A planning calibration method and system for an information analysis intelligent assistant

Abstract

The application discloses a planning calibration method and system for an information analysis intelligent assistant, and comprises the following steps: in each planning round, a candidate strategy set is generated by a planning agent, the candidate strategy set contains an information retrieval target, a retrieval keyword combination and an information processing flow; the candidate strategy set is evaluated by multiple agents with heterogeneous private memories, a strategy selection module based on the agents is constructed for information consistency, and a strategy with the most stable evaluation result is selected as an optimal strategy of the round; a difference between a self-selected strategy of the planning agent and the selected optimal strategy of the round is recorded, a calibration constraint module guided by consistency is constructed, and the difference is converted into a cognitive calibration constraint and stored in the memory of the planning agent; strategy selection optimization is realized in single-round planning by relying on the strategy selection module based on information consistency; and the cognitive calibration constraint generated by the calibration constraint module guided by consistency is combined, so that the constraint is transmitted between different rounds and gradually accumulated, and a constraint is formed on strategy generation of the planning agent; and the trained agent is applied to an open domain webpage retrieval task to generate accurate retrieval results.

Description

Technical Field

[0001] This invention relates to the fields of artificial intelligence, large language models and multi-agent collaboration, and in particular to a planning and calibration method and system for an intelligent assistant for information analysis. Background Technology

[0002] In recent years, intelligent assistants for information analysis have been gradually applied to practical scenarios such as scientific research intelligence analysis, enterprise decision support, and professional information services. [1-3] In systems such as technology intelligence analysis, users often pose complex tasks, such as: "Comprehensively analyze the research hotspots of a certain emerging technology over the past five years and assess its future development trends." These tasks involve multi-source heterogeneous data retrieval, cross-document information integration, and long-chain reasoning, which are difficult for a single agent to complete efficiently. Therefore, a multi-agent system consisting of planning agents, execution agents, etc., is typically used for collaborative processing. [1,4] .

[0003] In practical deployments, existing intelligent information analysis assistants are widely used in patent analysis platforms, scientific research trend assessment systems, and enterprise technology decision support systems to assist users in mining high-value information. However, in these specific application scenarios, the stability and reliability of information analysis assistants still face significant challenges. [5] .

[0004] Existing troubleshooting methods for intelligent information analysis assistants primarily focus on execution-level errors, such as incorrect tool call formats, failed API calls, or interrupted retrieval, and are corrected through execution log analysis or post-event reflection mechanisms. However, in practical applications, a more subtle and impactful failure mode has been discovered: even if all execution operations are performed correctly, issues such as incorrect tool usage or low-quality search keywords may result in the assistant's final output containing low-quality information, thus failing to meet user needs. [6,7] .

[0005] This problem is particularly prominent in scenarios such as technical intelligence analysis. For example, when conducting trend analysis on an emerging technology, if the information analysis AI assistant incorrectly judges in the initial stage that "existing knowledge is sufficient to cover the key directions," it may formulate an insufficient retrieval strategy, leading to subsequent analysis focusing only on local information and ultimately generating biased or even erroneous conclusions.

[0006] The aforementioned problems essentially stem from epistemic miscalibration in the planning phase of information analysis intelligent assistants. Specifically, during task decomposition and strategy formulation, these assistants become overconfident in their ability to retrieve and analyze existing or potential information, failing to accurately identify information gaps or uncertainties. This miscalibration typically manifests as logically consistent and executable strategies in the generated policies, thus avoiding explicit error signals and rendering existing remedial mechanisms based on execution anomaly detection ineffective.

[0007] Furthermore, during dynamic information retrieval, as new data is continuously introduced and the internal state of the intelligent assistant is constantly updated, early cognitive biases may be masked, causing the intelligent assistant to repeatedly execute inefficient or even ineffective retrieval and analysis processes, which seriously affects the reliability and efficiency of the system in practical applications.

[0008] Therefore, there is an urgent need for a planning and calibration method and system for information retrieval and analysis scenarios. This method would identify and correct cognitive inaccuracies in the planning phase of information analysis intelligent assistants, thereby improving the decision-making quality and application reliability of these assistants in complex information processing tasks. References [1]Becattini, M., Verdecchia, R.,&Vicario, E. (2025, April). Sallma:A software architecture for llm-based multi-agent systems. In 2025 IEEE / ACMInternational Workshop New Trends in Software Architecture (SATrends) (pp. 5-8). IEEE. [2]Sun, C., Huang, S.,&Pompili, D. (2025). Llm-based multi-agent decision-making: Challenges and future directions. IEEE Robotics and Automation Letters . [3]He, J., Treude, C., & Lo, D. (2025). Llm-based multi-agent systems for software engineering: Literature review, vision, and the road ahead.ACM Transactions on Software Engineering and Methodology , 34 (5), 1-30. [4]Ronanki, K. (2025, June). Facilitating Trustworthy Human-AgentCollaboration in LLM-based Multi-Agent Systemoriented Software Engineering.In Proceedings of the 33rd ACM International Conference on the Foundations of Software Engineering (pp. 1333-1337). [5]Cemri, M., Pan, MZ, Yang, S., Agrawal, LA, Chopra, B., Tiwari, R., ...&Stoica, I. (2025). Why do multi-agentllm systemsfail?. arXiv preprint arXiv:2503.13657 . [6]Liang, C., Gan, J., Hong, K., Tian, ​​Q., Wu, Z.,&Li, R. (2025).COCO: Cognitive Operating System with Continuous Oversight for Multi-AgentWorkflow Reliability. arXiv preprint arXiv:2508.13815 . [7]Ma, M., Zhang, J., Yang, F., Kang, Y., Lin, Q., Rajmohan, S.,&Zhang, D. (2025). DoVer: Intervention-Driven Auto Debugging for LLM Multi-Agent Systems. arXiv preprint arXiv:2512.06749 . Summary of the Invention

[0009] This invention provides a planning and calibration method and system for intelligent information analysis assistants. This invention addresses the problems of repeated invalid searches, low-quality information processing and analysis caused by cognitive inaccuracies in intelligent assistants during information retrieval, integration, and analysis, thereby improving the response success rate and reliability of intelligent assistants in complex information analysis tasks. Details are described below: Firstly, a planning and calibration method for an information analysis intelligent assistant, the method comprising: In each planning round, the planning agent generates a set of candidate strategies, which includes information retrieval objectives, combinations of retrieval keywords, and information processing flows. The candidate policy set is evaluated by multiple agents with heterogeneous private memories. Each agent is based on different information sources. A policy selection module with consistent information is constructed based on these agents, and the policy with the most stable evaluation results is selected as the optimal policy for this round. Record the discrepancies between the self-selected strategy and the selected optimal strategy for this round of planning, construct a consistency-guided calibration constraint module, and transform the discrepancy into a cognitive calibration constraint and store it in the memory of the planning agent; By using a strategy selection module that relies on information consistency, strategy selection optimization is achieved during single-round planning; and by combining the cognitive calibration constraints generated by the consistency-guided calibration constraint module, these constraints are passed down and gradually accumulated across different rounds, thus constraining the strategy generation of the planning agent. The trained agent is applied to open-domain web page retrieval tasks to generate accurate search results.

[0010] The information consistency strategy selection module is used to solve the strategy selection bias problem caused by a single cognitive perspective in complex multi-step tasks of information analysis intelligent assistants. The information consistency strategy selection module introduces a planning agent, an execution agent, and a diagnostic agent. The agents share the system-level memory but maintain their own private memory, forming a heterogeneous information state. The planning agent focuses on generating candidate strategies based on historical experience, the execution agent focuses on evaluating the actual feasibility of candidate strategies, and the diagnostic agent focuses on judging the rationality of candidate strategies in terms of information sufficiency and risk.

[0011] The information consistency strategy selection module is used to calculate the evaluation consistency score of each candidate strategy among different agents, which is used to measure the stability and reliability of each candidate strategy under multiple cognitive perspectives.

[0012] The consistency-guided calibration constraint module is used to continuously monitor whether there is a discrepancy between the planning agent's self-selected planning strategy based on its private information and the globally optimal planning strategy selected based on the information consistency scoring mechanism. When the two are inconsistent, it indicates that the planning agent has a bias in its grasp of task information in the current cognitive state, and a potential cognitive calibration misalignment event has occurred. Upon detecting such discrepancies, the consistency-guided calibration constraint module generates a lightweight cognitive constraint rule, which is written into the persistent memory of the planning agent and participates in policy generation as a constraint in subsequent planning processes.

[0013] The optimal strategy for this round is as follows:

[0014] in, Let be the set of candidate strategies executed in round t. For strategy Global information consistency score;

[0015] in, For intelligent agents The k-th candidate strategy generated during the t-th round of execution Individual consistency score, It is the set of all intelligent agents.

[0016] The lightweight cognitive constraint rule is as follows:

[0017] in, For planning intelligent agents Based on its own evaluation vector It chooses a strategy that it considers optimal. This is the optimal strategy for this round. This is the current system memory.

[0018] Second aspect, a planning and calibration system for an information analysis intelligent assistant, the system comprising: a processor and a memory, the memory storing program instructions, the processor calling the program instructions stored in the memory to cause the device to perform the method described in any of the first aspects.

[0019] Third aspect, a computer-readable storage medium storing a computer program, the computer program including program instructions that, when executed by a processor, cause the processor to perform the method described in any one of the first aspects.

[0020] The beneficial effects of the technical solution provided by this invention are: 1) Overcomes the task failure caused by the low feasible strategy formulated by the information analysis intelligent assistant in the planning stage: This invention does not rely on the explicit error signals of the underlying tools of the information analysis intelligent assistant. By directly intervening in the planning stage, it effectively identifies and repairs the invalid information retrieval (such as low accuracy of web page retrieval) and dead loop planning problems (such as repeated retrieval - no usable information) caused by the "overconfidence" of the information analysis intelligent assistant in the planning stage. 2) The robustness of the planning is improved by adding heterogeneous information from different agents in the information analysis intelligent assistant and optimizing the cross-evaluation mechanism: The strategy selection mechanism based on information consistency proposed in this invention utilizes the information difference generated by the private memory of different agents within the information analysis intelligent assistant (introducing the logarithmic difference formula between predicted score and direct score) to screen out execution strategies that are highly feasible from multiple cognitive perspectives (such as high-precision target information, high-success-rate information retrieval keywords, etc.), effectively avoiding the analysis bias of the information analysis intelligent assistant caused by the local perspective of a single agent. 3) Achieved continuous calibration under the dynamic information needs of users: The consistency-guided cognitive state refinement mechanism proposed in this invention transforms the evaluation disagreement of multiple agents in a single round within the information analysis intelligent assistant into a persistent memory constraint across rounds, preventing the information analysis intelligent assistant from repeating the same cognitive errors after absorbing new search results, and ensuring the stability of complex dialogues with long logical chains. 4) Significantly improves the overall success rate of the information analysis intelligent assistant in meeting users' information collection and analysis requirements: Compared with existing methods such as direct execution, partial retries and global rollback, this invention can effectively eliminate failed plans in high-difficulty reasoning and multi-step network search tasks, greatly enhancing the engineering usability and user requirement fulfillment rate of the information analysis intelligent assistant. Attached Figure Description

[0021] Figure 1 A flowchart of a planning and calibration method for an intelligent information analysis assistant; Figure 2 A schematic diagram illustrating the detailed principles of strategy selection based on information consistency and cognitive state guidance based on consistency. Detailed Implementation

[0022] To make the objectives, technical solutions, and advantages of the present invention clearer, the embodiments of the present invention will be described in further detail below.

[0023] Given that existing information analysis intelligent assistants based on multi-agent systems are prone to failing to meet user needs when handling complex tasks due to cognitive misalignment during the planning phase, even if all planning actions (such as information retrieval, information organization and analysis) are executed correctly, the task may still fail to achieve its expected goals. This failure stems from the intelligent assistant's inaccurate assessment of its own knowledge boundaries and overconfidence in the feasibility evaluation of the strategy during the planning phase, resulting in a strategy that is fundamentally unfeasible. More seriously, these errors are insidious and dynamic: they are difficult to directly observe during the assistant's operation and recur with the acquisition of new information, rendering traditional methods that rely on execution error signals ineffective.

[0024] To address the aforementioned problems, embodiments of the present invention provide a cognitive calibration method and apparatus for the planning phase of an information analysis intelligent assistant. This method optimizes the multi-agent collaborative workflow within the information analysis agent, introduces evaluation consistency from heterogeneous information perspectives as a calibration signal during the planning phase, and utilizes a cross-round memory constraint mechanism to prevent the recurrence of cognitive errors, thereby fundamentally improving the reliability of the intelligent assistant's planning.

[0025] Example 1 A planning and calibration method for an information analysis intelligent assistant, see [link to relevant documentation]. Figure 1 The method includes: 101: In each planning round, the planning agent generates a set of candidate strategies; The candidate strategy set may include information retrieval targets (such as company, time range), combinations of search keywords (such as "AI investment + company + year"), and information processing flow (such as "retrieval → summary → comparative analysis"). 102: A set of candidate strategies is evaluated by multiple agents with heterogeneous private memories, each based on different information sources (such as historical retrieval success rates, failure cases, and current context). A strategy selection module with consistent information is constructed based on these agents, selecting the strategy with the most stable evaluation results as the optimal strategy for this round. 103: Record the discrepancies between the planning agent's self-selected strategy and the selected optimal strategy for this round, construct a consistency-guided calibration constraint module, and transform the discrepancy into a cognitive calibration constraint stored in the planning agent's memory to guide subsequent planning decisions; For example: "Prioritize expanding keywords when new terms are involved" and "Prohibit repeated search paths when there is no valid information in two consecutive rounds".

[0026] 104: Finally, through the strategy selection module that relies on information consistency, strategy selection optimization is achieved during single-round planning; and combined with the cognitive calibration constraints generated by the consistency-guided calibration constraint module, the constraints are transmitted and gradually accumulated between different rounds, thereby constraining the strategy generation of the planning agent and effectively alleviating the cognitive calibration misalignment problem in the planning stage.

[0027] For example, prematurely stopping the search when information is insufficient may lead to the misjudgment that "sufficient information is available"; generating low-quality search strategies (such as keyword bias or unclear target definition) may result in invalid search results. This can improve the feasibility of the planning strategy, the efficiency of information acquisition, and the stability of long-chain task execution.

[0028] 105: Apply the trained agent to open-domain web page retrieval tasks to generate accurate retrieval results.

[0029] Specifically, the strategy selection module based on information consistency in steps 102 and 104 is as follows: This module operates within a single planning round during the operation of the information analysis intelligent assistant, and is mainly used to solve the problem of strategy selection bias caused by a single cognitive perspective in complex multi-step tasks.

[0030] In practical applications, such as tasks like enterprise research, public opinion analysis, or technical intelligence retrieval, intelligent assistants often need to generate search keywords and analysis paths based on existing information. However, due to inaccurate judgment of their own knowledge boundaries and overconfidence in the feasibility assessment of strategies, the generated strategies are fundamentally unfeasible, easily leading to problems such as keyword selection bias and unclear definition of search targets, resulting in irrelevant search results or insufficient information coverage.

[0031] In further practical applications, especially in open-domain web page retrieval tasks (such as information retrieval based on search engines, cross-source web page content aggregation and analysis, etc.), the above problems become more prominent. In such tasks, the intelligent assistant needs to automatically generate search queries, select search strategies (such as keyword combinations, Boolean expressions, site restrictions, etc.), and perform multiple rounds of iterative optimization based on the returned results. Without an effective strategy verification mechanism, the system is prone to problems such as query expression bias (too broad or too narrow), incorrect search path (selection of non-authoritative or low-relevance information sources), strategy non-executability (generating queries that do not conform to search engine syntax), and insufficient information coverage (premature convergence to local results), which seriously affect the quality of information retrieval.

[0032] To address this, the information consistency-based strategy selection module introduces a planning agent. Execution agent and diagnostic intelligent agents These agents share a system-level memory but maintain their own private memories, resulting in a heterogeneous information state. Planners focus on generating candidate strategies based on historical experience, executors focus on evaluating the practical feasibility of the strategies (e.g., whether they can effectively acquire information), and diagnosticians focus on judging the rationality of the strategies in terms of information sufficiency and risk. Each agent independently evaluates the same set of candidate strategies and provides a score based on its own cognitive perspective.

[0033] In web retrieval scenarios, candidate strategies can be specifically manifested as different retrieval query expressions and information retrieval paths. For example, the planning agent can generate multiple sets of candidate queries (including different keyword combinations, time range limitations, or site constraints); the execution agent evaluates the feasibility of each query in a real search environment (including whether it conforms to search engine syntax and whether it can return valid results); and the diagnostic agent evaluates the rationality of the strategy from the perspectives of information coverage, potential biases, and result diversity.

[0034] Building upon this foundation, a strategy selection module based on information consistency calculates the evaluation consistency score of each candidate strategy across different agents. This score measures the stability and reliability of the strategy across multiple cognitive perspectives. Strategies with higher consistency scores not only perform well from a single perspective but also receive support in terms of execution feasibility and information sufficiency. Therefore, they are considered cognitively more robust and selected as the optimal strategy for this round of execution. This mechanism effectively avoids the problem of selecting low-quality strategies due to "overconfidence" of a single agent, thus improving the overall robustness of the information retrieval and analysis process.

[0035] Specifically, the consistency-guided calibration constraint module in steps 103 and 104 is as follows: This consistency-guided calibration constraint module operates across multiple planning rounds, primarily addressing the problem that information analysis intelligent assistants fail to learn from historical errors and repeatedly generate invalid plans during multi-round interactions. In practical applications, when user needs are complex or constantly changing, intelligent assistants typically require multiple rounds of retrieval and analysis. However, existing methods often lack explicit modeling of "planning errors," causing the system to repeat similar erroneous strategies in subsequent rounds after encountering failures (e.g., continuous retrieval of no valid information, repeated access to low-quality information sources, or premature generation of conclusions when information is insufficient), forming a cycle of "invalid retrieval—insufficient information—re-retrieve."

[0036] To address the aforementioned issues, this consistency-guided calibration constraint module continuously monitors for discrepancies between the planning agent's self-selected planning strategy based on its private information and the globally optimal planning strategy selected based on the Information Consistency Scoring (IPS) mechanism. When these two strategies are inconsistent, it indicates a deviation in the planning agent's understanding of task information at its current cognitive state, signifying a potential cognitive calibration misalignment event. For example, the planning agent might mistakenly believe that a certain retrieval path can obtain crucial information, while the multi-agent consensus evaluation results indicate that this path has a low probability of success.

[0037] Upon detecting such discrepancies, the consistency-guided calibration constraint module further analyzes the causes of the discrepancies, such as whether they are due to inappropriate keyword selection, incorrect information coverage estimation, or over-reliance on existing information. Based on this, it generates a lightweight cognitive constraint rule (such as "expand the keyword space after two consecutive failed searches" or "prohibit entry into the analysis phase when the information source is singular"). This constraint is written into the persistent memory of the planning agent and participates in policy generation as a constraint in subsequent planning processes, thereby suppressing the recurrence of similar errors.

[0038] Through the aforementioned cross-cycle cognitive constraint accumulation mechanism, the intelligent assistant can gradually correct its judgment biases regarding "known and unknown information," achieve continuous adaptation to dynamic information environments, avoid repeatedly falling into inefficient or ineffective planning paths, and significantly improve the execution stability and result quality in complex long-chain tasks.

[0039] In summary, the embodiments of the present invention, through steps 101-105, can effectively eliminate failed plans in high-difficulty reasoning and multi-step network search tasks, greatly enhancing the engineering usability and user requirement fulfillment rate of the information analysis intelligent assistant.

[0040] Example 2 The scheme in Example 1 will be further described below with reference to the specific calculation steps: 1. Overall Framework The overall framework of this invention adopts a strategy that combines "planning-time calibration" and "cross-round refinement". First, within each planning round, the IPS module uses the evaluation consistency of heterogeneous perspectives of multiple agents to screen out cognitively robust strategies. Second, between rounds, the consistency-guided calibration constraint module (CESR) transforms historical calibration signals into persistent cognitive constraints, thereby achieving long-term guidance and refinement of the cognitive state of the planning agent.

[0041] 2. Strategy selection module based on information consistency 2.1) Candidate Strategy Generation In the In the planning round, planning intelligent agents Based on current system-level memory and its private information status Generate a set of candidate strategies with diverse structures. :

[0042] in, Let K be the k-th candidate strategy executed in round t, where K is the number of candidate strategies.

[0043] 2.2) Strategy Evaluation under Information Conditions Define a generic evaluation interface. It can map the agent's information state, system memory, and candidate policy set into a real-valued score vector.

[0044] Each agent This interface is used for evaluation:

[0045] in, Represents intelligent agents Candidate strategies Feasibility score, To execute the intelligent agent, To diagnose intelligent agents, For the corresponding intelligent agent Private information.

[0046] 2.3) Cross-agent evaluation and prediction Since each agent's direct evaluation may be influenced by its private information bias, each agent... It will try to predict other intelligent agents How would the agent evaluate these strategies in its informational state? By combining the role descriptions of other intelligent agents and one's own private memories To approximate construction Information status :

[0047] in, This indicates a convention definition.

[0048] Then, the intelligent agent Predict using the evaluation interface Rating:

[0049] in, For intelligent agents Predictive agents The score for the candidate strategy, E, represents the evaluation interface based on the large model call.

[0050] intelligent agent The average of all prediction scores for other agents is used to obtain the policy evaluation. Group expectation assessment:

[0051] in, For intelligent agents Average prediction score for other agents It is the set of all intelligent agents.

[0052] 2.4) Information Consistency Aggregation and Strategy Selection To ensure consistency in quantification of evaluation, each agent... Calculate the logarithmic difference between its own rating and the predicted group rating as its individual consistency score:

[0053] in, For intelligent agents The k-th candidate strategy generated during the t-th round of execution Individual consistency score, For intelligent agents Candidate strategies Direct rating, To remove intelligent agents Other agents besides the candidate policy The direct score.

[0054] The strategy is obtained by averaging the individual scores of all agents. Global information consistency score :

[0055] Ultimately, the IPS module selects the strategy with the highest consistency score as the execution strategy for this round. :

[0056] in, Let be the set of candidate strategies to be executed in round t.

[0057] 3. Consistency-Guided Calibration Constraint Module (CESR) 3.1) Cross-round strategy divergence detection In the Wheel, planning intelligent agent It will be based on its own evaluation vector Choose a strategy that it considers optimal. :

[0058] CESR module continuous monitoring Selected by IPS Are there any disagreements? If If so, it is determined that a cognitive calibration disorder event has occurred.

[0059] 3.2) Cognitive Constraints Induction and Memory Update When a discrepancy is detected, the diagnostic agent... We will analyze these differing strategies. and the current system memory This leads to a lightweight cognitive constraint. :

[0060] This constraint abstracts the key information or structural features that led to this calibration misalignment. Subsequently, this constraint is stored in the planning agent. In private memories:

[0061] in, For planning intelligent agents In the Private memory during round execution, For planning intelligent agents Private memory in the t-th round of execution.

[0062] These accumulated cognitive constraints will continue to influence the planning agent in all future planning rounds. The strategy generation and evaluation process effectively filters out unsafe planning paths that are prone to calibration misalignment, thereby achieving long-term refinement and guidance of the cognitive state of the planning agent.

[0063] In summary, the embodiments of the present invention can effectively eliminate failed plans in high-difficulty reasoning and multi-step network search tasks through the above operations, which greatly enhances the engineering usability and user requirement fulfillment rate of the information analysis intelligent assistant.

[0064] Example 3 To verify the effectiveness of the cognitive calibration planning method (EPC-AW) proposed in this embodiment of the invention, it was experimentally compared with three other methods: the no-repair method (i.e., the original AgentFlow framework), the local retry method (Retry, based on metacognitive retry strategy), and the system-level rollback method (Rollback, based on COCO state rollback strategy). Experiments were conducted on six benchmark sets (GAIA, Bamboogle, 2Wiki, HotpotQA, Musique, and MedQA) covering different decision dimensions, search complexity, and inference depth. The evaluation metric was system-level accuracy, i.e., the consistency between the system's final output answer and the ground truth. The experimental results are shown in Table 1.

[0065] The overall experimental results show that the EPC-AW method proposed in this invention achieves optimal performance under all experimental settings. On datasets handling complex trajectories, EPC-AW achieves an absolute improvement of 9.75% in average success rate compared to the baseline method No-Repair, and a performance gain of 4.64% compared to the second-best performing Rollback method. The improvement of EPC-AW is particularly significant in tasks such as GAIA and 2Wiki, which heavily rely on external information retrieval and long-range planning, indicating that this invention has a significant advantage in accurately locating cognitive mismatch and avoiding cascade planning failures.

[0066] Further analysis of the ablation performance of each component (IPS and CESR) reveals the following: 1. In reasoning-intensive datasets (such as Bamboogle), the IPS module alone can effectively identify “fragile plans” that are executable but lack evidence support through cross-agent evaluation stability detection, thereby making the system more inclined to choose paths with more rigorous logic and more robust support; 2. The CESR module demonstrates its core value in datasets involving complex long-range dependencies (such as GAIA). Large language models are prone to context degradation when dealing with unstructured long trajectories, leading to the recurrence of the same cognitive error across multiple rounds of interaction. The CESR module successfully suppresses the propagation of this error by sensing explicit cognitive constraints and updating its private memory, enabling the system to achieve cognitive self-correction without relying on external supervision signals.

[0067] In contrast, existing repair methods such as Retry and Rollback perform poorly. This is because they heavily rely on explicit observations at the execution layer (such as tool errors or formatting mistakes) as triggering conditions for repair, while the "cognitive mismatch" failures targeted by this invention often manifest as normal trajectories at the execution layer, exhibiting strong latency. Traditional methods, unable to detect implicit mismatches at the planning level, often execute the correct actions on the wrong path, resulting in low repair efficiency.

[0068] Experimental results demonstrate that, compared to holistic analysis methods, EPC-AW significantly improves the success rate of attribution and planning from the baseline level in complex multi-agent collaborative scenarios. This fully verifies the significant effect of combining heterogeneous information consistency calibration and memory-driven cognitive constraint refinement on improving the robustness and engineering usability of intelligent information analysis assistants.

[0069] Table 1

[0070] Example 4 A planning and calibration device for an information analysis intelligent assistant includes a processor and a memory. The memory stores program instructions, and the processor invokes the program instructions stored in the memory to cause the device to execute the following method steps in Embodiment 1: Initializing a heterogeneous storage system: System-level memory Stored Query Execution history and descriptions of each agent's role Planning intelligent agents Execution agent and diagnostic intelligent agents Maintain private memory separately , , Initial set of evidence It is an empty set.

[0071] Execute the cognitive calibration planning loop: in each round of interaction iteration In this process, the processor driver performs the following detailed steps: 1. Candidate Plan Generation: The planning agent generates plans based on system memory and private policies. Generate candidate planning set .

[0072] 2. IPS Planning and Calibration Selection: By calling the IPS module, the device selects from the candidate planning set. In the process of identifying the program with the highest log-consistency score in heterogeneous information states, the target program is selected. .

[0073] 3. Planning, Execution, and Diagnosis: Instantiation of the execution agent. And obtain the execution result. The diagnostic agent evaluates the execution results and generates diagnostic feedback. .

[0074] 4. Dynamic updating of evidence set: If the diagnostic results show that the execution meets the plan, then verification evidence is extracted from the execution results. And stored in the evidence set .

[0075] 5. CESR Cognitive Refinement Update: Device Computational Planning Agent Optimal Preferences and Disagreement signals between them; if a disagreement exists, cognitive constraints are sensed through the CESR module. And update it to the planning agent's private memory. It is used to constrain the generation of the next round of planning.

[0076] 6. Termination Judgment and Output: The diagnostic agent, based on accumulated evidence... Determine if the termination criteria are met. If met, the processor outputs the final answer. If the conditions are not met and the maximum number of rounds has not been reached; Then proceed to the next cycle.

[0077] The complete interaction sequence includes unstructured text information such as multi-turn dialogue content, internal reasoning chains, tool call parameters, and environmental feedback.

[0078] The aforementioned processor and memory can be devices with computing capabilities, such as computers, microcontrollers, microcontrollers, or cloud servers. In actual implementation, this embodiment of the invention does not limit the execution device and selects it based on the computational load in the actual application (such as the need for large language model inference). Data signals are transmitted between the memory and the processor via a system bus.

[0079] Based on the same inventive concept, embodiments of the present invention also provide a computer-readable storage medium, the storage medium including a stored program, which, when the program is running, controls the device where the storage medium is located to execute the method steps in the above embodiments.

[0080] The computer-readable storage medium includes, but is not limited to, flash memory, hard disk, solid-state drive, etc.

[0081] It should be noted that the description of the readable storage medium in the above embodiments corresponds to the description of the method in the embodiments, and the embodiments of the present invention will not be repeated here.

[0082] In the above embodiments, implementation can be achieved, in whole or in part, through software, hardware, firmware, or any combination thereof. When implemented in software, it can be implemented, in whole or in part, as a computer program product. A computer program product includes one or more computer instructions. When the computer program instructions are loaded and executed on a computer, all or part of the flow or function according to the embodiments of the present invention is generated.

[0083] A computer can be a general-purpose computer, a special-purpose computer, a computer network, or other programmable device. Computer instructions can be stored in or transmitted through a computer-readable storage medium. A computer-readable storage medium can be any available medium accessible to a computer or a data storage device such as a server or data center that integrates one or more available media. The available medium can be magnetic or semiconductor, etc.

[0084] Unless otherwise specified, the model numbers of the various devices in this embodiment of the invention are not limited, and any device that can perform the above functions is acceptable.

[0085] Those skilled in the art will understand that the accompanying drawings are merely schematic diagrams of a preferred embodiment, and the sequence numbers of the above embodiments of the present invention are for descriptive purposes only and do not represent the superiority or inferiority of the embodiments.

[0086] The above description is only a preferred embodiment of the present invention and is not intended to limit the present invention. Any modifications, equivalent substitutions, improvements, etc., made within the spirit and principles of the present invention should be included within the protection scope of the present invention.

Claims

1. A planning and calibration method for an information analysis intelligent assistant, characterized in that, The method includes: In each planning round, the planning agent generates a set of candidate strategies, which includes information retrieval objectives, combinations of retrieval keywords, and information processing flows. The candidate policy set is evaluated by multiple agents with heterogeneous private memories. Each agent is based on different information sources. A policy selection module with consistent information is constructed based on these agents, and the policy with the most stable evaluation results is selected as the optimal policy for this round. Record the discrepancies between the self-selected strategy and the selected optimal strategy for this round of planning, construct a consistency-guided calibration constraint module, and transform the discrepancy into a cognitive calibration constraint and store it in the memory of the planning agent; By using a strategy selection module that relies on information consistency, strategy selection optimization is achieved during single-round planning; and by combining the cognitive calibration constraints generated by the consistency-guided calibration constraint module, these constraints are passed down and gradually accumulated across different rounds, thus constraining the strategy generation of the planning agent. The trained agent is applied to open-domain web page retrieval tasks to generate accurate search results.

2. The planning and calibration method for an information analysis intelligent assistant according to claim 1, characterized in that, The information consistency strategy selection module is used to solve the problem of strategy selection bias caused by a single cognitive perspective in complex multi-step tasks by the information analysis intelligent assistant. The information consistency strategy selection module introduces a planning agent, an execution agent, and a diagnostic agent. The agents share the system-level memory but maintain their own private memory, forming a heterogeneous information state. The planning agent focuses on generating candidate strategies based on historical experience, the execution agent focuses on evaluating the actual feasibility of candidate strategies, and the diagnostic agent focuses on judging the rationality of candidate strategies in terms of information sufficiency and risk.

3. The planning and calibration method for an intelligent information analysis assistant according to claim 1, characterized in that, The information consistency strategy selection module is used to calculate the evaluation consistency score of each candidate strategy among different agents, which is used to measure the stability and reliability of each candidate strategy under multiple cognitive perspectives.

4. The planning and calibration method for an intelligent information analysis assistant according to claim 1, characterized in that, The consistency-guided calibration constraint module is used to continuously monitor whether there is a discrepancy between the planning agent's self-selected planning strategy based on its private information and the globally optimal planning strategy selected based on the information consistency scoring mechanism. When the two are inconsistent, it indicates that the planning agent has a bias in its grasp of task information in the current cognitive state, and a potential cognitive calibration misalignment event has occurred. Upon detecting such discrepancies, the consistency-guided calibration constraint module generates a lightweight cognitive constraint rule, which is written into the persistent memory of the planning agent and participates in policy generation as a constraint in subsequent planning processes.

5. The planning and calibration method for an intelligent information analysis assistant according to claim 1, characterized in that, The optimal strategy for this round is: ； in, Let be the set of candidate strategies executed in round t. For strategy Global information consistency score; ； in, For intelligent agents The k-th candidate strategy generated during the t-th round of execution Individual consistency score, It is the set of all intelligent agents.

6. The planning and calibration method for an intelligent information analysis assistant according to claim 4, characterized in that, The lightweight cognitive constraint rule is as follows: ； in, For planning intelligent agents Based on its own evaluation vector It chooses a strategy that it considers optimal. This is the optimal strategy for this round. This is the current system memory.

7. A planning and calibration system for an information analysis intelligent assistant, the system comprising: A processor and a memory, the memory storing program instructions, the processor calling the program instructions stored in the memory to cause the apparatus to perform the method of any one of claims 1-6.

8. A computer-readable storage medium, characterized in that, The computer-readable storage medium stores a computer program, the computer program including program instructions that, when executed by a processor, cause the processor to perform the method described in any one of claims 1-6.

Images