AI task operating system with built-in dual-model thinking ability evaluation

Abstract

The present invention discloses an AI task operating system with an embedded dual-model for thinking ability evaluation, which supports natural language and graph structure language input, and drives the AI ​​model to execute the reasoning path through the Prompt+JSON+TP control language, so as to achieve structuring of task information and clear expression of goals. It includes a TP scoring model and a TI scoring enhancement module, which respectively score the basic adaptability and structural adjustment ability of the task path, forming a dual-model task scheduling mechanism. The scoring fusion generates an enhanced thinking ability score TP′, and the task execution controller TE‑Control performs path judgment and scheduling. Combined with the output structure of the T‑Level output control module and the feedback mechanism of the CSI cognitive scoring and feedback module, an intelligent task execution closed loop is constructed to improve the adaptability and output matching of complex tasks and enhance the consistency of AI's response to user goals. The system has closed the scoring and scheduling closed loop and has the ability to self-evolve.

Description

Technical Field

[0001] The present invention relates to the field of artificial intelligence application systems, and more specifically, to an intelligent task management operating system with an embedded dual-model for thinking ability evaluation, namely, an AI task management operating system (AIOS), which is used to enhance the cognitive reasoning ability, execution control ability, and dynamic adaptability of AI in complex task environments. Background Art

[0002] Existing AI systems often rely on cloud-based model training and task execution, which poses privacy risks and model instability. This invention utilizes a three-in-one architecture design to run the AIOS task control system locally, enabling rapid adaptation and autonomous training of open-source AI models.

[0003] With the rapid development of large AI models in areas such as natural language processing, task reasoning, and decision support, traditional AI task management systems are facing a series of bottlenecks and challenges, mainly reflected in the following aspects:

[0004] (1) Uncertainty and unstructured problems of task input: Most existing AI task management systems rely on users to describe task requirements in natural language. However, natural language has problems such as ambiguous expression, implicit structure, and logical jumps, which makes it easy for AI systems to have misunderstandings when parsing task intentions, thereby affecting the accuracy of reasoning path generation and execution control.

[0005] (2) Lack of standardization and dynamic adjustment capabilities in reasoning path generation: Existing systems often generate static reasoning paths based on single inputs, making it difficult to dynamically optimize reasoning flows based on task execution feedback. They lack the ability to adjust reasoning paths, reallocate resources, and call modules based on execution status, resulting in insufficient system execution flexibility and weak ability to adapt to complex environmental changes.

[0006] (3) The cognitive reasoning process lacks an internal quality control mechanism: During the execution of traditional AI tasks, there is a lack of an internal cognitive reasoning quality assessment system, which makes it impossible to evaluate the reasoning depth, reasoning speed, and reasoning accuracy in real time. As a result, the system is prone to accumulate reasoning deviations in multiple rounds of reasoning, which ultimately affects the reliability of the output results and user experience.

[0007] (4) Lack of standardized structural input and reasoning optimization support tools: Currently, there is a lack of a unified tool or language system that can input task information in a structured and standardized manner, and cooperate with natural language parsing to support reasoning path generation, dynamic adjustment and feedback optimization.

[0008] To address the above problems, the present invention proposes an AI task management operating system (AIOS) with an embedded thinking ability evaluation dual model. Summary of the Invention

[0009] The present invention aims to address the shortcomings of existing technologies by proposing an AI task operating system (AIOS) with an embedded dual-model for thinking ability evaluation. The AIOS system utilizes a task input mechanism that combines natural language and standardized logical graph language. Integrating an internal NLP semantic parsing engine, a task scheduling engine module, and a thinking ability scheduling and feedback optimization module, the system uses structured task language to guide the AI ​​to activate its three core thinking dimensions: thinking speed (the efficiency of task response execution), thinking depth (the logical hierarchy of path construction), and thinking accuracy (the degree to which the output matches the user's goal). These dimensions are quantitatively evaluated through a TP thinking ability scoring mechanism, enabling intelligent closed-loop control from task understanding, reasoning path generation, dynamic feedback adjustment, to final task output. The logical graph language input mechanism serves as an auxiliary tool to enhance the structure and standardization of the task input stage, providing the system with clear task module relationships, execution paths, and feedback control logic, thereby optimizing the overall task reasoning efficiency and cognitive control capabilities of the AIOS system.

[0010] The object of the present invention is achieved through the following technical solution: an AI task operating system with an embedded thinking ability evaluation dual model, the system comprising:

[0011] A task input module is used to receive a natural language description and / or a logical graph language structure input by a user and generate standardized task input information;

[0012] The NLP semantic parsing module is used to perform semantic parsing on natural language input, extract task objectives, logical relationships, and execution conditions, and integrate the parsing with the logical graph structure to generate a unified standardized task input structure;

[0013] The task scheduling engine module is used to generate control language instructions based on the parsed task structure and build the reasoning path;

[0014] The thinking power scheduling and feedback optimization module is used to comprehensively evaluate thinking depth TD, thinking speed TS, and thinking accuracy TA based on the thinking power evaluation model TP, and dynamically adjust the reasoning path and module call according to the scoring results;

[0015] The structured output processing module is used to generate output content of different accuracy and cognitive depth based on the cognitive output results after reasoning is completed;

[0016] User behavior database module, used to record user operation behavior, feedback information and cognitive preference data during task execution;

[0017] The cognitive collaboration module is used to collaborate with the underlying AI model to perform task reasoning and cognitive output through structured input, reasoning control and dynamic feedback optimization.

[0018] Furthermore, the task input module further includes:

[0019] The natural language input module is used to receive the task description information expressed by the user in natural language and extract the task objectives, logical relationships and execution conditions through the NLP semantic parsing module;

[0020] The logic diagram language input module is used to receive task nodes, execution paths, control logic and parameter conditions expressed in the form of standardized logic diagrams.

[0021] Furthermore, the task scheduling engine module further includes:

[0022] The control language generator TML is used to generate internal control language instructions based on the standardized task input structure. The control language instructions define task node calls, execution order, condition judgment, exception handling and backup path settings;

[0023] The reasoning path construction module is used to construct a reasoning path according to the control language instruction, wherein the reasoning path includes a main reasoning path and a backup reasoning path and assigns corresponding path weights.

[0024] Furthermore, the thinking power evaluation model TP in the thinking power scheduling and feedback optimization module includes:

[0025] Thinking Depth TD score, used to evaluate the hierarchical depth of task node calls and the rigor of reasoning logic during the reasoning process;

[0026] Thinking speed TS score is used to evaluate the time consumption of reasoning path execution and reasoning response efficiency;

[0027] Thinking accuracy TA score, used to evaluate the consistency between reasoning output results and task goals or user expectations;

[0028] Dynamic weight adjustment mechanism, used to dynamically adjust the weight coefficients of TD, TS, and TA indicators according to task type, execution stage, or user preferences, so that the total thinking power score TP_Score optimally matches the task execution requirements;

[0029] The total thinking ability score TP_Score is calculated according to the following formula:

[0030] TP_Score=λ1×TD+λ2×TS+λ3×TA

[0031] Where, λ1+λ2+λ3=1 is the weight coefficient corresponding to the dynamically adjusted TD, TS, and TA;

[0032] The dynamic adjustment mechanism is used to dynamically adjust the flow of reasoning paths during task execution based on reasoning feedback results, changes in thinking ability scores, or abnormal trigger conditions. This includes switching to alternative paths, reallocating reasoning resources, or adjusting the execution order.

[0033] Furthermore, the thinking power scheduling and feedback optimization module further includes:

[0034] The thinking ability enhancement mechanism introduces a thinking increment ΔTI based on the score enhancement module TI, and incorporates the thinking increment into the total thinking ability score to form an enhanced thinking ability score TP′, where:

[0035] When insufficient reasoning depth, decreased reasoning accuracy, or a bottleneck in the reasoning path is detected during the reasoning process, the TI module is automatically called;

[0036] The TI module generates new auxiliary reasoning paths, alternative reasoning schemes, or reasoning branches, and sets the corresponding thinking increment ΔTI as supplementary cognitive value to be included in the thinking ability score adjustment; the calculation formula for the enhanced thinking ability score TP′ is:

[0037] TP′ = TP + ΔTI

[0038] Among them, TP represents the original total thinking ability score TP_Score, ΔTI serves as an enhancement scoring factor, and combines the system task complexity and historical task scoring trends to participate in the dynamic adjustment of the weight coefficients of the TD, TS and TA indicators and path optimization in real time during the task structure identification or path reasoning process, forming a dynamic enhancement mechanism for the system thinking ability score TP′.

[0039] Furthermore, the structured output processing module further includes a T-Level output control module, which dynamically determines the cognitive depth and accuracy level of the output content based on the thinking power score TP' result, task complexity and user requirements. The T-Level output control module includes:

[0040] T1 level: Rapid response output based on preliminary reasoning results, with concise output content and response speed as priority;

[0041] T2 level: Adds logical details and conditional analysis based on T1 and provides standard reasoning output;

[0042] T3 level: Combines multi-angle reasoning and auxiliary path analysis to generate deep reasoning output;

[0043] T4 level: Introducing new reasoning content generated by the scoring enhancement module TI to form multi-dimensional comprehensive reasoning output;

[0044] T5 level: Based on historical data, feedback evolution and reasoning results, it outputs a comprehensive analysis solution that combines deep reasoning and multiple dimensions.

[0045] Among them, the T-Level output control module dynamically upgrades or downgrades the output level according to the TP′ score during task execution to adaptively optimize the task output quality and response speed.

[0046] Furthermore, the user behavior database module is used to record the user's input content, execution feedback, operation behavior trajectory and cognitive preference information during the task execution process, extract the user's personalized preference characteristics for reasoning depth, reasoning speed, and output granularity, and use the preference characteristics for task output control and execution path optimization.

[0047] Furthermore, the cognitive collaboration module is used to collaborate with the AI ​​model to perform task reasoning and cognitive output, specifically including:

[0048] A structured reasoning control mechanism is used to transmit standardized task inputs, reasoning paths, and dynamic feedback instructions to the AI ​​model through the control language TML, guiding the AI ​​model to perform reasoning tasks;

[0049] A dynamic thinking power guidance mechanism is used to adjust the reasoning path, switch to the backup path, activate the score enhancement module, and transmit the adjustment instructions to the AI ​​model based on the real-time TP′ score or TP_Score and reasoning feedback;

[0050] An innovative reasoning collaboration mechanism is used to call upon the innovative reasoning capabilities of AI models during the reasoning process to generate new reasoning branches or auxiliary solutions;

[0051] The cognitive feedback optimization mechanism is used to dynamically optimize the reasoning strategy based on the reasoning execution results and user feedback, and continuously improve the overall cognitive reasoning effect.

[0052] Furthermore, the control language generator TML includes a logic diagram language control module, specifically as follows:

[0053] (1) TML-Node, a graph language task node structure, represents the structural action or judgment unit of the task;

[0054] (2) TML-Path, a graph language path connection structure, represents the logical jump and execution order between task steps;

[0055] (3) TML-Control, a graph language control structure, is used to express conditional judgment, exception handling, and path switching;

[0056] (4) Graph language parameter tag structure TML-Tag, which binds task parameters, user expectations and output goals.

[0057] Furthermore, the AI ​​task operating system works with personal PCs and open-source AI model platforms to form a three-in-one intelligent execution architecture;

[0058] The AI ​​task operating system is responsible for task input parsing, thinking power scheduling, control path generation, and feedback closed-loop control;

[0059] Personal PCs have local computing resources, structured task execution memory space, and task management and scheduling units;

[0060] The open source AI model platform conducts two-way communication and task takeover with the AI ​​task operating system through an interface protocol.

[0061] Beneficial effects of the present invention:

[0062] 1. This paper proposes an AI task operating system (AIOS) with an embedded dual-model for thinking ability evaluation. This system supports both natural language and logical graph-structured task input, is compatible with a hybrid mechanism of prompt, JSON, and control language, and integrates a natural language parsing engine, an inference path generation mechanism, and dynamic thinking ability scheduling and feedback optimization mechanisms. This system establishes a complete cognitive closed loop from task input to inference execution, feedback adjustment, thinking ability evaluation, and task output, significantly improving the AI ​​system's task comprehension, inference control, and execution adaptability in complex environments. The logical graph language system, as an auxiliary module, addresses the shortcomings of natural language expression in terms of structure and standardization. By clearly defining task nodes, execution paths, control logic, and parameter labels, it effectively supports the AIOS system's task parsing and dynamic inference control processes.

[0063] 2. This invention adheres to the design principle of "minimum intrusion, maximum activation," without interfering with the underlying model structure or refactoring the user interface. It maximizes the ability of large AI models to autonomously invoke their native thinking resources, achieving intelligent and controllable human-machine task collaboration. It is compatible with mainstream large-scale model systems, including DeepSeek. The core of this invention lies in establishing a mechanism for invoking, evaluating, and optimizing AI's internal "thinking resources," making AIOS a transferable framework that transforms general models into domain-specific intelligent entities, providing enterprises and institutions with a new type of intelligent operating infrastructure offering "thinking as a service." This invention is applicable to a variety of application scenarios requiring complex task management, cognitive reasoning optimization, and dynamic execution adjustment, such as intelligent manufacturing, smart healthcare, education and training systems, and financial intelligent analysis and decision-making.

[0064] 3. This invention utilizes a dual-model system consisting of a TP (Thinking Power Scoring Model) and a TI (Thinking Power Scoring Enhancement Model). TP scoring guides basic path scheduling. When the score is insufficient or a structural bottleneck is encountered, the TI module is dynamically activated to generate an incremental ΔTI. Based on the structural path performance and feedback behavior, an enhanced score TP′ is generated. The TE-Control module performs path judgment and module scheduling. This, combined with the T-Level output control module and the CSI cognitive feedback channel, creates an intelligent closed-loop task execution system, providing a path foundation for optimizing complex task structures and continuously enhancing model capabilities. This mechanism enhances the flexibility of task path generation and the matching of results, forming the intelligent core of this system's thinking power control and execution scheduling.

[0065] Therefore, the "enhancement" in the TI module is a path regeneration mechanism triggered by a structural scheduling mechanism that can be recognized and executed by controllable language logic. It is controllable, explainable and reproducible, and constitutes the technical core of the system of the present invention to achieve structural optimization and improve the adaptability of complex tasks. BRIEF DESCRIPTION OF THE DRAWINGS

[0066] In order to more clearly illustrate the embodiments of the present invention or the technical solutions in the prior art, the following briefly introduces the drawings required for use in the embodiments or the description of the prior art. Obviously, the drawings described below are only some embodiments of the present invention. For ordinary technicians in this field, other drawings can be obtained based on these drawings without creative work.

[0067] Figure 1 This is a schematic diagram of the principle architecture of the AI ​​task operating system with embedded thinking ability evaluation dual models.

[0068] Figure 2 Generate a schematic diagram of the TP scoring mechanism and pathway.

[0069] Figure 3 This is a schematic diagram of the basic structure of the logic diagram language system.

[0070] Figure 4 A standard flow chart for inputting logic diagram language into AI systems.

[0071] Figure 5 Schematic diagram of the conversion mechanism between logical graph language and hybrid coding language.

[0072] Figure 6 Figure 2 is the control structure diagram for TP′ enhancement scoring and TE path scheduling.

[0073] Figure 7 This is the logic diagram of the output quality evaluation and intelligent callback system.

[0074] Figure 8Output and user configuration mechanism logic diagram for structured tasks.

[0075] Figure 9 Logic diagram of the joint control system for local AI cognition execution.

[0076] Figure 10 Local adaptive training logic diagram driven by user business behavior. DETAILED DESCRIPTION

[0077] In order to make the purpose, technical solution and advantages of the present invention more clearly understood, the present invention is further described below with reference to the accompanying drawings and examples. It should be understood that the specific examples described herein are only used to explain the present invention and are not intended to limit the present invention.

[0078] like Figure 1 As shown, the system's full-process control logic in key links such as user task input, semantic parsing, structural scoring, path scheduling, AI model calling and cognitive feedback is demonstrated. In the task input stage of the AI ​​task operating system AIOS with an embedded thinking ability evaluation dual model provided by the present invention, in addition to the natural language input method, the system also supports task construction and execution process expression in the form of a structured logic diagram language. Users can use the logic diagram language to construct a task graph, including task target nodes, judgment condition nodes, operation execution nodes and path connections and control symbols between them, and attach personalized task parameters such as execution level and output mode through parameter labels. The system of the present invention uses a dual thinking ability evaluation mechanism composed of a TP scoring module and a TI scoring enhancement module (a second thinking ability evaluation model) as the central scoring engine, combined with the task execution controller TE-Control module, the Prompt generation module and the cognitive feedback module to realize intelligent selection of task execution paths, structural control and output optimization. Figure 1 It reflects the complete closed loop of the system "input-scoring-control-execution-feedback". The dual-model scoring mechanism is the basic scoring component for reasoning and scheduling. Its detailed structure is shown in Figure 2 . Figure 2This paper demonstrates the task scoring and path control mechanism of this system based on the dual thinking power evaluation model. After the user input is processed by the NLP semantic parsing engine, the initial execution strategy is generated and scored by the first thinking power model TP. The scoring results are preliminarily scheduled by the TE-Control module based on the "entry path judgment logic". If the task complexity or goal is unclear, the second thinking power model TI will be activated for enhanced scoring to form the TP' result, guiding the Prompt module to generate more accurate task instructions, thereby realizing the closed-loop execution of path control and feedback scheduling. The TE-Control module contains two control algorithms: (1) a multi-round dynamic adjustment algorithm, which is mainly used for path scoring and feedback scheduling during task execution; (2) an adaptive algorithm, which is used for structure identification and path initialization in the task input stage. The two jointly constitute the task execution controller (TE-Control), which realizes the closed-loop control of early task understanding and mid-term scheduling.

[0079] The specific implementation of the present invention is as follows:

[0080] After receiving the logic graph language structure, the system uses a logic graph language parsing engine to identify and analyze the node structure and path logic. First, the parsing module extracts the node types and their connections within the graph, identifying control symbols (such as sequence, branch, and loop) to form a logical execution sequence. The parameter translator then processes the parameter tags into a system-recognizable format (such as a JSON fragment) and feeds them, along with the structure path, into the hybrid coding language generator.

[0081] The control language generator generates a Prompt fragment based on the graph structure content, and generates a composite structure language of Prompt+JSON after fusing the parameter fields. The structured control language is sent to the AI ​​large model for inference execution through the Prompt generation module.

[0082] The core structure of the Task Management Language (TML) generator in the system of the present invention includes:

[0083] Structural parsing unit: accepts logic graph language structure input and identifies task nodes, paths, and parameters;

[0084] Prompt template selector: dynamically selects prompt templates based on task type, target output level (T-Level) and TP score results;

[0085] Parameter field fusion module: inserts user input parameters, default task constraints, and graph structure path information into the Prompt template;

[0086] Structured output generator: Outputs a composite control language structure in the form of "Prompt instruction + JSON control parameters"; the generated result is transferred from the Prompt generation module to the AI ​​model interface for executing structured reasoning tasks.

[0087] Logic diagram language control module, as follows:

[0088] (1) TML-Node, a graph language task node structure, represents the structural action or judgment unit of the task;

[0089] (2) TML-Path, a graph language path connection structure, represents the logical jump and execution order between task steps;

[0090] (3) TML-Control, a graph language control structure, is used to express conditional judgment, exception handling, and path switching;

[0091] (4) Graph language parameter tag structure TML-Tag, which binds task parameters, user expectations and output goals.

[0092] During system operation, the CSI cognitive scoring and feedback module scores the AI ​​output results, and the scoring results are fed back to the logical graph language module to update the mapping rules and parameter translation logic, thereby achieving dynamic optimization of the graph structure execution strategy.

[0093] Through this implementation, the logic diagram language module forms a closed-loop execution path with the TP scoring mechanism, TE-Control module, and Prompt generation module in the AIOS backbone system, realizing the structured task control logic of the entire process from task input to execution feedback.

[0094] The logic diagram language mechanism in this invention, as a means of expressing task execution manuals, primarily helps users structure task objectives, execution modules, and control paths. It is a quick and easy language tool for AI task management systems. By combining logic diagram language with natural language, complex task logic relationships can be quickly and accurately converted into a system-recognizable structured control language (TML). This enables the AI ​​system to quickly understand and accurately mobilize its thinking resources, generating comprehensive solutions that meet user expectations.

[0095] The core innovation of this invention lies in: dynamically evaluating task complexity, reasoning depth and output quality through the TP thinking power scoring mechanism, and constructing the thinking path and scheduling process by the TE-Control module to achieve intelligent control and quality feedback closed loop of the task execution process.

[0096] Compared with existing technical solutions centered on model structure control, graph learning reasoning, or model scrambling protection, the present invention has essential differences in design principles and execution logic.

[0097] Example 1: Description of the overall architecture of the AIOS system:

[0098] This embodiment provides an overall architecture of an AI task management operating system (AIOS) with an embedded thinking ability evaluation dual model, aiming to achieve integrated intelligent closed-loop control of task input parsing, reasoning path generation, cognitive reasoning optimization, dynamic feedback adjustment, and thinking ability evaluation. The AIOS system mainly includes the following core modules: Figure 1 As shown:

[0099] (1) Task input module:

[0100] It supports users to describe task requirements through natural language, while also introducing a standardized logic diagram language input method. The logic diagram language defines task nodes, execution paths, control logic, and parameter labels, assisting natural language input to form a structured and standardized task input system.

[0101] (2) NLP semantic parsing module:

[0102] Perform semantic analysis on natural language input to extract key elements such as task objectives, logical relationships, and execution conditions, and form a fusion analysis with the logic graph language input to generate a preliminary structured task description.

[0103] (3) Task scheduling engine module:

[0104] Based on the parsed task structure, the Task Management Language (TML) generator generates internal executable control instructions and constructs an inference path. The inference path includes a primary path, backup paths, and weight settings, supporting dynamic feedback adjustment.

[0105] (4) Thinking power scheduling and feedback optimization module:

[0106] A built-in TP thinking ability evaluation model monitors in real time the reasoning depth (TD) associated with the logical hierarchy of path construction, the reasoning speed (TS) associated with the execution efficiency of task responses, and the reasoning accuracy (TA) associated with the degree to which output results match user goals. The reasoning path is dynamically adjusted based on the TP score, and the score enhancement module (TI module) is invoked when necessary to increase the reasoning depth or create a new reasoning path.

[0107] (5) Structured output processing module:

[0108] Based on the cognitive output results after reasoning is completed, combined with the cognitive output hierarchy mechanism of the T-Level output control module, output content with corresponding accuracy and cognitive depth is generated according to the task complexity and user demand level.

[0109] (6) User behavior database module (UBB module):

[0110] Record the user's operational behavior and feedback correction information during task execution as an important data source for the system's adaptive cognitive evolution.

[0111] (7) Cognitive collaboration mechanism module (AIOS + AI > 2):

[0112] The "cognitive collaboration module" refers to a module system in the system that is collaboratively composed of task structure expression (graph language module), reasoning path scheduling (TP module), cognitive output generation (T-Level), cognitive feedback evaluation (CSI cognitive scoring and feedback module) and parameter adjustment mechanism. It is used to support the AI ​​model in the entire process of structural understanding, reasoning execution and output feedback of cognitive tasks.

[0113] Through the collaboration of the above modules, the AIOS system not only passively responds to user commands, but also actively calls the thinking resources of the AI ​​model based on the task structure to achieve cognitive intelligence improvement, and the overall cognitive ability exceeds the level of a single AI model.

[0114] The overall architecture design forms the following closed-loop process:

[0115] [Task Input] → [Semantic Analysis] → [Inference Path Construction] → [Dynamic Scheduling of Thinking Power] → [Cognitive Output Generation] → [User Feedback Collection and Adaptive Evolution]

[0116] The system uses a built-in dynamic control mechanism for thinking power and a local cognitive adaptation mechanism to ensure intelligent task processing, controllable reasoning, and high-quality output results in changing environments. The local cognitive adaptation mechanism uses the UBB user business behavior database as a foundation, combining output control parameters with graph language mapping rules to achieve continuous adaptation and structural optimization of user cognitive preferences.

[0117] Example 2: Task input and structured parsing mechanism implementation:

[0118] This embodiment describes in detail the specific implementation of the task input and structured parsing mechanism in the AIOS system. In order to solve the problems of vague description, logical jumps and parsing difficulties in the task input stage of traditional AI systems, this invention adopts a dual-channel task input mode that combines a natural language input module with a standardized logic diagram language input module, which effectively improves the structuring degree and parsing accuracy of task descriptions. Figure 3, showing the core structural system of the logic graph language, including four basic units in task expression: task node unit (M1), path connection unit (M2), control symbol unit (M3) and parameter label unit (M4). Figure 3 Block A defines the modules that make up the logical diagram language structure, and clearly presents the origin and classification principles of each unit. Figure 3 Block B provides a typical example of logic diagram language expression, demonstrating the path flow structure of the task process from data analysis and condition judgment to operation execution. Output level L2 is marked with a parameter label, indicating the output control level corresponding to the structural statement. Figure 3 As the structural basis of the AIOS logic graph language module, core mechanisms such as reasoning path planning and control language (TML) generation constitute the underlying expression specifications of the logic graph language system in the system of the present invention.

[0119] (1) Natural language task input

[0120] Users submit task requirements through natural language, including but not limited to:

[0121] 1. Task Goal

[0122] 2. Execution Conditions

[0123] 3. Input Parameters

[0124] 4. Output Requirements

[0125] 5. Special Constraints

[0126] The system's internal NLP semantic parsing engine performs the following processing on natural language text:

[0127] Keyword extraction: Identifying mission-critical words and concepts.

[0128] Logical relationship identification: analyzing the causal, sequential, and dependency relationships between tasks.

[0129] Structure extraction: abstract the task description into a parsable data structure (such as a task tree or task graph).

[0130] (2) Logical graph language task input

[0131] To make up for the shortcomings of natural language in structured expression, the system also supports logical graph language input, including:

[0132] Node Definition: Define each subtask or operation unit.

[0133] Execution Path: describes the logical flow relationship between task nodes, including sequential execution, parallel execution, conditional jumps, etc.

[0134] Control Logic: Set trigger conditions, judgment conditions, execution constraints, etc.

[0135] Parameter Tags: mark the parameter information required or output by the task node.

[0136] Logical graph language is input into the system in the form of structured data (such as JSON or graph structure model) and integrated with the natural language parsing results. Figure 4 , demonstrating the standard conversion process of logical graph language from structural input to hybrid language generation. The system uses three types of input elements, "graph language structure + parameter label + control logic", to uniformly construct the task structure through the structure builder, symbol generator and parameter embedder. The constructed graph language structure will enter the mapping rule library and parser and other processing modules to generate Prompt fragments and JSON control parameters, which will be integrated to form a structured control language, that is, a hybrid language format of Prompt + JSON. The final control language is sent to the local AI model execution module through the Prompt + JSON output module to trigger the task reasoning process. The reasoning results are fed back to the CSI cognitive scoring and feedback module for rule modification and parameter optimization, forming a closed-loop "structure generation → reasoning execution → output evaluation" path.

[0137] (3) Fusion parsing mechanism

[0138] The system uses a fusion parsing mechanism to align and integrate the natural language parsing structure with the logical graph language structure to form a unified initial reasoning structure. The specific steps are as follows:

[0139] 1. Modularize the natural language parsing results and map them into the four elements of a graph language structure: node, path, control logic, and parameter tag, which serve as the basic input for structured control language.

[0140] 2. Map the parsed modules to logic graph nodes and paths;

[0141] 3. Provide supplementary analysis and automatic reasoning for ambiguities in the mapping process;

[0142] 4. Generate a standardized task structure as the basis for subsequent reasoning path generation and execution control.

[0143] (4) Standard input example

[0144] For example, a user submits a natural language description: "Analyze customer data, classify customers by age and purchase frequency, and predict future purchase potential." At the same time, they submit the following node and path relationships defined in a logical graph language:

[0145] Node 1: Data Collection → Node 2: Customer Segmentation (Based on Age) → Node 3: Customer Segmentation (Based on Purchase Frequency) → Node 4: Purchase Potential Prediction

[0146] The modular split results are as follows:

[0147] Node: Customer data, age, purchase frequency, purchase potential

[0148] Path: Starting from "customer data", analyze "age" and "purchase frequency" in turn to form a classification path, and finally point to "predicted purchase potential"

[0149] Control logic: If age < 30 and purchase frequency is high → high potential customer; otherwise, average customer

[0150] Parameter tags: classification threshold = age 30, frequency threshold = 5 times per month

[0151] After system fusion and analysis, a standardized initial reasoning structure is formed for subsequent processing by the control language generator (TML).

[0152] Example 3: Control language generation and reasoning path construction:

[0153] This embodiment describes in detail the specific implementation of the control language (TML) generation and reasoning path construction mechanism in the AIOS system.

[0154] In order to achieve efficient conversion of task analysis results into AI internal executable instructions, this invention introduces a control language generator (TML: Task Management Language) and uses an inference path construction mechanism to ensure that the task execution process has high flexibility, high adaptability and dynamic adjustment capabilities. Figure 5, demonstrating the complete mechanism path for converting a logical graph language structure into an AI-executable hybrid language format (Prompt + JSON). After the user enters the logical graph language content, the system first constructs a structured task graph through the node builder, control symbol generator, and parameter label embedder. It then enters the graph structure parser, mapping rule library, and parameter translator to generate the structural semantics, prompt fragments, and JSON control fields corresponding to the logical graph. The generated Prompt + JSON structure is transmitted to the local AI model (such as DeepSeek) via the Prompt + JSON output module for inference execution. The inference results are fed back to the CSI cognitive scoring and feedback module, which is further used for rule optimization and parameter correction, building a two-way closed-loop mechanism of "graph structure → hybrid language → model inference → CSI feedback."

[0155] (1) Control language generation (TML generator) is specifically implemented as follows: The core functions of the control language generator include:

[0156] Instruction structuring: The task structure after fusion analysis is converted into a standardized control instruction set. The instructions include basic operation units such as task call, condition judgment, loop execution, and exception handling.

[0157] Semantic fidelity: Ensure that the control language strictly retains the original semantics and execution intent of the natural language and logic diagram language input during the conversion process.

[0158] Extensible design: The control language supports module expansion, facilitating the subsequent addition of new task modules and reasoning module instructions.

[0159] The control language format example is as follows (simplified version):

[0160] json

[0161] "TaskID": "T001",

[0162] "TaskName": "Customer Data Classification",

[0163] "ExecutionFlow": [

[0164] {"Node": "Data Collection", "Action": "Collect", "Parameters": {"DataSource": "CRM database"}},

[0165] {"Node": "Age Classification", "Action": "Classify", "Criteria": "Age"},

[0166] {"Node": "Purchase Frequency Classification", "Action": "Classify", "Criteria": "PurchaseFrequency"},

[0167] {"Node": "Purchase Potential Prediction", "Action": "Predict", "Model": "Potential Prediction Model"}

[0168] ]}

[0169] After the control language instructions are formed, the inference path construction stage begins.

[0170] (2) The specific implementation of the reasoning path construction mechanism is as follows:

[0171] The main task of the reasoning path construction mechanism is to generate a reasoning path diagram that meets the task logic flow and execution conditions based on the control language instruction set. The construction process includes:

[0172] Main path generation: Generates the main inference path based on the logical order of the tasks.

[0173] Backup path design: Configure backup paths for critical mission nodes to automatically switch execution when reasoning fails, anomalies occur, or feedback deviations occur, ensuring high system robustness.

[0174] Path weight allocation: Weights are assigned to multiple possible paths. The system dynamically selects the optimal path based on the thinking ability score during the reasoning process.

[0175] Dynamic Adjustment Point Setting: Dynamic adjustment points (Adjustment Points) are set up in the inference path, allowing the system to adjust the inference flow in real time based on feedback information during execution. "Dynamic Adjustment Points" in this system refer to the structural node positions preset in the logic graph language path structure. These nodes are used to perform conditional jumps or parameter adjustments on the execution flow after receiving feedback input (such as TP scores, CSI results, and user behavior preferences) during inference execution. Adjustment points are usually set at key nodes or parameter branch nodes in the path, and their functions include:

[0176] 1. When the scoring result deviates from the preset threshold, the original path is interrupted and the alternative branch path (such as the TP′ path) is switched.

[0177] 2. Adjust the T-Level output structure based on CSI feedback and restructure the output method;

[0178] 3. When the task structure does not fully match the user goal, inject UBB business behavior characteristics and reconstruct the path parameters.

[0179] The reasoning path is managed within the system in the form of a directed graph, where each node (task unit) and edge (execution relationship) is marked with a condition and a scoring indicator.

[0180] (3) Reasoning path example (simplified version)

[0181] [Data Collection] → [Age Classification] → [Purchase Frequency Classification] → [Purchase Potential Prediction]

[0182] ↘ (Alternative) [User Interest Category]

[0183] The main path performs data collection → classification → prediction in sequence;

[0184] If an exception occurs in the purchase frequency classification node, the backup path is switched to execute the user interest classification module.

[0185] (4) Reasoning Path and Dynamic Feedback Linkage

[0186] The reasoning path is not static. The system will make real-time adjustments during execution based on dynamic feedback (such as changes in TP thinking ability scores and task execution deviations). This includes adjusting the order of nodes, dynamically calling backup modules, reconfiguring parameters and conditions, and increasing or decreasing the depth or breadth of reasoning.

[0187] This mechanism ensures that the AIOS system can maintain flexibility and efficiency in reasoning even in complex environments.

[0188] Example 4: Thinking Ability Evaluation Model (TP Model) and Dynamic Reasoning Guidance:

[0189] This embodiment describes in detail the thinking ability evaluation model (TP model) embedded in the AIOS system and its application mechanism in dynamic reasoning guidance.

[0190] In order to achieve real-time cognitive evaluation and reasoning quality control during the reasoning process, the present invention designs a thinking ability evaluation system (TP model), and dynamically guides the reasoning path and module scheduling based on the scoring results, thereby improving the overall cognitive reasoning ability and execution adaptability of the system.

[0191] (1) Design of TP thinking ability evaluation model

[0192] The TP (Thinking Power) model is the core cognitive evaluation engine of the AIOS system and consists of the following three basic indicators:

[0193] TD (Thought Depth) measures the hierarchical depth and logical rigor of node calls in the reasoning path.

[0194] TS (Thought Speed) measures the time or computing resource consumption required to complete a complete reasoning path, reflecting reasoning efficiency.

[0195] TA (Thought Accuracy) measures the consistency between the reasoning output results and the task objectives and user expectations, reflecting the quality of reasoning.

[0196] Each indicator is scored separately and assigned different weights according to the nature of the task, which together form the TP score.

[0197] (2) TP comprehensive scoring formula

[0198] The formula for calculating the total TP score is as follows:

[0199] TP_Score=λ1×TD+λ2×TS+λ3×TA

[0200] Among them: λ1, λ2, λ3 are dynamically adjustable weight factors, satisfying: λ1+λ2+λ3=1

[0201] The weight factors can change dynamically according to task requirements, execution stage, and user preferences.

[0202] For example, in the "customer behavior prediction task", if the task is in the initial data modeling stage, the system will pay more attention to the depth of reasoning (TD), and the weight factors can be set to λ1 = 0.6, λ2 = 0.3, and λ3 = 0.1;

[0203] When the system enters a real-time response scenario (such as an online recommendation system), the system pays more attention to the inference speed (TS), and the weights can be dynamically adjusted to λ1 = 0.2, λ2 = 0.6, and λ3 = 0.2;

[0204] If the user explicitly requires readable output, the output accuracy (TA) ratio is increased and adjusted to λ1 = 0.3, λ2 = 0.2, and λ3 = 0.5.

[0205] The above weight factors change dynamically according to the task stage, preference records in the user behavior database, and output feedback indicators, thereby improving the fit and interpretability of the TP score to task performance.

[0206] The TP weight factor is fine-tuned through the CSI cognitive scoring and feedback module:

[0207] CSI=α×PMR+β×(1-MIS)+γ×OHR

[0208] Among them, α, β, and γ are dynamic weights that can be adjusted according to the task type. PMR (Prompt Matching Rate) measures the degree to which the output matches the task objective input by the user; MIS (Missing Information Score) measures the degree to which key information is omitted from the output; and OHR (Output Heuristic Richness) evaluates the heuristic, depth, and diversity of the output. The system also incorporates a 3W module (What, Why, When) to analyze the task objective, motivational context, and time requirements contained in user input. This module, embedded as a preprocessing mechanism during the task input phase, improves the contextual adaptability of TP evaluation and prompt generation, ensuring accurate semantic alignment of the task.

[0209] The output results of the 3W module will be written into the input preprocessing buffer of the TP model as one of the reference factors for the integrity of the task semantic structure in the scoring calculation, thereby completing the detection and enhancement of semantic clarity and intention integrity before TD, TS, and TA scoring, forming one of the important pre-links of the TP model task scoring mechanism.

[0210] (3) Dynamic reasoning guidance mechanism

[0211] The scoring results of the TP model are directly applied to the inference execution control, and the specific guidance method is as follows:

[0212] Dynamic adjustment of inference paths

[0213] When the TP_Score drops below the preset threshold, the system automatically triggers the reasoning path adjustment mechanism, including: switching to an alternative reasoning path; adjusting the order of reasoning nodes or adding auxiliary nodes; activating the score enhancement module (TI module) to introduce a new reasoning branch.

[0214] Dynamic scheduling of module tasks

[0215] Based on the TP scores of different inference stages, the most suitable module or model resources for the current state are dynamically selected to optimize inference efficiency and output quality.

[0216] Adaptive parameter adjustment

[0217] Automatically adjust reasoning parameters (such as node weights, path priorities, and resource allocation ratios) to improve the cognitive adaptability of the reasoning process.

[0218] Enhanced thought triggers

[0219] When reasoning reaches a bottleneck (such as multiple continuous declines in scores), the system activates the TI module based on the TP score, calls on the innovative reasoning capabilities of the AI ​​model, and generates new reasoning directions or supplementary solutions.

[0220] (4) Dynamic guidance example process

[0221] 1. The initial reasoning path is executed according to the standard task structure;

[0222] 2. Real-time monitoring of TD, TS, and TA indicators during the execution process;

[0223] 3. When the TS (speed) score drops below the threshold, it indicates that the reasoning efficiency is insufficient;

[0224] 4. The system adjusts the reasoning path, skipping low-priority nodes or executing some tasks in parallel;

[0225] 5. During the reasoning process, if the TA (accuracy) decreases, the system calls the backup reasoning module or the scoring enhancement module;

[0226] 6. Dynamically adjust until TP_Score reaches the set threshold again to ensure the quality of the output results.

[0227] Supplementary Notes on System Intelligent Evolution and Adaptive Enhancement Capabilities

[0228] The proposed AIOS task management operating system not only establishes a controllable task guidance system but also embodies the developmental characteristics of "self-evolution" and "intelligence enhancement" in its system capability design. The core of this capability is based on the dynamic scheduling mechanism of the TP thinking model. This enables the system to identify reasoning bottlenecks, cognitive biases, and path deviations in real time during task execution, and automatically invoke the scoring enhancement module (TI) for supplementary reasoning.

[0229] Furthermore, AIOS provides a "dynamically schedulable task-thinking resource guidance system" operating paradigm. By collecting behavioral feedback through the UBB module and using the TP model's scoring to drive path reconstruction, it enables the evolution from a static execution system to an intelligent feedback system. This system architecture features sustainable optimization, autonomous adjustment, and intelligent evolution. This concept is reflected in the system's support for multi-module coordinated control and a self-optimization mechanism based on cognitive quality.

[0230] Description of the Intelligent Evolution Control Module

[0231] In order to enhance the system's ability to control and schedule the inference path structure during task execution, the present invention adds an intelligent evolution control module.

[0232] The intelligent evolutionary control module, located between task input parsing and module scheduling, is responsible for optimizing the structured control language (TML) and managing path priorities. This module receives scoring feedback from the TP model and behavioral preference data from the UBB business behavior database. Based on the scoring weights, task feedback data, and policy rules, it dynamically constructs multiple execution paths and prioritizes them. Policy rules include built-in path optimization rules (such as prioritizing the highest-scoring path based on TP scores), user preference guidance policies (such as prioritizing fast branch paths based on UBB data), and task history feedback adjustment policies (such as re-prioritizing execution based on CSI scores). This policy set can be set and dynamically updated by developers and forms the core basis for system path scheduling.

[0233] The intelligent evolution control module has the following key functions:

[0234] 1. Path combination construction capability: Automatically construct multiple combinations of main paths, backup paths, and innovative paths;

[0235] 2. Execution order optimization mechanism: Dynamically adjusts the path execution order based on TP / CSI score feedback;

[0236] 3. Scheduling central control logic: coordinates module task calls, resource allocation, and path return strategies.

[0237] The intelligent evolution control module serves as the core scheduling module for the AIOS system's path strategy evolution. Working with the TP and TI modules, it builds a complete intelligent reasoning chain control mechanism, significantly enhancing the system's adaptability and path evolution capabilities.

[0238] Description of the User Evolution Data Management Module

[0239] The User Evolution Data Management Module, as a component supporting intelligent evolutionary scheduling embedded within the AIOS system, is specifically designed to long-term preserve, schedule, and manage behavioral data and path strategy data during the system's task evolution. This module covers both user behavior evolution information (output from the local guidance driver module) and path optimization scheduling data (output from the intelligent evolutionary control module). Its functions are as follows:

[0240] 1. The user evolution data management module uses a built-in lightweight local database (such as SQLite) to create a unified data storage platform and manage multiple sub-table structures, including the user behavior log table (user_behavior_log), the path combination record table (path_evolution_log), and the system space status table (meta_db_status).

[0241] 2. The module has the ability to monitor evolving data storage, automatically prompting users to perform cleanup operations when the database space reaches a set threshold, and regularly executes redundancy strategies and invalid record cleanup mechanisms to ensure the durability and efficiency of system operation.

[0242] 3. The module implements an access control mechanism for all evolution data. Users can view its historical evolution records, but have no right to directly edit or delete the original records, ensuring the objectivity and traceability of intelligent evolution results.

[0243] 4. This module coordinates data writing and reading between the local guidance drive module and the intelligent evolution control module. As the back-end support module for the execution of intelligent scheduling strategies, it ensures the continuous optimization of the system's long-term intelligent behavior, the stable migration of user preferences, and the gradual evolution and accumulation of task strategy paths.

[0244] The user evolution data management module effectively fills the gap in the lack of persistent learning and memory mechanism in the existing AI task scheduling system, enabling the AIOS system to have long-term intelligence accumulation capabilities, providing a personalized evolution support path for large models, and significantly improving the system's intelligent task execution efficiency and adaptability, with significant innovation and practical value.

[0245] Description of the task input integrity and output expected matching control module

[0246] To ensure that the AIOS system can obtain complete and accurate task input information from users during task execution and ensure that the AI ​​output results are highly consistent with user expectations, the system introduces a task input integrity and output expectation matching control module.

[0247] This module uses natural language interaction and structured input templates to guide users during the task initiation phase, filling out key fields such as the task name, description, objectives, and background information. Furthermore, the system incorporates an intelligent prompt mechanism that automatically determines whether input completion or objective updates are necessary at key points in the task execution process. It then generates appropriate prompts to assist users in refining their task expression and establishing a clear structure for their "user intent."

[0248] The task input integrity and output expectation matching control module integrates the TP thinking ability evaluation mechanism, the task execution feedback module, and the system intelligent evolution control module to form a closed-loop input-output feedback chain. By real-time evaluating and correcting output deviations at the source, this module establishes a dynamic tuning mechanism within AIOS, centered on tasks and anchored by user expectations. This design is one of the key technical paths to achieving a "highly matched comprehensive solution" and serves as a foundational guarantee for AIOS's autonomous optimization and alignment with user cognition.

[0249] Task input integrity and output expected matching control module and Figure 2 The TE-Control module in the system jointly builds a task scheduling closed loop, which is a key component for realizing task structure evolution and model path regeneration, and also constitutes one of the essential differences between AIOS and other static AI control systems.

[0250] The present invention constructs a dual thinking ability evaluation model system composed of TP and TI. The TP model uses three indicators, namely task thinking depth TD, thinking speed TS and thinking accuracy TA, for weighted scoring to form a basic score TP. When the system identifies a structural bottleneck or insufficient score, the TI model is dynamically activated to generate an incremental ΔTI for the enhanced score, which supplements the original score and ultimately forms an enhanced score TP′ = TP + ΔTI. The TP′ score result is input into the TE-Control module to drive path judgment and task scheduling. TP provides a stable main score, and TI provides structural enhancement correction to form an intelligent collaborative mechanism for scoring and execution. The cognition of the CSI module is not a general cognition, but is concretely manifested as: whether AIOS can accurately and efficiently mobilize the AI ​​model's own thinking resources (the three elements of TA / TD / TS) based on the nature of the user's task, and ultimately generate a comprehensive solution that meets user expectations. CSI is the feedback bridge for this process. (See for details. Figure 2 and Figure 6 ).

[0251] Example 5: Thinking Ability Enhancement and Dynamic Scoring System:

[0252] This embodiment describes in detail the design and implementation of the thinking ability enhancement and dynamic scoring system in the AIOS system.

[0253] In order to further optimize the quality of reasoning, enhance the system's cognitive depth and creative reasoning ability, this invention designs a thinking ability enhancement mechanism based on the TP basic thinking ability model, and constructs a dynamic scoring system throughout the entire reasoning process, realizing real-time adaptive control during the reasoning process. Figure 6This paper demonstrates the synergistic relationship between the basic scoring model TP and the thinking enhancement module TI, and how the conditional scoring enhancement path (TP′) drives the execution process of the task path judgment and language generation modules. The TP model provides a comprehensive weighted score based on three dimensions: task depth (TD), execution speed (TS), and thinking accuracy (TA), which serves as the primary basis for task execution scoring. If the TP score is insufficient, the system dynamically activates the TI module to generate a structural enhancement score ΔTI, which is integrated with the TP to form an enhanced score TP′, which serves as the conditional input to the path judgment logic. It is important to note that the TI module and the ΔTI path are not invoked for every task scoring. The TP′ score it generates is a "conditional enhancement path" and is only triggered when the TP score is insufficient, thus implementing an intelligent scoring reinforcement mechanism. The TP or TP′ scoring results are used by the TE-Control module to make path decisions, guiding the prompt generation module to generate a structural or innovative output task path.

[0254] (1) The design of the thinking ability enhancement mechanism is as follows:

[0255] When the system detects that the reasoning depth is insufficient, the reasoning path reaches a bottleneck, or the accuracy decreases, it automatically calls the scoring enhancement module (TI module) to inject control language containing alternative path suggestions, nonlinear thinking parameters, or multi-round mutation prompts into the AI ​​model, thereby guiding the AI ​​model to generate more creative new paths or complete the reasoning chain.

[0256] The cognitive increments generated by the scoring enhancement module are received and evaluated by the TP scoring system to form an enhanced scoring TP′, thereby achieving an intelligent improvement in the system's output capabilities.

[0257] (2) TP′ (Enhanced Thinking Skills Score) Formula

[0258] The system thinking ability enhancement mechanism introduces the cognitive increment ΔTI to modify and enhance the original TP score to form the TP′ score:

[0259] TP′ = TP + ΔTI

[0260] Where TP = λ1·TD + λ2·TS + λ3·TA is the original scoring structure, and ΔTI is dynamically calculated by the TI module when the path is generated, reflecting the scoring bonus of the newly added cognitive value.

[0261] In the extended structure, it can also be expressed as:

[0262] TP′ = λ1·TD + λ2·TS + λ3·TA + λ4·ΔTI, where λ1+λ2+λ3+λ4=1

[0263] When the TI module successfully introduces an increment in thinking and reasoning power, the system adjusts the TD and TA scores based on the impact of the increment to form a new scoring standard for TP′; the more effective the TI, the more obvious the improvement in TP′ scores, thereby encouraging the system to boldly call on thinking power resources when necessary.

[0264] (3) Dynamic scoring system runs through the entire process

[0265] The dynamic scoring system runs continuously during inference execution and mainly includes the following functions:

[0266] Real-time rating monitoring

[0267] Each time a reasoning stage is completed (such as when a reasoning path or a task module is executed), the system immediately updates the TD, TS, TA, and TI indicators and calculates a new TP′ score.

[0268] Rating trend analysis

[0269] Not only does it monitor single-point scores, but it also analyzes the trend of score changes to determine whether the reasoning process is continuously improving, stagnating, or deteriorating.

[0270] Adaptive Inference Tuning

[0271] According to the TP′ score and trend analysis results, the reasoning depth, path complexity, backup path call frequency and scoring enhancement module activation frequency are dynamically adjusted.

[0272] Output accuracy control

[0273] The TP′ score also serves as the basis for the subsequent T-Level output control module (see Example 6), directly affecting the depth and completeness of the output content.

[0274] (4) Example process for enhancing thinking ability

[0275] 1. Real-time calculation of TD, TS, and TA during inference;

[0276] 2. Detecting a decrease in the reasoning depth TD score, activate the TI module;

[0277] 3. The scoring enhancement module generates new auxiliary reasoning paths to supplement the reasoning depth;

[0278] 4. Count the TI increment into the TP′ score and update the reasoning status;

[0279] 5. Dynamically adjust the reasoning strategy to ensure that the final output meets the requirements of cognitive depth and accuracy.

[0280] Example 6: Application of T-Level output control module:

[0281] This embodiment describes in detail the design and application of the cognitive output level mechanism of the T-Level output control module in the AIOS system. Figure 7 , demonstrating the complete operating logic and control flow of the AI ​​intelligent task management operating system with an embedded thinking ability evaluation model. The system takes the task goal input by the user as the starting point, and enters the initial stage of the execution path after passing the structured information extraction and NLP semantic parsing modules. The system uses the TP thinking ability evaluation model to comprehensively evaluate the complexity, depth, speed and accuracy of the task, and decides whether to call the TI thinking ability enhancement module for structural supplementation based on the "satisfaction judgment path". The core of the system is driven by the TE-Control module, which supports multi-round dynamic task scheduling and optimized path selection. In the execution path, the system can automatically switch between ordinary task processing and innovative task processing mechanisms to ensure that the generated Prompt structured prompts meet user expectations. Figure 7 It embodies the closed loop of the entire process of the AI ​​task system, from language analysis to thinking ability assessment, and then to decision control and prompt generation, emphasizing the system's adaptive and innovative capabilities.

[0282] In order to adapt to different task complexities, user demand levels and cognitive depth requirements of reasoning results, this paper proposes a T-Level output control module based on TP' score drive to achieve dynamic hierarchical control of reasoning output and cognitive depth adjustment. Figure 8 , demonstrating the overall process logic of structured task output and user configuration mechanism. The system first receives the user's task input content, including task objectives and specific requirements, and then uses the AIOS-3W user call control mechanism module and T-Level output control module to perform phased analysis of the task objectives and output level matching. On this basis, the system constructs a PSCP output model to generate multi-dimensional structured content according to different task scenarios and user goal types (such as product design, strategic planning, research projects, etc.), including key items such as product design, function introduction, design plan, cost estimation, etc., and presents and adapts them in different levels (L1~L4). Ultimately, the system outputs a comprehensive solution that meets user expectations, achieving accurate correspondence and quality assurance between task content and target requirements.

[0283] (1) Hierarchical system design of T-Level output control module

[0284] The T-Level output control module is a module used for structured output hierarchical scheduling in the system. It controls output paths and screens results based on cognitive priorities, serving as the execution carrier of the cognitive output hierarchical mechanism. The T-Level output control module divides task outputs into different cognitive depth levels, with each level representing an improvement in the output content in terms of logical integrity, reasoning details, and analytical perspectives. The standard division is as follows:

[0285] T1 level: fast response output

[0286] Based on preliminary reasoning, it provides quick but coarse-grained answers. Its fast output makes it suitable for simple tasks or those requiring real-time feedback.

[0287] T2 level: standard reasoning output

[0288] Based on T1, it adds logical details and conditional analysis. The output is more complete and suitable for tasks of general complexity.

[0289] T3 level: deep reasoning output

[0290] Combines multi-angle analysis with auxiliary reasoning paths. Suitable for highly complex tasks or scenarios requiring comprehensive analysis.

[0291] T4 level: multi-dimensional comprehensive reasoning output

[0292] Introducing the reasoning results of the Scoring Enhancement Module (TI). The output covers predictive analysis, hypothetical deduction, and innovative deconstruction, suitable for decision support and complex system modeling.

[0293] T5 level: extremely high cognitive depth output

[0294] Combining historical mission evolution data (UBB module), multiple rounds of feedback adjustments, and innovative reasoning, it generates an optimal comprehensive solution across all dimensions. This approach is suitable for fields with extremely high cognitive requirements, such as high-end think tank analysis and complex strategic planning.

[0295] (2) T-Level determination basis

[0296] T-Level selection is dynamically determined based on the following three factors:

[0297] TP′ score level The higher the TP′, the system has the ability to generate higher-level outputs;

[0298] Task complexity labeling tasks mark complexity during the parsing phase. High complexity tasks default to high T-Level outputs.

[0299] User output requirement setting

[0300] The user can specify the desired T-Level during the task input phase, or the system can automatically infer user preferences through historical interaction behaviors (UBB database).

[0301] (3) Output generation and control process

[0302] 1. The system monitors the changes in TP′ scores during the reasoning process in real time;

[0303] 2. Comprehensively consider task complexity and user requirements to determine the current target T-Level;

[0304] 3. According to the T-Level standard, control the logical depth, detail richness, and number of reasoning branches of the output content;

[0305] 4. If the TP′ score drops during the reasoning process, the T-Level will be automatically downgraded to ensure timely output;

[0306] 5. If the reasoning quality improves or the thinking is enriched, the T-Level will be automatically upgraded to provide higher quality output.

[0307] (4) T-Level application examples

[0308] Example task: Developing a digital transformation plan for a medium-sized enterprise

[0309] T1 output: Provide a list of general processes for digital transformation;

[0310] T2 output: Based on the current situation of the enterprise, list the key steps and precautions;

[0311] T3 output: Analyze the current status of each department of the enterprise and formulate detailed phased transformation goals;

[0312] T4 output: Introduce innovative transformation model cases and predict potential risks and opportunities;

[0313] T5 output: Integrate historical corporate data, industry trends, and innovative thinking results to develop a dynamic and adjustable strategic blueprint.

[0314] Example 7: User Business Behavior Database (UBB) and Local Cognition Adaptive Evolution:

[0315] This embodiment describes in detail the design of the user business behavior database (UBB) module in the AIOS system and the specific implementation of the local cognitive adaptive evolution mechanism. Figure 9-10 , Figure 9 This is the logic diagram of the local AI cognitive execution joint control system. The system parses, schedules and intelligently responds to task content through the collaborative execution system of the locally deployed AI big model and the AIOS task operating system. Figure 9The core mechanism revolves around the TP thinking ability evaluation model. The system uses the TP model to assess the depth, speed, breadth, and accuracy of tasks. Based on this, the CSI cognitive scoring and feedback module further identifies user intent and output accuracy. The CSI cognitive scoring and feedback module generates a cognitive score by weighting three key dimensions (PMR, MIS, and OHR). Internally, the system integrates the Prompt generation module, the Transformers execution module, the TE-Control module, and the T-Level output control module to ensure precise matching and quality control of output content. Furthermore, the system integrates output behavior data through the output integration module and writes it to the UBB user business behavior database, forming a closed feedback loop for subsequent optimization iterations and local training support. The overall process emphasizes deep integration of locally deployed models with the AIOS system to ensure data security, real-time response, and model control, ultimately generating a comprehensive solution that meets user expectations. Figure 10 The logic diagram of local adaptive training driven by user business behavior shows the logic mechanism of AI local adaptive training driven by user business behavior. Based on user task input, the system performs semantic analysis and TP thinking ability scoring through the AIOS task management system, and provides feedback and evaluation of task execution results through the CSI cognitive scoring and feedback module. During the execution process, the system calls the local AI model in real time, dynamically adjusts the task execution instructions through the Prompt generation module and the Transformers execution module, and continuously writes the task behavior data and system response data into the UBB user business behavior database. The local system continuously collects user business behavior data and establishes a cognitive model of user business behavior characteristics through the cognitive behavior model generation module. Based on this model, it realizes personalized adaptation of task scheduling, output preferences and reasoning paths, thereby improving the accuracy of system response and user satisfaction, forming a cognitive behavior data closed loop. Figure 10 The lower-right center section shows how the system performs operations such as "task feature modeling," "cognitive evaluation model generation," and "behavioral data comparative analysis" based on local UBB data. The optimized operating parameters are then injected back into the AIOS system, forming a local self-learning and self-optimizing intelligent training channel. This mechanism ensures the AIOS system's adaptive adjustment capabilities, enabling continuous evolution of model performance based on actual user behavior, improving AI's response accuracy and task completion capabilities in specific enterprise or user scenarios.

[0316] In order to improve the system's cognitive adaptation capabilities to individual user needs and realize the dynamic evolution of task execution strategies, this paper introduces the UBB (User Behavior Base) user business behavior database module. By continuously recording, analyzing and applying user interaction behavior data, it drives the system to perform cognitive adaptive optimization in the local environment.

[0317] (1) User behavior database (UBB module) design

[0318] The UBB module is specifically designed to collect, store, and manage the following types of user behavior data:

[0319] Task Input Record

[0320] Task subject, description, goal and parameter information submitted by the user.

[0321] Task execution feedback

[0322] User evaluation of system output, correction suggestions, and task reconstruction information.

[0323] Operational behavior data

[0324] It includes the user's interface operation trajectory, option selection, path preference, etc. during task execution.

[0325] Cognitive preference information

[0326] Through multiple rounds of interaction, the system automatically infers the user's preferences for inference depth, response speed, and output accuracy.

[0327] (2) Data collection and privacy protection

[0328] When collecting data, the UBB module follows the following design principles:

[0329] Principle of minimum necessity: only collect data necessary to complete cognitive adaptation, and do not accumulate redundant data.

[0330] Local storage priority: Behavioral data is stored in the user's local device or dedicated server first to ensure privacy and security.

[0331] User controllability: Users can view, export, or clear their own behavior data records at any time.

[0332] (3) Local cognitive adaptation evolution mechanism

[0333] Based on the behavioral data accumulated by the UBB module, the AIOS system can perform the following adaptive optimization process:

[0334] Cognitive Preference Modeling

[0335] By analyzing historical task execution behaviors, a user's cognitive preference model is established (e.g., preference for concise answers or detailed deductions, preference for speed or accuracy).

[0336] Adaptive adjustment of inference parameters

[0337] Before executing a new task, the reasoning parameters are dynamically adjusted according to the user's cognitive preferences, including: T-Level default output level; reasoning path complexity; calling frequency of the scoring enhancement module (TI module); and dynamic feedback weight allocation.

[0338] Personalized evolution of execution strategies

[0339] The AIOS system continuously fine-tunes the reasoning process and module calling sequence based on behavioral evolution data, gradually forming an execution style and cognitive logic that is highly consistent with specific user needs.

[0340] Cognitive Evolution Feedback Loop

[0341] After each task is completed, the execution results and user feedback are written into the UBB database again, forming a self-evolving cognitive closed loop, and the system's cognitive adaptability is continuously improved.

[0342] (4) Example application process

[0343] 1. The user completes multiple task submissions and outputs feedback;

[0344] 2. Systematically analyze behavioral data to identify users who tend to have high cognitive depth output (preferring T4 and T5 levels);

[0345] 3. When executing new tasks, the system will automatically improve the depth of reasoning and give priority to introducing enhanced thinking modules;

[0346] 4. After the task is completed, continue to collect user feedback and refine the cognitive preference modeling;

[0347] 5. As the number of uses increases, the system becomes highly synchronized with the user's cognitive style, improving output conformity and user satisfaction.

[0348] Example 8: AIOS and AI model collaboration system:

[0349] Sample Application

[0350] This embodiment describes in detail the implementation and example applications of the AIOS system and AI model collaboration mechanism (i.e., cognitive collaboration system).

[0351] To overcome the problems of insufficient cognitive reasoning depth and limited autonomous adjustment capabilities of traditional AI models in the execution of complex tasks, the present invention achieves a nonlinear leap in overall cognitive intelligence through close collaboration with the underlying AI model through the AIOS system's structured input, dynamic guidance of thinking power, and cognitive evolution mechanism.

[0352] (1) Overall design of the collaboration system

[0353] When the AIOS system collaborates with AI models, the division of labor is as follows:

[0354] AIOS system:

[0355] Responsible for task analysis, reasoning structure construction, dynamic guidance of thinking power and feedback optimization; manage dynamic adjustments in the reasoning process, enhance thinking triggering and cognitive evolution.

[0356] AI models (underlying models, such as DeepSeek, GPT, etc.):

[0357] As a cognitive execution resource, it performs specific reasoning, inference, and content generation according to the control instructions (TML instructions) generated by the AIOS system; it also responds to the AIOS system's module calls, reasoning adjustments, and enhanced thinking trigger requests. The two collaborate to form a complete closed loop, no longer a one-way instruction input, but a structured cognitive process control and dynamic reasoning interaction.

[0358] (2) Collaborative operation process

[0359] 1. Task input stage

[0360] Users input task requirements through natural language + logical graph language, and the AIOS system parses and generates standardized reasoning structures.

[0361] 2. Reasoning preparation stage

[0362] The AIOS system formulates the reasoning path, control instruction set (TML) and sets dynamic feedback monitoring points based on the task structure.

[0363] 3. Inference execution phase

[0364] The AI ​​model gradually completes inference execution according to TML instructions, and the AIOS system monitors the inference depth, speed and accuracy in real time.

[0365] 4. Dynamic adjustment stage

[0366] If an inference deviation or bottleneck is detected, the AIOS system dynamically adjusts the inference path, switches modules, or calls enhanced thinking functions to guide the AI ​​model to continue inference.

[0367] 5. Cognitive output stage

[0368] After inference is completed, the AIOS system determines the output T-Level level based on the TP′ score and controls the output accuracy and depth.

[0369] 6. Feedback learning stage

[0370] User feedback information and task execution data are written into the UBB user business behavior database to optimize the inference strategy before the next inference.

[0371] Through the above process, the AIOS system forms cognitive structure control and reasoning optimization guidance for AI models, making the overall cognitive intelligence level exceed that of a single AI model;

[0372] (3) Example application scenarios

[0373] Example 1: Financial risk control strategy formulation

[0374] Traditional model: Directly ask the AI ​​model for risk control strategies and get a single recommendation or fixed process.

[0375] AIOS collaboration model:

[0376] Analyze the specific financial business background and risk indicator requirements input by the user; dynamically infer and generate multiple alternative risk control strategy paths; screen the optimal risk control combination based on real-time score adjustments; introduce a score enhancement module to generate a predictive risk evolution model; output a multi-dimensional, dynamically adjustable risk management plan, and output at the T4 / T5 level.

[0377] Example 2: Smart medical decision-making support

[0378] Traditional model: The AI ​​model directly provides standardized treatment recommendations.

[0379] AIOS collaboration model:

[0380] Analyze patient medical records, treatment history, and individual differences; build a multi-path diagnosis and treatment reasoning structure; dynamically adjust treatment plan combinations and optimize treatment recommendations based on real-time data feedback; introduce innovative treatment ideas and explore the possibility of new therapies; output multi-plan comparative analysis and dynamic adjustment recommendations, with high cognitive depth to support doctor decision-making.

[0381] (4) Summary of the advantages of the collaborative system

[0382] Improve the quality of task parsing and reasoning structure generation; enhance the dynamic adaptability and innovation of the reasoning process; significantly improve the cognitive depth, logical integrity and user fit of the final output; and form a truly human-computer collaborative and cognitively evolutionary intelligent system.

[0383] Supplementary Note: AIOS system deployment method and interactive interface adaptation

[0384] The deployment strategy of the AIOS system fully reflects the design concept of "minimum intrusion, maximum activation". In the entire system architecture, AIOS does not replace the original functional interface or interaction channel of the AI ​​model. The present invention designs a model adaptation guidance module to guide the system to adapt and connect with different types of AI models after the task structure is generated. Through the structural conversion and input reconstruction mechanism, the universality and low intrusion of model access are achieved, and a unified access channel is provided to support the deployment of multiple models (including open source models and private models). It is a bridge module for the conversion of task execution logic to AI models; the present invention uses an embedded guidance mechanism to dynamically activate the thinking resources within the AI ​​model and dispatch its native reasoning capabilities, thereby outputting a comprehensive solution that better meets user expectations.

[0385] Therefore, during deployment, the AIOS system can seamlessly integrate with large open-source AI models (such as DeepSeek, ChatGLM, and Mistral), eliminating the need to develop new graphical user interfaces (UIs). The system recommends using the native interactive interfaces provided by the AI ​​models themselves (such as web UIs or CLIs) to complete user task interaction and execution at the prompt structure and thought path control level.

[0386] On the one hand, this approach simplifies the AIOS system deployment path and improves system operation stability; on the other hand, it fully reflects the independence of the AIOS system: the core it protects is the intelligent task guidance system based on the "Prompt+Steering+TP evaluation mechanism". The present invention mainly focuses on the method path and scheduling mechanism, and does not involve changes to the UI or model structure itself.

[0387] The AI ​​task operating system proposed in this invention provides controllable, secure, and intelligent collaborative underlying support for future intelligent work platforms and new-quality productivity infrastructure through the integrated application of task control systems, guided open source AI model platforms, and enhanced computing environments. It has broad personalized application potential and social promotion value.

[0388] In summary, the AIOS system can achieve functional enhancements in intelligent task management and reasoning guidance without building an independent UI portal, which is in line with the core idea of ​​"guided intelligent enhancement" of this system.

[0389] The system of the present invention has closed the scoring and scheduling loop and has the ability to self-evolve. The specific indicators are shown in Table 1:

[0390] Table 1

[0391]

[0392] The above embodiments are used to illustrate the present invention rather than to limit the present invention. Any modifications and changes made to the present invention within the spirit of the present invention and the protection scope of the claims shall fall within the protection scope of the present invention.

Claims

1. An AI task operating system with a built-in dual-model for thinking ability evaluation, characterized by: The system includes: A task input module is used to receive a natural language description and / or a logical graph language structure input by a user and generate standardized task input information; The NLP semantic parsing module is used to perform semantic parsing on natural language input, extract task objectives, logical relationships, and execution conditions, and integrate the parsing with the logical graph structure to generate a unified standardized task input structure; The task scheduling engine module is used to generate control language instructions based on the parsed task structure and build the reasoning path; The thinking power scheduling and feedback optimization module is used to comprehensively evaluate thinking depth TD, thinking speed TS, and thinking accuracy TA based on the thinking power evaluation model TP, and dynamically adjust the reasoning path and module call according to the scoring results; The thinking power scheduling and feedback optimization module further includes a thinking power enhancement mechanism, which introduces a thinking increment ΔTI based on the score enhancement module TI, and incorporates the thinking increment into the total thinking power score to form an enhanced thinking power score TP′, wherein: When insufficient reasoning depth, decreased reasoning accuracy, or a bottleneck in the reasoning path is detected during the reasoning process, the TI module is automatically called; The TI module generates new auxiliary reasoning paths, alternative reasoning schemes, or reasoning branches, and sets the corresponding thinking increment ΔTI as supplementary cognitive value to be included in the thinking ability score adjustment; the calculation formula for the enhanced thinking ability score TP′ is: TP′ = TP + ΔTI TP represents the original total thinking ability score TP_Score, and ΔTI serves as an enhancement factor. It combines the system task complexity and historical task score trends to dynamically adjust the weight coefficients of the TD, TS, and TA indicators and optimize the path in real time during task structure identification or path reasoning, forming a dynamic enhancement mechanism for the system thinking ability score TP′. The structured output processing module is used to generate output content of different accuracy and cognitive depth based on the cognitive output results after reasoning is completed; User behavior database module, used to record user operation behavior, feedback information and cognitive preference data during task execution; The cognitive collaboration module is used to collaborate with the underlying AI model to perform task reasoning and cognitive output through structured input, reasoning control and dynamic feedback optimization.

2. The AI ​​task operating system according to claim 1, characterized in that: The task input module further includes: The natural language input module is used to receive the task description information expressed by the user in natural language and extract the task objectives, logical relationships and execution conditions through the NLP semantic parsing module; The logic diagram language input module is used to receive task nodes, execution paths, control logic and parameter conditions expressed in the form of standardized logic diagrams.

3. The AI ​​task operating system according to claim 1, wherein: The task scheduling engine module further includes: The control language generator TML is used to generate internal control language instructions based on the standardized task input structure. The control language instructions define task node calls, execution order, condition judgment, exception handling and backup path settings; The reasoning path construction module is used to construct a reasoning path according to the control language instruction, wherein the reasoning path includes a main reasoning path and a backup reasoning path and assigns corresponding path weights.

4. The AI ​​task operating system according to claim 1, wherein: The thinking power evaluation model TP in the thinking power scheduling and feedback optimization module includes: Thinking Depth TD score, used to evaluate the hierarchical depth of task node calls and the rigor of reasoning logic during the reasoning process; Thinking speed TS score, used to evaluate the time consumption of reasoning path execution and reasoning response efficiency; Thinking accuracy TA score, used to evaluate the consistency between reasoning output results and task goals or user expectations; Dynamic weight adjustment mechanism, used to dynamically adjust the weight coefficients of TD, TS, and TA indicators according to task type, execution stage, or user preferences, so that the total thinking power score TP_Score optimally matches the task execution requirements; The total thinking ability score TP_Score is calculated according to the following formula: TP_Score=λ1×TD+λ2×TS+λ3×TA Where, λ1+λ2+λ3=1 is the weight coefficient corresponding to the dynamically adjusted TD, TS, and TA; The dynamic adjustment mechanism is used to dynamically adjust the flow of reasoning paths during task execution based on reasoning feedback results, changes in thinking ability scores, or abnormal trigger conditions. This includes switching to alternative paths, reallocating reasoning resources, or adjusting the execution order.

5. The AI ​​task operating system according to claim 1, wherein: The structured output processing module further includes a T-Level output control module, which dynamically determines the cognitive depth and accuracy level of the output content based on the thinking power score TP' result, task complexity and user requirements. The T-Level output control module includes: T1 level: Rapid response output based on preliminary reasoning results, with concise output content and response speed as priority; T2 level: Adds logical details and conditional analysis based on T1 and provides standard reasoning output; T3 level: Combines multi-angle reasoning and auxiliary path analysis to generate deep reasoning output; T4 level: Introducing new reasoning content generated by the scoring enhancement module TI to form multi-dimensional comprehensive reasoning output; T5 level: Based on historical data, feedback evolution and reasoning results, it outputs a comprehensive analysis solution that combines deep reasoning and multiple dimensions; Among them, the T-Level output control module dynamically upgrades or downgrades the output level according to the TP′ score during task execution to adaptively optimize the task output quality and response speed.

6. The AI ​​task operating system according to claim 1, characterized in that: The user behavior database module is used to record the user's input content, execution feedback, operation behavior trajectory and cognitive preference information during the task execution process, extract the user's personalized preference characteristics for reasoning depth, reasoning speed, and output granularity, and use these preference characteristics for task output control and execution path optimization.

7. The AI ​​task operating system according to claim 1, characterized in that: The cognitive collaboration module is used to collaborate with the AI ​​model to perform task reasoning and cognitive output, specifically including: A structured reasoning control mechanism is used to pass standardized task inputs, reasoning paths, and dynamic feedback instructions to the AI ​​model through the control language generator TML, guiding the AI ​​model to perform reasoning tasks; A dynamic thinking power guidance mechanism is used to adjust the reasoning path, switch to the backup path, activate the score enhancement module, and transmit the adjustment instructions to the AI ​​model based on the real-time TP′ score or TP_Score and reasoning feedback; An innovative reasoning collaboration mechanism is used to call upon the innovative reasoning capabilities of AI models during the reasoning process to generate new reasoning branches or auxiliary solutions; The cognitive feedback optimization mechanism is used to dynamically optimize the reasoning strategy based on the reasoning execution results and user feedback, and continuously improve the overall cognitive reasoning effect.

8. The AI ​​task operating system according to claim 3, characterized in that: The control language generator TML includes a logic diagram language control module, which is as follows: (1) TML-Node, a logical graph language task node structure, represents the structural action or judgment unit of the task; (2) TML-Path, a logical graph language path connection structure, represents the logical jump and execution order between task steps; (3) TML-Control, a logic diagram language control symbol structure, is used to express conditional judgment, exception handling, and path switching; (4) Logic diagram language parameter tag structure TML-Tag, which binds task parameters, user expectations and output goals.

9. The AI ​​task operating system according to claim 1, wherein: The AI ​​task operating system, personal PC, and open source AI model platform work together to form a three-in-one intelligent execution architecture; The AI ​​task operating system is responsible for task input parsing, thinking power scheduling, control path generation and feedback closed-loop control; Personal PCs have local computing resources, structured task execution memory space, and task management and scheduling units; The open source AI model platform conducts two-way communication and task takeover with the AI ​​task operating system through an interface protocol.

Images