A Skill-Based Iterative Triggering Method Based on Asymmetric Credit Score and Joint Threshold

By combining asymmetric credit scores and a joint threshold mechanism with a large language model to revise skills, the problems of inaccurate skill quality assessment and untimely iteration in existing technologies are solved, and automated quality management and continuous optimization of the skill base are realized.

CN123117203APending Publication Date: 2026-11-17QISHU DATA INTELLIGENT SYST CO LTD
View PDF 1 Cites 0 Cited by

Patent Information

Application Number
CN202611618687.4
Authority / Receiving Office
CN · China
Patent Type
Applications(China)
Current Assignee / Owner
Filing Date
2026-10-20
Publication Date
2026-11-17

AI Technical Summary

Technical Problem

Existing technologies lack precise quantitative evaluation and automatic iterative optimization mechanisms for skill performance, resulting in the inability to identify and revise low-quality skills in a timely manner, and the iteration process is hampered by issues such as accidental triggering and quality assessment.

Method used

An asymmetric credit score mechanism is used to quantify the performance of skills. Iteration is triggered by combining the number of times used and the credit score threshold. Low-quality skills are revised through a large language model. After each iteration, the credit score and version number are reset to form an independent observation cycle. Combined with semantic retrieval, a negative feedback loop is formed.

Benefits of technology

It enables accurate identification and automatic iterative optimization of low-quality skills, prevents accidental triggering, ensures independent evaluation of new versions, and promotes the autonomous and continuous improvement of the skill library quality.

✦ Generated by Eureka AI based on patent content.

Smart Images

  • Figure CN123117203A_ABST
    Figure CN123117203A_ABST
Patent Text Reader

Abstract

This invention discloses a method for triggering skill iteration based on asymmetric credit score and joint threshold, comprising the following steps: obtaining the execution results of skills automatically extracted from interaction experience; performing asymmetric bidirectional adjustment on the credit score of the skill, increasing the credit score by a first increment when the execution is successful and decreasing the credit score by a second decrement when the execution fails; marking the skill as needing iteration after each recorded usage result when the cumulative usage reaches a usage threshold and the credit score is lower than a score threshold; revising the description and execution logic of the marked skill by calling a large language model in conjunction with recent relevant experience records to generate a new version of the skill; after the new version is generated, forcibly resetting the credit score to a neutral initial value, clearing the iteration mark and incrementing the iteration version number, so that the skill enters a new independent observation cycle; this method has the characteristics of autonomous quality management applicable to the skill library of a large language model intelligent agent.
Need to check novelty before this filing date? Find Prior Art

Description

Technical Field

[0001] This invention relates to the fields of artificial intelligence and large language model intelligent agents, specifically to a method for iteratively triggering generation skills based on asymmetric credit scores and joint thresholds. Background Technology

[0002] In recent years, Large Language Model (LLM) agent systems have begun to possess the ability to automatically extract reusable skills from historical interaction experiences. For example, NVIDIA's Voyager system, proposed in 2023, can automatically extract skill codes from game exploration and store them in a skill library; Trace2Skill realizes the paradigm of "skills automatically growing from experience" by executing trajectory distillation of skill definitions. In addition, Chinese patent CN118333086A discloses a general framework for agents to automatically extract and update memories from experience, covering the entire lifecycle of memory creation, retrieval, updating, and elimination.

[0003] However, all of the aforementioned existing technologies have varying degrees of shortcomings. The Voyager system's skill library only supports additions; once skill content is fixed, it cannot be automatically revised. Even if a skill repeatedly fails in actual use, the system cannot automatically identify and trigger revisions, relying solely on manual intervention or new skills to overwrite old ones. Trace2Skill lacks a quantitative feedback mechanism for skill usage effectiveness; skill quality assessment relies on manual review and lacks automated iteration triggering capabilities. The memory update mechanism disclosed in Chinese patent CN118333086A is based on time decay (replacing old with new), rather than quantitative feedback based on the actual execution effect of the skill, making it unable to identify low-quality skills based on usage performance. Furthermore, if a symmetrical scoring method is used for skill execution results, occasional successes may mask persistent failures, preventing low-quality skills from being identified in a timely manner. If iteration is triggered solely by a single score threshold, the first random failure of a skill may lead to false triggers. If historical scores of older versions are retained after iteration, historically low scores will continue to drag down the quality assessment of new versions. Therefore, existing technologies lack a closed-loop mechanism capable of accurately assessing the quality of automatically extracted skills based on execution performance, accurately triggering iterative optimization, and ensuring independent evaluation of new versions after iteration. Summary of the Invention

[0004] Therefore, it is necessary to provide a skill generation iteration triggering method based on asymmetric credit score and joint threshold. By asymmetrically quantifying the usage effect of automatically extracted skills and combining the joint triggering conditions of usage and score, low-quality skills that need iterative optimization can be accurately identified. After the iteration is completed, the score is reset and a new observation cycle is entered, so as to realize the automatic closed-loop improvement of skill quality.

[0005] To achieve the above objectives, the inventors provide a method for iteratively triggering generation skills based on asymmetric credit scores and joint thresholds, comprising the following steps:

[0006] S1, obtain the execution result of the skill automatically extracted from the interaction experience, and perform asymmetric bidirectional adjustment on the credit score of the skill according to the execution result. When the execution is successful, the credit score is increased by a first increment, and when the execution fails, the credit score is decreased by a second decrement. The first increment is less than the second decrement, and the initial value of the credit score is a preset neutral initial value.

[0007] S2, after each recording of the skill's usage result, determine whether the cumulative number of times the skill has been used has reached a preset usage threshold and whether the credit score is lower than a preset score threshold. When the cumulative number of times the skill has been used has reached the usage threshold and the credit score is lower than the score threshold, mark the skill as needing iteration.

[0008] S3, for skills marked as needing iteration, the large language model is invoked to revise the description and execution logic of the skills in conjunction with recent relevant experience records, generating a new version of the skills;

[0009] S4. After the new version of the skill is generated, the credit score of the skill is forcibly reset to the neutral initial value, the iteration mark is cleared, and the iteration version number is incremented, so that the skill enters a new independent observation period.

[0010] As a preferred embodiment of the present invention, it further includes a credit score and retrieval injection linkage step: during the semantic retrieval stage, the current credit score of the skill is read, and the current credit score is compared with a preset retrieval threshold. When the current credit score is lower than the retrieval threshold, the skill is excluded from the tool call list for injection into the large language model context, making the skill invisible during the semantic retrieval stage; after the skill completes iteration and the credit score is reset to the neutral initial value, the visibility of the skill in the semantic retrieval stage is restored; the retrieval threshold and the semantic retrieval similarity threshold are independent of each other, and the two are combined to take effect.

[0011] In a preferred embodiment of the present invention, step S3, which involves revising the large language model for skills marked as requiring iteration, includes the following steps:

[0012] S301, Read the current description, triggering conditions and execution logic of the marked skill;

[0013] S302, retrieve recent experience records related to the marked skill from the experience database;

[0014] S303, construct a revision suggestion that includes the current description, triggering conditions, execution logic and recent experience records, and submit the revision suggestion to the large language model;

[0015] S304, parse the revised content output by the large language model and write the revised content into the new version of the skill file;

[0016] S305, automatically increment the version number of the skill and re-register the skill to the skill registry center.

[0017] As a preferred embodiment of the present invention, the ratio of the first increment to the second reduction ranges from 1:1.5 to 1:3.

[0018] In a preferred embodiment of the present invention, the credit score is truncated after each asymmetric bidirectional adjustment, and the updated credit score is truncated to a preset minimum and maximum value range.

[0019] In a preferred embodiment of the present invention, an explicit null value judgment is used when reading the credit score. When the storage field corresponding to the credit score is null, the credit score is taken as the neutral initial value; when the value of the storage field is the preset minimum value, the minimum value is retained and it is not reset to the neutral initial value.

[0020] As a preferred embodiment of the present invention, it further includes an asynchronous decoupling execution step of marking the skill as requiring iteration and performing an iterative operation to revise the marked skill: the marking operation is completed synchronously each time the usage result of the skill is recorded, and the iterative operation is asynchronously triggered by the background scheduler after periodically scanning the marked skill. During the window period from the completion of the marking operation to the completion of the iterative operation, the credit score update generated by the continued use of the skill does not affect the established iterative marking.

[0021] In a preferred embodiment of the present invention, the tracking of cumulative usage counts and the determination of iteration markers are performed only on skills automatically extracted from interaction experience; manually written skills are not included in the tracking of cumulative usage counts and the determination of iteration markers.

[0022] In a preferred embodiment of the present invention, the skill usage effect data is persistently stored with the unique name of the skill as the primary key, and the usage effect data includes the cumulative number of uses, the cumulative number of successes, the current credit score, the number of completed iterations, the iteration flag, the most recent usage timestamp, and the most recent iteration timestamp.

[0023] In a preferred embodiment of the present invention, the credit score ranges from 0 to 1, the neutral initial value is 0.5, the first increment is 0.08, the second decrement is 0.15, the usage threshold is 5 times, and the score threshold is 0.3.

[0024] Unlike existing technologies, the above technical solution achieves the following beneficial effects:

[0025] (1) This method designs the credit score adjustment as an asymmetric two-way mechanism (the first increment is less than the second decrement), so that the credit score decreases faster when the skill fails than it increases when it succeeds, thereby ensuring that the credit score of continuously failed skills is quickly exposed. This overcomes the problem of missed iterations caused by occasional successes masking continuous failures and low-quality skills remaining in the skill library for a long time under the symmetric scoring method.

[0026] (2) This method designs the usage threshold and credit score threshold as joint triggering conditions. Both conditions must be met simultaneously to trigger iteration marking, ensuring that only skills that have been fully used and consistently perform poorly are identified as needing iteration. This prevents the accidental triggering of iteration for new skills with insufficient usage samples due to occasional failures, and also prevents high-scoring skills from being mistakenly marked due to reaching the usage threshold, thus achieving dual precise control of iteration triggering.

[0027] (3) This method establishes an independent observation period for each new version skill by forcibly resetting the skill credit score to a neutral initial value and incrementing the version number after the iteration is completed. This eliminates the drag effect of the historical low score of the old version on the quality evaluation of the new version. At the same time, since the failure penalty is greater than the success reward and the system is forcibly reset after the iteration, the system has a natural resistance to the behavior of manipulating credit scores by making a large number of calls in a short period of time.

[0028] (4) This method links credit score with semantic retrieval injection. When the skill credit score is lower than the retrieval threshold, it is automatically excluded from the tool call list of the large language model context, forming a negative feedback loop of skill quality. After the credit score of low quality skills drops, the exposure opportunities are automatically reduced to avoid continuing to produce erroneous execution results. At the same time, high-quality skills get more usage opportunities because the credit score is consistently high. The retrieval threshold and semantic similarity threshold are independently superimposed and take effect, respectively filtering the skill injection from the two dimensions of "whether it is trustworthy" and "whether it is relevant", which further improves the reliability of the tool call of the large language model.

[0029] (5) This method decouples the marking operation from the iteration operation asynchronously. The marking operation is completed synchronously each time a record is used, while the iteration operation is triggered asynchronously by the background scheduler at regular intervals. This avoids the iteration operation blocking the main process. The entire quality management closed loop from use to scoring, triggering, iteration, reset, and re-observation does not require manual intervention, thus realizing the autonomous and continuous improvement of the quality of the skill library. Attached Figure Description

[0030] Figure 1 This is a schematic diagram of the closed-loop process of the method described in a specific implementation. Detailed Implementation

[0031] To explain in detail the technical content, structural features, objectives, and effects of the technical solution, the following description is provided in conjunction with specific embodiments and accompanying drawings.

[0032] like Figure 1 As shown, this embodiment provides a method for triggering skill iteration based on asymmetric credit scores and joint thresholds. The core processing mechanism of this method includes: asymmetric bidirectional credit score adjustment of skill execution results (failure penalty is greater than success reward), iterative triggering based on joint conditions of usage frequency and credit score, invoking a large language model combined with recent experience to perform self-iterative revision of the labeled skill, and forcibly resetting the credit score to a neutral initial value and starting a new independent observation cycle after iteration. Furthermore, credit scores can be linked with semantic retrieval injection to form a negative feedback loop for skill quality. The synergistic effect of these mechanisms ensures that the entire quality management process, from skill usage to scoring, triggering, iteration, resetting, and re-observation, requires no manual intervention. Figure 1 The closed-loop process shown below will be explained in detail through multiple implementation methods.

[0033] In one implementation, the complete closed-loop process of the skill iteration triggering method based on asymmetric credit score and joint threshold is as follows. Taking the skill "auto_device_priority_sort" (device priority sorting) automatically extracted from historical interaction experience in a large language model intelligent agent system as an example, the goal of this skill is to prioritize capture devices according to device type and location. When the skill is entered into the database, its usage effect data is persistently stored in the effect record table with the skill's unique name as the primary key. The initial state is: the current credit score is set to a neutral initial value of 0.5, the cumulative number of uses is 0, the number of completed iterations is 0, and the iteration mark is unmarked. The storage fields of the effect record table include the skill's unique name, cumulative number of uses, cumulative number of successes, current credit score, number of completed iterations, iteration mark, recent use timestamp, and recent iteration timestamp. These fields together support the continuous updating of the credit score, the determination of joint triggering conditions, and the tracing of iteration history.

[0034] Each time the skill is adopted and executed by the large language model, the system performs asymmetric bidirectional adjustment on the credit score based on the execution result. Successful execution increases the credit score by a first increment, while failure decreases it by a second decrement, with the first increment strictly less than the second decrement. In this implementation, the first increment is 0.08, and the second decrement is 0.15, with a ratio of 1:1.875, satisfying the constraint that the ratio of the first increment to the second decrement ranges from 1:1.5 to 1:3. After each credit score update, a truncation process is immediately performed, cutting the updated credit score to a preset minimum value of 0 and a maximum value of 1, ensuring that the credit score always remains within the valid value range.

[0035] Taking the example of this skill being used 6 times consecutively, we can illustrate the process of accumulating asymmetric credit scores.

[0036] Initial state:

[0037] Current credit score ([0,1]): effectiveness_score = 0.5 (neutral initial score);

[0038] Total number of uses: usage_count=0;

[0039] Number of iterations completed: iteration_count=0;

[0040] Whether it is marked as needing iteration: need_iteration=false;

[0041] During the usage phase, the accumulation of asymmetric credit scores is shown in Table 1:

[0042] Table 1: Cumulative Table of Asymmetric Credit Scores

[0043] order result Credit score changes Current credit score usage_count 1 success +0.08 0.58 1 2 fail -0.15 0.43 2 3 fail -0.15 0.28 3 4 success +0.08 0.36 4 5 fail -0.15 0.21 5 6 fail -0.15 0.06 6

[0044] The first execution was successful, increasing the credit score from 0.5 to 0.58, with a total of 1 use. The second execution failed, decreasing the credit score from 0.58 to 0.43, with a total of 2 uses. The third execution failed, decreasing the credit score from 0.43 to 0.28, with a total of 3 uses. The fourth execution was successful, increasing the credit score from 0.28 to 0.36, with a total of 4 uses. The fifth execution failed, decreasing the credit score from 0.36 to 0.21, with a total of 5 uses. The sixth execution failed, decreasing the credit score from 0.21 to 0.06, with a total of 6 uses.

[0045] In the six uses mentioned above, there were two successes (cumulative increase of 0.16) and four failures (cumulative decrease of 0.60), resulting in a final credit score of 0.5 + 0.16 - 0.60 = 0.06. If a symmetrical scoring method were used (0.08 for success, 0.08 for failure), the final credit score would be 0.5 + 0.16 - 0.32 = 0.34, which is higher than the score threshold of 0.3, thus not triggering an iteration flag. This is precisely the problem of missed iterations caused by "occasional successes masking continuous failures" under the symmetrical scoring method. The asymmetric design makes the penalty for failure greater than the reward for success, causing the skill credit score for continuous failures to drop rapidly and be exposed, thereby solving the aforementioned missed iteration problem.

[0046] After each skill usage result is recorded, the system synchronously performs a joint trigger condition judgment: determining whether the cumulative number of uses has reached a preset usage threshold and whether the current credit score is lower than a preset score threshold. Only when both conditions are met simultaneously is the skill's iteration flag set to an iterative state requiring iteration. In this implementation, the usage threshold is 5 times, and the score threshold is 0.3. After the 4th use, the cumulative usage is 4 times, which does not reach the usage threshold of 5 times, so the condition is not met and the iteration flag is not triggered. After the 5th use, the cumulative usage is 5 times, reaching the usage threshold (condition A is met), and the current credit score is 0.21, which is lower than the score threshold of 0.3 (condition B is met). Both conditions are met simultaneously, and the system marks the skill as requiring iteration. The 6th use occurs within the window period after the iteration flag is established and before the background scheduler executes the iteration operation. The credit score continues to accumulate and decrease to 0.06, but the established iteration flag remains unchanged and is unaffected by changes in the credit score during the window period.

[0047] The marking and iteration operations are executed asynchronously and decoupled: the marking operation is completed synchronously each time a skill usage result is recorded, and is part of the main process; the iteration operation is triggered asynchronously by the background scheduler after periodically scanning all skills marked as requiring iteration, without blocking the main process. During the window period between the completion of the marking operation and the completion of the iteration operation, the credit score updates generated by continued skill usage do not affect the already established iteration markings. This asynchronous decoupling design ensures that while the system is performing the relatively time-consuming iteration revision operation, the main process of skill usage and scoring recording can still proceed normally.

[0048] After the background scheduler scans and identifies a skill marked as requiring iteration, it triggers a self-iterative execution process for that skill. The self-iterative execution process includes the following steps:

[0049] The system calls the iterate_skill method:

[0050] Read the current SKILL.md content of this skill (description, trigger conditions, execution logic);

[0051] Retrieve 8 recent new experience records related to this skill;

[0052] LLM building tip: "The following skills are not effective in actual use (low credit score). Please revise their descriptions, triggering conditions and execution logic based on recent new experiences."

[0053] LLM outputs revised skill descriptions and execution logic;

[0054] Write a new version of SKILL.md, increasing the version number from 0.1.0 to 0.1.1;

[0055] Clear the embedding cache (because the content has changed);

[0056] Call _mark_iterated;

[0057] needs_iteration=false;

[0058] iteration_count=1;

[0059] effectiveness_score=0.5 (forced reset to neutral score);

[0060] last_iterated_at = current time.

[0061] First, the current description, trigger conditions, and execution logic of the marked skill are read. Second, recent experience records related to the marked skill are retrieved from the experience database. Then, a revision suggestion containing the current description, trigger conditions, execution logic, and recent experience records is constructed and submitted to the large language model. Next, the revision content output by the large language model is parsed and written into the new version of the skill file. Finally, the version number of the skill is automatically incremented, and the skill is re-registered with the skill registry center.

[0062] After each iteration, the system performs the following reset operations on the skill: clears the iteration marker to an unmarked state; increments the number of completed iterations by 1; forcibly resets the credit score to a neutral initial value of 0.5; and records the timestamp of the most recent iteration. The credit score does not retain historical low scores from older versions but is forcibly reset to a neutral initial value. This is to ensure that historical low scores from older versions do not negatively impact the fair evaluation of the revised new version; simultaneously, because the penalty for failure is greater than the reward for success and a forced reset occurs after each iteration, the system has a natural resistance to manipulation of the credit score through a large number of calls in a short period; furthermore, the forced reset to a neutral initial value allows the new version to have an independent observation period, accumulating usage feedback again starting from 0.5. After each iteration, if the quality of the new version improves, the credit score will gradually increase with subsequent use; if quality issues persist, an iteration marker will be triggered again after at least 5 more uses and when the credit score falls below the score threshold again.

[0063] The aforementioned tracking and iteration marking of cumulative usage counts are only applied to skills automatically extracted from interactive experiences. Manually written skills are not included in the tracking and iteration marking of cumulative usage counts, thus limiting the scope of the automatic quality management mechanism to automatically generated skills and avoiding interference with skills that have undergone manual review.

[0064] When reading credit scores, the system employs explicit null value checks: the credit score is only initialized to a neutral initial value of 0.5 when the storage field corresponding to the credit score is null; when the value of the storage field is the preset minimum value of 0, this minimum value is retained and the score is not reset to a neutral initial value. The necessity of this mechanism is evident in the following scenario: after a skill experiences consecutive failures, the credit score is truncated to the lower limit of 0.0. If a truth check is used when reading the credit score, treating 0.0 as a false value and setting the neutral default value to 0.5, then valid 0.0 scores would be mistakenly judged as null and reset to 0.5, causing low-quality skills that have fallen to the lower limit to unexpectedly revert to a neutral score, thus escaping iteration triggers. Explicit null value checks only use a neutral initial value when the field is indeed null, meaning the skill is first entered into the database before any score values ​​are written, ensuring the consistency and reliability of the credit score system at boundary values.

[0065] In some implementations, based on the above basic implementation, the credit score can be linked with semantic retrieval injection to further form a negative feedback loop for skill quality. Specifically, during the semantic retrieval phase, the skill's current credit score is read and compared with a preset retrieval threshold. If the current credit score is lower than the threshold, the skill is excluded from the tool call list injected into the large language model context, making the skill invisible during the semantic retrieval phase. After the skill completes iteration and the credit score is reset to a neutral initial value, the skill's visibility during the semantic retrieval phase is restored.

[0066] The retrieval threshold, applied to the credit score dimension, and the similarity threshold (applied to the vector cosine similarity dimension) of semantic retrieval are two independent filtering conditions that work together: the similarity threshold determines whether a skill is "relevant" to the current user request, while the retrieval threshold determines whether a skill is "trustworthy." Only skills that pass both filtering conditions are injected into the tool invocation list of the large language model context. In this implementation, the retrieval threshold is set to 0.3. Taking the skill "auto_device_priority_sort" as an example, when its credit score drops to 0.21, which is lower than the retrieval threshold of 0.3, even if the semantic similarity between the user request and the skill meets the standard, the skill will not be injected into the tool invocation list and will remain temporarily invisible. This prevents low-quality skills from being discovered and invoked by the large language model, thus blocking the continuous accumulation of erroneous execution results. After the skill completes its iteration and its credit score is reset to 0.5, which is higher than the retrieval threshold of 0.3, the skill regains visibility and re-participates in semantic retrieval injection.

[0067] It should be noted that the credit score threshold filtering mentioned above only applies to automatically extracted skills with existing usage performance records. Since manually written skills are not included in the usage performance data tracking, they do not have corresponding records in the performance record table. Therefore, they are not included in the credit score threshold determination when constructing the tool call list and can always be injected regardless of semantic similarity. Only automatically extracted skills with existing usage performance records and a current credit score lower than the retrieval threshold will be excluded after semantic retrieval and before being injected into the tool call list. Explicit null value judgment is used when reading the credit score. Automatically extracted skills that have fallen to the lower limit of 0.0 are correctly judged as being below the retrieval threshold and excluded, and will not be allowed entry due to false null value judgments.

[0068] The aforementioned linkage mechanism between credit score and semantic retrieval injection forms a negative feedback loop for skill quality: high-quality skills, due to their consistently high credit scores, are frequently retrieved and injected, gaining more usage opportunities and further solidifying their high scores; low-quality skills, due to their continuously declining credit scores, are automatically removed from the tool call list, their usage opportunities decrease, and they enter a waiting-for-iteration state. After iteration, their credit scores are reset and they are re-evaluated. This loop allows the skill library to naturally form a dynamic quality balance of survival of the fittest.

[0069] In another implementation, the condition "cumulative usage count reaches a preset usage count threshold" in the joint triggering condition can be replaced with the condition "consecutive failure count reaches a preset consecutive failure count threshold". Specifically, the system maintains a consecutive failure counter for each skill. Each time a skill execution fails, the consecutive failure counter is incremented by 1; each time a skill execution succeeds, the consecutive failure counter is reset to 0. When the consecutive failure count reaches the preset consecutive failure count threshold and the current credit score is lower than the score threshold, the skill is marked as needing iteration. This alternative implementation can also prevent false triggering based on a single random failure, but it identifies skills with "consecutive and stable failures," and its ability to identify low-quality skills with "alternating successes and failures" is relatively weak. It is suitable for application scenarios where skill execution results have a strong temporal correlation. In contrast, using a cumulative usage count threshold method has a better ability to identify various failure distribution patterns.

[0070] It should be noted that although the above embodiments have been described herein, this does not limit the scope of patent protection of the present invention. Therefore, any changes and modifications made to the embodiments described herein based on the innovative concept of the present invention, or equivalent structural or procedural transformations made using the content of the present invention's specification and drawings, directly or indirectly applying the above technical solutions to other related technical fields, are all included within the scope of patent protection of the present invention.

Claims

1. A method for iteratively triggering skill generation based on asymmetric credit score and joint threshold, characterized in that, Includes the following steps: S1, obtain the execution result of the skill automatically extracted from the interaction experience, and perform asymmetric bidirectional adjustment on the credit score of the skill according to the execution result. When the execution is successful, the credit score is increased by a first increment, and when the execution fails, the credit score is decreased by a second decrement. The first increment is less than the second decrement, and the initial value of the credit score is a preset neutral initial value. S2, after each recording of the skill's usage result, determine whether the cumulative number of times the skill has been used has reached a preset usage threshold and whether the credit score is lower than a preset score threshold. When the cumulative number of times the skill has been used has reached the usage threshold and the credit score is lower than the score threshold, mark the skill as needing iteration. S3, for skills marked as needing iteration, the large language model is invoked to revise the description and execution logic of the skills in conjunction with recent relevant experience records, generating a new version of the skills; S4. After the new version of the skill is generated, the credit score of the skill is forcibly reset to the neutral initial value, the iteration mark is cleared, and the iteration version number is incremented, so that the skill enters a new independent observation period.

2. The skill generation iteration triggering method based on asymmetric credit score and joint threshold according to claim 1, characterized in that, It also includes a credit score and retrieval injection linkage step: during the semantic retrieval stage, the current credit score of the skill is read, and the current credit score is compared with a preset retrieval threshold. When the current credit score is lower than the retrieval threshold, the skill is excluded from the tool call list for injection into the large language model context, so that the skill is not visible during the semantic retrieval stage. After the skill completes its iteration and the credit score is reset to the neutral initial value, the visibility of the skill in the semantic retrieval stage is restored; the retrieval threshold and the semantic retrieval similarity threshold are independent of each other, and the two are combined to take effect.

3. The skill generation iteration triggering method based on asymmetric credit score and joint threshold according to claim 1, characterized in that, In step S3, the step of revising the large language model for skills marked as needing iteration includes the following steps: S301, Read the current description, triggering conditions and execution logic of the marked skill; S302, retrieve recent experience records related to the marked skill from the experience database; S303, construct a revision suggestion that includes the current description, triggering conditions, execution logic and recent experience records, and submit the revision suggestion to the large language model; S304, parse the revised content output by the large language model and write the revised content into the new version of the skill file; S305, automatically increment the version number of the skill and re-register the skill to the skill registry center.

4. The skill generation iteration triggering method based on asymmetric credit score and joint threshold according to claim 1, characterized in that: The ratio of the first increment to the second decrease ranges from 1:1.5 to 1:

3.

5. The skill generation iteration triggering method based on asymmetric credit score and joint threshold according to claim 1, characterized in that: The credit score is truncated after each asymmetric bidirectional adjustment, and the updated credit score is truncated to a preset minimum and maximum value range.

6. The skill generation iteration triggering method based on asymmetric credit score and joint threshold according to claim 5, characterized in that, When reading the credit score, an explicit null value judgment is used. When the storage field corresponding to the credit score is null, the credit score is taken as the neutral initial value; when the value of the storage field is the preset minimum value, the minimum value is retained and it is not reset to the neutral initial value.

7. The skill generation iteration triggering method based on asymmetric credit score and joint threshold according to claim 1, characterized in that, It also includes an asynchronous decoupling execution step between the marking operation that marks the skill as needing iteration and the iterative operation that revises the marked skill: the marking operation is completed synchronously each time the usage result of the skill is recorded, and the iterative operation is asynchronously triggered by the background scheduler after periodically scanning the marked skill. During the window period from the completion of the marking operation to the completion of the iterative operation, the credit score update generated by the continued use of the skill does not affect the established iterative marking.

8. The skill generation iteration triggering method based on asymmetric credit score and joint threshold according to claim 1, characterized in that: The tracking of cumulative usage counts and the determination of iteration markers are only performed on skills automatically extracted from interaction experience; manually written skills are not included in the tracking of cumulative usage counts and the determination of iteration markers.

9. The skill generation iteration triggering method based on asymmetric credit score and joint threshold according to claim 1, characterized in that: The skill usage effect data is persistently stored with the skill's unique name as the primary key. The usage effect data includes the cumulative number of uses, the cumulative number of successes, the current credit score, the number of completed iterations, the iteration flag, the most recent usage timestamp, and the most recent iteration timestamp.

10. The skill generation iteration triggering method based on asymmetric credit score and joint threshold according to any one of claims 1 to 9, characterized in that: The credit score ranges from 0 to 1, the initial neutral value is 0.5, the first increment is 0.08, the second decrement is 0.15, the usage threshold is 5 times, and the score threshold is 0.3.

Citation Information

Patent Citations

  • Agent program memory system capable of self-evolution and agent

    CN118333086A