Structured representation methods, systems, devices, and storage media for reasoning processes

Abstract

This application relates to the fields of artificial intelligence and data processing technology, and provides a method, system, computer device, and storage medium for the structured display of the reasoning process in interface generation. The method acquires user-input interface design prompts and decomposes them to generate a structured analysis result containing requirement types, style constraints, and functional requirements; acquires the original reasoning data stream of a large model and identifies stage transition markers, automatically mapping the reasoning content to preset problem analysis, creative direction thinking, or design scheme organization stages; extracts feature items based on a domain feature library and refines them into core points using a semantic condensation algorithm; synchronously presents the key points and progress animations in logical order, and triggers local reasoning updates based on adjustment instructions by constructing incremental prompts. This application achieves the structured reconstruction and visualization of the reasoning logic of a large language model, solving the problems that limit interactive response performance and resource utilization.

Description

Technical Field

[0001] This application relates to the field of artificial intelligence technology, and provides a structured display method, system, computer device, and storage medium for the reasoning process in interface generation. Background Technology

[0002] In the field of AIGC (Artificial Intelligence Generated Content) interface generation, large language models typically employ a continuous streaming mechanism when outputting inference data. However, due to the lack of fine-grained state management and logical demarcation in existing systems during the inference process, the following technical shortcomings exist in practical engineering applications:

[0003] Currently, the reasoning process of large language models is typically viewed as an unstructured and indivisible holistic data stream. When logical deviations or parameter mismatches occur in the early stages of inference, existing technologies usually can only truncate and discard the entire inference chain that has already been generated. This processing mode forces users to trigger a full restart of the inference process from the initial state by re-entering interface design prompts. This full restart correction mechanism not only causes serious waste of computing resources but also significantly increases the latency of the system during state iteration and result confirmation, severely limiting the interactive response performance and resource utilization of the interface generation system. Summary of the Invention

[0004] This application relates to the field of artificial intelligence technology, and provides a structured display method, system, computer device and storage medium for the reasoning process in interface generation, which solves the technical problems that limit the interactive response performance and resource utilization of the interface generation system.

[0005] A structured display method for the reasoning process in interface generation, the method comprising:

[0006] Obtain user-inputted interface design prompts, perform semantic analysis and requirement decomposition on the interface design prompts, and generate a structured parsing result containing at least one of the requirement type, style constraints, and functional requirements;

[0007] Obtain the raw inference data stream output by the large language model for the interface design prompt words, and identify the stage transition marker bits in the raw inference data stream;

[0008] Based on the structured parsing results and the stage transition markers, the original reasoning data stream is automatically mapped to the preset problem analysis stage, creative direction thinking stage, or design scheme organization stage.

[0009] For each stage, feature items are extracted from the original inference data stream based on the feature library of the interface design domain, and the feature items are refined into a preset number of core points of the stage using a semantic condensation algorithm.

[0010] According to the progressive order of the stages, the core points of each stage and the corresponding stage progress animation are presented synchronously on the user interface.

[0011] Obtain the user's adjustment instructions for the core points of the currently displayed stage, construct incremental prompt words based on the adjustment instructions, and send them to the large language model to trigger the update of the inference data stream for the corresponding stage and subsequent stages.

[0012] A structured display system for the reasoning process in interface generation, the system comprising:

[0013] The parsing module is used to obtain the interface design prompts input by the user, perform semantic analysis and requirement decomposition on the interface design prompts, and generate a structured parsing result containing at least one of the requirement type, style constraints and functional requirements.

[0014] The splitting module is used to obtain the raw inference data stream output by the large language model for the interface design prompt words, and to identify the stage transition marker bits in the raw inference data stream; based on the structured parsing results and the stage transition marker bits, the raw inference data stream is automatically mapped to the preset problem analysis stage, the creative direction thinking stage, or the design scheme organization stage.

[0015] The extraction module is used to extract feature items from the original inference data stream based on the feature library of the interface design domain for each stage, and to extract the feature items into a preset number of stage core points using a semantic condensation algorithm.

[0016] The rendering module is used to synchronously present the core points of each stage and the corresponding stage progress animation on the user interface in the order of the stages.

[0017] The interactive feedback module is used to obtain the user's adjustment instructions for the core points of the currently displayed stage, construct incremental prompt words based on the adjustment instructions, and send them to the large language model to trigger the update of the inference data stream for the corresponding stage and subsequent stages.

[0018] A computer device includes a memory, a processor, and a computer program stored in the memory and executable on the processor, wherein the processor, when executing the computer program, implements the steps of the structured presentation method as described in any of the preceding claims.

[0019] A computer-readable storage medium storing a computer program that, when executed by a processor, implements the steps of the structured representation method as described in any of the preceding claims.

[0020] In one of the solutions provided in this application, by acquiring the original inference data stream and identifying stage transition markers, and automatically mapping the original inference data stream to preset problem analysis, creative direction thinking, or design scheme organization stages, the original linear, continuous, and indivisible inference process of the large model is transformed into structured stages with clear logical boundaries and functional definitions at the technical level. This staged mapping mechanism enables the system to locate the current logical node of the model in real time (e.g., the instant from understanding requirements to conceptualizing a solution), thus providing the necessary physical basis and data benchmark for subsequent local intervention and precise state control. Based on adjustment instructions, incremental prompts are constructed and sent to the large model. By synchronously presenting the core points of each stage and corresponding stage progress animations, and triggering updates for the corresponding and subsequent stages, the system achieves a dynamic interactive mode of inference, feedback, and optimization. Compared to the background technology's mode of overall judgment after solution generation and rerunning if inconsistent, this application allows users to complete logical calibration according to the stage generation progress. This mechanism significantly reduces the runtime of invalid states and greatly shortens the overall response latency from receiving instructions to outputting adaptation results, thereby improving interactive response performance and resource utilization. Attached Figure Description

[0021] To more clearly illustrate the technical solutions of the embodiments of this application, the drawings used in the description of the embodiments of this application will be briefly introduced below. Obviously, the drawings described below are only some embodiments of this application. For those skilled in the art, other drawings can be obtained based on these drawings without creative effort.

[0022] Figure 1 This is a flowchart illustrating a structured display method for the reasoning process in interface generation, as described in this application.

[0023] Figure 2 This is a schematic diagram of a structured display system for the reasoning process in interface generation according to an embodiment of this application;

[0024] Figure 3 This is a schematic diagram of a problem analysis stage in the structured display process of the reasoning process in the interface generation of this application.

[0025] Figure 4 This is a schematic diagram of the creative direction thinking stage in the structured display process of the reasoning process in the interface generation of this application.

[0026] Figure 5 This is a schematic diagram of another display process in the stage of thinking about creative directions during the structured display process of reasoning in the interface generation of this application.

[0027] Figure 6 This is a schematic diagram of the design scheme organization stage in the structured display process of the reasoning process in the interface generation of this application. Detailed Implementation

[0028] The technical solutions of the embodiments of this application will be clearly and completely described below with reference to the accompanying drawings. Obviously, the described embodiments are only some embodiments of this application, not all embodiments. Based on the embodiments of this application, all other embodiments obtained by those skilled in the art without creative effort are within the scope of protection of this application.

[0029] It should be understood that the sequence number of each step in the following embodiments does not imply the order of execution. The execution order of each process should be determined by its function and internal logic, and should not constitute any limitation on the implementation process of the embodiments of this application.

[0030] As mentioned in the background, traditional AIGC systems suffer from inefficient feedback presentation and difficulty in extracting key information. In traditional AIGC interface generation software, after the user inputs interface design prompts, the feedback from the large model consists mainly of continuous, lengthy textual reasoning, updated rapidly, and lacking structured breakdown. This type of information presentation does not consider the user's information reception habits. Long blocks of text are mixed with redundant thinking and repetitive expressions, requiring users to sift through rapidly scrolling text to find key information. This not only consumes considerable time and energy but also makes it easy to miss crucial content, resulting in low information acquisition efficiency. Furthermore, this indiscriminate text display mode merely presents the large model's thought process, failing to highlight its core capabilities of analysis, deduction, and creative divergence. Users cannot intuitively perceive the depth of the large language model's understanding of the requirements and the rationality of its design logic. Ultimately, they can only passively focus on the final generated result, unable to judge the suitability of the result through the thought process, or make targeted adjustments based on the thought process.

[0031] Traditional feedback methods lack structure and visual representation, resulting in insufficient user engagement and trust. Traditional large-scale model-based feedback processes lack clear stage divisions, presenting an overall chaotic reasoning approach. Users cannot clearly understand the complete logical chain from receiving requirements to outputting results, leading to a lack of user awareness of the model's workflow, difficulty in predicting the direction of the generated outcome, and consequently, reduced trust in the results. Furthermore, purely text-based static feedback is too monotonous, failing to reinforce key information through visualization or guide users to synchronously understand the requirement processing process along with the model's thought process. Users remain in a passive receiving state with weak engagement. Even if the generated result does not meet expectations, it is difficult to pinpoint the problem in the understanding of requirements, creative direction, or solution integration, hindering subsequent optimization of prompts or adjustments to the result.

[0032] Furthermore, existing technologies fail to address the issue of visualizing the alignment between the thought process and user needs: traditional AIGC software's feedback focuses solely on the reasoning trajectory of the large model itself, neglecting to address core user needs such as how to quickly verify whether the large model accurately understands the requirements and whether the creative direction meets expectations. Users cannot promptly identify misinterpretations of requirements or deviations in creative direction by the large model during the thought process; they can only make an overall judgment after the final result is generated. If the result does not meet the requirements, they must re-enter prompts to restart the generation process, resulting in high costs and low efficiency from repeated trial and error. Simultaneously, the lack of structured and visualized presentation of the thought process prevents the effective transfer of the large model's creative and analytical capabilities to the user, limiting the efficiency of collaborative design between the user and the large model and hindering the interactive effect of thinking, adjusting, and optimizing simultaneously.

[0033] The above situation not only causes serious waste of computing resources, but also significantly increases the latency of the system during state iteration and result confirmation, severely limiting the interactive response performance and resource utilization of the interface generation system.

[0034] To address the aforementioned issues that limit the interactive response performance and resource utilization of interface generation systems, this application provides a solution. Specific embodiments are described below to illustrate the technical solution of this application.

[0035] In one embodiment, such as Figure 1 This paper provides a structured display method for the reasoning process in interface generation, specifically a structured display method for the reasoning process in AIGC interface generation. The method includes the following steps:

[0036] S10. Obtain the user-input interface design prompts, perform semantic analysis and requirement decomposition on the interface design prompts, and generate a structured parsing result containing at least one of the requirement type, style constraints, and functional requirements.

[0037] S20. Obtain the raw inference data stream output by the large language model for the interface design prompt words, and identify the stage transition marker bits in the raw inference data stream;

[0038] S30. Based on the structured parsing results and stage transition markers, the original reasoning data stream is automatically mapped to the preset problem analysis stage, creative direction thinking stage, or design scheme organization stage.

[0039] S40. For each stage, extract feature items from the original inference data stream based on the feature library of the interface design domain, and use semantic condensation algorithm to refine the feature items into a preset number of core points for the stage.

[0040] S50. In accordance with the progressive order of the stages, the core points of each stage and the corresponding stage progress animation are presented synchronously on the user interface.

[0041] S60. Obtain the user's adjustment instructions for the core points of the currently displayed stage, construct incremental prompt words based on the adjustment instructions, and send them to the large language model to trigger the update of the inference data stream for the corresponding stage and subsequent stages.

[0042] In this embodiment, the interface design prompts refer to the user's input natural language design requirements, such as: "I want a card for ordering takeout, with a liquid glass style, generate it for me." Then, semantic analysis and requirement decomposition are performed on the interface design prompts to generate a structured parsing result containing at least one of the following: requirement type, style constraints, and functional requirements. Semantic analysis and requirement decomposition refer to the process of using Natural Language Processing (NLP) technology combined with a domain-specific thesaurus to segment and annotate the prompts, understand their semantics, and parse out the core parameters. The requirement type refers to the terminal or presentation form of the interface, such as a mobile app homepage, UI component cards, or a responsive webpage; style constraints refer to restrictions on visual tone, color, and materials, such as flat design, minimalism, or liquid glass materials; functional requirements refer to the interactive elements or information items that the interface must have, such as a search bar, a carousel, or delivery fee information. The structured parsing result is a data set that encapsulates the above core parameters into a standard format (such as JSON).

[0043] The raw inference data stream refers to the continuous text token sequence output in real time by the large language model before generating the final interface design result. It records the thought path or process of the large language model from receiving instructions to forming a solution. The stage transition marker is the anchor point used in the continuous inference stream to identify the shift in the large language model's thinking focus from one pre-set logical stage (such as understanding requirements) to another logical stage (such as conceptualizing a solution). The pre-set stages include: the problem analysis stage (focusing on requirement identification and constraint recognition), the creative direction thinking stage (focusing on creative divergence and logical deduction), and the design solution organization stage (focusing on solution integration and component parameter determination).

[0044] Semantic condensation algorithms are text compression techniques based on deep learning semantic representations. Their core definition lies in calculating the relationship between candidate text blocks and the target state of the design result to filter out logical fragments that decisively contribute to interface generation as feature terms. The UI design domain feature library is the core knowledge benchmark for transforming raw inference data from large models into structured design points. The UI design domain feature library refers to a pre-built, structured knowledge base specifically designed for UI (user interface) design, containing professional terminology, logical rules, and semantic patterns. It covers professional terms in UI design (such as mobile layout, flat design style, responsive adaptation, etc.) and extraction rules corresponding to each logical stage. It is a core component for subsequent refinement, used to identify design-guiding feature terms in the unstructured inference data stream output by large models.

[0045] In this embodiment, feature items are extracted from the original inference data stream based on a feature library in the interface design domain. A semantic condensation algorithm is then used to refine these feature items into a predetermined number of core points for each stage. Feature items refer to key sentences with interface design attributes identified from the inference stream of the large language model, such as: using a frosted glass effect to enhance the sense of depth in the background. The core points for each stage refer to the refined core conclusions, preferably within 15 characters in length. Stage progress animations refer to dynamic effects that guide the user's visual progression along with the model's thought process, such as loading progress bars, icon lighting, or card appearance.

[0046] Incremental prompts refer to new instructions dynamically constructed by the user based on the adjustment instructions proposed by the user for the core key points of the current stage of the large model being displayed, by combining the context snapshot of the current session.

[0047] To facilitate understanding, the processing procedure of this application embodiment is illustrated below with reference to the above terminology: The system first receives the user's input of interface design prompts, such as: "I want a card for ordering takeout, with a liquid glass style, generate it for me." In this example, the system first identifies the demand type as a takeout card, the style constraint as liquid glass, and the functional requirements as including elements such as product price display, and outputs a structured parsing result. Next, the system monitors the raw inference data stream output by the large language model in real time, scans the stage transition marker bits in the token sequence of the raw inference data stream, and identifies the transition signal when the large language model shifts from restating the user's intent to conceiving visual details. The system uses a semantic association algorithm, combined with the structured parsing result, to categorize the inference content: content about takeout business logic and constraints is mapped to the problem analysis stage; color schemes for summer mango chiffon and coconut latte are mapped to the creative direction thinking stage; specific configurations at the level of features, display, interaction, and visuals (e.g., features: 5 items in the limited-time flash sale area) are mapped to the design scheme organization stage.

[0048] Subsequently, a feature library from the interface design domain is invoked. For example, during the creative phase, feature items are extracted from the description output by the model, such as: the original price is displayed with a strikethrough to highlight the discount. The semantic condensation algorithm filters logical transitional text such as: "Next, we suggest..." and extracts it into the core points of the stage with a length of no more than 15 characters.

[0049] Finally, execute the progressive animation in logical order. Assign unique icons to each stage, such as a magnifying glass for problem analysis, a light bulb for creative direction, and a blueprint for design solutions. Key points are presented on the information card with a word-by-word animation. If the user inputs an adjustment command for the currently displayed key point, such as "add a business feel to the style," a snapshot of the current conversation is captured, corresponding incremental prompts are built, and the large language model is triggered to regenerate the inference flow from the corresponding breakpoint.

[0050] In traditional AIGC (AI-Generated Content) interface generation software, after users input interface design prompts, the feedback from the large model mainly consists of continuous, lengthy textual reasoning, updated rapidly, and lacking structured breakdown. This information presentation format does not consider users' information reception habits. Long blocks of text are interspersed with redundant thinking and repetitive expressions, requiring users to sift through rapidly scrolling text to find key information (such as core design logic and core elements of the solution). This not only consumes a lot of time and energy but also makes it easy to miss crucial content, resulting in low information acquisition efficiency. Furthermore, this indiscriminate text display mode merely presents the large model's thought process, failing to highlight its core capabilities of analysis, deduction, and creative divergence. Users cannot intuitively perceive the depth of the large model's understanding of the requirements and the rationality of its design logic. Ultimately, they can only passively focus on the final generated result, unable to judge the suitability of the result through the thought process, and unable to make targeted adjustments based on the thought process. Traditional feedback methods lack structure and visual representation, resulting in insufficient user engagement and trust: The thinking and feedback process of traditional large models lacks clear stage divisions, presenting an overall chaotic reasoning, making it difficult for users to clearly understand the complete logical chain from receiving requirements to outputting results.

[0051] In this embodiment, the originally chaotic and redundant lengthy reasoning process of the large language model is transformed into a structured analytical result containing requirement types, style constraints, and functional requirements. This result is then forcibly mapped to three preset stages with clear logical boundaries, establishing a perceptible chain of reasoning logic. This approach addresses the pain points of traditional feedback methods, such as the difficulty in information extraction and the easy burying of key elements. Users only need to view the core points extracted at each stage and the corresponding synchronous animation to quickly grasp the large language model's deep understanding and deductive logic of the design. This significantly improves information acquisition efficiency and enhances the predictability and trust in the AI-generated results, effectively solving the problems of inefficient information presentation and difficulty in extracting key information in traditional AIGC software feedback.

[0052] It's also worth noting that traditional AIGC software's feedback focuses solely on the large model's own reasoning trajectory, neglecting to address core user needs such as how to quickly verify whether the large model accurately understands the requirements or whether the creative direction meets expectations. Users cannot promptly identify misinterpretations of requirements or deviations in creative direction by the large model during the thinking process; they can only make an overall judgment after the final result is generated. If the result doesn't meet the requirements, they must re-enter prompts to restart the generation process, leading to high costs and low efficiency from repeated trial and error. This limits the efficiency of collaborative design between users and the large model, making it difficult to achieve the interactive effect of thinking, adjusting, and optimizing simultaneously. In this embodiment, following the progressive order of stages, the core points of each stage and corresponding stage progress animations are simultaneously presented on the user interface; user adjustment instructions for the currently displayed stage's core points are obtained, incremental prompts are constructed based on these instructions and sent to the large language model to trigger updates to the inference data stream for the corresponding and subsequent stages, effectively reducing interaction time.

[0053] As can be seen, in this embodiment, by acquiring the original inference data stream and identifying stage transition markers, and automatically mapping the original inference data stream to preset problem analysis, creative direction thinking, or design scheme organization stages, the original linear, continuous, and indivisible inference process of the large model is transformed into structured stages with clear logical boundaries and functional definitions at the technical level. This staged mapping mechanism enables the system to locate the current logical node of the model in real time (e.g., the instant from understanding requirements to conceptualizing a solution), thus providing the necessary physical basis and data benchmark for subsequent local intervention and precise state control. Furthermore, by constructing incremental prompts based on adjustment instructions and sending them to the large model, and by synchronously presenting the core points of each stage and corresponding stage progress animations, as well as triggering updates for the corresponding and subsequent stages, the system achieves a dynamic interactive mode of inference, feedback, and optimization simultaneously. Compared to the background technology's mode of overall judgment after solution generation and rerunning if inconsistent, this application allows users to complete logical calibration according to the stage generation progress. This mechanism significantly reduces the running time of the system in invalid states and greatly shortens the overall response delay from receiving instructions to outputting adaptation results, thereby improving interactive response performance and resource utilization.

[0054] In one embodiment, step S20, namely identifying the stage transition marker bits in the original inference data stream, includes the following steps:

[0055] S201. Scan the token sequence in the original inference data stream and identify the preset explicit delimiter as the stage transition marker.

[0056] S202. If no explicit delimiter is identified, calculate the vector representation of the continuous token sequence and determine the position where the semantic centroid offset of the vector representation exceeds the preset threshold as the stage transition marker.

[0057] In this embodiment, vector representation refers to the process of mapping text characters to a high-dimensional numerical vector space, enabling semantic similarity to be measured by geometric distance. The semantic centroid refers to the center point of a set of continuous token vectors in space. The semantic centroid offset refers to the Euclidean distance between the center points of two adjacent text windows.

[0058] In this embodiment, a real-time large language model scans the token sequence in the raw inference data stream. If the large language model outputs explicit delimiters such as [STAGE_BREAK] under control, these are directly used as stage transition markers. For large language models with non-standard outputs, implicit monitoring can be initiated: continuously generated text blocks are converted into high-dimensional vectors and their centroids are calculated using a sliding window technique. When the discussion shifts from addressing user background pain points to describing specific component shadow parameters, the semantic meaning of the words undergoes a fundamental change, causing the semantic centroid offset to exceed a preset threshold. At this point, the system identifies this abrupt change as a marker. For example, when shifting from the problem analysis stage to the creative direction thinking stage, the semantic meaning of the output words will change, causing the semantic centroid offset of the vector representation to exceed a preset threshold. In this case, the position where the offset exceeds the limit is identified as a stage transition marker.

[0059] As can be seen, in this embodiment, by combining explicit flag recognition and implicit semantic centroid offset detection, the ability to be compatible with large language models from different manufacturers and with different levels of control is ensured. This design not only improves the accuracy of stage segmentation, but also enhances the ability to structure uncontrolled inference flows, providing a robust data foundation for subsequent logic presentation.

[0060] In other embodiments, a timestamp-based auxiliary determination mechanism can be introduced when identifying stage transition markers in the original inference data stream. Specifically, during the acquisition of the original inference data stream, the real-time timestamp of each token generation is recorded synchronously, and the output time interval between adjacent tokens is calculated. Because the underlying computational attention distribution of a large language model undergoes an instantaneous reconstruction when its thought process shifts fundamentally (e.g., from demand logic analysis to visual parameter conception), this shift in computational load often manifests as a slight generation pause at the output. By monitoring the generation rate of the token stream, when a significant abrupt change in the time interval between adjacent tokens is detected (e.g., exceeding a preset multiple of the average generation time interval), that time point is determined as the stage transition marker.

[0061] In this embodiment, by introducing timestamp markers as a physical-level monitoring dimension, the physiological pauses that occur when the large model switches thoughts can be captured more sensitively. This approach, combined with semantic centroid offset determination, constructs a dual recognition mechanism in the time and frequency domains, greatly improving the accuracy of stage division and ensuring the accuracy of subsequent structured display in logical segmentation.

[0062] In one embodiment, step S30, namely, automatically mapping the original inference data stream to a preset stage based on the structured parsing results and stage transition marker bits, includes the following steps:

[0063] S301. Calculate the matching degree between the reasoning content at each stage and the different requirements in the structured parsing results using a semantic association algorithm;

[0064] S302. Based on the matching degree and stage transition marker, the original inference data stream is semantically segmented and classified into the corresponding preset stages; among which: the output of the problem analysis stage includes core requirement positioning, potential requirement mining and design constraint identification; the output of the creative direction thinking stage includes core creative direction, design logic and adaptability analysis; the output of the design scheme organization stage includes the final scheme selection reasons, core architecture of the scheme and key component design logic.

[0065] The semantic association algorithm mentioned in this embodiment refers to a deep learning model used to calculate the semantic similarity between two text segments. Matching degree refers to the relevance score between the inference segment and the requirement elements (such as style constraints, functional requirements, etc.). Specifically, firstly, the generated JSON structured parsing result is obtained, which clearly defines the boundaries of the interface generation. When obtaining the raw inference data stream of the large language model, the aforementioned semantic association algorithm is used to monitor the matching degree between the generated token sequence and the JSON fields in real time. For example:

[0066] When the inference flow is identified to contain discussions about the target user group, business background, or terminal environment, this content, due to its highest match with the demand types and functional requirements in the structured results, is semantically decomposed and categorized into the problem analysis stage, combined with stage transition markers. As an example, this stage outputs the large language model's response to the prompt word parsing results. For instance, it clarifies that the core demand is a high-frequency interactive food delivery card and uncovers potential demands such as enhanced visual contrast in nighttime delivery modes, while also identifying design constraints such as adaptation to narrow-screen mobile devices.

[0067] When the reasoning flow shifts to describe color psychology, visual materials, or conceptual inspiration (such as mentioning liquid glass or summer mango chiffon cake), the match with the style constraint field significantly improves. Therefore, this is mapped to the creative direction thinking stage. At this stage, for example, 3 to 5 core creative directions are output, along with a detailed explanation of the design logic behind each direction (e.g., projection is used to address the issue of hierarchical depth) and an analysis of the compatibility of the concept with user brand preferences.

[0068] When the content involves specific parameter selection, layout proportions, or component logic, it should be categorized under the design scheme development stage. For example, output the reasons for the final scheme selection and define the core architecture of the scheme, covering the design logic of key components such as displaying the original price with strikethrough, configuring 5 data items in the flash sale area, and the automatic switching logic of the carousel.

[0069] In this embodiment, a semantic association algorithm is used to map the original inference data stream to the structured parsing results point-to-point, achieving precise categorization from a chaotic text flow to an ordered logical stage. This processing method makes the model's thinking a clear logical chain, allowing users to judge the depth of the model's understanding through potential needs mining and to judge the rationality of the idea through adaptability analysis. This fine-grained mapping mechanism significantly improves the interpretability of the reasoning process, enabling users to complete logical calibration before the final solution is generated, greatly enhancing users' trust in the generated results.

[0070] In one embodiment, step S40, which is the core point of the stage of refining feature items into a preset number of items using a semantic condensation algorithm, includes the following steps:

[0071] S401. Identify and filter redundant information and logical transition text in feature items, and retain keywords and phrases related to interface design parameters.

[0072] S402. Calculate the semantic relevance between the filtered keywords and the final generated interface design parameters, and select a preset number of key points as the core key points of the stage based on the semantic relevance.

[0073] In this embodiment, redundant information refers to decorative descriptions or repetitive expressions that exist in the original inference data stream of the large language model and do not contribute substantially to design decisions. Logical transition text refers to phrases that serve a connecting function during the text logic transformation process, such as: "First, let's look at...", "Next, let's conceive of...", "In summary...", "We can see...", etc. Interface design parameters refer to parameter variables that can be directly mapped to the UI rendering development stage, specifically including color codes, rounded corner values, component spacing, layout proportions, font sizes, and the number of elements. Semantic relevance is used to quantify the similarity of two sets of textual information at the design logic level, typically measured based on the cosine similarity of high-dimensional semantic vectors.

[0074] The specific processing procedure is as follows: For the output of each logical stage, the above steps S401-S402 are executed to extract the logic. However, each stage, based on its own feature library and output standard, produces different levels of business depth in its core points.

[0075] As an example, during the problem analysis phase, the system uses S401 to filter out redundant statements that restate the requirements in the large language model, retaining only the features related to the constraints. The core output points are such as: the core requirement is positioned as a high-frequency interactive takeaway card, and design constraints such as narrow-screen mobile device adaptation are identified.

[0076] As another example, during the creative direction-thinking phase, identify the logical deduction of design inspiration within the larger language model and extract short phrases that represent the direction. For example... Figure 4 As shown, the core output points at this stage include: summer mango chiffon cake, displaying the original price with a strikethrough to highlight the discount, and tagging prompts. These points directly reflect the large language module's concept for interface style and basic visual logic.

[0077] As another example, during the design phase, focus on key component parameters directly related to the final result. Utilize S402 to calculate the correlation between core elements and the final UI code generation, ensuring the selected information explains the composition of the interface components. For example... Figure 5 As shown, the core output at this stage includes: 5 items in the limited-time flash sale section, 4 items in the recommended product section, 6 items with added shopping attributes, and automatic switching of the carousel. This information precisely defines the functional layout and dynamic interaction logic of the interface.

[0078] As can be seen, this embodiment achieves a standardized transformation of the reasoning flow from unstructured text to high-density design decisions by applying a semantic condensation algorithm to each specific logical stage. This phased, highly relevant extraction mechanism not only eliminates redundant linguistic interference but also ensures that the 3 to 5 key points presented to the user at each stage are the most critical logical support for that stage. Through these extracted short sentences of no more than 15 characters (such as displaying the original price with a strikethrough and automatically switching carousels), users can instantly perceive the design logic of the large language model in different dimensions, thereby significantly reducing the time cost of filtering key information from long texts, significantly improving the review efficiency of human-computer collaborative design, and significantly reducing cognitive load and information processing effort.

[0079] In one embodiment, step S50, which involves synchronously presenting the core points of each stage and the corresponding stage progress animation on the user interface, specifically includes the following steps:

[0080] S501. Execute the stage transition animation in the order of problem analysis, thinking about creative directions, and organizing design schemes. The stage transition animation includes the exit effect of the current stage and the entrance effect of the next stage.

[0081] S502. When showcasing the core points of the stage, display the independent information cards corresponding to the core points of the stage simultaneously, and highlight the design keywords in the core points of the stage.

[0082] In this embodiment, the stage transition animation refers to a processing mechanism that guides the user's visual focus to flow naturally between different thought stages through dynamic visual feedback. The exit effect is preferably a smooth disappearance animation such as fade-in or fade-out, while the entrance effect is preferably a dynamic path with a sense of progression, such as sliding combined with scaling. Independent information cards refer to UI containers in the user interface that carry the core points of each stage and have independent boundaries and visual hierarchy; they typically have a slight floating shadow or border highlight effect. Design keywords refer to professional terms in the key points that can represent the core variables of the interface design, such as mobile e-commerce, minimalist style, and three-column layout. Highlighting refers to the visual enhancement processing of keywords through differences in color saturation, bolding, or background color filling.

[0083] In this embodiment, the presentation logic of the problem analysis stage is initiated first, following a preset logical progression. A dedicated magnifying glass icon can be configured for this stage to enhance its recognizability. Once the feature extraction for this stage is complete, the stage transition animation in S501 is executed. For example, the content of the problem analysis stage gradually fades away with a fade-in / fade-out effect, while the subsequent creative direction thinking stage quickly enters the user's field of vision with a sliding and zooming effect, and a light bulb icon lights up simultaneously. When the final design scheme stage enters, a blueprint icon is displayed.

[0084] During the display of specific information at each stage, the display logic in S502 is executed. Each semantically condensed core point does not pop up all at once, but is presented as an animated text-by-text on an independent information card. The design keywords in the points are automatically identified and highlighted. For example, in the creative direction stage, style terms such as "summer mango chiffon" are highlighted. Simultaneously, as the information cards appear, the interface displays the percentage of completion for each stage, allowing users to anticipate the progress of their reasoning in real time.

[0085] As can be seen, in this embodiment, by executing progressively switching animations according to the preset logical order of problem analysis, creative direction thinking, and design scheme organization, the originally chaotic and static text reasoning process of the large language model is transformed into a temporal visual flow that conforms to the laws of human cognition, allowing users to clearly perceive the progress of the thinking process. Furthermore, the card-based display of independent information cards and the highlighting of design keywords transform passive reading into active perception, significantly improving the visual hierarchy and the recognizability of core information. This animation presentation mechanism, tailored to the specific stages of this case, effectively guides the user's attention to follow the model's thinking rhythm and synchronously understand the requirement processing process, greatly enhancing the user's sense of control and participation in the AI ​​design logic, and significantly optimizing the overall efficiency of information acquisition.

[0086] In one embodiment, step S60, namely constructing incremental prompt words based on adjustment instructions and sending them to the large language model, includes the following steps:

[0087] S601. Obtain the current session snapshot when the user triggers the adjustment command. The session snapshot records the inference results of the completed stages.

[0088] S602. Associate the adjustment instruction with the completed stage results in the session snapshot, construct a breakpoint redirection instruction and send it to the large language model so that the large language model calls the cached inference state and regenerates the subsequent inference data stream from the position corresponding to the instruction.

[0089] In this embodiment, the current session snapshot refers to the real-time mirror record of all current context states of the large language model at the moment the user triggers a feedback action. This snapshot includes the token sequence generated by the model, the internal hidden layer state vector, and key-value cache (KVCache) data. The breakpoint redirection instruction is a prompt structure that instructs the large language model to change the reasoning direction after a specific logical node while preserving some valid historical context. The cached reasoning state refers to the intermediate computational values ​​stored in memory by the model, corresponding to confirmed correct reasoning stages. By calling these states, the large language model does not need to recalculate from the initial input.

[0090] To facilitate understanding, the processing procedure is illustrated using a specific application scenario as an example: Taking the design of a liquid glass-style takeout card as an example, when showcasing key creative elements such as summer mango chiffon cake during the creative direction thinking stage, if the user feels that this direction does not meet the brand's expectations, they can click the "Need Adjustment" button on the interface. At this point, S601 is executed, immediately capturing a snapshot of the current session. This snapshot fully records the mobile adaptation requirements and core functional positioning of the takeout card, which were confirmed during the problem analysis stage. Subsequently, the user enters specific adjustment instructions in the pop-up input box, for example: the creative direction needs to be changed from a mango color to a retro-textured dark brown tone. S602 is executed, semantically associating this instruction with the data in the snapshot, identifying the logical breakpoint where the adjustment action occurred at the beginning of the creative direction thinking stage. Based on this, a breakpoint redirection instruction is constructed and sent to the large language model. This instruction explicitly guides the model to call the cached problem analysis stage inference state in the snapshot. After receiving the instruction, the large language model directly skips the already reached consensus on requirements and starts anew from the creative conception stage to generate a new original inference data stream.

[0091] The specific processing procedure is as follows: When a user discovers a deviation in the model's direction during the creative direction stage (e.g., deems the summer style too childish), clicks "adjustment needed" and enters an adjustment command to enhance the business feel, the system immediately captures a snapshot of the current session. The system logically aligns this incremental command with the historical inference results in the snapshot, constructs a breakpoint redirection command, and sends it to the model. The model, by calling the cached inference state, directly generates a data stream that meets the new requirements from the creative direction stage, and automatically refreshes the display points for the corresponding and subsequent stages.

[0092] In this embodiment, by acquiring session snapshots and constructing breakpoint redirection instructions, precise intervention and partial updates to the large model inference process are achieved. Since the large language model is allowed to directly access the cached inference state, the waste of computational resources caused by regenerating the entire model from scratch due to minor modifications is effectively avoided, significantly reducing user waiting time. This breakpoint redirection mechanism establishes a dynamic closed loop of thought process, user feedback, and model optimization, enabling users to achieve a collaborative interaction mechanism of thinking, feedback, and optimization simultaneously during the design process, significantly improving the flexibility of human-computer collaborative design and the adaptability of the final interface solution.

[0093] In terms of the security and consistency design of interactive feedback, the interactive feedback module sets strict logical range limits on the adjustment commands input by users.

[0094] It's also worth noting that when a user clicks the adjustment button for a specific stage (such as the creative direction thinking stage), the pop-up simplified input box will undergo front-end validation and back-end semantic filtering to ensure that the input box only supports adjustment commands related to the current stage. For example, during the creative conception stage, users will be restricted from inputting commands to change the terminal type (which belongs to the problem analysis stage).

[0095] In this embodiment, by limiting the scope of the instruction's effectiveness, the disorder of the reasoning chain caused by cross-stage interference from user instructions is effectively avoided. This controlled interaction mode of partial intervention and full-chain update ensures that the generation of incremental prompts always remains within legal logical boundaries, significantly improving the stability and controllability of content generation.

[0096] In one embodiment, a structured display system for the reasoning process in interface generation is provided, with reference to... Figure 2 As shown, system 70 includes: a parsing module 710, a splitting module 720, a refining module 730, a rendering module 740, and an interactive feedback module 750, wherein:

[0097] The parsing module 710 is used to acquire user-inputted UI design prompts, perform semantic analysis and requirement decomposition on these prompts, and generate a structured parsing result containing at least one of the following: requirement type, style constraints, and functional requirements. In this process, UI design prompts refer to the user's UI design intent input through natural language, such as: design a liquid glass-style takeout card. Semantic analysis and requirement decomposition refer to the process of segmenting and semantically annotating the prompts using natural language processing (NLP) technology combined with a professional thesaurus in the field of UI design (such as mobile layout, flat design, responsive design, etc.).

[0098] Specifically, the parsing module 710 can construct a requirement priority determination model and use machine learning algorithms to identify core requirements (such as interface type and core functions) and secondary requirements (such as color preferences and detailed decorations) in prompts, and label constraints such as terminal type and compatible size. The structured parsing result refers to the parsing product outputting the above elements in standard data formats such as JSON, including fields such as requirement type (such as delivery card), style constraints (such as liquid glass material), and functional requirements (such as delivery fee display logic). By parsing ambiguous natural language instructions into structured parsing results containing requirement type, style constraints, and functional requirements, the parsing module 710 establishes a benchmark for the model's understanding of requirements. This structured transformation can accurately capture users' explicit and implicit needs, providing a digital fulcrum for the calibration of subsequent reasoning logic, significantly improving the accuracy of requirement understanding, and reducing generation bias caused by misinterpretation of requirements.

[0099] The splitting module 720 is used to obtain the raw inference data stream output by the large language model for interface design prompts and to identify the stage transition markers in the raw inference data stream. Based on the structured parsing results and the stage transition markers, the raw inference data stream is automatically mapped to the preset problem analysis stage, the creative direction thinking stage, or the design scheme organization stage.

[0100] The original inference data stream refers to the continuous token sequence generated before the model generates a solution, recording the model's thought process. Stage transition markers define the boundary points where the model's logical thought jumps. During processing, the splitting module 720 first identifies preset explicit separators (such as specific transition symbols) by scanning the token sequence. If not identified, it calculates the vector representation and semantic centroid offset of the sequence. When the offset exceeds a preset threshold, it is identified as a marker. The splitting module 720 then uses a semantic association algorithm to categorize the split inference content into preset stages based on the matching degree: the problem analysis stage outputs requirement positioning and design constraint identification; the creative direction thinking stage outputs core creative ideas (such as summer mango chiffon cake) and their suitability analysis; the design scheme organization stage outputs selection reasons and key component design logic (such as displaying the original price with strikethrough). By identifying the stage transition markers in the original inference data stream and automatically mapping them to preset logical stages, the splitting module 720 can reconstruct the originally chaotic and invisible black-box inference process of the large model into a clear logical chain. This mechanism allows users to intuitively perceive the details of the model's processing in understanding needs, generating ideas, and organizing solutions, greatly enhancing the explainability of the AI ​​reasoning process and user trust.

[0101] The extraction module 730 is used to extract feature items from the raw inference data stream for each stage based on the feature library of the interface design domain, and to refine the feature items into a preset number of core points for the stage using a semantic condensation algorithm. The feature library of the interface design domain refers to a knowledge base that stores professional design vocabulary and parameters (such as spacing, rounded corners, and layout proportions). Feature items refer to key sentences in the inference stream that have practical design guidance significance.

[0102] In specific processing, the extraction module 730 identifies and filters redundant information and logical transition text (such as "in summary," "next we will consider," etc.) in the feature items, retaining keywords and phrases related to the interface design parameters. It then calculates the semantic relevance between the keywords and phrases and the final generated interface parameters, selecting the top 3 to 5 points with the highest relevance as the core points of the stage, with each point preferably not exceeding 15 characters in length. By extracting feature items based on a feature library of the interface design domain and using a semantic condensation algorithm for extraction, the extraction module 730 achieves high-density compression of massive amounts of reasoning text. This strategy eliminates redundant verbal noise, retaining only the core information that substantially contributes to design decisions, allowing users to instantly grasp the core design logic and significantly reducing the time cost and cognitive load of sifting through large blocks of text.

[0103] The rendering module 740 is used to synchronously present the core points of each stage and the corresponding stage progress animation on the user interface according to the progressive order of the stages. The progressive order of the stages refers to the logical sequence of problem analysis → thinking about creative directions → organizing design solutions. The stage progress animation includes the exit effect of the current stage (such as fade in and fade out) and the entrance effect of the next stage (such as sliding and zooming).

[0104] The rendering module 740 assigns unique icons (such as magnifying glasses, light bulbs, and blueprint icons) to each stage to enhance recognizability. Key points for each stage are presented in a word-for-word format, highlighted, and in a card-like structure. Key design terms (such as mobile e-commerce and liquid glass) are displayed in prominent colors, and the percentage of completion for each stage is shown simultaneously. By presenting key points and progress animations in a sequential manner according to the stages, the rendering module 740 can visualize abstract thought processes into dynamic, interactive flows. This presentation method aligns with human cognitive logic. Combined with unique icons and keyword highlighting, it transforms passive reading into visual perception, significantly enhancing user engagement and guiding users to understand the requirement processing process in sync with the model's rhythm.

[0105] In other words, as an example, the rendering module 740 mainly implements the following:

[0106] First, the stage transition animation: a progressive animation effect is used, with the three stages presented in the order of problem analysis → thinking about creative direction → organizing design solutions. The previous stage exits in a fade-in and fade-out manner, and the next stage enters in a sliding and zooming manner, while displaying the stage name and progress bar simultaneously, allowing users to clearly perceive the progress of the thinking process.

[0107] Second, the core points are presented with animation: each core point is presented with an animation that appears word by word and is highlighted. Keywords (such as mobile e-commerce, minimalist style, and three-column layout) are highlighted in different colors to attract users' attention. At the same time, the information is displayed in the form of cards, with each core point corresponding to an independent information card. The cards have a slight floating effect to enhance the visual hierarchy.

[0108] Third, visual aids for design: each stage is assigned a unique icon (e.g., a magnifying glass icon for the problem analysis stage, a light bulb icon for the creative direction stage, and a blueprint icon for the design scheme stage), which enhances the recognizability of each stage by combining icons and text; at the same time, as the thinking process progresses, the percentage of completion of each stage is displayed synchronously, allowing users to anticipate the progress of their thinking.

[0109] The interactive feedback module 750 is used to obtain user adjustment instructions for the core points of the currently displayed stage. Based on the adjustment instructions, incremental prompts are constructed and sent to the large language model to trigger updates to the inference data stream for the corresponding stage and subsequent stages. Adjustment instructions refer to the feedback entered by the user after clicking the button requiring adjustment at a certain stage (e.g., adding a retro style for creative direction). Incremental prompts are prompts constructed based on the instructions and session snapshots (model state records).

[0110] The Interactive Feedback Module 750 utilizes a breakpoint redirection mechanism to allow the model to recall cached inference states and regenerate from the corresponding instruction position, eliminating the need to start from scratch. This module also provides features such as a stage navigation bar, details expansion / collapse, and progress control (pause / continue). By acquiring adjustment instructions and constructing incremental prompts to send to the model, the Interactive Feedback Module 750 achieves an instantaneous dynamic closed loop between the user and the large model's thought process. This breakpoint redirection approach allows users to intervene in real-time during the generation process and perform targeted optimizations, avoiding the waste of resources from starting over when the final result does not meet expectations, and significantly improving the flexibility and efficiency of human-computer collaborative design.

[0111] In summary, the interactive feedback module 750 mainly includes the following functions:

[0112] First, the stage jump function: users can directly jump to any completed stage through the stage navigation bar on the interface, view the corresponding key information, and support expanding the details of the stage information (to view the complete reasoning text of the stage) or collapsing the details (to retain only the core points);

[0113] Second, real-time feedback function: Each stage has two feedback buttons: "Meets expectations" and "Needs adjustment". When the user clicks "Needs adjustment", a simplified input box pops up, supporting the input of targeted adjustment instructions (such as "XX requirement was missed in the problem analysis" or "Retro style needs to be added to the creative direction"). The adjustment instructions are synchronized to the large language model in real time, triggering the optimization of the thinking process at the corresponding stage.

[0114] Third, progress control function: Supports pause and continue buttons, allowing users to pause the animation during the thinking process, carefully review key information at the current stage, and avoid missing information due to the animation speed being too fast.

[0115] In one embodiment, the interactive feedback module 750 is specifically configured to: provide a stage navigation bar to allow users to jump to any completed stage to review the corresponding key points of the stage in response to user operations, and to support expanding or collapsing the key points of the stage; set a "meeting expectations" button and a "needs adjustment" button for each stage; pop up a simplified input box to receive user input of specific adjustment instructions for the current stage in response to the triggering of the "needs adjustment" button; and provide a progress control button with pause and resume functions to control the playback progress of the original inference data stream display process.

[0116] In this embodiment, the stage navigation bar refers to a positioning index component set parallel to the reasoning flow in the user interface. It is used to record and guide users to switch views between logical nodes such as problem analysis, thinking about creative directions, and organizing design schemes. The core points of a stage refer to high-density short sentences of information that have been semantically condensed, representing the model's design conclusions at a specific stage. The expand or collapse operation refers to the user's dynamic adjustment mechanism for information; when expanded, it displays the complete text flow of the original reasoning from the large language model; when collapsed, it retains only the extracted key points, such as 15 characters or less. The "Meets Expectations" and "Needs Adjustment" buttons are interactive touchpoints used to collect user subjective evaluations and trigger subsequent logical branches. The input box is a text input component specifically designed for the current logical stage, used to receive incremental instructions. The progress control button is a global switch that controls the execution state of the animation rendering engine.

[0117] It should be understood that during the interface generation process, the interactive feedback module essentially provides users with a full-lifecycle monitoring and intervention environment. The system displays the current thought process progress via a stage navigation bar on the left or top of the interface. For example, when the system enters the stage of thinking about creative directions, the light bulb icon in the navigation bar lights up. If users have questions about previously completed problem analysis stages, they can click the corresponding magnifying glass icon in the navigation bar to rewind the view. During the rewind process, users can click the "Expand Details" button for a specific key point (such as identifying mobile adaptation limitations). The system will then retrieve the corresponding raw inference data stream from the background cache for display, allowing users to review the underlying logic of the larger model.

[0118] Below the key points display area for each stage, the system simultaneously presents two function buttons: "As Expected" and "Needs Adjustment." If the creative direction proposed by the large model is a summer mango chiffon cake, but the user feels that the style is too lively and does not match the brand's tone, they can click the "Needs Adjustment" button. At this time, the system will respond to this trigger action and pop up a concise input box at the current card position. The user can enter specific adjustment instructions in the box, such as: the creative direction should be changed from a mango color tone to a dark blue tone with a business feel. After this instruction is received, it will be synchronized to the large language model in real time, triggering the large language model to use the conversation snapshot to redirect the breakpoint and regenerate the subsequent inference content.

[0119] Furthermore, considering that the rendering module uses progressive animation to present information, to prevent users from missing key design parameters (such as corner radius values ​​or spacing ratios) due to excessively rapid information updates, the system provides a progress control button in a prominent position on the interactive interface. After the user clicks the pause button, the current token stream display and progress bar scrolling will be immediately suspended. The system will resume the display of the subsequent original inference data stream only after the user has carefully reviewed the data and clicked the continue button.

[0120] In this embodiment, by providing a stage navigation bar, details expand / collapse, and real-time feedback buttons, the system breaks the deadlock of traditional AIGC software where users can only passively receive the final generated results, constructing a highly transparent and interactive collaborative design environment. The stage jump function meets the user's need to quickly switch between different logical stages; the combination of the adjustment button and the simplified input box enables precise intervention in the intermediate thought process, avoiding the huge cost of starting over due to minor deviations. The progress control button design fully considers the rhythm of human information acquisition, ensuring the efficient transmission of high-value design parameters. Through the above interactive mechanisms, the system realizes a dynamic closed loop of thought process, user feedback, and model optimization, greatly enhancing the user's sense of control and participation in the AI ​​design process, and significantly optimizing the overall efficiency of human-computer collaborative design.

[0121] Based on the description of the above method and system embodiments, the specific interactive performance and processing logic of the present application solution will be introduced next as an example, in conjunction with the interface display sequence in the accompanying drawings.

[0122] Please see Figures 3-6 As shown, on the current dialogue page, when the user enters the interface design prompt "I want a takeout card with a liquid glass style" in the bottom input box (e.g., input is limited to 0 / 5000 characters and the corresponding deep thinking mode can be enabled), the system immediately starts the parsing module.

[0123] In the initial stage of logical processing, as shown in the appendix Figure 3 As shown, the system is currently in the result return stage: problem analysis. At this time, the problem analysis icon in the progress sidebar on the left is dynamically loading. The system is identifying the business attributes in the prompt words based on the feature library of the interface design domain to establish the design benchmark for the takeaway catering card.

[0124] Subsequently, as shown in the attached image. Figure 4 As shown, when the problem analysis marker in the sidebar switches to the completed state with a checkmark, the status bar changes to the "Return to Results" stage: "Thinking about Creative Directions." At this stage, the system's extraction module extracts key tags from the reasoning flow, such as "Summer Mango Chiffon," "Original Price Displayed with Strikethrough," "Discount Highlighted," "Labeled Hints," and "Product Description Emphasizing Ingredients," and displays them in the form of card-like animations.

[0125] Next, as shown in the attached diagram Figure 5As shown, the thought process moves into the structured preparation of the results. The system's refinement module further structures and maps creative tags into five hierarchical dimensions: features (e.g., 5 items in the limited-time flash sale section), display (e.g., 4 items in the recommended product section), interaction (e.g., 6 items with added shopping attributes), visual (e.g., automatic switching of carousel images), and price. This division of dimensions makes the layout of the card structure in the large model highly readable.

[0126] During the rendering process of the solution, as shown in the attached diagram... Figure 6 As shown, the main display area shows a dynamic progress bar indicating 10% progress in card generation during the result image generation process. Ultimately, the system will present a complete liquid glass-style takeout card with detailed text descriptions: The design adopts a modern minimalist style, combining flat takeout icons with soft illustrative strokes, resulting in a professional yet approachable overall look. The card specifically showcases a product called "Coconut Latte," for example, indicating a price of $7.85, and includes an orange group-buying button. Key delivery and rating metrics such as 25 minutes, 1.6km, 4.8 stars, and 200+ sold are also displayed at the bottom.

[0127] It's also worth noting that a history panel is located on the right side of the interface. The system can categorize and store past records according to time (e.g., today, yesterday, within 7 days, within 30 days). Users can search for records by entering folder names in the search box or rename specific records by clicking the edit button.

[0128] In summary, this application embodiment restructures and visualizes the thought process of the large model feedback, breaking it down into three core stages: problem analysis, creative direction thinking, and design scheme organization. Only the most critical core information of each stage is presented, and animated display and interactive feedback are used. This reduces the user's information processing cost and clearly presents the thought logic chain of the large model, making the thought process more organized, visualized, and efficient. It improves the user's information acquisition efficiency, sense of participation, and trust, reduces the time cost of repeated trial and error, reduces the running time of the system in an invalid state, and significantly shortens the overall response delay from receiving instructions to outputting adaptation results, thereby improving interactive response performance and resource utilization.

[0129] It should be understood that the sequence number of each step in the above embodiments does not imply the order of execution. The execution order of each process should be determined by its function and internal logic, and should not constitute any limitation on the implementation process of the embodiments of this application.

[0130] For specific limitations of the above system, please refer to all functions or steps in the method embodiment section above, which will not be repeated here.

[0131] In one embodiment, a computer device is provided, wherein the processor of the computer device is used to provide computing and control capabilities, the computer device including a memory, a processor, and a computer program stored in the memory and executable on the processor, wherein the processor executes the computer program to implement the steps of the structured representation method as described in any of the preceding embodiments.

[0132] In one embodiment, a computer-readable storage medium is provided having a computer program stored thereon, which, when executed by a processor, implements the functions of the user filtering control or the steps of the target user determination method described above.

[0133] Those skilled in the art will understand that all or part of the processes in the methods of the above embodiments can be implemented by a computer program instructing related hardware. The computer program can be stored in a non-volatile computer-readable storage medium, and when executed, it can include the processes of the embodiments of the methods described above. Furthermore, any references to memory, storage, user databases, or other media used in the embodiments provided in this application can include non-volatile and / or volatile memory.

[0134] Those skilled in the art will clearly understand that, for the sake of convenience and brevity, the above-described division of functional units and modules is used as an example. In practical applications, the above functions can be assigned to different functional units and modules as needed, that is, the internal structure of the system can be divided into different functional units or modules to complete all or part of the functions described above.

[0135] The above-described embodiments are only used to illustrate the technical solutions of this application, and are not intended to limit them. Although this application has been described in detail with reference to the foregoing embodiments, those skilled in the art should understand that modifications can still be made to the technical solutions described in the foregoing embodiments, or equivalent substitutions can be made to some of the technical features. Such modifications or substitutions do not cause the essence of the corresponding technical solutions to deviate from the spirit and scope of the technical solutions of the embodiments of this application, and should all be included within the protection scope of this application.

Claims

1. A structured display method for the reasoning process in interface generation, characterized in that, The method includes: Obtain user-inputted interface design prompts, perform semantic analysis and requirement decomposition on the interface design prompts, and generate a structured parsing result containing at least one of the requirement type, style constraints, and functional requirements; Obtain the raw inference data stream output by the large language model for the interface design prompt words, and identify the stage transition marker bits in the raw inference data stream; Based on the structured parsing results and the stage transition markers, the original reasoning data stream is automatically mapped to the preset problem analysis stage, creative direction thinking stage, or design scheme organization stage. For each stage, feature items are extracted from the original inference data stream based on the feature library of the interface design domain, and the feature items are refined into a preset number of core points of the stage using a semantic condensation algorithm. According to the progressive order of the stages, the core points of each stage and the corresponding stage progress animation are presented synchronously on the user interface. Obtain the user's adjustment instructions for the core points of the currently displayed stage, construct incremental prompt words based on the adjustment instructions and send them to the large language model to trigger the update of the inference data stream for the corresponding stage and subsequent stages; The identification of stage transition marker bits in the original inference data stream includes: Scan the token sequence in the original inference data stream and identify the preset explicit delimiter as the stage transition marker; if the explicit delimiter is not identified, calculate the vector representation of the continuous token sequence and determine the position where the semantic centroid offset of the vector representation exceeds the preset threshold as the stage transition marker. The step of automatically mapping the original inference data stream to a preset stage based on the structured parsing result and the stage transition marker includes: The semantic association algorithm is used to calculate the matching degree between the reasoning content at each stage and the different requirements in the structured parsing results; The original inference data stream is semantically segmented based on the matching degree and the stage transition flag, and then categorized into the corresponding preset stage; wherein: The output of the problem analysis phase includes core requirement identification, potential requirement discovery, and design constraint identification. The output of the creative direction thinking stage includes the core creative direction, design logic, and adaptability analysis. The outputs of the design scheme stage include the rationale for the final scheme selection, the core architecture of the scheme, and the design logic of key components.

2. The structured display method as described in claim 1, characterized in that, The key points of the stage of extracting the feature items into a preset number of items using a semantic condensation algorithm include: Identify and filter redundant information and logical transition text in the feature items, and retain keywords and phrases related to interface design parameters; Calculate the semantic relevance between the filtered keywords and the final generated interface design parameters, and select a preset number of key points as the core key points of the stage based on the semantic relevance.

3. The structured display method as described in claim 1, characterized in that, The synchronous presentation of the core points of each stage and the corresponding stage progress animation on the user interface includes: The stage transition animation is executed in the order of problem analysis, thinking about creative directions, and organizing design schemes. The stage transition animation includes the exit effect of the current stage and the entrance effect of the next stage. When displaying the key points of the stage, the independent information cards corresponding to the key points of the stage are displayed simultaneously, and the design keywords in the key points of the stage are highlighted.

4. The structured display method as described in claim 1, characterized in that, The step of constructing incremental prompt words based on the adjustment instructions and sending them to the large language model includes: Obtain a snapshot of the current session when the user triggers the adjustment command; the snapshot records the inference results of the completed phases. The adjustment instruction is associated with the completed stage results in the session snapshot, a breakpoint redirection instruction is constructed and sent to the large language model, so that the large language model calls the cached inference state and regenerates the subsequent inference data stream from the position corresponding to the instruction.

5. A structured display system for the reasoning process in interface generation, the system being used to implement the method as described in claim 1, characterized in that, The system includes: The parsing module is used to obtain the interface design prompts input by the user, perform semantic analysis and requirement decomposition on the interface design prompts, and generate a structured parsing result containing at least one of the requirement type, style constraints and functional requirements. The splitting module is used to obtain the raw inference data stream output by the large language model for the interface design prompt words, and to identify the stage transition marker bits in the raw inference data stream; based on the structured parsing results and the stage transition marker bits, the raw inference data stream is automatically mapped to the preset problem analysis stage, the creative direction thinking stage, or the design scheme organization stage. The extraction module is used to extract feature items from the original inference data stream based on the feature library of the interface design domain for each stage, and to extract the feature items into a preset number of stage core points using a semantic condensation algorithm. The rendering module is used to synchronously present the core points of each stage and the corresponding stage progress animation on the user interface in the order of the stages. The interactive feedback module is used to obtain the user's adjustment instructions for the core points of the currently displayed stage, construct incremental prompt words based on the adjustment instructions, and send them to the large language model to trigger the update of the inference data stream for the corresponding stage and subsequent stages.

6. The system according to claim 5, characterized in that, The interactive feedback module is specifically used for: Provide a stage navigation bar to respond to user operations and jump to any completed stage to review the corresponding stage's key points, and support expanding or collapsing the key points of the stage. Set up buttons that meet expectations and buttons that need adjustment at each stage; In response to the triggering of the adjustment button, a simplified input box pops up to receive the user's specific adjustment instructions for the current stage; A progress control button with pause and resume functions is provided to control the playback progress of the original inference data stream display process.

7. A computer device comprising a memory, a processor, and a computer program stored in the memory and executable on the processor, characterized in that, When the processor executes the computer program, it implements the steps of the structured presentation method as described in any one of claims 1-4.

8. A computer-readable storage medium storing a computer program, characterized in that, When the computer program is executed by a processor, it implements the steps of the structured presentation method as described in any one of claims 1-4.

Images