] Secondly, based on the task input information, the executing agent determines the target inference pattern and several target task networks. According to the received voice command "identify the chair in the image" and the image information, the executing agent determines a target inference pattern for "object recognition" through pre-set simple rules. The constraint of this inference pattern is set to focus only on the category of "chair". Simultaneously, the executing agent determines a target task network for "image recognition," which is the part of the pre-trained world model specifically used for image content analysis.

[0050] Next, the executing entity inputs the task input information into the target task network to obtain prediction information. The executing entity inputs the image uploaded by user A into the "image recognition" target task network. The network analyzes the image and outputs prediction information, such as "the image contains a table, chair, and lamp."

[0051] Subsequently, based on this target inference pattern, the executing entity infers from the predicted information to obtain inferred information. The "object recognition" target inference pattern receives the predicted information "The image contains a table, chair, and lamp." Since the constraint of this inference pattern is "focusing only on chairs," the predicted information is filtered out, removing information about "table" and "lamp," ultimately resulting in the inferred information "The image contains a chair."

[0052] Finally, based on this inference information, the executing entity responds to the task request corresponding to the task input information. For example, based on the inference information "the image contains a chair," the executing entity responds to user A's task request. For instance, the executing entity can highlight the chair in the image on the screen and simultaneously provide a voice prompt to the user, "The chair has been identified."

[0053] The above technical solution effectively addresses the limitations of existing world models in application by introducing the concepts of target inference patterns and target task networks. In the example above, a traditional world model might only be able to recognize all objects in an image, but cannot accurately filter out the specific target "chair" based on user A's voice command. The method in this embodiment, however, can dynamically determine a reasoning pattern focused on "chair" recognition and an image recognition task network based on multimodal task input information (image and voice), thereby performing targeted inference on the predicted information. Therefore, this method can accurately respond to user A's specific task request, rather than simply providing a general recognition result. This method enables the world model to adapt to a wider range of application scenarios and multimodal tasks, improving its versatility and adaptability, and ensuring an accurate match between the final response and the user's request.

[0054] See Figure 3 In one embodiment, the specific method for determining the target reasoning pattern includes, but is not limited to, steps S301 to S304.

[0055] Step S301: Extract request instructions from the task input information to obtain request instruction information.

[0056] Step S302: Calculate the matching degree between the request instruction information and multiple inference patterns to obtain the pattern matching degree.

[0057] Step S303: Based on the pattern matching degree, determine the candidate inference patterns.

[0058] Step S304: Configure constraint parameters for the candidate reasoning patterns to obtain the target reasoning pattern.

[0059] Extracting request instructions from task input aims to identify and extract the core instructions or intentions of the user or their desired intent from raw task input information that may contain multiple modalities and complex semantics. This can be achieved through various methods. For example, Natural Language Processing (NLP) techniques can be used to segment, syntactically analyze, and recognize named entities from text information, combined with speech recognition technology to convert speech information into text, or image recognition technology can be used to perform content understanding on image information, thereby extracting key request instructions from inputs of different modalities. Alternatively, a pre-trained multimodal intent recognition model can be used, which can directly process the fused multimodal input features and output request instruction information representing the user's core intent.

[0060] After obtaining the request instruction information, it is necessary to calculate the matching degree between the request instruction information and multiple inference patterns to quantify the correlation or similarity between the request instruction information and various preset inference patterns. One implementation is to convert the request instruction information and each inference pattern into vector representations (e.g., through word embedding, sentence embedding, or more advanced semantic embedding), and then calculate the similarity between these vectors, such as cosine similarity, to obtain the pattern matching degree. Another implementation is to construct a classification model that takes the request instruction information as input and outputs the probability or confidence level of its belonging to each inference pattern; these probabilities or confidence levels are the pattern matching degree.

[0061] Based on pattern matching, candidate inference patterns can be determined to filter out those highly relevant to the current request instruction from all possible inference patterns. For example, a preset matching threshold can be set, and all inference patterns with a matching degree higher than that threshold can be selected as candidate inference patterns. Alternatively, the N inference patterns with the highest matching degree can be selected as candidate inference patterns.

[0062] Constraint parameters are configured for candidate inference patterns to personalize and refine the selected patterns based on the specific details of the request instruction information, making them more accurately suited to the current task. For example, specific entity, time, location, style, and other information can be parsed from the request instruction information and used as parameters to populate the candidate inference pattern, thereby forming a target inference pattern with specific constraints. Furthermore, rule-based systems or machine learning models can be used to automatically infer and configure the diffusion constraint parameters, denoising constraint parameters, etc., required for the inference pattern based on the semantic features and contextual information of the request instruction information.

[0063] This application's solution extracts request instructions from task input information, refining the original complex input information into core intents, thereby reducing the complexity of subsequent matching. Next, by calculating the matching degree between the request instruction information and multiple inference patterns, the applicability of each inference pattern can be quantitatively evaluated, avoiding the bias of subjective judgment. Based on this, candidate inference patterns are determined, further focusing on the most relevant set of inference patterns. Finally, by configuring constraint parameters for the candidate inference patterns, the selected inference pattern can be finely adjusted according to the specific requirements of the task input information, thus forming a highly customized and accurate target inference pattern. This step-by-step, refined processing flow ensures accurate and efficient identification of user intent from diverse task input information and precise determination of the most suitable inference pattern for the current task, providing a solid foundation for subsequent world-model-based inference processes.

[0064] The following example illustrates this. Suppose a user issues a command through a voice assistant: "Generate an image of future city transportation, in a sci-fi style, including flying cars." The executing entity first performs speech recognition and natural language understanding on the voice input, extracting request command information such as "generate image," "future city transportation," "sci-fi style," and "flying car." Then, the executing entity calculates the matching degree between these request command information and several preset inference modes, including "image generation mode," "text summarization mode," and "code generation mode." Since the command explicitly requires "generating an image," the "image generation mode" will obtain the highest mode matching degree. Based on this high matching degree, the executing entity determines the "image generation mode" as a candidate inference mode. Next, the executing entity parses the specific details in the request command information, configuring "future city transportation" as a theme parameter for image generation, "sci-fi style" as a style parameter, and "flying car" as a containment element parameter, thus obtaining a specific and constrained target inference mode to guide the subsequent image generation task.

[0065] The above technical solution effectively addresses the problem of inappropriate reasoning pattern selection caused by the diversity and complexity of task input information. Through a structured process of request instruction extraction, matching, filtering, and configuration, it ensures that the determined target reasoning pattern accurately reflects the user's intent and possesses precise constraints for specific tasks, thereby significantly improving the accuracy and efficiency of world-model-based task processing methods when handling complex tasks.

[0066] See Figure 4 In one embodiment, the specific method for determining the target task network includes, but is not limited to, steps S401 to S404.

[0067] Step S401: Extract the context from the task input information to obtain the context information.

[0068] Step S402: Merge the task input information and context information to obtain fused information.

[0069] Step S403: Calculate the matching degree between the fused information and the task network to obtain the network matching degree.

[0070] Step S404: Determine the target task network based on network matching degree.

[0071] Context extraction of task input information refers to identifying and extracting background, context, or auxiliary information relevant to the current task from the original task input information. Its purpose is to enrich the semantics of the task input information, providing a more comprehensive understanding foundation for more accurate matching of the task network. For example, natural language processing (NLP) techniques, such as named entity recognition, keyword extraction, and topic modeling, can be used to identify task-related entities, concepts, or topics from textual information as context. Furthermore, for image-based information, image recognition, scene understanding, and other techniques can be used to identify the subject, background, and environmental features in the image as contextual information.

[0072] Fusion of task input and contextual information refers to the effective combination of the original task input information and extracted contextual information to form a more representative and semantically rich information representation. Its function is to associate the direct content of the task with its background knowledge, providing a more comprehensive input for subsequent matching. Specifically, feature concatenation can be used to directly connect the feature vectors of the task input information and the feature vectors of the context information to form a longer fused feature vector. Alternatively, multimodal fusion techniques, such as gating mechanisms and fusion networks, can be used to deeply interact and fuse task input and contextual information from different modalities.

[0073] Calculating the matching degree between the fused information and the task network refers to quantifying the similarity or correlation between the fused information and the various task networks in the pre-trained world model. Its purpose is to evaluate the fit between each task network and the current task, providing a basis for selecting the most suitable target task network. For example, measures such as cosine similarity and Euclidean distance can be used to calculate the similarity between the feature vector of the fused information and the corresponding feature vector (or its representative embedding) of each task network. Alternatively, a matching network or classifier can be trained, taking the representations of the fused information and the task network as input, and outputting a score representing the degree of matching.

[0074] Determining the target task network based on network matching degree refers to selecting one or more task networks from the pre-trained world model that best meet the requirements of the current task, based on the calculated network matching degree. Its purpose is to provide a correct model foundation for subsequent prediction information generation. For example, a matching degree threshold can be set, and all task networks with matching degrees higher than this threshold can be selected as target task networks. Alternatively, the K task networks with the highest matching degrees can be selected as target task networks, where K is a preset integer.

[0075] This application's solution extracts contextual information from task input, enriching the semantic representation of the task by acquiring relevant background and contextual information from the original input. Subsequently, the original task input is fused with this contextual information to generate a more comprehensive and representative fused information. This fused information not only includes the direct content of the task but also incorporates its deeper meaning and relevant background, significantly enhancing the depth of task understanding. Based on this, by calculating the matching degree between this fused information and various task networks in the pre-trained world model, the fit between each task network and the current task can be more accurately evaluated. Finally, based on these precisely calculated network matching degrees, the executing agent can reliably determine the target task network that best meets the current task requirements. Compared to relying solely on the original task input for matching, this solution, by introducing and effectively fusing contextual information, significantly improves the accuracy and robustness of target task network selection. This provides a more solid foundation for subsequently inputting task input information into the target task network to obtain prediction information, and for inferring the prediction information based on the target inference pattern, ensuring the effectiveness and accuracy of the entire task processing flow.

[0076] The following is a concrete example. Suppose the task input is a question posed by a user via voice or text, such as "What are the features of this product?". As a specific implementation, when extracting the context of the task input, the executing entity can identify that the "product" referred to in the question is a specific product on the product page currently being viewed by the user. At this point, information such as the product's detailed description, user reviews, and historical purchase records can be extracted as contextual information. When fusing the task input and the contextual information, the user's question text can be converted into a feature vector using a text encoder, and the extracted product description and other contextual information can also be converted into feature vectors using a corresponding encoder. These two feature vectors are then concatenated or deeply interacted through a multimodal fusion module to obtain fused information containing user intent and product background knowledge. Subsequently, the matching degree between this fused information and multiple task networks (e.g., "product detail query network," "user review analysis network," "recommendation system network," etc.) stored in the pre-trained world model is calculated. For example, the cosine similarity between the fused information vector and the representative vector of each task network can be calculated. Finally, based on the calculated network matching degree, one or more task networks with the highest matching degree are selected as target task networks. For example, if the fused information has the highest matching degree with the "product details query network", then that network will be identified as the target task network for subsequent processing.

[0077] By employing the aforementioned technical solution, this application, when determining the target task network, no longer limits itself to the original task input information. Instead, it introduces contextual information and effectively fuses it, greatly enriching the semantic understanding of the task. This more comprehensive and accurate task representation enables the executing agent to more precisely calculate the matching degree between the fused information and each task network in the pre-trained world model, thereby significantly improving the accuracy and robustness of target task network selection. This effectively solves the problem that relying solely on original input information may lead to insufficient matching accuracy, ensuring that the selected task network can respond to task requests more accurately, thus improving the efficiency and effectiveness of the entire world model-based task processing method.

[0078] In some embodiments, fusing task input information and context information includes: determining the task input information as a cross-attention query vector; transforming the context information to obtain a cross-attention key vector and a cross-attention value vector; and performing cross-attention processing on the cross-attention query vector, the cross-attention key vector, and the cross-attention value vector to obtain fused information.

[0079] This application's solution achieves deep fusion of task input information and contextual information by introducing a cross-attention mechanism. Specifically, the task input information is used as a query vector, enabling it to actively search for relevant content within the context. Simultaneously, the contextual information is transformed to generate key and value vectors. The key vector is used to calculate similarity with the query vector, while the value vector carries the specific content of the contextual information. Through cross-attention processing, the query vector can dynamically assign attention weights to different parts of the contextual information based on its own content, thus selectively extracting the contextual features most relevant to the task input information. This mechanism ensures that the fused information not only includes the features of the task input information itself but also incorporates weighted and filtered contextual information, effectively capturing the complex relationship between the two and providing richer and more accurate input for subsequent calculations of network matching. This deep fusion approach allows for a more comprehensive consideration of the interaction between task input information and contextual information when determining the target task network, thereby improving the accuracy and robustness of matching.

[0080] As a specific implementation method, the above-mentioned technical means can be implemented with reference to the following example. For example, when fusing task input information and context information, the task input information (e.g., encoded text embedding vectors or image feature vectors) can be directly used as cross-attention query vectors. For context information (e.g., sequence features or user profile features extracted from historical interaction records), it can first be mapped into cross-attention key vectors and cross-attention value vectors respectively through two independent linear transformation layers (e.g., fully connected layers). Subsequently, these three vectors are input into a multi-head cross-attention module. Inside this module, the query vector and key vector are subjected to a dot product operation and then processed by scaling and a softmax function to generate attention weight matrices. These weight matrices are then weighted and summed with the value vectors to finally obtain the fused information. This approach enables the task input information to interact with context information in multiple different representation subspaces, thereby capturing richer semantic associations and generating fused information that can effectively represent the complex relationship between the two.

[0081] By employing the cross-attention mechanism to fuse task input and contextual information, the aforementioned technical solution effectively addresses the challenge of simple fusion methods failing to capture the deep correlation between the two. This mechanism enables the task input to selectively extract the most relevant features from the contextual information, generating more representative and discriminative fused information. This significantly improves the accuracy of subsequent calculations of the matching degree between the fused information and the task network, allowing the executing agent to more accurately identify the world model task network that best matches the current task request, thereby enhancing the efficiency and robustness of the entire task processing method. This deep fusion capability, especially when handling multimodal or complex contextual information, effectively avoids information loss or underutilization, ensuring the accuracy of target task network selection.

[0082] In some embodiments, reasoning about the predicted information based on the target reasoning pattern includes: denoising and spreading the predicted information based on the constraint parameters of the target reasoning pattern, so as to reconstruct the corresponding task requirement information based on the predicted information and obtain the reasoning information.

[0083] Denoising diffusion refers to the process of progressively adding noise to data and then learning an inverse process to gradually remove the noise, thereby recovering clear data from the noise. In this application, the role of denoising diffusion is to extract and generate high-quality information that meets task requirements from prediction information that may contain noise or uncertainty. As one possible implementation, a score-matching-based denoising diffusion model can be used, which guides the denoising process by learning the gradient of the data distribution and uses the constraint parameters of the target inference pattern as conditions to guide the diffusion process in a direction that meets task requirements. Another possible implementation is to use a variational inference-based denoising diffusion model, which approximates the posterior distribution by optimizing the variational lower bound and incorporates the constraint parameters of the target inference pattern into the variational objective function to achieve conditional denoising and reconstruction.

[0084] The proposed solution inputs the predicted information obtained from the target task network into a denoising diffusion process guided by the constraint parameters of the target inference model. This denoising diffusion process does not unconditionally remove noise, but rather refines and reconstructs the predicted information in a directional manner based on the specific constraints and objectives set by the target inference model. Specifically, any ambiguous, redundant, or incomplete parts of the predicted information are gradually corrected and improved through iterative processing by the denoising diffusion model, combined with the semantic or structural constraints provided by the target inference model. For example, if the target inference model requires the output of structured data in a specific format, the denoising diffusion process tends to generate information conforming to that format. Ultimately, this process transforms the predicted information into clear, accurate, and task-responsive inference information. This mechanism cleverly combines the original predicted data with the specific requirements of the task, ensuring the accuracy and usability of the inference results and effectively addressing the impact of uncertainty or noise in the predicted information on the accuracy and robustness of the task response.

[0085] The following example illustrates this. Suppose a task is to control smart home devices based on user-provided voice commands (task input information). First, the voice command is processed and used by the target task network to obtain preliminary prediction information. For example, the voice recognition result for "turn on the living room lights" might contain some ambiguous words or environmental noise, leading to the prediction information "turn on the living room lights, there might be some background noise." At this point, the target inference mode is determined to be "smart home control mode," whose constraint parameters might include "identify device type," "identify operation command," and "confirm device location." To obtain accurate inference information, the executing entity will perform denoising and diffusion on the predicted information based on these constraint parameters. For example, the denoising and diffusion model will, based on the constraint of "identify device type," clearly define "lights" as "lights," and based on the constraint of "confirm device location," define "living room" as the device location. Simultaneously, the model will remove uncertain information such as "there might be some background noise." Finally, a clear task requirement information of "turn on the living room lights" is reconstructed and output as the inference information, thus accurately responding to the user's task request.

[0086] Through the above technical solution, this application can effectively handle the noise and uncertainty in the prediction information from the target task network. By introducing constraint parameters of the target inference pattern to guide the denoising diffusion process, the inference process not only purifies the information but also ensures that the reconstructed task requirement information highly matches the actual goal of the task. This significantly improves the accuracy, reliability, and usability of the inference information, enabling the executing entity to respond more accurately and robustly to various complex task requests, overcoming the challenges that direct inference may face.

[0087] In some embodiments, the denoising diffusion of the prediction information based on the constraint parameters of the target inference mode includes: adding Gaussian distributed noise to the prediction information to obtain initial noisy information; performing forward diffusion of the initial noisy information based on the diffusion constraint parameters of the target inference mode to obtain noisy state information; and performing reverse denoising of the noisy state information based on the denoising constraint parameters of the target inference mode to obtain inference information.

[0088] Adding Gaussian noise to the prediction information refers to superimposing random noise following a Gaussian distribution onto the prediction information. By adding Gaussian noise, the structure of the original data can be gradually destroyed, eventually turning it into pure noise, providing a basis for subsequent denoising processes. This can be achieved by generating random numbers that follow a Gaussian distribution with a specific mean (e.g., 0) and variance (e.g., 1) and adding them element-wise to the prediction information. Alternatively, the built-in random number generation function in deep learning frameworks can be used, combined with Gaussian distribution parameters, to directly generate a noise tensor matching the dimension of the prediction information and then superimposing it.

[0089] Based on the diffusion constraint parameters of the target inference model, forward diffusion of initial noisy information refers to gradually adding noise to the data through a series of predefined steps until the data becomes completely random noise. The diffusion constraint parameters define the noise addition rate, step size, and noise type, ensuring that the diffusion process is controllable and reversible. Its role is to gradually transform the original information into noisy states of varying degrees, providing training data or an inference starting point for reverse denoising. This can be achieved by iteratively adding a small amount of Gaussian noise to the current noisy information and adjusting the amount of noise added at each step according to the diffusion constraint parameters (such as a noise scheduling table) until a preset number of diffusion steps or noise level is reached. Alternatively, a forward diffuser in a pre-trained diffusion model can be used. This diffuser automatically performs a multi-step noise addition process based on the diffusion constraint parameters (e.g., a preset noise variance sequence), gradually transforming the initial noisy information into a more highly noisy state.

[0090] Denoising constraint parameters based on target inference patterns, used for inverse denoising of noisy state information, refer to the process of learning to remove noise from noisy data and gradually recovering the original data. Denoising constraint parameters guide the behavior of the denoising model, such as the denoising step size, number of iterations, and weights of the denoising model, ensuring that the denoising process can effectively recover meaningful inference information from noisy state information. Its role is to gradually restore clear and accurate task requirement information from highly noisy information. This can be achieved using a pre-trained denoising network (such as a U-Net structure), which takes noisy state information and the current time step as input, predicts the noise at the current time step, or directly predicts the denoised information. Denoising constraint parameters can include the model weights, learning rate, and optimizer configuration of the denoising network. Alternatively, it can be achieved by iteratively applying a denoising function that subtracts the estimated noise from the noisy state information based on the denoising constraint parameters (e.g., a preset denoising step size and denoising intensity factor), gradually recovering the original information.

[0091] This application's solution first adds Gaussian noise to the predicted information, transforming it into initial noisy information, thus simulating the degradation process of information in a complex environment. Subsequently, based on the diffusion constraint parameters provided by the target inference model, this initial noisy information is forward-divered in a controlled manner, gradually evolving into noisy state information of varying degrees. This process ensures that the introduction of noise is systematic and reversible, providing multi-scale noise samples for subsequent denoising operations. Next, using the preset denoising constraint parameters in the target inference model, a reverse denoising operation is performed on this noisy state information. The reverse denoising process, by gradually removing noise, accurately recovers the original, clear inference information from the highly noisy information, i.e., the reconstructed task requirement information. Through this constrained denoising diffusion mechanism, this application can effectively extract and reconstruct high-quality task requirement information that meets the task requirements from predicted information that may contain uncertainty or noise, thereby solving the problem of how to accurately control the denoising diffusion process to effectively extract and reconstruct task requirement information.

[0092] As a specific implementation, assuming the prediction information is an image tensor, a noise tensor with the same size as the image tensor and following a Gaussian distribution with a mean of 0 and a standard deviation of 0.1 can be generated. This noise tensor is then superimposed pixel-by-pixel with the prediction information tensor to obtain an initial noisy image. Subsequently, the diffusion constraint parameters of the target inference mode can define a noise schedule, such as a linearly increasing noise variance sequence from time step t=0 to t=1000. During the forward diffusion process, according to this schedule, Gaussian noise with the corresponding variance can be added to the current noisy image at each time step t, iterating until a preset final time step is reached, resulting in highly noisy image state information. Next, the denoising constraint parameters of the target inference mode can be a pre-trained U-Net denoising model. This model takes the noisy state information (e.g., the noisy image at a certain time step t) and the current time step t as input, and predicts the noise that should be removed at the current time step. By subtracting the predicted noise from the noisy image and iteratively performing this process, gradually backtracking from t=1000 to t=0, the denoised image, i.e., the inference information, is finally obtained.

[0093] Through the above technical solution, this application provides a structured and controllable denoising diffusion mechanism. By clearly defining the steps of noise addition, forward diffusion, and reverse denoising, and using the diffusion constraint parameters and denoising constraint parameters of the target inference pattern for precise guidance, high-quality task requirement information can be effectively reconstructed step by step and accurately from noisy or uncertain prediction information. This significantly improves the robustness and accuracy of task processing, ensuring that the executing entity can stably output reliable inference results in complex and ever-changing task scenarios.

[0094] In some embodiments, before determining the target inference pattern and several target task networks based on the task input information, the method further includes: preprocessing the task input information to obtain preprocessed task input information; and encoding the preprocessed task input information to obtain encoded task input information.

[0095] Preprocessing task input aims to improve its quality, making it more suitable for subsequent encoding and model processing. Preprocessing can include noise removal, data format standardization, and missing value imputation. For example, for image-based information, preprocessing may include image denoising (such as using Gaussian filtering or median filtering), size normalization, color space conversion (such as RGB to grayscale), and cropping. For speech-based information, preprocessing may include noise reduction, silence removal, speech enhancement, and sampling rate conversion. For text-based information, preprocessing may include stop word removal, lemmatization, word segmentation, case conversion, and special character cleanup. Furthermore, preprocessing can be customized according to the task type. For example, for structured data in a specific domain, outlier detection and correction, data type conversion, and feature scaling (such as Min-Max normalization or Z-score standardization) can be performed to ensure the consistency and validity of the input data.

[0096] Encoding the preprocessed task input information involves converting it into a numerical representation that the model can understand, typically a high-dimensional vector. This allows different types of information (images, speech, text, etc.) to be processed and fused within a unified feature space. For example, convolutional neural networks (CNNs) can be used to extract feature vectors for image-based information; recurrent neural networks (RNNs), Transformer encoders, or acoustic models (such as Wav2Vec 2.0) can be used to extract speech embeddings for speech-based information; and pre-trained language models such as word embeddings (such as Word2Vec, GloVe), BERT, and GPT can be used to generate text embeddings. The encoding process can also be selected based on the complexity of the task and the specific architecture of the world model. For example, for multimodal tasks, a multimodal fusion encoder can be used to map feature vectors from different modalities into a shared embedding space. The encoded information is usually a fixed-length vector, facilitating subsequent matching degree calculations and inference.

[0097] The solution of this application significantly improves the robustness and accuracy of the entire task processing method by first preprocessing and encoding the task input information before determining the target inference mode and several target task networks. Specifically, the preprocessing step can effectively remove noise and redundancy in the original task input information and unify the data format, ensuring the purity and consistency of the input data. On this basis, the encoding step converts the preprocessed task input information into a high-dimensional and semantically rich vector representation, enabling information of different modalities to be efficiently and meaningfully represented in a unified feature space. This encoded task input information after preprocessing and encoding has higher quality and stronger expressive power compared to the original, unprocessed task input information. When subsequent steps such as request instruction extraction, context extraction, and calculation of the matching degree between the fused information and the task network use this high-quality encoded task input information, the true intention and task requirements of the user can be identified more accurately and efficiently. For example, when determining the target inference mode, request instruction extraction is performed on the encoded task input information, and the request instruction can be more precisely identified, thereby calculating a more accurate mode matching degree and finally determining a target inference mode that better meets the actual needs. Similarly, when determining the target task network, context extraction is performed on the encoded task input information and fused with the encoded task input information, forming a more representative fused information, thereby calculating a more accurate network matching degree and finally selecting the task network in the world model that is most suitable for the current task. This synergy avoids misjudgment or inefficiency caused by problems with the quality of the original data, enabling the entire task processing flow to run more stably and precisely.

[0098] The following is illustrated by a specific example. Suppose the task input information includes a picture and a text describing the content of the picture. First, preprocess the picture, for example, scale its size to 224x224 pixels, perform Gaussian blur denoising, and then perform color normalization. At the same time, preprocess the text, for example, use a word segmentation tool (such as the jieba segmentation library) for word segmentation, remove stop words (such as "de", "shi"), and convert all letters to lowercase. Then, encode the preprocessed picture. A pre-trained ResNet-50 model can be used as the encoder. Input the picture and extract a 2048-dimensional feature vector. Encode the preprocessed text. A pre-trained BERT model can be used as the encoder. Input the text and extract a 768-dimensional feature vector. Subsequently, these two feature vectors can be fused through a linear layer or an attention mechanism to obtain a unified encoded task input information vector, for example, 1024-dimensional. This encoded task input information vector is then used for subsequent operations such as request instruction extraction and context extraction to determine the target inference mode and target task network.

[0099] By preprocessing the task input information using the above technical solutions, noise and redundant information can be effectively removed, and the data format can be standardized, significantly improving the quality and consistency of the input data. Encoding the preprocessed task input information transforms it into a high-dimensional, semantically rich vector representation, enabling efficient processing of information from different modalities within a unified feature space. This high-quality, standardized encoded task input information makes the subsequent determination of the target inference pattern and target task network more accurate and robust. For example, in request instruction extraction and context extraction, user intent and relevant background information can be more clearly identified, thereby improving the accuracy of pattern matching and network matching. Ultimately, this ensures that the determined target inference pattern and target task network can respond to task requests more accurately, improving the performance and reliability of the entire task processing method.

[0100] See Figure 5 This application also provides a task processing apparatus based on a world model, which can implement the above-described task processing method based on a world model. The apparatus includes: The first module 501 is used to acquire task input information; the task input information includes image information, voice information, text information, structured information and / or action information; The second module 502 is used to determine the target inference pattern and several target task networks based on the task input information; the target inference pattern is a constraint process with the task input information as a condition, and the target task network is a task network in the pre-trained world model that meets the matching conditions. The third module 503 is used to input task input information into the target task network to obtain prediction information; The fourth module 504 is used to reason about the predicted information based on the target reasoning pattern to obtain reasoning information; Module 505 is used to respond to task requests corresponding to task input information based on reasoning information.

[0101] The specific implementation of this world model-based task processing device is basically the same as the specific implementation of the world model-based task processing method described above, and will not be repeated here.

[0102] Figure 6 This is a schematic diagram of the hardware structure of the electronic device provided in the embodiments of this application.

[0103] The following reference Figure 6 To describe an electronic device 600 according to such an embodiment of the present disclosure. Figure 6 The electronic device 600 shown is merely an example and should not impose any limitation on the functionality and scope of use of the embodiments disclosed herein.

[0104] like Figure 6 As shown, the electronic device 600 is presented in the form of a general-purpose computing device. The components of the electronic device 600 may include, but are not limited to: at least one processing unit 610, at least one storage unit 620, a bus 630 connecting different system components (including storage unit 620 and processing unit 610), a display unit 640, etc.

[0105] The storage unit stores program code, which can be executed by the processing unit 610, causing the processing unit 610 to perform the steps described in the above-described section of the task processing method based on the world model according to various exemplary embodiments of this disclosure.

[0106] Storage unit 620 may include a readable medium in the form of a volatile storage unit, such as random access memory (RAM) 6201 and / or cache memory 6202, and may further include a read-only memory (ROM) 6203.

[0107] Storage unit 620 may also include a program / utility 6204 having a set (at least one) program module 6205, such program module 6205 including but not limited to: operating system, one or more application programs, other program modules and program data, each or some combination of these examples may include an implementation of a network environment.

[0108] Bus 630 can represent one or more of several types of bus structures, including a memory cell bus or memory cell controller, a peripheral bus, a graphics acceleration port, a processing unit, or a local bus using any of the various bus structures.

[0109] Electronic device 600 can also communicate with one or more external devices 600' (e.g., keyboard, pointing device, Bluetooth device, etc.), and with one or more devices that enable a user to interact with electronic device 600, and / or with any device that enables electronic device 600 to communicate with one or more other computing devices (e.g., router, modem, etc.). This communication can be performed via input / output (I / O) interface 650. Furthermore, electronic device 600 can also communicate with one or more networks (e.g., local area network (LAN), wide area network (WAN), and / or public networks, such as the Internet) via network adapter 660. Network adapter 660 can communicate with other modules of electronic device 600 via bus 630. It should be understood that, although not shown in the figures, other hardware and / or software modules can be used in conjunction with electronic device 600, including but not limited to: microcode, device drivers, redundant processing units, external disk drive arrays, RAID systems, tape drives, and data backup storage systems.

[0110] This application also provides a computer-readable storage medium storing a computer program that, when executed by a processor, implements the above-described method.

[0111] The task processing method, apparatus, device, and medium based on a world model provided in this application embodiment pre-integrates a world model with multiple task networks. By acquiring multimodal task input information and dynamically determining the target inference mode and target task network, the target task network generates corresponding prediction information based on the task input information. Then, based on the target inference mode, the prediction information is inferred to obtain inference information. Finally, the task request corresponding to the task input information is responded to based on the inference information. Therefore, by pre-integrating a world model with multiple task networks and dynamically determining the target inference mode and target task network based on multimodal task input information, the appropriate task network in the world model can be invoked and the appropriate inference mode determined for the task input information. This enables targeted prediction and inference of multimodal task input information, flexibly responding to different task requirements, ensuring the universality and diversity of the world model's capabilities to adapt to various task scenarios.

[0112] From the above description of the embodiments, those skilled in the art will readily understand that the exemplary embodiments described herein can be implemented by software or by combining software with necessary hardware. Therefore, the technical solutions according to the embodiments of this disclosure can be embodied in the form of a software product, which can be stored in a non-volatile storage medium (such as a CD-ROM, USB flash drive, external hard drive, etc.) or on a network, including several instructions to cause a computing device (such as a personal computer, server, or network device, etc.) to execute the methods described above according to the embodiments of this disclosure.

[0113] The program product may employ any combination of one or more readable media. A readable medium may be a readable signal medium or a readable storage medium. A readable storage medium may be, for example, but not limited to, an electrical, magnetic, optical, electromagnetic, infrared, or semiconductor system, apparatus, or device, or any combination thereof. More specific examples (a non-exhaustive list) of readable storage media include: electrical connections having one or more wires, portable disks, hard disks, random access memory (RAM), read-only memory (ROM), erasable programmable read-only memory (EPROM or flash memory), optical fiber, portable compact disk read-only memory (CD-ROM), optical storage devices, magnetic storage devices, or any suitable combination thereof.

[0114] Computer-readable storage media may include data signals propagated in baseband or as part of a carrier wave, carrying readable program code. Such propagated data signals may take various forms, including but not limited to electromagnetic signals, optical signals, or any suitable combination thereof. A readable storage medium may also be any readable medium other than a readable storage medium that can transmit, propagate, or transfer a program for use by or in connection with an instruction execution system, apparatus, or device. The program code contained on the readable storage medium may be transmitted using any suitable medium, including but not limited to wireless, wired, optical fiber, RF, etc., or any suitable combination thereof.

[0115] Those skilled in the art will understand that the above modules can be distributed in the device as described in the embodiments, or they can be modified accordingly and placed in one or more devices that are unique to this embodiment. The modules in the above embodiments can be combined into one module, or they can be further divided into multiple sub-modules.

[0116] Exemplary embodiments of this disclosure have been specifically shown and described above. It should be understood that this disclosure is not limited to the detailed structures, arrangements, or implementations described herein; rather, this disclosure is intended to cover various modifications and equivalent arrangements contained within the spirit and scope of the appended claims.

Claims

1. A task processing method based on a world model, characterized in that, include: Obtain task input information; The task input information includes image information, voice information, text information, structured information, and / or action information; Based on the task input information, a target reasoning pattern and several target task networks are determined; the target reasoning pattern is a constraint process with the task input information as a condition, and the target task network is a task network in a pre-trained world model that meets the matching conditions. The task input information is input into the target task network to obtain prediction information; Based on the target reasoning pattern, the predicted information is reasoned to obtain reasoning information; Based on the reasoning information, respond to the task request corresponding to the task input information.

2. The task processing method based on a world model according to claim 1, characterized in that, The method for determining the target reasoning pattern includes: The task input information is processed to extract request instructions, resulting in request instruction information. Calculate the matching degree between the request instruction information and multiple inference patterns to obtain the pattern matching degree; Based on the pattern matching degree, candidate reasoning patterns are determined; The candidate reasoning patterns are configured with constraint parameters to obtain the target reasoning pattern.

3. The task processing method based on a world model according to claim 1, characterized in that, The method for determining the target task network includes: Context extraction is performed on the task input information to obtain context information; The task input information and the context information are combined to obtain fused information; Calculate the matching degree between the fused information and the task network to obtain the network matching degree; The target task network is determined based on the network matching degree.

4. The task processing method based on a world model according to claim 3, characterized in that, The fusion of the task input information and the context information includes: The task input information is determined as a cross-attention query vector; The context information is transformed to obtain the cross-attention key vector and the cross-attention value vector; The cross-attention query vector, the cross-attention key vector, and the cross-attention value vector are subjected to cross-attention processing to obtain the fused information.

5. The task processing method based on a world model according to claim 1, characterized in that, The reasoning based on the target reasoning pattern on the predicted information includes: Based on the constraint parameters of the target reasoning mode, the prediction information is denoised and diffused to reconstruct the corresponding task requirement information and obtain the reasoning information.

6. The task processing method based on a world model according to claim 5, characterized in that, The denoising and diffusion of the predicted information based on the constraint parameters of the target inference pattern includes: Gaussian noise is added to the predicted information to obtain initial noisy information; Based on the diffusion constraint parameters of the target reasoning mode, the initial noisy information is forward diffused to obtain noisy state information; Based on the denoising constraint parameters of the target reasoning mode, the noisy state information is denoised in reverse to obtain the reasoning information.

7. The task processing method based on a world model according to any one of claims 1 to 6, characterized in that, Before determining the target inference pattern and several target task networks based on the task input information, the method further includes: The task input information is preprocessed to obtain preprocessed task input information; The preprocessed task input information is encoded to obtain encoded task input information.

8. A task processing device based on a world model, characterized in that, include: The first module is used to obtain task input information; The task input information includes image information, voice information, text information, structured information, and / or action information; The second module is used to determine the target reasoning pattern and several target task networks based on the task input information. The target reasoning mode is a constraint process based on the task input information, and the target task network is a task network in a pre-trained world model that meets the matching conditions. The third module is used to input the task input information into the target task network to obtain prediction information; The fourth module is used to reason about the predicted information based on the target reasoning pattern to obtain reasoning information; The fifth module is used to respond to the task request corresponding to the task input information based on the reasoning information.

9. An electronic device, characterized in that, The electronic device includes a memory and a processor, the memory storing a computer program, and the processor executing the computer program to implement the task processing method based on the world model as described in any one of claims 1 to 7.

10. A computer-readable storage medium storing a computer program, characterized in that, When the computer program is executed by the processor, it implements the task processing method based on the world model as described in any one of claims 1 to 7.

Images