A method and system for verifying office tasks using modal adaptive constraint binding
By adopting a modal adaptive constraint binding method and utilizing a coordinating agent and a verification agent architecture, office tasks are decomposed into sub-tasks and the verification intensity is adaptively adjusted. This solves the defects of modal verification in existing technologies, realizes critical verification and resource optimization, and improves the completion rate and verification accuracy of multimodal tasks.
Patent Information
- Application Number
- CN202611609667.0
- Authority / Receiving Office
- CN · China
- Patent Type
- Applications(China)
- Current Assignee / Owner
- Filing Date
- 2026-10-19
- Publication Date
- 2026-11-17
AI Technical Summary
Existing office task verification methods lack modality specificity, suffer from confirmation bias, subjective ambiguity in verification standards, and lack of adaptability in verification strategies, resulting in the omission of key quality defects and waste of resources.
A modality-adaptive constraint binding method is adopted. By coordinating agents, executing agents, and verifying agents, office tasks are decomposed into sub-tasks. Task constraints are bound according to modality type, and the verification intensity is adaptively adjusted through reinforcement learning to achieve critical verification and resource optimization.
It effectively prevented key quality defects from being missed, enabled critical verification of office tasks and balanced use of resources, and improved the completion rate and verification accuracy of multimodal tasks.
Smart Images

Figure SMS_77 
Figure SMS_78
Abstract
Description
Technical Field
[0001] This invention belongs to the field of intelligent office system technology, specifically relating to a modal adaptive constraint binding method and system for verifying office tasks. Background Technology
[0002] The rapid development of Large Language Models (LLMs) has given rise to an agent-based autonomous task execution paradigm. Modern agent frameworks, such as AutoGPT, MetaGPT, and LangChain Agent, can decompose user natural language instructions into executable steps, invoke tools to complete task processing, and output multimodal result files such as documents, spreadsheets, images, web pages, audio, and video.
[0003] In daily office scenarios, users frequently need to perform cross-modal tasks, such as converting PDF financial statements into Excel files with formulas, generating Word reports or PPTs based on online surveys, generating visualization charts from data, and transcribing audio recordings into meeting minutes. The inputs and outputs of these tasks span different file modalities and have clear professional quality requirements for the outputs. Currently, LLM agents have the ability to perform these tasks independently, but there are systematic deficiencies in the quality verification process. The current verification methods mainly include: (1) implicit self-verification, where the agent reads back the file and makes its own judgment in the same dialogue context; (2) reflection mechanism, where multiple rounds of reflection and improvement are performed within the same inference chain; (3) tool-enhanced verification, where external tools are called for assistance but a general strategy is used without distinguishing modalities; and (4) multi-agent role division, where role separation is achieved, but the verification process relies on programmable signals such as code testing.
[0004] None of the above solutions perform differentiated verification for the specific modal types of the output, resulting in the following core defects: Defect 1: Weak validation lacking modality-specific targeting. Existing validation strategies do not perform targeted matching based on the modality type of the output, only performing shallow, general checks, such as whether the file exists and whether the content is not empty, while ignoring the truly critical quality dimensions under that modality. For example: for Excel output, only confirming that the file can be opened is not checked to see if formulas are used instead of hard-coded values; for image output, only verifying the format is valid, but not checking whether the semantic content of the image is consistent with the task requirements; for document output, only confirming that there is text is not verified to see the accuracy of the chapter structure and content source.
[0005] Defect 2: Confirmation bias. When an agent validates its output within the same dialogue context, it has access to its own reasoning process and tends to confirm its own decisions rather than critically examining the output;
[0006] Defect 3: Subjective and ambiguous verification criteria. Existing technologies use open-ended evaluation problems, which leads to inconsistent judgments. Multiple rounds of verification attempts may yield contradictory conclusions for the same output.
[0007] Defect 4: Fixed verification strategy wastes resources. Existing solutions apply the same verification intensity to all tasks, failing to adapt to task difficulty and historical success rate, resulting in over-verification of simple tasks and under-verification of difficult tasks. Summary of the Invention
[0008] The purpose of this invention is to provide a modal adaptive constraint binding method and system for verifying office tasks, in order to solve the problems of existing office tasks lacking multimodal verification, lacking critical verification, having vague verification standards, and lacking adaptive verification strategies.
[0009] To achieve the above objectives, the technical solution of the present invention is as follows: A modal adaptive constraint binding method for verifying office tasks, comprising the following steps: S1. Input the office task and analyze the input task modality and the task modality to be output; the executing agent adaptively sets risk factors according to the input modality; and then decomposes the office task into several sub-tasks to be executed sequentially; S2. Analyze the modality of each subtask, and bind task constraints to the subtask from the predefined modality-constraint rule table according to the subtask's modality; adaptively add task constraints according to the risk factors of the office task; and convert all task constraints of the subtask into a binary format. S3. Execute all subtasks sequentially, verify the strength level of each subtask, and perform task constraint verification based on the strength level. If any constraint of the current subtask fails verification, the corresponding constraint handling strategy is adaptively triggered according to the subtask modality to complete the subtask's retry, early termination, or task rollback strategy until all subtask constraint verifications are completed and the office task ends.
[0010] Preferably, the subtask execution task constraint verification in S3 includes the following three types: Consistency verification, bypass prevention verification, and correlation verification are performed. Consistency verification involves multiple independent verification agents verifying the subtask constraints. If the number of failed verifications exceeds a preset threshold, the constraint is marked as failed; otherwise, it passes. Bypass prevention verification determines whether the executing agent can still pass the constraint verification even without actually completing all the subtask requirements. If it is determined that the constraint can be circumvented, the constraint verification result is marked as failed; otherwise, it passes. Correlation verification checks whether the current constraint can truly cover the quality dimensions that the task instruction is concerned with, marking irrelevant constraints as failed.
[0011] Preferably, the verification strength level in S3 is denoted as ,in Indicates skipping; This indicates lightweight validation, which means randomly selecting two constraints for validation; This indicates standard verification, which means verifying all constraints; This indicates deep validation, which means validating each constraint and the script file, and objectively validating the output file.
[0012] Preferably, if the verification strength level is or Further, a consistency score is calculated for each subtask. If the consistency score is greater than or equal to the learning parameter, the current output is determined to be in a stable state, and the policy controller can select the current verification strength level to verify the task constraints. If the consistency score is less than the learning parameter, a deeper verification is forced.
[0013] Preferably, continuing to retry in S3 includes the following operations: Extract the failure constraint identifier and failure reason summary corresponding to this failure to form the failure mode feature. Compare the current failure mode feature with the historical failure mode set archived locally in the system. If no matching record is found, it is determined to be a new mode. For the new mode, the system creates a new execution agent session instance with complete feedback information and restarts the current subtask for retry correction.
[0014] Preferably, the strategy rollback in S3 includes the following operations: If the process fails after three retries, the current subtask iteration process will be terminated, and the process will revert to the initial input task state to re-execute the task decomposition.
[0015] Preferably, the modal-constraint rules in S2 include: Structured data modalities are bound to structural constraints and data accuracy constraints; formatted document modalities are bound to document hierarchy constraints and content accuracy constraints; chart modalities are bound to visual content semantic constraints; and audio-visual modalities are bound to media technology parameter constraints.
[0016] Preferably, the adaptive addition of task constraints for risk factors in office tasks in S2 includes the following operations: Add source verification constraints and anti-fabrication verification constraints to the first risk factor; add content coverage verification constraints to the second risk factor; add content authenticity constraints to the third risk factor; and add data loading verification and code executability constraints to the fourth risk factor.
[0017] Preferably, the task mode in S1 includes one of the following modes: Structured data modality; formatted document modality; web interface modality; chart modality; plain text modality; code / script modality; audio / video modality; same-format editing modality, i.e., the output format is consistent with the input.
[0018] An office task verification system with modal adaptive constraint binding includes the following modules: The task decomposition module takes an office task as input and analyzes the input task modality and the task modality to be output; the executing agent adaptively sets risk factors based on the input modality; and then decomposes the office task into several sub-tasks to be executed sequentially. The constraint binding module analyzes the modality of each subtask, binds task constraints to the subtask from a predefined modality-constraint rule table based on the subtask's modality, adaptively adds task constraints based on the risk factors of the office task, and converts all task constraints of the subtask into a binary format. The constraint verification module executes all subtasks sequentially, verifies the strength level of each subtask, and performs task constraint verification based on the strength level. If any constraint in the current subtask fails verification, the module adaptively triggers the corresponding constraint handling strategy based on the subtask modality, completing the subtask's retry, early termination, or task rollback strategy until all subtask constraint verifications are completed and the office task ends.
[0019] Compared with the prior art, the technical solution provided by this invention has the following advantages: This invention addresses the issue of binding differentiated task constraints to multimodal office tasks, effectively preventing key quality defects from being missed. By adopting a multi-agent collaborative strategy, the functions of task generation, task verification, and task coordination are decoupled from each other. When the verification agent verifies the task constraints, context isolation is achieved, the reasoning context information of the executing agent is not reused, the inherent bias of agent self-affirmation is reduced, and critical verification of the task is achieved.
[0020] By adaptively adjusting the constraint strength through reinforcement learning, over-validation and under-validation can be avoided, thus achieving a balance between validation accuracy and computational resources. Detailed Implementation
[0021] This invention adopts a coordinating agent-executing agent-verifying agent architecture, where the coordinating agent is responsible for task scheduling, information routing, and process control; the executing agent, created on demand by the coordinating agent, is responsible for executing tasks; and the verifying agent, invoked on demand by the coordinating agent, is responsible for verifying tasks using independent large language model query instances. To further understand this invention, a detailed description is provided with reference to embodiments. An office task verification method based on modal adaptive constraint binding includes the following steps: S1. Input the office task and analyze the input task modality and the task modality to be output; the executing agent adaptively sets risk factors according to the input modality; and then decomposes the office task into several sub-tasks to be executed sequentially.
[0022] After receiving a natural language task instruction, the coordinating agent creates an execution agent session instance. The execution agent parses the task instruction content and analyzes the task modality to be output. Each task modality is mutually exclusive. Optionally, the task modality includes one of the following modalities: Structured data modality; formatted document modality; web interface modality; chart modality; plain text modality; code / script modality; audio / video modality; same-format editing modality, i.e., the output format is consistent with the input.
[0023] The agent executes the task instructions while parsing the feature words with modal semantics. When it is recognized that the current task requires obtaining information from web searches or accessing web links, or extracting content from input documents such as PDF / Word, or the task contains images that require semantic understanding, or requires writing and executing program code to process data, it is determined that the current task has risks. Risk factors are configured for the current task. If the current task has multiple risk situations at the same time, the risk factors can be superimposed on each other.
[0024] Furthermore, a first risk factor is configured for tasks that require obtaining information from web searches or webpage links; a second risk factor is configured for tasks that require extracting content from input documents such as PDFs / Word documents; a third risk factor is configured for tasks that include images requiring semantic understanding; and a fourth risk factor is configured for tasks that require writing and executing program code to process data.
[0025] Furthermore, the executing agent decomposes the natural language task into several subtasks, each with an independent output modality and a clearly defined completion boundary. The dependencies between the subtasks are modeled as a directed acyclic graph. Preferably, the present invention sets the number of subtasks to 2-5.
[0026] For example, a natural language task could be: converting a PDF annual financial report into an Excel file containing charts; the executing agent would break down the natural language task into: Subtask 1: Parse the PDF and extract data; Subtask 2: Organize the data into an Excel spreadsheet and fill in the formulas; Subtask 3: Generate a chart to embed in Excel.
[0027] Subtask 1 is parsed into a plain text modality, subtask 2 into a structured data modality, and subtask 3 into a graph modality. The executing agent writes the subtask list, subtask modalities, and task risks into a structured file.
[0028] S2. Analyze the modality of each subtask, and bind task constraints to the subtask from the predefined modality-constraint rule table according to the subtask's modality; adaptively add task constraints according to the risk factors of the office task; and convert all task constraints of the subtask into a binary format.
[0029] For each subtask, based on its task modality, query the predefined modality-constraint binding rule table and bind task constraints to the subtask.
[0030] Preferably, the modal constraint rules include: Structured data modalities are bound to structural constraints and data accuracy constraints; formatted document modalities are bound to document hierarchy constraints and content accuracy constraints; chart modalities are bound to visual content semantic constraints; and audio-visual modalities are bound to media technology parameter constraints.
[0031] Specifically, structural constraints check whether the output structured data files have the correct document structure. For example, does an Excel file have the correct header row, does the number of columns match expectations, and are the data rows divided into cells as required? Does a JSON file have the correct nesting level and key names?
[0032] Data precision constraints check whether the specific values in the output data are accurate. For example, whether the numbers extracted from PDF to Excel are consistent with the original text, whether the formula calculation results are correct, whether the decimal places are retained to the required precision, and whether there are any issues such as data truncation or rounding errors.
[0033] The document hierarchy constraint check ensures that the chapter structure of the output document is complete and orderly. For example, does the Word document contain all required first-level and second-level headings, are the heading hierarchy relationships correct, and is the chapter order logical?
[0034] The content accuracy constraint checks whether the main text of the document is true and accurate, and whether there is any fabrication or distortion of facts. For example: whether the data cited in the report is consistent with the original source, whether the summary is faithful to the meaning of the original text, and whether the translation accurately conveys the original meaning.
[0035] The visual content semantic constraint check ensures that the generated images meet the task requirements at the semantic level. This doesn't involve checking the image resolution or file format, but rather the content within the image. For example, if the task requires generating a bar chart showing quarterly sales figures for 2023, the check should verify that the chart is indeed a bar chart, that it contains bars for all four quarters, and that the numerical labels are present and appropriate.
[0036] The media technical parameter constraint check ensures that the output audio and video files meet the technical specifications. For example, whether the video resolution, frame rate, encoding format, and duration meet the requirements; and whether the audio sampling rate, bit rate, and number of channels are correct.
[0037] Furthermore, the adaptive addition of task constraints based on risk factors for office tasks includes the following operations: Add source verification constraints and anti-fabrication verification constraints to the first risk factor; add content coverage verification constraints to the second risk factor; add content authenticity constraints to the third risk factor; and add data loading verification constraints and code executability constraints to the fourth risk factor.
[0038] Specifically, the source verification constraint verifies whether the referenced webpage link is genuine and accessible, and whether the actual content of the webpage matches the content referenced by the agent, when the task requires information from a web search or webpage link. This prevents the agent from fabricating a seemingly reasonable but non-existent URL, or from distorting the meaning of the original webpage text.
[0039] Fabrication prevention verification constraints ensure that agents do not fabricate facts or data out of thin air. When the task involves retrieving information from the internet, all information claiming to come from external sources must be verifiable, and fictitious content is not allowed.
[0040] Content coverage validation constraints verify whether all required data has been completely extracted when a task needs to extract content from an input document, such as a PDF or Word document, without omitting any pages, paragraphs, or tables. For example, when converting a 20-page PDF financial report to Excel, does it ensure that all pages are covered, rather than just processing the first 5 pages?
[0041] The content authenticity constraint verifies whether the agent's description or extraction of the image content faithfully reflects the original image when the input contains an image that needs to be understood. This prevents the agent from embellishing or misrepresenting the image. For example, if the image shows three people standing in front of a building, the agent should not describe it as five people sitting in a park.
[0042] The data loading validation constraint verifies whether the code correctly loads the input data file when the task involves writing code to process data. For example, it checks whether the code correctly specifies the file path, uses the correct encoding to read the table file, whether the number of rows and columns in the read table matches the original file, and whether there are any empty values or garbled characters.
[0043] Code executability constraints verify whether the code written by the agent can run normally in the target environment without syntax errors, missing dependencies, or runtime crashes.
[0044] Furthermore, all constraints are transformed into a binary standardized format, meaning that binary judgments are performed on all constraint cases. For each constraint, only two judgment results are output: the constraint is satisfied or the constraint is not satisfied, removing intermediate judgment results. The verification output of each constraint is uniformly standardized into a binary Boolean expression.
[0045] S3. Execute all subtasks sequentially, verify the strength level of each subtask, and perform task constraint verification based on the strength level. If any constraint of the current subtask fails verification, the corresponding constraint handling strategy is adaptively triggered according to the subtask modality to complete the subtask's retry, early termination, or task rollback strategy until all subtask constraint verifications are completed and the office task ends.
[0046] For all subtasks, execution follows the order of the directed acyclic graph. The coordinating agent creates a session instance for the currently executing subtask, which has no historical dialogue context. The executing agent receives only the current subtask description, task constraints, and input file path, outputs the corresponding output file and execution summary file, writes the output file and summary file to disk, and destroys the current session instance.
[0047] The coordinating agent invokes the verifier agent via a stateless, single-round query. The verifier agent cannot access the inference history of the executing agent; it only receives task constraints and output file paths, and provides a decision criterion for each binary task constraint. The key to this design lies in the physical isolation between the executing and verifying agents—that is, there is no shared dialogue context or inference state between them, and the only channel for information transmission is a file on the computer device's disk.
[0048] Specifically, the importance of each subtask constraint is determined by the reinforcement learning policy controller. The decision-making process of the policy controller is as follows: First, construct the state space. , recorded as ,in Indicates the output modality type of the current subtask; The task complexity score is calculated using the formula: s2 = min(10, number of subtasks × 2 + total number of task constraints × 0.5 + total number of risk factors × 1.5). This represents the success rate of historical tasks with the same modality, and its value is [0,1]. This is the current number of retries; The confidence score represents the agent's ability to complete the current task, and is calculated using the following formula: , The value of is [0,1], where The number of execution exception signals, including the total number of tool call errors, retries / rollbacks, and execution path deviations from the plan during execution, can be obtained from the execution log. This represents the total number of execution steps and the total number of tool calls made by the agent to complete the subtask, which can be obtained from the execution log.
[0049] The policy controller selects the verification strength level for the sub-task based on the current state vector. Specifically, the verification strength level corresponds to the action space. , recorded as ,in Indicates skipping; This indicates lightweight validation, which means randomly selecting two constraints for validation; This indicates standard verification, which means verifying all constraints; This indicates deep validation, which means validating each constraint and the script file, and objectively validating the output file.
[0050] If the selected verification strength level is or Perform verification directly; if the selected verification strength level is or Further calculate the consistency score for each subtask. , represented as: , in, The number of samples is the default value in this invention. ; Indicates the total number of constraints for the current subtask; This represents the inconsistency decision number, i.e. The number of times inconsistent conclusions are given for the same constraint in each sampling.
[0051] like If the current output is determined to be in a stable state, the strategy controller can select the current verification strength level to verify the task constraints.
[0052] like Then, a deeper verification will be enforced, i.e. or By dynamically adjusting the verification intensity according to the task difficulty, resource waste in simple tasks and insufficient verification in difficult tasks are avoided. Among other things, This is the dynamic policy learning parameter in reinforcement learning, with an initial value of 0.8.
[0053] Based on the intensity level, the task constraint verification is performed. If any constraint of the current subtask fails to be verified, the corresponding constraint handling strategy is adaptively triggered according to the subtask mode, and the subtask is retried, terminated early, or rolled back until all subtask constraint verifications are completed and the office task ends.
[0054] Specifically, this invention employs a deep Q-network algorithm to determine the validation strength. The Q-function is fitted by a lightweight neural network with an input dimension of 12, two hidden layers of 64 nodes each, and an output dimension of 4. An empirical replay buffer stores historical samples, and updates are performed in batches every 32 new samples. The Q-function is defined as follows: , in, Represents the reward function; It expresses expectation.
[0055] reward function Represented as: , in, This indicates a task success flag, with a value of 0 or 1, indicating whether the subtask ultimately passed verification. The verification cost is calculated by multiplying the number of calls made by the verification agent by the average number of tokens consumed per verification and then dividing by the normalization constant. This indicates the penalty for missed detections, when the verification strength level is... or However, the value is 1 if the subtask ultimately fails, and 0 otherwise. , , As hyperparameters, this invention sets , Much larger This reflects a design preference that prioritizes avoiding missed detections over saving resources. Setting it to 0.3 results in a perceptible penalty of approximately -0.3 for the verification cost item in deep verification scenarios, ensuring that the policy controller can distinguish the cost differences between different verification levels from the reward signal.
[0056] The policy controller has an internal constraint priority tuning submodule, which maintains the constraint priority weight matrix. The matrix dimension is , This represents the sum of the total number of output modalities and the total number of risk factors. Represents the total number of task constraint types; each element in the matrix... Representing modes The importance weight of the task constraint y.
[0057] The weight update uses the Thompson sampling strategy, that is, for each element... Maintaining the Beta distribution ; The distribution parameter represents the cumulative number of times the task constraint effectively captures errors in the modality, and its initial value is 1; The distribution parameter represents the cumulative number of times the task constraint failed to effectively capture errors, and its initial value is 1. Before performing task constraint verification, for each constraint, samples are taken from its corresponding Beta distribution to obtain the priority weight of the task constraint in this round.
[0058] Once the subtask completes its full lifecycle—that is, the subtask passes verification, terminates early, or rolls back the strategy—update the parameters in the corresponding Beta distribution based on whether an error was actually captured during the verification process, and then update the priority weights.
[0059] The constraint priority tuning submodule continuously collects verification result data during long-term system operation: Record the judgment result (yes / no) of each constraint in each verification and the final result of the corresponding task (success / failure), forming a triplet {task constraint, judgment result, task final result}.
[0060] Furthermore, the priority of constraints is adjusted based on their correlation with the success of the overall task.
[0061] Specifically, calculate the discrimination index for each task constraint. : , in, This indicates that the task was successful. This indicates that the constraint verification has passed. This indicates that the constraint validation failed. The higher the value, the stronger the correlation between the constraint and task success. Specifically, To indicate high discriminative power, retain this constraint and increase its weight; Indicate the level of discrimination and maintain observation; This indicates low discriminative power; the constraint is marked as a candidate for elimination.
[0062] By optimizing the weights of task constraints and adaptively increasing the verification strength, the system can output more flexible constraint verification results.
[0063] Specifically, to ensure the execution quality of each subtask, task constraints are verified based on verification strength. Preferably, the subtask execution task constraint verification includes the following three types: The system includes consistency checks, bypass checks, and relevance checks. Consistency checks involve multiple independent verification agents validating subtask constraints. If the number of failed checks exceeds a preset threshold, the constraint is marked as failed; otherwise, it passes. Bypass checks determine if the executing agent can pass a constraint even without actually completing all subtask requirements. If the constraint can be circumvented, its validation result is marked as failed; otherwise, it passes. Relevance checks determine if the current constraint truly covers the quality dimensions relevant to the task instructions, marking irrelevant constraints as failed. The quality dimensions are determined by explicit or implicit requirements in the task instructions. The logic is: if removing the constraint still results in the user believing the output meets the task requirements, then the constraint is considered irrelevant.
[0064] If all task constraint checks pass, mark the subtask as complete.
[0065] If not all task constraint checks pass, the verification agent generates targeted modification suggestions for the corresponding failed task constraints based on the importance of the constraints. The coordinating agent then sends the targeted modification suggestions back to the executing agent for task correction. This loop is executed at most three times.
[0066] During the task constraint correction and iteration process, the system synchronously executes the preset failure state handling logic. The targeted modification suggestions include three categories: continue retrying, early termination, and strategy rollback, which are used to dynamically manage the iteration process and avoid invalid loops.
[0067] Preferably, continuing to retry includes the following operations: Extract the failure constraint identifier and failure reason summary corresponding to this failure to form the failure mode feature. Compare the current failure mode feature with the historical failure mode set archived locally in the system. If no matching record is found, it is determined to be a new mode. For the new mode, the system creates a new execution agent session instance with complete feedback information and restarts the current subtask for retry correction.
[0068] Early termination: If the same failure constraint and failure mode occur in two consecutive retries, it is determined to be a fixed defect iteration, no further correction is needed, the subtask is marked as failed and skipped; Strategy rollback: If the process fails after three retries, the current subtask iteration process will be terminated, and the process will roll back to the initial input task state and re-execute the task decomposition.
[0069] To further verify the effectiveness of this invention, experimental verification was conducted. This invention uses the Claude Opus 4.6 agent as the base model. The control group is the general single-agent verification process of the Claude Code framework, where the agent executes the task and performs its own verification within the same context.
[0070] The experimental group used the method of this invention, employing the same task set and tool permissions as the control group, with the only difference being the verification method. Specifically, the modality-aware step in algorithm S1 of this invention was removed, i.e., the input task modality and output task modality were not distinguished, forming test group 1; the binary constraint step in the algorithm of this invention was removed, forming test group 2. Evaluation was conducted by an independent large language model LLM evaluator according to task-specific scoring criteria. The experiment used 220 professional office tasks from the GDPVAL dataset as the dataset, covering real-world professional work scenarios across multiple industries, to quantitatively evaluate the execution accuracy and closed-loop completion capability of the multi-agent system in high-economic-value professional tasks. Table 1 shows the ablation experiment and performance comparison. In Table 1, the GDPVAL completion rate characterizes the overall delivery quality and verification stability of the agent for complex professional office tasks.
[0071] Table 1 Ablation Experiments and Performance Comparison
[0072] As shown in Table 1, the GDPVAL completion rate of the control group was 64.01%; using the test group 1 method, the GDPVAL completion rate was 74.09%, which was 10.08% higher than that of the control group; using the test group 2 method, the GDPVAL completion rate was 73.50%, which was 9.49% higher than that of the control group; using the complete scheme of this invention, the GDPVAL completion rate was 77.36%, which was 13.35% higher than that of the control group.
[0073] Furthermore, 104 daily life tasks were added to the AgentIF-OneDay dataset for multi-dataset modal validation, and the results are shown in Table 2.
[0074] Table 2 Comparison of Completion Rates under Different Output Modes
[0075] Compared to the control group, this invention improved the completion rate to varying degrees across different output modalities. Specifically, for the formatted document modality, the completion rate increased from 67.70% to 81.64% across 145 tasks in the GDPVAL dataset, an improvement of 13.94%; for the structured data modality, the completion rate increased from 59.79% to 68.93% across 65 tasks in the GDPVAL dataset, an improvement of 9.14%; for the chart modality, the completion rate increased from 64.95% to 80.23% across 29 tasks in the AgentIF-OneDay dataset, an improvement of 15.28%; and for the web interface modality, the completion rate increased from 61.73% to 68.05% across 23 tasks in the AgentIF-OneDay dataset, an improvement of 6.32%.
[0076] Furthermore, the present invention also includes a modal adaptive constraint binding office task verification system for implementing the aforementioned modal adaptive constraint binding office task verification method. This system includes the following modules: The task decomposition module takes an office task as input and analyzes the input task modality and the task modality to be output; the executing agent adaptively sets risk factors based on the input modality; and then decomposes the office task into several sub-tasks to be executed sequentially. The constraint binding module analyzes the modality of each subtask, binds task constraints to the subtask from a predefined modality-constraint rule table based on the subtask's modality, adaptively adds task constraints based on the risk factors of the office task, and converts all task constraints of the subtask into a binary format. The constraint verification module executes all subtasks sequentially, verifies the strength level of each subtask, and performs task constraint verification based on the strength level. If any constraint in the current subtask fails verification, the module adaptively triggers the corresponding constraint handling strategy based on the subtask modality, completing the subtask's retry, early termination, or task rollback strategy until all subtask constraint verifications are completed and the office task ends.
[0077] The above description is only a preferred embodiment of the present invention, but the scope of protection of the present invention is not limited thereto. Any equivalent substitutions or modifications made by those skilled in the art within the scope of the technology disclosed in the present invention, based on the technical solution and inventive concept of the present invention, should be covered within the scope of protection of the present invention.
Claims
1. A method for office task validation of modal adaptive constraint binding, characterized in that, Includes the following steps: S1. Input the office task and analyze the input task modality and the task modality to be output; the executing agent adaptively sets risk factors according to the input modality; and then decomposes the office task into several sub-tasks to be executed sequentially; S2. Analyze the modality of each subtask and bind task constraints to the subtask from the predefined modality-constraint rule table based on the subtask's modality; It also adaptively adds task constraints based on the risk factors of the office tasks; and converts all task constraints of the subtasks into a binary format. S3. Execute all subtasks sequentially, verify the strength level of each subtask, and perform task constraint verification based on the strength level. If any constraint of the current subtask fails verification, the corresponding constraint handling strategy is adaptively triggered according to the subtask modality to complete the subtask's retry, early termination, or task rollback strategy until all subtask constraint verifications are completed and the office task ends.
2. The method of claim 1, wherein the method is a modal adaptive constraint binding office task verification method. In S3, subtask execution constraint verification includes the following three types: Consistency check, bypass check, and correlation check; consistency check is performed by multiple independent verification agents to check the subtask constraints. If the number of failures exceeds a preset threshold, the constraint is marked as failing the check; otherwise, the check passes. The bypass check determines whether the executing agent can still pass the constraint check even if all subtask requirements have not been actually completed. If it is determined that the constraint can be bypassed, the check result of the constraint is marked as failing; otherwise, the check passes. The correlation check determines whether the current constraint can truly cover the quality dimensions that the task instruction is concerned with, and marks irrelevant constraints as failing.
3. The method for verifying office tasks using modal adaptive constraint binding according to claim 1, characterized in that: In S3, the verification strength level is denoted as ,in Indicates skipping; This indicates lightweight validation, which means randomly selecting two constraints for validation; This indicates standard verification, which means verifying all constraints; This indicates deep validation, which means validating each constraint and the script file, and objectively validating the output file.
4. The method for verifying office tasks using modal adaptive constraint binding according to claim 3, characterized in that: If the verification strength level is Further, a consistency score is calculated for each subtask. If the consistency score is greater than or equal to the learning parameter, the current output is determined to be in a stable state, and the policy controller can select the current verification strength level to verify the task constraints. If the consistency score is less than the learning parameter, a deeper verification is forced.
5. The method for verifying office tasks using modal adaptive constraint binding according to claim 1, characterized in that... Retrying in S3 includes the following operations: Extract the failure constraint identifier and failure reason summary corresponding to this failure to form the failure mode feature; compare the current failure mode feature with the historical failure mode set archived locally in the system. If there is no matching record, it is determined to be a new mode; for the new mode, the system creates a brand new execution agent session instance with complete feedback information and restarts the current subtask for retry correction.
6. The method for verifying office tasks with modal adaptive constraint binding according to claim 1, characterized in that... Strategy rollback in S3 includes the following operations: If the process fails after three retries, the current subtask iteration process will be terminated, and the process will revert to the initial input task state to re-execute the task decomposition.
7. The method for verifying office tasks with modal adaptive constraint binding according to claim 1, characterized in that... The modal constraint rules in S2 include: Structured data modalities are bound to structural constraints and data accuracy constraints; formatted document modalities are bound to document hierarchy constraints and content accuracy constraints; chart modalities are bound to visual content semantic constraints; and audio-visual modalities are bound to media technology parameter constraints.
8. The method for verifying office tasks with modal adaptive constraint binding according to claim 1, characterized in that... The adaptive addition of risk factors to office tasks in S2 includes the following operations: Add source verification constraints and anti-fabrication verification constraints to the first risk factor; add content coverage verification constraints to the second risk factor; add content authenticity constraints to the third risk factor; and add data loading verification and code executability constraints to the fourth risk factor.
9. The method for verifying office tasks with modal adaptive constraint binding according to claim 1, characterized in that... The task mode in S1 includes one of the following modes: Structured data modality; formatted document modality; web interface modality; chart modality; plain text modality; code / script modality; audio / video modality; same-format editing modality, i.e., the output format is consistent with the input.
10. An office task verification system with modal adaptive constraint binding, characterized in that, It includes the following modules: a task decomposition module, which takes an office task as input and analyzes the input task modality and the task modality to be output; the executing agent adaptively sets risk factors based on the input modality; and then decomposes the office task into several sequentially executed sub-tasks; The constraint binding module analyzes the modality of each subtask and binds task constraints to the subtask from a predefined modality-constraint rule table based on the subtask's modality. It also adaptively adds task constraints based on the risk factors of the office tasks; and converts all task constraints of the subtasks into a binary format. The constraint verification module executes all subtasks sequentially, verifies the strength level of each subtask, and performs task constraint verification based on the strength level. If any constraint in the current subtask fails verification, the module adaptively triggers the corresponding constraint handling strategy based on the subtask modality, completing the subtask's retry, early termination, or task rollback strategy until all subtask constraint verifications are completed and the office task ends.
