Content arrangement and multi-dimensional closed-loop correction method, system, device and storage medium

Abstract

The application discloses a content arrangement and multi-dimensional closed-loop correction method, system, device and storage medium, which are corresponding solutions, in the solutions: the best path can be selected according to the user intention, for example, the structure is guaranteed to be correct by using the code when a webpage / PPTX is generated, and the visual impact force is guaranteed by using a model when a poster is generated, so that the optimal solution of the scene is realized; and, different from an existing single generation solution, the application introduces a feedback loop mechanism, through continuous self-auditing and modification, the generation result can be continuously optimized, and through testing, the result output by the application is superior to the existing solution in three dimensions of data accuracy, text correctness and aesthetic degree.

Description

Technical Field

[0001] This invention relates to the field of automated content generation technology, and in particular to a content arrangement and multi-dimensional closed-loop correction method, system, device and storage medium. Background Technology

[0002] With the increasing demand for data analytics, the challenge of quickly transforming user intent or structured data into high-quality visual products (such as posters, data charts, web interfaces, and presentations) has become a major industry pain point. Existing automation solutions typically involve two distinct technical approaches, both with significant drawbacks: 1. Rule-based / Code-based Generation: This approach utilizes LLM (Large Language Model) to generate domain-specific code, such as Python (a computer programming language) drawing code, HTML (Hypertext Markup Language) / CSS (Cascading Style Sheets) front-end code, and LaTeX (a typesetting system) / PPTX (Presentation Script) typesetting code, which is then used to render the corresponding content. However, it suffers from the following drawbacks: rigid aesthetics and a lack of creativity. For example, the generated web pages are not visually appealing, and the generated PPTX lacks illustrations and artistic flair, failing to meet the high aesthetic demands of commercial advertising. Furthermore, once the code encounters errors or layout anomalies (such as text obscuring the text), the system often lacks automatic "visual debugging" capabilities. Debugging encompasses the entire process of discovering and correcting errors.

[0003] 2. Generation-based approach based on visual models: This relies on text-based graph models, such as Diffusion. However, it suffers from the following drawbacks: poor logical controllability and a high susceptibility to visual illusions. Examples include spelling errors on posters, unclickable buttons on web pages, discrepancies between numerical values ​​and bar heights in data charts, and chaotic text layout in PowerPoint presentations. Most importantly, current models are typically open-loop, meaning they cannot automatically correct erroneous results.

[0004] Therefore, there is an urgent need for a unified intelligent orchestration solution that can automatically select the best execution tool based on the task type (focusing on logical structure or visual creativity) and establish a closed-loop feedback mechanism of "observation-reflection-action". This allows the machine (automatic execution tool / software) to iteratively optimize the visual results by continuously issuing new correction instructions, just like a human designer, until it meets multi-dimensional quality standards.

[0005] In view of this, the present invention is hereby proposed. Summary of the Invention

[0006] The purpose of this invention is to provide a content arrangement and multi-dimensional closed-loop correction method, system, device and storage medium that can correct the generated results through a closed-loop feedback mechanism and improve the quality of the generated results.

[0007] The objective of this invention is achieved through the following technical solution: A content arrangement and multi-dimensional closed-loop correction method includes: Step 1: Analyze the user's intent based on the user's input information, and select the logical construction path and / or visual perception path according to the user's intent; Step 2: Transform user input information into intermediate representations, that is: for logical construction paths, transform user input information into executable code; for visual perception paths, transform user input information into visual prompts. Step 3: Perform the content generation task based on the intermediate representation, output the generation result, and combine the generation result with the intermediate representation as the normalized result; Step 4: Perform multi-dimensional auditing on the normalization result to determine if there are any errors in the generated result. If so, output a structured error report; otherwise, use the normalization result as the final output. Step 5: Correct the intermediate representation based on the error report and output the corrected generation result, then go back to step 4; repeat steps 4 and 5 until the corrected generation result is error-free or the set number of iterations is reached.

[0008] A content orchestration and multi-dimensional closed-loop correction system includes: an orchestrator, a heterogeneous executor, and an audit layer; The orchestrator is used to execute steps 1 and 2; the heterogeneous executor is used to execute step 3; and the orchestrator and the heterogeneous executor cooperate to execute step 5; the audit layer is used to execute step 4, wherein: Step 1 includes: analyzing user intent based on user input information, and selecting logical construction path and / or visual perception path based on user intent; Step 2 includes: transforming user input information into intermediate representations: for logical construction paths, user input information is transformed into executable code; for visual perception paths, user input information is transformed into visual prompts. Step 3 includes: performing a content generation task based on the intermediate representation, outputting the generation result, and combining the generation result with the intermediate representation as a normalized result; Step 4 includes: performing multi-dimensional audits on the normalization results to determine whether there are any errors in the generated results. If so, a structured error report is output; otherwise, the normalization results are used as the final output. Step 5 includes: the orchestrator corrects the intermediate representation based on the error report, the heterogeneous actuator outputs the corrected generation result based on the corrected intermediate representation, and then proceeds to step 4; Steps 4 and 5 are repeatedly executed through the orchestrator, audit layer, and heterogeneous executor until the corrected generated result is free of errors or the set number of iterations is reached.

[0009] A processing device includes: one or more processors; and a memory for storing one or more programs; When the one or more programs are executed by the one or more processors, the one or more processors implement the aforementioned method.

[0010] A readable storage medium storing a computer program that, when executed by a processor, implements the aforementioned method.

[0011] As can be seen from the technical solution provided by the present invention, the optimal path can be selected based on the user's intent. For example, when generating a webpage / PPTX (corresponding to the logical construction path), code is used to ensure structural correctness, and when generating a poster (corresponding to the visual perception path), a generation model is used to ensure visual impact, thus achieving the optimal solution for the scene. Furthermore, unlike existing single-generation schemes, the present invention introduces a feedback loop, which can continuously optimize the generation results through continuous self-auditing and modification. After testing, the output results of the present invention are superior to existing schemes in terms of data accuracy, text correctness, and aesthetics. Attached Figure Description

[0012] To more clearly illustrate the technical solutions of the embodiments of the present invention, the drawings used in the following description of the embodiments will be briefly introduced. Obviously, the drawings described below are only some embodiments of the present invention. For those skilled in the art, other drawings can be obtained based on these drawings without creative effort.

[0013] Figure 1 This is a flowchart of a content arrangement and multi-dimensional closed-loop correction method provided in an embodiment of the present invention.

[0014] Figure 2 This is a schematic diagram of a content arrangement and multi-dimensional closed-loop correction system provided in an embodiment of the present invention.

[0015] Figure 3 This is a schematic diagram of a processing device provided in an embodiment of the present invention. Detailed Implementation

[0016] The technical solutions of the embodiments of the present invention will be clearly and completely described below with reference to the accompanying drawings. Obviously, the described embodiments are only some embodiments of the present invention, and not all embodiments. Based on the embodiments of the present invention, all other embodiments obtained by those skilled in the art without creative effort are within the protection scope of the present invention.

[0017] First, the following explanations are provided for the terms that may be used in this article: The term "and / or" means that either or both can be achieved simultaneously. For example, X and / or Y means that it includes both "X" or "Y" as well as the three cases of "X and Y".

[0018] The terms "comprising," "including," "containing," "having," or other similar semantic descriptions should be interpreted as non-exclusive inclusion. For example, including a technical feature element (such as raw material, component, ingredient, carrier, dosage form, material, size, part, component, mechanism, device, step, process, method, reaction conditions, processing conditions, parameter, algorithm, signal, data, product or article of manufacture, etc.) should be interpreted as including not only the expressly listed technical feature element, but also other technical feature elements that are not expressly listed and are well-known in the art.

[0019] The term "composed of" excludes any technical features not expressly listed. When used in a claim, it closes the claim to exclude all technical features other than those expressly listed, except for associated conventional impurities. If the term appears only in a clause of a claim, it limits the claim to the elements expressly listed in that clause; elements recited in other clauses are not excluded from the overall claim.

[0020] The following provides a detailed description of a content arrangement and multi-dimensional closed-loop correction method, system, device, and storage medium provided by this invention. Contents not described in detail in the embodiments of this invention are prior art known to those skilled in the art. Where specific conditions are not specified in the embodiments of this invention, they are performed according to conventional conditions in the art or conditions recommended by the manufacturer. Where the manufacturers of reagents or instruments used in the embodiments of this invention are not specified, they are all conventional products that can be purchased commercially.

[0021] Example 1 This invention provides a content arrangement and multi-dimensional closed-loop correction method, such as... Figure 1 As shown, it mainly includes the following steps: Step 1: Intent Understanding and Two-Path Routing.

[0022] In this embodiment of the invention, the user's intent is analyzed based on the user's input information, and a logical construction path and / or a visual perception path is selected based on the user's intent.

[0023] Step 2: Transform user input information into an intermediate representation.

[0024] In this embodiment of the invention, for logical construction paths, user input information is converted into executable code, and for visual perception paths, user input information is converted into visual prompts.

[0025] Step 3: Execute the generation task and output the normalized result.

[0026] In this embodiment of the invention, a content generation task is performed based on the intermediate representation, the generation result is output, and the generation result is combined with the intermediate representation as a normalized result.

[0027] Specifically: for logical construction paths, the code interpreter renders visual structural data based on the executable code; for visual perception paths, the generative model generates image data based on visual cues.

[0028] In each path, the generated results and corresponding intermediate representations are combined into normalized results (e.g., visual structural data and executable code, image data and visual cues) to facilitate subsequent auditing and correction.

[0029] Step 4: Multidimensional audit.

[0030] In this embodiment of the invention, the normalization result is audited in multiple dimensions to determine whether there are any errors in the generated result. If so, a structured error report is output; otherwise, the normalization result is used as the final output.

[0031] In this embodiment of the invention, a multi-modal large language model is used to audit the generated results in the normalized results from multiple dimensions, including: auditing three dimensions: data consistency, text accuracy, and layout logic correctness. Specifically, data consistency auditing checks whether the visual elements in the generated results match the data content; text accuracy auditing checks whether the text in the generated results is correct; and layout logic correctness auditing checks for label overlap and / or occlusion. If any dimension fails the audit, the generated results are deemed to have errors; if all three dimensions pass the audit, the generated results are deemed to have no errors.

[0032] Step 5: Feedback and correction loop.

[0033] In this embodiment of the invention, the intermediate representation is corrected according to the error report, and the corrected generation result is output, and then the process proceeds to step 4; steps 4 and 5 are repeated until the corrected generation result is free of errors or the set number of iterations is reached.

[0034] In this embodiment of the invention, for the logical construction path, the corresponding content of the executable code is corrected according to the error report, and the corrected generation result is generated (by calling the code interpreter to re-render); for the visual perception path, the visual prompts are corrected according to the error situation in the error report, a local redraw instruction is generated, the generation result is locally redrawn, and the corrected generation result is output.

[0035] The above-mentioned solution provided by the embodiments of the present invention can select the best path according to the user's intent. For example, when generating a webpage / PPTX (corresponding to the logical construction path), code is used to ensure the correct structure, and when generating a poster (corresponding to the visual perception path), a generation model is used to ensure visual impact, thus achieving the optimal solution for the scene. Furthermore, unlike the existing single generation schemes, the present invention introduces a feedback loop, which can continuously optimize the generation results and output high-quality generation results through continuous self-auditing and modification.

[0036] The above-described solution provided in this embodiment of the invention can be loaded onto a computer or server for multimodal structured data orchestration. Based on user input and instructions, iterative orchestration and multidimensional closed-loop correction are performed to generate high-quality results.

[0037] To more clearly demonstrate the technical solution and its effects provided by the present invention, the method provided by the embodiments of the present invention will be described in detail below with reference to specific examples.

[0038] I. Overall Overview of the Plan

[0039] The method provided in this invention is an iterative multimodal structured content orchestration and multidimensional closed-loop correction method based on a multimodal large language model (MLLM). The entire orchestration task is modeled as a multi-round "instruction-execution-verification" loop, with the main process as follows: (1) Intelligent routing: Based on the user's intent, determine the task requirements (poster / chart / webpage / PPTX), automatically decide the execution path, and then call the corresponding code interpreter or generation model (e.g., diffusion model) to output normalized results.

[0040] (2) Unified auditing: For each execution path, the "visual object" is audited based on the multimodal large language model. The visual object here refers to the aforementioned generated result, which includes visual structure data and image data.

[0041] (3) Closed-loop correction: Based on audit feedback, automatically plan correction strategies (modify code parameters or issue editing instructions) to achieve fully automatic quality iteration.

[0042] II. Detailed introduction of the plan.

[0043] 1. Intent Analysis & Dual-path Routing.

[0044] First, the program receives structured data or text from the user (such as tables, copy, or requirements documents) and the generated intent. The orchestrator analyzes the intent and selects the execution path. Path A (Logical Construction Path, also known as Code Path): Suitable for scenarios with extremely high requirements for structural accuracy and interactivity.

[0045] Typical scenarios: Web page front-end generation (HTML / CSS), single-page PTX layout, precise statistical charts, etc.

[0046] Path B (Visual Perception Path, also known as Generation Path): Suitable for scenarios with extremely high requirements for atmosphere, creativity, and lighting details.

[0047] Typical scenarios: creative promotional posters, artistic data visualization, background image generation, etc.

[0048] 2. Heterogeneous Instruction Construction.

[0049] The orchestrator transforms user input into a machine-executable intermediate representation based on the selected path.

[0050] If path A is selected, the orchestrator generates an executable code script. For example, it might generate HTML code with Flexbox and Tailwind CSS styles, or a script that calls the Python-PPTX library. TailwindCSS: Atomic Cascading Style Sheets / Atomic CSS. Flexbox: Flexible layout that allows child elements to be arranged in rows or columns.

[0051] If path B is selected, the orchestrator performs Rationale reasoning, transforming abstract requirements into detailed visual prompts, which contain layout constraints.

[0052] 3. Heterogeneous execution and result normalization.

[0053] This part is accomplished through a heterogeneous executor, which includes a code interpreter and a generative model.

[0054] For path A: Run the code script through the code interpreter to render a webpage screenshot, PPTX preview image, or vector graphic.

[0055] For path B: Generate a bitmap image based on the prompt words using a generative model (e.g., a diffusion model).

[0056] Normalization: Regardless of the path, the final output is a combination of "visual object (Image)" and "source file (Code / Prompt)". The source file here is the input information of the heterogeneous executor.

[0057] 4. Unified Multi-modal Auditing.

[0058] The audit layer acts as an all-around quality inspector, executing inspection protocols for specific scenarios, such as checking the following dimensions: 1) Data consistency: whether visual elements match the data. 2) Text accuracy: whether there are garbled characters or spelling errors. 3) Layout logic: whether there are overlapping or occlusion of labels (for path A), etc.

[0059] Taking layout logic as an example, for web pages / PPTX: check for layout overflow, text obscuration, and color contrast. For charts / posters: check for accurate data points (illusion detection), correct text spelling (OCR verification), and style matching the description.

[0060] Output: A structured error report (JSON) indicating the specific location, type, and severity of the error.

[0061] In this embodiment of the invention, both the orchestrator and the audit layer can be implemented based on a multimodal large language model.

[0062] 5. Heterogeneous Refinement Loop.

[0063] This step is the core of the invention. When the audit fails, the orchestrator automatically performs corrective operations via heterogeneous actuators based on the error report until the audit is passed. For path A, code debugging is performed: If the error originates from the code, for example, a misaligned div (distinguishing mark) on the webpage, which is a layout logic error, the editor reads the error report and then regenerates the corrected code snippet through the code interpreter.

[0064] For path B, visual refinement is performed: if the error originates from the generated image (such as a misspelled character on a poster), the editor generates an in-painting instruction, calling the image editing model to repair the specific area. The image editing model here is a heterogeneous executor; it can be the aforementioned generation model or other image generation models.

[0065] The following example uses chart drawing, which involves the processing of two types of paths, to introduce the process of steps 4 and 5 above.

[0066] (1) Correction Strategy Planning (Reasoning): The orchestrator reads the error report and analyzes the cause of the error. For example: The error report shows that "the height of bar A is visually lower than that of B, but numerically A>B".

[0067] (2) Correction Instruction Issuance: The editor generates targeted tool call instructions based on the current path type: Case 1 (Currently in the code path): The editor generates code repair instructions. Instruction example: "Modify matplotlib parameters to expand the y-axis range and rotate the x-axis labels to resolve overlap." Action: Call the code interpreter to rerun the new code. Case 2 (Currently in the generation path): The editor generates visual editing instructions. Instruction example: "Keep the background unchanged, select the red bar representing 'A' in the image, and increase its height to the 600 mark of the guide line." Action: Call the image editing model to locally redraw the image.

[0068] (3) Execution and state update: The heterogeneous executor produces the corrected visual object based on the new instructions.

[0069] (4) Loop check: Audit the corrected visual object again. If the error is resolved, output it; otherwise, continue to the next round of correction until the maximum number of iterations is reached; if the audit is still not passed after the maximum number of iterations is reached, the result of the last correction is used as the output result.

[0070] III. Example of the solution.

[0071] Example 1: Task: Draw a cyberpunk-style pie chart about coffee bean production.

[0072] Routing: The requirement is a "cyberpunk style" with a focus on visuals, so path B is selected.

[0073] Execution: The orchestrator generates descriptive terms, and the Diffusion model generates images.

[0074] Audit and Correction: The audit layer found that the sector ratio of "Brazil" was too small → The orchestrator issued an editing instruction "Expand the angle of the Brazil sector" → The image editing model redrew the image.

[0075] Example 2: Task: Generate a Spring Festival promotion poster with the text "50% off全场".

[0076] Routing: Focusing on the appeal of the画面, select path B.

[0077] Execution: The orchestrator designs the画面(red background, lanterns) → Diffusion generates an image.

[0078] Audit and Correction: The audit layer found that there was an extra stroke in the character "折" (text artifact) → The orchestrator issued an editing instruction "Repair the text area" → The image editing model redrew the image.

[0079] Example 3: Task: Write a responsive landing page that includes a Hero Section and a price list.

[0080] Routing: Focusing on the layout structure and interaction, select path A.

[0081] Execution: The orchestrator generates HTML / Tailwind code → The browser (code interpreter) renders a screenshot.

[0082] Audit and Correction: The audit layer compared the design specifications and found that the layout of the price list was incorrect in the mobile mode → Reported the layout error → The orchestrator modified the CSS code (added flex-wrap) → Rendered again.

[0083] In this example: The Hero Section refers to the core display area on the first screen of the home page, usually including the main title, subtitle, main button, and main visual, which is used to convey the product value immediately; flex-wrap is an automatic line break property in CSS flexible layout, which is used to make elements move to the next line when there are too many elements in one line.

[0084] Example 4: Task: Generate a one-page summary PPTX based on this financial report text, with three key data cards.

[0085] Routing: Focusing on the information hierarchy and typesetting, select path A (such as Python-PPTX or LaTeX Beamer); LaTeXBeamer is a presentation typesetting tool based on LaTeX (a professional typesetting system), which is often used to generate academic or business slides with standardized structures and unified styles.

[0086] Execution: The LLM generates typesetting code → Renders the PPTX page.

[0087] Audit and Correction: MLLM check found "the text on the third card is out of bounds" → reported an Overflow error -> LLM adjusted the font size or text box size in the code → re-rendered.

[0088] The above-mentioned solutions provided by the embodiments of the present invention mainly have the following effects: (1) Fully automatic closed-loop quality control: Unlike the existing single-generation scheme, this invention introduces a feedback loop. The arranger is not only a "painter" but also a "modifier". Through continuous self-examination and tool calls, it can automatically fix common problems such as overlapping labels and incorrect data drawing.

[0089] (2) Flexibility of tool invocation: This invention does not bind to a single tool, but drives heterogeneous tools through a unified MLLM interface. Whether it is modifying the parameters of the code or sending editing instructions to the generated model, it is all abstracted into the "instruction" action of the orchestrator, which is highly extensible.

[0090] (3) It solves the pain point of uncontrollability of generative models: For visual perception path, this invention uses an image editing model to gradually correct the initial generated illusion through multiple rounds of audit-edit loop, so that generative charts can also meet the rigorous standards of data visualization.

[0091] (4) It breaks the contradiction between accuracy and creativity in the generation task: Through intelligent routing, the code is used to ensure the correct structure when generating web pages / PPTX, and the model is used to ensure the visual impact when generating posters, thus achieving the optimal solution for the scene.

[0092] To verify the effectiveness of this invention, 100 data points were constructed for testing. The final content was scored from three dimensions: data accuracy, text correctness, and aesthetics. The results are shown in Table 1. Data accuracy was used to evaluate whether the data information presented in the final content strictly matches the original data or intent provided by the user, without errors, omissions, or misinterpretations. Text correctness was used to evaluate whether the text content in the final content is correct and consistent with the user's intent or context. Aesthetics was used to evaluate whether the various design elements of the final content meet the aesthetic requirements of the application scenario. Two benchmark models were used. For the logical construction path, the benchmark module selected was the Qwen3-VL-235B model, where Qwen ( (The preceding text) is the model name, 3 is the version number, VL represents Vision-Language, and 235B (B stands for billion) is the number of model parameters. During testing, data is input into the Qwen3-VL-235B model, which generates code and draws images, excluding subsequent auditing and correction processes. For the visual perception path, the benchmark module selects the Qwen-Image model, which is the basic model for image generation and editing. During testing, data is input into the Qwen-Image model, which generates images, excluding subsequent auditing and correction processes. The test cases contain two paths; therefore, the results of the benchmark model in Table 1 are a summary of the results of the two models.

[0093] Table 1: Comparison results between the present invention and the benchmark scheme

[0094] The results shown in Table 1 demonstrate that the present invention has significant advantages over the baseline model.

[0095] Through the above description of the embodiments, those skilled in the art can clearly understand that the above embodiments can be implemented by software, or by using software plus necessary general-purpose hardware platforms. Based on this understanding, the technical solutions of the above embodiments 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, mobile hard drive, etc.), including several instructions to cause a computer device (such as a personal computer, server, or network device, etc.) to execute the methods described in the various embodiments of the present invention.

[0096] Example 2 This invention also provides a content arrangement and multi-dimensional closed-loop correction system, which is mainly used to implement the methods provided in the foregoing embodiments, such as... Figure 2 As shown, the system mainly includes: an orchestrator, heterogeneous executors, and an audit layer; The orchestrator is used to execute steps 1 and 2; the heterogeneous executor is used to execute step 3; and the orchestrator and the heterogeneous executor cooperate to execute step 5; the audit layer is used to execute step 4, wherein: Step 1 includes: analyzing user intent based on user input information, and selecting logical construction path and / or visual perception path based on user intent; Step 2 includes: transforming user input information into intermediate representations: for logical construction paths, user input information is transformed into executable code; for visual perception paths, user input information is transformed into visual prompts. Step 3 includes: performing a content generation task based on the intermediate representation, outputting the generation result, and combining the generation result with the intermediate representation as a normalized result; Step 4 includes: performing multi-dimensional audits on the normalization results to determine whether there are any errors in the generated results. If so, a structured error report is output; otherwise, the normalization results are used as the final output. Step 5 includes: the orchestrator corrects the intermediate representation based on the error report, the heterogeneous actuator outputs the corrected generation result based on the corrected intermediate representation, and then proceeds to step 4; Steps 4 and 5 are repeatedly executed through the orchestrator, audit layer, and heterogeneous executor until the corrected generated result is free of errors or the set number of iterations is reached.

[0097] Steps 4 and 5 are repeatedly executed through the orchestrator, audit layer, and heterogeneous executor until the corrected generated result is free of errors or the set number of iterations is reached.

[0098] In this embodiment of the invention, the heterogeneous executor includes: a code interpreter and a generation model; The code interpreter is applied to the logical build path to render visual structured data based on the executable code; Generative models are applied to visual perception paths to generate image data based on visual cues.

[0099] In this embodiment of the invention, the multi-dimensional auditing of the normalization result to determine whether there are errors in the generated result includes: The generated results in the normalized results are audited in multiple dimensions using a multimodal large language model, including: auditing data consistency, text accuracy, and code or layout logic correctness. Among them, data consistency auditing refers to checking whether the visual elements in the generated results match the data content; text accuracy auditing refers to checking whether the text in the generated results is correct; and layout logic correctness auditing refers to checking whether there is label overlap and / or occlusion. If any dimension fails the audit, the generated result is considered to contain errors; if all three dimensions pass the audit, the generated result is considered to contain no errors.

[0100] In this embodiment of the invention, the step of correcting the intermediate representation based on the error report and outputting the corrected generation result includes: For the logical build path, the orchestrator corrects the corresponding content of the executable code based on the error report, and generates the corrected result through the code interpreter; For the visual perception path, the orchestrator corrects the visual cue words based on the error report, generates a local redraw instruction, calls the image editing model to perform local redrawing on the generated result, and outputs the corrected generated result. Both the code interpreter and the image editing model belong to heterogeneous executors.

[0101] Those skilled in the art will understand that, for the sake of convenience and brevity, the above-described division of functional modules is used as an example. In practical applications, the above functions can be assigned to different functional modules as needed, that is, the internal structure of the system can be divided into different functional modules to complete all or part of the functions described above.

[0102] Example 3 The present invention also provides a processing device, such as Figure 3 As shown, it mainly includes: one or more processors; a memory for storing one or more programs; wherein, when the one or more programs are executed by the one or more processors, the one or more processors implement the method provided in the foregoing embodiments.

[0103] Furthermore, the processing device also includes at least one input device and at least one output device; in the processing device, the processor, memory, input device, and output device are connected via a bus.

[0104] In this embodiment of the invention, the specific types of the memory, input device, and output device are not limited; for example: Input devices can be touchscreens, image acquisition devices, physical buttons, or mice, etc. The output device can be a display terminal; The memory can be random access memory (RAM) or non-volatile memory, such as disk storage.

[0105] Example 4 The present invention also provides a readable storage medium storing a computer program that, when executed by a processor, implements the method provided in the foregoing embodiments.

[0106] In this embodiment of the invention, the readable storage medium is a computer-readable storage medium and can be disposed in the aforementioned processing device, for example, as a memory in the processing device. Furthermore, the readable storage medium can also be any medium capable of storing program code, such as a USB flash drive, portable hard drive, read-only memory (ROM), magnetic disk, or optical disk.

[0107] The above description is merely a preferred embodiment of the present invention, but the scope of protection of the present invention is not limited thereto. Any variations or substitutions that can be easily conceived by those skilled in the art within the scope of the technology disclosed in the present invention should be included within the scope of protection of the present invention. Therefore, the scope of protection of the present invention should be determined by the scope of the claims. The information disclosed in the background section is intended only to enhance the understanding of the overall background technology of the present invention and should not be construed as an admission or implication in any way that such information constitutes prior art known to those skilled in the art.

Claims

1. A content arrangement and multi-dimensional closed-loop correction method, characterized in that, include: Step 1: Analyze the user's intent based on the user's input information, and select the logical construction path and / or visual perception path according to the user's intent; Step 2: Transform user input information into intermediate representations: For logical construction paths, transform user input information into executable code; for visual perception paths, transform user input information into visual prompts. Step 3: Perform the content generation task based on the intermediate representation, output the generation result, and combine the generation result with the intermediate representation as the normalized result; wherein, the step of performing the content generation task based on the intermediate representation and outputting the generation result includes: for logical construction paths, rendering visual structure data based on executable code; for visual perception paths, generating image data based on visual cues. Step 4: Perform multi-dimensional auditing on the normalized result to determine if there are any errors in the generated result. If so, output a structured error report; otherwise, use the normalized result as the final output. This multi-dimensional auditing of the normalized result to determine if there are any errors includes: auditing the generated result in the normalized result using a multimodal large language model, including auditing three dimensions: data consistency, text accuracy, and code or layout logic correctness. Data consistency auditing checks whether the visual elements in the generated result match the data content; text accuracy auditing checks whether the text in the generated result is correct; layout logic correctness auditing checks for label overlap and / or occlusion. If any dimension fails the audit, the generated result is deemed to have an error; if all three dimensions pass the audit, the generated result is deemed to have no errors. Step 5: Correct the intermediate representation based on the error report and output the corrected generation result, then return to step 4; repeat steps 4 and 5 until the corrected generation result is error-free or the set number of iterations is reached; wherein, correcting the intermediate representation based on the error report and outputting the corrected generation result includes: for the logical construction path, correcting the corresponding content of the executable code based on the error report and thereby generating the corrected generation result; for the visual perception path, correcting the visual prompts based on the error situation in the error report, generating a local redraw instruction, performing a local redraw on the generation result, and outputting the corrected generation result.

2. A content arrangement and multi-dimensional closed-loop correction system, characterized in that, include: Orchestrator, heterogeneous executors, and audit layer; The orchestrator is used to execute steps 1 and 2, the heterogeneous executor is used to execute step 3, and the orchestrator and the heterogeneous executor cooperate to execute step 5. The audit layer is used to perform step 4, wherein: Step 1 includes: analyzing user intent based on user input information, and selecting logical construction path and / or visual perception path based on user intent; Step 2 includes: transforming user input information into intermediate representations: for logical construction paths, user input information is transformed into executable code; for visual perception paths, user input information is transformed into visual prompts. Step 3 includes: generating a task based on the intermediate representation, outputting the generated result, and combining the generated result with the intermediate representation as a normalized result; the heterogeneous executor includes: a code interpreter and a generation model; the code interpreter is applied to the logic construction path and is used to render visual structure data based on the executable code; the generation model is applied to the visual perception path and is used to generate image data based on visual cues. Step 4 includes: performing multi-dimensional audits on the normalized results to determine if there are any errors in the generated results. If so, a structured error report is output; otherwise, the normalized results are used as the final output. The multi-dimensional audit of the normalized results to determine if there are any errors in the generated results includes: performing multi-dimensional audits on the generated results in the normalized results using a multimodal large language model, including audits of three dimensions: data consistency, text accuracy, and code or layout logic correctness. Data consistency audit refers to checking whether the visual elements in the generated results match the data content; text accuracy audit refers to checking whether the text in the generated results is correct; layout logic correctness audit refers to checking whether there are overlapping and / or occlusion of labels. If any dimension fails the audit, the generated results are determined to have errors; if all three dimensions pass the audit, the generated results are determined to have no errors. Step 5 includes: the orchestrator corrects the intermediate representation based on the error report, the heterogeneous executor outputs the corrected generation result based on the corrected intermediate representation, and then proceeds to step 4; wherein, the correction of the intermediate representation based on the error report and the output of the corrected generation result includes: for the logical construction path, correcting the corresponding content of the executable code based on the error report, and thereby generating the corrected generation result; for the visual perception path, correcting the visual prompts based on the error situation in the error report, generating a local redraw instruction, performing a local redraw on the generation result, and outputting the corrected generation result; wherein, steps 4 and 5 are repeatedly executed by the orchestrator, the audit layer, and the heterogeneous executor until the corrected generation result is free of errors, or the set number of iterations is reached.

3. A processing device, characterized in that, include: One or more processors; Memory, used to store one or more programs; Wherein, when the one or more programs are executed by the one or more processors, the one or more processors cause the one or more processors to implement the method as described in claim 1.

4. A readable storage medium storing a computer program, characterized in that, The method described in claim 1 is implemented when the computer program is executed by the processor.

Images