Method and apparatus for controlling content generation using generative ai

The content generation control system addresses repetitive and policy-violating issues in generative AI by using a closed-loop method with prompt configuration and verification, enhancing content diversity and compliance.

KR103043684B1Active Publication Date: 2026-11-06EN HANCE
View PDF 4 Cites 0 Cited by

Patent Information

Application Number
KR1020260125975
Authority / Receiving Office
KR · KR
Patent Type
Patents
Current Assignee / Owner
Filing Date
2026-07-09
Publication Date
2026-11-06
Estimated Expiration
2046-07-09

AI Technical Summary

Technical Problem

Generative AI systems often generate repetitive content, violate policies unintentionally, fail to connect with specific details of the target, and repeat verification failures, especially when generation parameters change.

Method used

A content generation control method and apparatus that includes a prompt configuration unit to avoid repetition, a verification unit to check policy compliance and repetition, and a regeneration process to address failure reasons, using a closed-loop system to ensure content diversity and adherence to policies.

Benefits of technology

Simultaneously suppresses repetitive and policy-violating content, reduces repeated failures, and diversifies generated content by integrating failure reasons into regeneration prompts, ensuring content relevance and compliance.

✦ Generated by Eureka AI based on patent content.

Smart Images

  • Figure R1020260125975_ABST
    Figure R1020260125975_ABST
Patent Text Reader

Abstract

The present invention relates to a method and apparatus for controlling content generation using generative artificial intelligence. A content generation control apparatus according to one embodiment of the present invention comprises: a prompt configuration unit that configures a generation prompt including instructions to avoid repetition with previously generated content based on target content and previously generated content; a content acquisition unit that acquires candidate content from a generative artificial intelligence model into which the generation prompt is input; and a verification unit that verifies whether the candidate content satisfies a policy and whether the candidate content has a repeating pattern with respect to previously generated content. If verification fails, the prompt configuration unit configures a regeneration prompt including the reason for failure, and the content acquisition unit acquires the regenerated candidate content. Accordingly, policy violations of the content and convergence into a repeating pattern can be suppressed.
Need to check novelty before this filing date? Find Prior Art

Description

Technology Field

[0001] The present invention relates to a content generation control technology using generative artificial intelligence, and more specifically, to a method and apparatus for controlling generation so that the generated content does not converge into a repetitive pattern while conforming to a preset policy. Background Technology

[0003] Generative AI can automatically generate text content for an account. For example, generative AI can generate content such as replies, comments, quotes, messages, or community responses to a target post based on the body and context of the target post.

[0004] However, the method of generating content by relying solely on a single prompt has the following problems.

[0005] First, the writing style of the generated content may be repetitive, giving a mechanical impression that distinguishes it from human-written content. For example, specific opening phrases, sentence structures, or vocabulary may appear repeatedly across multiple pieces of content.

[0006] Second, expressions that are not permitted by policy may be unintentionally included in the generated content. For example, account names, product names, links, calls to one's own profile, follow requests, or direct promotional phrases may be included in the generated content.

[0007] Third, the generated content may fail to connect with the specific details of the target and remain limited to generalizations. In this case, the generated content cannot respond to the specific clues of the target post.

[0008] Fourth, the reason for the failure to validate the generated content is not reflected in subsequent generation, so the same type of error may occur repeatedly.

[0009] Fifth, if generation is attempted using sampling parameters not supported by the generation model, generation may fail, and the likelihood of such failure may increase if the generation model or its version is changed.

[0010] As such, relying solely on the content generation capabilities of generative AI makes it difficult to reliably generate content that complies with policies and does not converge into repetitive patterns; therefore, technology to control generation at the pre- and post-generation stages is required. Prior art literature

[0012] Korean Published Patent No. 10-2025-0113091 (July 25, 2025), Generative AI-based automatic question-answer system and method for executing the same The problem to be solved

[0013] The objective of the present invention is to provide a method and apparatus for controlling content generation using generative artificial intelligence.

[0014] Another objective of the present invention is to provide a content generation control method and apparatus that controls generation so that the generated content conforms to a preset policy but does not converge into a repetitive pattern.

[0015] Another objective of the present invention is to provide a content generation control method and apparatus that reflects the reason for the failure of content verification in subsequent generation, thereby preventing the same type of error from being repeated while the generated content is connected to specific clues of the target.

[0016] However, the technical problems that the present invention aims to solve are not limited to those described above, and other unmentioned problems can be clearly understood by those skilled in the art from the description of the invention below. means of solving the problem

[0018] According to one embodiment of the present invention for achieving the above objective, a content generation control device is disclosed, comprising: a prompt configuration unit that configures a generation prompt including instructions to avoid repetition with said previous generation content based on a target content and said previous generation content; a content acquisition unit that acquires candidate content from a generational artificial intelligence model into which the generation prompt is input; and a verification unit that verifies whether said candidate content satisfies a preset policy and whether said candidate content has a repeating pattern with respect to said previous generation content, wherein the prompt configuration unit configures a regeneration prompt including a reason for failure when the verification by the verification unit fails, and the content acquisition unit acquires regenerated candidate content from the generational artificial intelligence model into which the regeneration prompt is input.

[0019] In one embodiment, the verification unit may calculate a repetition score by combining two or more of the partial interval similarity, n-gram overlap, embedding similarity, frequent vocabulary inclusion, and sentence structure similarity between the candidate content and the previously generated content, and determine whether the repetition pattern is present based on the repetition score.

[0020] In one embodiment, the prompt component may identify a factor that increased the repetition score and include instructions to avoid the identified factor in the regeneration prompt.

[0021] In one embodiment, the verification unit classifies the candidate content as a regeneration target if the repetition score is greater than or equal to a first threshold, and classifies the candidate content as a human approval target if the repetition score is less than the first threshold and greater than or equal to a second threshold.

[0022] In one embodiment, the prompt configuration unit may configure a failure reason vector by combining a failure reason due to policy verification failure of the verification unit and a failure reason due to repeated pattern detection, and may configure the regeneration prompt using the failure reason vector.

[0023] The content acquisition unit acquires a skip signal from the generative artificial intelligence model and transmits the skip signal to the verification unit, and when the skip signal is acquired, the verification by the verification unit for the target content may not be performed.

[0024] In one embodiment, the prompt configuration unit may configure the generation prompt based on one generation mode selected from a plurality of content generation modes.

[0025] In one embodiment, the generation mode is selected in a performance-based manner, and the performance-based method can adjust the selection weight for each generation mode based on the verification result of the previous generation content.

[0026] In one embodiment, the prompt configuration unit may configure the generation prompt that instructs to respond to one or more specific clues among numbers, names, claims, events, and terms included in the target content.

[0027] In one embodiment, the previously generated content may be stored in association with one or more of the content body, the generation mode used for generation, the reason for failure, the number of regenerations, whether it has been published, and the target content identifier.

[0028] According to one embodiment of the present invention, to achieve the above objective, a method for controlling content generation using a generative artificial intelligence is disclosed, comprising: a step of configuring a generation prompt including instructions to avoid repetition with said previous generated content based on a target content and said previous generated content; a step of obtaining candidate content from a generative artificial intelligence model into which the generation prompt is input; a step of verifying whether said candidate content satisfies a preset policy and whether said candidate content has a repeating pattern with respect to said previous generated content; a step of configuring a regeneration prompt including reasons for failure if said verification fails; and a step of obtaining regenerated candidate content from the generative artificial intelligence model into which the regeneration prompt is input.

[0029] In one embodiment, the verification step may include calculating a repetition score by combining two or more of partial interval similarity, n-gram overlap, embedding similarity, frequent vocabulary inclusion, and sentence structure similarity between the candidate content and the previously generated content, and determining whether the repetition pattern is present based on the repetition score.

[0030] In one embodiment, the step of configuring the regeneration prompt may include identifying a factor that increased the repetition score and including instructions to avoid the identified factor in the regeneration prompt.

[0031] In one embodiment, the verification step may include classifying the candidate content as a regeneration target if the repetition score is greater than or equal to a first threshold, and classifying the candidate content as a human approval target if the repetition score is less than the first threshold and greater than or equal to a second threshold.

[0032] In one embodiment, the step of configuring the regeneration prompt may include configuring a failure reason vector by combining a failure reason due to a failure in verification of the policy and a failure reason due to the detection of a repeating pattern, and configuring the regeneration prompt using the failure reason vector.

[0033] In one embodiment, the content generation control method may further include the step of obtaining a skip signal from the generative artificial intelligence model; and the step of terminating the content generation process for the target content when the skip signal is obtained.

[0034] In one embodiment, the step of configuring the generation prompt may include the step of configuring the generation prompt based on one generation mode selected from a plurality of content generation modes.

[0035] In one embodiment, the generation mode is selected in a performance-based manner, and the performance-based method can adjust the selection weight for each generation mode based on the verification result of the previous generation content.

[0036] In one embodiment, the step of configuring the generation prompt may include the step of configuring the generation prompt to instruct to respond to one or more specific clues among numbers, names, claims, events, and terms included in the target content.

[0037] In one embodiment, the previously generated content may be stored in association with one or more of the content body, the generation mode used for generation, the reason for failure, the number of regenerations, whether it has been published, and the target content identifier. Effects of the invention

[0039] According to one embodiment of the present invention, whether the generated content violates a preset policy and whether the generated content has a repeating pattern with respect to previous generated content can be verified together within a single generation control flow. Accordingly, repetitive content that is not filtered by a method that checks only the policy compliance of the content itself, and policy-violating content that is not filtered by a method that checks only for repetition, can be suppressed simultaneously. Furthermore, by including an instruction in the generation prompt to avoid repetition with previous generated content, the generation of repetitive content can be suppressed in advance at the generation stage, while repeating patterns in the generated content can be detected afterward.

[0040] According to one embodiment of the present invention, when verification fails, the likelihood of the same type of verification failure occurring repeatedly can be reduced by configuring a regeneration prompt that includes the reason for the failure and regenerating candidate content. In particular, when configuring a regeneration prompt using a failure reason vector that combines the reason for failure based on policy verification and the reason for failure based on repeat pattern detection, heterogeneous reasons for failure can be integrated and reflected into a single regeneration instruction. Furthermore, by identifying the factor that increased the repetition score and including instructions to avoid it in the regeneration prompt, closed-loop control can be implemented that feeds the cause of the repetition back to subsequent generation.

[0041] According to one embodiment of the present invention, the style and form of the generated content can be diversified by configuring a generation prompt based on a generation mode selected from a plurality of content generation modes, and cyclic control can be implemented by feeding back the verification results to the mode selection by adjusting the selection weight for each generation mode based on the verification results of the previously generated content. In addition, the possibility of generating general content unrelated to the target content can be reduced by configuring a generation prompt that instructs a response to specific clues included in the target content, and the processing path of the candidate content can be differentiated according to the degree of repetition by classifying the candidate content as a target for regeneration or a target for human approval based on a plurality of thresholds for repetition scores.

[0042] However, the effects obtainable through the present invention are not limited to those described above, and other unmentioned technical effects will be clearly understood by those skilled in the art from the description of the invention below. Brief explanation of the drawing

[0044] FIG. 1 is a block diagram showing the configuration of a content creation control device according to one embodiment of the present invention. FIG. 2 is a flowchart showing the operation of a verification unit according to one embodiment of the present invention. FIG. 3 is a data flow diagram showing the configuration process of a regeneration prompt according to one embodiment of the present invention. FIG. 4 is a flowchart illustrating the process of saving and reusing previously generated content according to one embodiment of the present invention. FIG. 5 is a flowchart illustrating a content creation control method according to an embodiment of the present invention. FIG. 6 is a diagram showing the operation between components in a verification and regeneration process according to an embodiment of the present invention. FIG. 7 is a block diagram showing the hardware configuration of a content creation control device according to one embodiment of the present invention. Specific details for implementing the invention

[0045] An embodiment of the present invention will be described in detail below with reference to the attached drawings. Prior to this, terms and words used in this specification and claims should not be interpreted as being limited to their ordinary or dictionary meanings. Instead, based on the principle that the inventor can appropriately define the concept of a term to best describe their invention, they should be interpreted in a meaning and concept consistent with the technical spirit of the present invention. Therefore, the embodiments described in this specification and the configurations illustrated in the drawings are merely partial embodiments of the present invention and do not represent all of the technical spirit of the present invention. It should be understood that various equivalents and modifications capable of replacing them may exist at the time of filing this application.

[0046] Additionally, as used herein, “comprise, include” and / or “comprising, including” specify the presence of the mentioned features, numbers, steps, actions, parts, elements, and / or groups thereof, and do not exclude the presence or addition of one or more other features, numbers, actions, parts, elements, and / or groups.

[0047] Additionally, to aid in understanding the invention, the attached drawings are not drawn to actual scale, and the dimensions of some components may be exaggerated. Furthermore, the same reference numerals may be assigned to identical components in different embodiments.

[0048] The statement that two subjects of comparison are 'identical' means that they are 'substantially identical.' Therefore, substantial identity may include deviations considered low in the industry, for example, deviations within 5%. Additionally, the statement that a parameter is uniform in a given area may mean that it is uniform from an average perspective.

[0049] Although terms such as "first," "second," etc., are used to describe various components, it goes without saying that these components are not limited by these terms. These terms are used merely to distinguish one component from another, and unless specifically stated otherwise, the first component may also be the second component.

[0050] Throughout the specification, unless specifically stated otherwise, each component may be singular or plural.

[0051] The fact that any configuration is placed on the "upper (or lower)" of a component or on the "upper (or lower)" of a component may mean not only that any configuration is placed in contact with the upper (or lower) surface of said component, but also that another configuration may be interposed between said component and any configuration placed on (or below) said component.

[0052] Furthermore, where one component is described as being "on," "connected to," or "coupled to" another component, it should be understood that while the components may be directly connected or coupled to each other, another component may be "interposed" between each component, or that each component may be "connected," "coupled," or "coupled" through another component.

[0053] As used herein, the term “and / or” includes any and all combinations of one or more of the associated listed items. Additionally, when describing embodiments of the invention, the use of “may” relates to “one or more embodiments of the invention.” Expressions such as “one or more” preceding a list of elements modify the entire list of elements and do not modify individual elements of the list.

[0054] Throughout the specification, "A and / or B" means A, B, or A and B unless specifically stated otherwise, and "C to D" means C or more and D or less, unless specifically stated otherwise.

[0055] When syntax such as "at least one of A, B, and C", "at least one of A, B, or C", "at least one selected from the group of A, B, and C", or "at least one selected from A, B, and C" is used to specify a list of elements A, B, and C, the syntax can refer to any suitable combination.

[0056] The term "use" may be considered synonymous with the term "utilize." As used herein, terms such as "substantially," "about," and similar terms are used as terms of approximation rather than degree, and are intended to account for the inherent variation of a measured or calculated value that would be recognized by a person skilled in the art.

[0057] In this specification, terms such as first, second, third, etc. may be used to describe various elements, components, regions, layers, and / or sections, but these elements, components, regions, layers, and / or sections should not be limited by these terms. These terms are used to distinguish one element, component, region, layer, or section from another element, component, region, layer, or section. Accordingly, the first element, component, region, layer, or section discussed below may be named the second element, component, region, layer, or section without departing from the teachings of the exemplary embodiments.

[0058] Spatial relative terms such as "beneath," "below," "lower," "above," and "upper" may be used in the specification for ease of description to explain the relationship between one element or feature and other element(s) or feature(s) as illustrated in the drawings. Spatially relative positions are to be understood as encompassing different orientations of the device in use or operation, in addition to the orientations depicted in the drawings. For example, if the device in the drawings is inverted, the other element is understood as "below" or "below," and the described element is understood as "above" or "upper" of the other element. Thus, the term "below" may encompass both the up and down directions.

[0059] The terms used in this specification are intended to describe embodiments of the invention and are not intended to limit the invention.

[0060] In this specification, "content" refers to a text-based result that can be generated by generative artificial intelligence. The content may include, but is not limited to, content in the form of replies, comments, quotes, messages, or community responses to the target content.

[0061] In this specification, target content refers to content that is the target of content creation. Target content may include not only the body of a post but also at least one of the author information, surrounding context, and reason for selection of said post.

[0062] In this specification, "previously generated content" refers to content previously generated by the control of the generative artificial intelligence and the present invention. Previously generated content is not limited to a specific single piece of content and may be a set of multiple pieces of content accumulated over a predetermined period or within a predetermined number of items. Previously generated content is distinct from the target content and does not imply the generation history of the target content.

[0063] In this specification, "policy" refers to information defining conditions that the content must satisfy. The policy may include conditions regarding at least one of the following: the length of the content, the number of sentences, the target language, prohibited words, subjects whose direct mention is prohibited, restrictions on the use of symbols, the prohibition of self-promotion, and the prohibition of calls to action.

[0064] In this specification, a repetition pattern refers to the repetitive similarity that candidate content has with respect to previously generated content. The repetition pattern may include at least one of similarity of partial segments of content, duplication of sequences of consecutive words or tokens, similarity of meaning, repetition of vocabulary, and similarity of sentence structure, and may be determined based on a repetition score between the candidate content and the previously generated content.

[0065] In this specification, a specific clue refers to an individual information element uniquely included in the target content that distinguishes the target content from other content. A specific clue may include at least one of a number, name, claim, event, and term included in the target content, and is distinguished from a general expression that can be applied to any content without relying on the specific target content.

[0066] In this specification, a generation prompt refers to information input to a generative artificial intelligence to generate content. A generation prompt may include at least one of target content, a policy, a selected generation mode, and instructions to avoid repetition with previously generated content.

[0067] In this specification, a skip signal refers to a signal indicating that the generation of content is suspended, which is output by a generative artificial intelligence instead of generating candidate content.

[0068] FIG. 1 is a block diagram showing the configuration of a content creation control device according to one embodiment of the present invention.

[0069] Referring to FIG. 1, the content creation control device (100) can control the creation of content by linking with a generative artificial intelligence model (200). The content creation control device (100) does not directly create content, but rather configures a creation prompt so that the generative artificial intelligence model (200) creates content that conforms to a policy, verifies the content obtained from the generative artificial intelligence model (200), and performs control to reflect the verification results in subsequent creation. At this time, the generative artificial intelligence model (200) may be configured to be located outside the content creation control device (100) and link with the content creation control device (100), and according to an embodiment, it may be integrated into a single system with the content creation control device (100).

[0070] The content creation control device (100) may include a prompt configuration unit (110), a content acquisition unit (120), and a verification unit (130). In one embodiment, the content creation control device (100) may further include at least one of a publishing gate unit (140) and a history storage unit (150).

[0071] The prompt configuration unit (110) can configure a generation prompt including instructions to avoid repetition with the previous generation content based on the target content and the previous generation content. The prompt configuration unit (110) can provide the configured generation prompt to a generative artificial intelligence model (200). In one embodiment, the prompt configuration unit (110) can configure a generation prompt based on one generation mode selected from a plurality of content generation modes. Additionally, the prompt configuration unit (110) can configure a generation prompt instructing to respond to one or more specific clues among numbers, names, claims, events, and terms included in the target content.

[0072] The content acquisition unit (120) can acquire candidate content from a generative artificial intelligence model (200) into which a generation prompt has been input. In one embodiment, the content acquisition unit (120) can acquire a skip signal from the generative artificial intelligence model (200), in which case verification of the candidate content may not be performed. Additionally, if the sampling parameter used when calling the generative artificial intelligence model (200) is rejected, the content acquisition unit (120) can re-call the generative artificial intelligence model (200) by removing at least a portion of the said sampling parameter. Accordingly, the content generation control device (100) can increase compatibility with changes to a plurality of generative artificial intelligence models or versions thereof.

[0073] The verification unit (130) can verify whether the candidate content satisfies the policy and whether the candidate content has a repeating pattern with respect to the previously generated content. If the verification by the verification unit (130) fails, the prompt configuration unit (110) can configure a regeneration prompt including the reason for the failure, and the content acquisition unit (120) can acquire the regenerated candidate content from the generative artificial intelligence model (200) into which the regeneration prompt was input. In this way, the content generation control device (100) can form a closed loop in which the generation, verification, and regeneration of candidate content are cyclically performed by the prompt configuration unit (110), the content acquisition unit (120), and the verification unit (130).

[0074] The posting gate unit (140) can selectively transfer candidate content that has passed the verification of the verification unit (130) to one of automatic posting processing, scheduled posting processing, or human approval processing.

[0075] The history storage unit (150) can store previously generated content. In one embodiment, the history storage unit (150) can store previously generated content in association with one or more of the content body, the generation mode used for generation, the reason for failure, the number of regenerations, whether it was published, and the target content identifier. Previously generated content stored in the history storage unit (150) can be reused as input for the repetition avoidance instruction of the prompt configuration unit (110) and the repetition pattern verification of the verification unit (130) when generating subsequent content.

[0076] FIG. 2 is a flowchart showing the operation of a verification unit according to one embodiment of the present invention.

[0077] Referring to FIG. 2, the verification unit (130) can verify whether the policy is satisfied for the candidate content and whether there is a repetition pattern for the previously generated content. For this verification, the verification unit (130) can calculate a repetition score that quantifies the degree of repetition between the candidate content and the previously generated content, and determine the processing path of the candidate content based on the calculated repetition score.

[0078] The verification unit (130) can receive candidate content (S210). The candidate content may be content obtained by the content acquisition unit (120) from the generative artificial intelligence model (200).

[0079] The verification unit (130) can verify whether the candidate content satisfies the policy (S220).

[0080] The verification unit (130) can check whether the candidate content satisfies one or more of the conditions regarding the length of the content, the number of sentences, the target language, prohibited words, subjects that are prohibited from being directly mentioned, restrictions on the use of symbols, the prohibition of self-promotion, and the prohibition of a call to action. If the candidate content does not satisfy the policy, the verification unit (130) can generate a reason for failure due to the failure of policy verification.

[0081] The verification unit (130) can calculate multiple indicators between candidate content and previous generated content (S230).

[0082] The verification unit (130) can calculate two or more of the partial interval similarity, n-gram overlap, embedding similarity, frequent vocabulary inclusion, and sentence structure similarity between candidate content and previous generated content.

[0083] The above partial segment similarity may be a value indicating the degree of similarity between a partial segment of candidate content and a partial segment of previously generated content. A partial segment is a portion of the content whose location is specified, and may include at least one of the beginning part, the ending part, and a part at a predetermined location of the content.

[0084] In one embodiment, a partial segment may be a starting segment corresponding to a predetermined range from the beginning of the content, and said starting segment may be defined as a portion from the first token of the content to a predetermined number of tokens, a portion from the first word of the content to a predetermined number of words, or the first sentence of the content. n-gram overlap may be a value indicating the degree of overlap in a sequence of n consecutive words or tokens between the candidate content and the previously generated content. Embedding similarity may be a value indicating the degree of similarity between the semantic vector of the candidate content and the semantic vector of the previously generated content. Frequent vocabulary inclusion may be a value indicating the degree to which vocabulary that appeared with a frequency exceeding a predetermined standard in the previously generated content is included in the candidate content. Sentence structure similarity may be a value indicating the degree of similarity in the structural form of the content, including at least one of the number of sentences, the distribution of sentence lengths, and the termination pattern.

[0085] The verification unit (130) can calculate a repeat score by weighting and combining the calculated indicators (S240).

[0086] The repetition score can be calculated as a weighted sum by applying a predetermined weight to each of two or more calculated indicators. By combining multiple indicators regarding different phases in this way, the verification unit (130) can detect repetitions that are not captured by any single indicator alone.

[0087] The verification unit (130) can determine whether the repetition score is greater than or equal to a first threshold (S250). If the repetition score is greater than or equal to the first threshold, the verification unit (130) can classify the candidate content as a regeneration target (S252). For the candidate content classified as a regeneration target, a reason for failure due to failure of repetition pattern detection may be generated, and the configuration of a regeneration prompt by the prompt configuration unit (110) may be performed.

[0088] If the repetition score is less than the first threshold, the verification unit (130) can determine whether the repetition score is greater than or equal to the second threshold (S260). If the repetition score is less than the first threshold and greater than or equal to the second threshold, the verification unit (130) can classify the candidate content as a target for human approval (S262). Candidate content classified as a target for human approval can be passed to processing for human approval instead of being automatically published.

[0089] If the repetition score is less than the second threshold, the verification unit (130) may determine that the candidate content has passed verification, and the candidate content that has passed verification may be processed into a waiting state for publication (S270).

[0090] Meanwhile, FIG. 2 illustrates an example in which a repetition score is compared with two thresholds to classify candidate content into one of the following: a regeneration target, a human approval target, or a pass, but this is exemplary. In another embodiment, the verification unit (130) may compare the repetition score with one threshold to determine only whether the candidate content has a repetition pattern. Additionally, the order of execution of the verification of policy satisfaction (S220) and the verification of the repetition pattern (S230 to S270) is not limited to the order shown, and the order may be reversed or performed in parallel.

[0091] FIG. 3 is a data flow diagram showing the configuration process of a regeneration prompt according to one embodiment of the present invention.

[0092] Referring to FIG. 3, the prompt configuration unit (110) can configure a regeneration prompt including the reason for failure when the verification of the verification unit (130) fails. The prompt configuration unit (110) can configure this regeneration prompt to include instructions to avoid factors that increase the repetition score, by reflecting both the reason for failure due to policy verification and the reason for failure due to the detection of a repetition pattern.

[0093] The prompt configuration unit (110) can collect the reason for failure due to failure of policy verification and the reason for failure due to failure of repeat pattern detection (S310). The reason for failure due to failure of policy verification may be information indicating the reason why the candidate content does not satisfy the policy, and the reason for failure due to failure of failure of repeat pattern detection may be information indicating the reason why the candidate content has a repeat pattern with respect to previously generated content. In one embodiment, each reason for failure may be expressed as a reason for failure code that identifies the type of failure.

[0094] The prompt configuration unit (110) can construct a failure reason vector by combining the collected failure reasons (S320). The failure reason vector may be information that includes both the failure reason resulting from the failure of policy verification and the failure reason resulting from the failure of repeat pattern detection. In one embodiment, the failure reason vector may include one or more of the failure reason resulting from the failure of policy verification, the failure reason resulting from the failure of repeat pattern detection, the repeat score, and the request change. By integrating the failure reasons resulting from different verifications, such as policy verification and repeat pattern detection, into a single failure reason vector, the prompt configuration unit (110) can reflect heterogeneous failure reasons as a single regeneration instruction.

[0095] The prompt configuration unit (110) can identify the factors that increased the repetition score (S330).

[0096] The prompt configuration unit (110) can identify an indicator that contributed to increasing the repetition score among a plurality of indicators combined in the calculation of the repetition score as a factor that increased the repetition score. For example, if the partial segment similarity is high because a partial segment of candidate content is similar to a partial segment of previously generated content, the prompt configuration unit (110) can identify the partial segment similarity as a factor that increased the repetition score.

[0097] The prompt configuration unit (110) can generate instructions to avoid identified factors (S340).

[0098] The prompt configuration unit (110) can generate at least one of instructions to use a different starting part, instructions to use a different vocabulary, and instructions to use a different sentence structure in response to the identified factor.

[0099] The prompt configuration unit (110) can configure a regeneration prompt by combining the previous failed candidate content and target content (S350).

[0100] The regeneration prompt may include one or more of a failure reason vector, instructions to avoid identified factors, previously failed candidate content, and target content. By outputting a regeneration prompt configured in this manner, the regeneration prompt can be input into a generative artificial intelligence model (200) through a content acquisition unit (120), and the generative artificial intelligence model (200) can generate regenerated candidate content based on the regeneration prompt.

[0101] Meanwhile, FIG. 3 illustrates an embodiment in which the configuration of a failure reason vector (S320), the identification of a factor that increased the repetition score (S330), and the generation of an avoidance instruction (S340) are performed sequentially, but this is exemplary. In other embodiments, the prompt configuration unit (110) may configure a regeneration prompt that includes only a failure reason due to a failure of policy verification or only a failure reason due to a failure of repetition pattern detection.

[0102] FIG. 4 is a flowchart illustrating the process of saving and reusing previously generated content according to one embodiment of the present invention.

[0103] Referring to FIG. 4, the history storage unit (150) can store content generated and verified by the content generation control device (100) as previously generated content, and the stored previously generated content can be reused as an input for suppressing and detecting repetition when generating subsequent content. Thus, the content generation control device (100) can form a structure in which the storage and reuse of previously generated content cycle.

[0104] The content creation control device (100) can determine the verification result and whether to publish (S410).

[0105] The verification result may include results regarding whether the policy is satisfied and whether a repeating pattern exists by the verification unit (130), and the posting status may include information regarding whether the content has been posted.

[0106] The content creation control device (100) can create a history record by associating a plurality of fields (S420). In one embodiment, the history record may include one or more fields among the content body, the creation mode used for creation, the reason for failure, the number of regenerations, whether it is published, and the target content identifier. Here, the content body may be information representing the created content itself, the creation mode used for creation may be information representing the creation mode used for creating the content, and the reason for failure may be information representing the reason for a verification failure that occurred during the creation process of the content. Additionally, the number of regenerations may be information representing the number of times the content has been regenerated, and the target content identifier may be information identifying the target content that became the target of creation for the content.

[0107] The content creation control device (100) can store the generated history record in the history storage unit (150) (S430). At this time, the content stored in the history storage unit (150) is not limited to published content. In one embodiment, the history storage unit (150) can store at least one of the content that was not published because it failed verification and the content that was discarded because it failed verification within the maximum number of regenerations, along with the reason for failure or the reason for discarding.

[0108] The content creation control device (100) can retrieve the stored history record as the previously created content when creating subsequent content (S440).

[0109] The content creation control device (100) can retrieve history records corresponding to a predetermined period or a predetermined number of ranges as previously created content.

[0110] The content creation control device (100) can reuse previously created content retrieved as input for repeated detection and mode selection (S450).

[0111] The retrieved previous generated content can be reused as an input for the prompt configuration unit (110) to configure a generation prompt that includes instructions to avoid repetition, and can be reused as an input for the verification unit (130) to detect a repetition pattern of candidate content. Additionally, the verification result of the retrieved previous generated content can be reused as an input to adjust the selection weight per generation mode when the generation mode is selected in a performance-based manner.

[0112] Meanwhile, FIG. 4 sequentially illustrates the process from the confirmation of verification results and whether to publish (S410) to the reuse of previously created content (S450), but this is exemplary. Previously created content stored in the history storage unit (150) can be repeatedly reused for the creation of multiple subsequent content, and the history storage unit (150) may be located inside the content creation control device (100) or implemented as an external storage device linked to the content creation control device (100).

[0113] FIG. 5 is a flowchart illustrating a content creation control method according to an embodiment of the present invention.

[0114] Referring to FIG. 5, the content creation control device (100) can proceed with the creation of content only when there is a specific clue in the target content, and if verification of the created candidate content fails, it can repeat the creation based on the reason for the failure. In this way, the content creation control device (100) can suppress the creation of content unrelated to the target content, while simultaneously creating content that conforms to the policy and is not repeated for previously created content.

[0115] The content creation control device (100) receives the target content and can check specific clues included in the target content (S510).

[0116] The content creation control device (100) can identify one or more specific clues among numbers, names, claims, events, and terms included in the target content.

[0117] The content generation control device (100) can determine whether there is a specific clue in the target content (S520). If there is no specific clue in the target content, the content generation control device (100) can obtain a skip signal from the generative artificial intelligence model (200), stop processing the target content, and record it (S515). In this case, since candidate content is not generated, verification of candidate content is not performed.

[0118] Verification regarding whether the candidate content itself satisfies the policy and verification regarding repeating patterns between the candidate content and previously generated content do not determine whether the candidate content is relevant to the target content. Therefore, general content that does not respond to specific clues of the target content may pass the above verification if it does not violate the policy and is not repeated with respect to previously generated content. By stopping the processing of target content for which specific clues do not exist due to a skip signal before the generation stage, the generation of general content that is not filtered out by the above verification can be suppressed.

[0119] If there is a specific clue in the target content, the content creation control device (100) can select one creation mode from a plurality of creation modes (S530).

[0120] In one embodiment, a plurality of generation modes may include two or more of a mode that responds to target content, a mode that agrees with target content, a mode that presents questions about target content, a mode that connects with experience, and a mode that presents alternative perspectives. The generation modes may be selected in one or more ways among equal random, weighted random, target rate-based, recent usage-based, and performance-based methods.

[0121] The content creation control device (100) can configure a creation prompt by combining a policy, previous creation content, and instructions to avoid repetition (S540).

[0122] In one embodiment, the generation prompt may include one or more of instructions regarding a selected generation mode, instructions to respond to specific clues of target content, instructions to comply with a policy, and instructions to avoid repetition with previous generation content.

[0123] The content creation control device (100) can obtain candidate content from the generative artificial intelligence model (200) (S550).

[0124] If the sampling parameter used when calling the generative artificial intelligence model (200) is rejected, the content creation control device (100) can call the generative artificial intelligence model (200) again by removing at least a portion of the said sampling parameter.

[0125] The content creation control device (100) can perform policy verification and repeat pattern detection on candidate content (S560). Policy verification and repeat pattern detection can be performed as described with reference to FIG. 2.

[0126] The content creation control device (100) can determine whether the candidate content has passed verification (S570). If the candidate content has not passed verification, the content creation control device (100) can configure a regeneration prompt based on the reason for failure (S575). In this case, the content creation control device (100) can obtain the regenerated candidate content from the generative artificial intelligence model (200) to which the regeneration prompt was input (S550), and verification can be performed again on the regenerated candidate content. Regeneration may be limited to a predetermined maximum number of times, and candidate content that has not passed verification within the said maximum number of times may be discarded.

[0127] If the candidate content passes verification, the content creation control device (100) can output the candidate content in a waiting state for publication and record it as a history (S580). The candidate content output in a waiting state for publication can be delivered to any one of automatic publication processing, scheduled publication processing, and human approval processing.

[0128] Meanwhile, FIG. 5 illustrates an embodiment in which the verification of specific clues (S510) and the determination of their existence (S520) precede the selection of a generation mode (S530), but this is exemplary. In other embodiments, the determination regarding specific clues may be performed during the generation of candidate content by a generative artificial intelligence model (200), in which case the generative artificial intelligence model (200) may output a skip signal instead of generating candidate content if it fails to detect specific clues.

[0129] FIG. 6 is a diagram showing the operation between components in a verification and regeneration process according to an embodiment of the present invention.

[0130] Referring to FIG. 6, the prompt configuration unit (110), content acquisition unit (120), generative artificial intelligence model (200), verification unit (130), and history storage unit (150) can perform generation, verification, and regeneration of candidate content by interacting with each other. If verification of candidate content fails, regeneration based on the reason for failure may be performed, and if verification passes, the content may be stored.

[0131] The prompt configuration unit (110) can transmit a creation prompt to the content acquisition unit (120) (S610). In one embodiment, the creation prompt may include one or more of instructions regarding a selected creation mode, instructions to comply with a policy, and instructions to avoid repetition with previous creation content.

[0132] The content acquisition unit (120) can input a generation prompt into a generative artificial intelligence model (200) (S620).

[0133] The generative artificial intelligence model (200) can generate candidate content based on a generation prompt, and the content acquisition unit (120) can receive candidate content from the generative artificial intelligence model (200) (S630).

[0134] The content acquisition unit (120) can transmit the returned candidate content to the verification unit (130) (S640). The verification unit (130) can verify whether the candidate content satisfies the policy and whether the candidate content has a repeating pattern with respect to the previously generated content.

[0135] If the candidate content does not satisfy the policy or the candidate content has a repeating pattern with respect to the previously generated content, the verification unit (130) can transmit the reason for failure to the prompt configuration unit (110) (S650). In one embodiment, the verification unit (130) can transmit a failure reason vector to the prompt configuration unit (110) by combining the reason for failure due to failure of policy verification and the reason for failure due to failure of repeating pattern detection.

[0136] The prompt configuration unit (110) can configure a regeneration prompt based on the reason for failure and transmit the regeneration prompt to the content acquisition unit (120) (S660).

[0137] The content acquisition unit (120) can input a regeneration prompt into the generative artificial intelligence model (200) (S670), and can receive the regenerated candidate content from the generative artificial intelligence model (200) (S680).

[0138] The content acquisition unit (120) can transmit the regenerated candidate content to the verification unit (130) so that re-verification is performed (S690).

[0139] If the candidate content passes verification, the verification unit (130) may transmit the final content, the generation mode used to generate the content, and the verification result to the history storage unit (150) for storage (S700). The content stored in the history storage unit (150) can be used to generate subsequent content as previously generated content.

[0140] The history storage unit (150) can provide the stored previous generated content to the prompt configuration unit (110) when generating subsequent content (S710). Thereby, the prompt configuration unit (110) can configure a generation prompt that includes instructions to avoid repetition with the previous generated content based on the previous generated content.

[0141] Meanwhile, FIG. 6 illustrates an embodiment in which regeneration is performed when candidate content does not satisfy the policy or has a repeating pattern, and storage is performed when it passes verification, but this is exemplary. Regeneration may be limited to a predetermined maximum number of times, and candidate content that does not pass verification within the maximum number of times may be discarded, and the discarded candidate content and the reason for the discarding may also be stored in the history storage unit (150).

[0142] FIG. 7 is a block diagram showing the hardware configuration of a content creation control device according to one embodiment of the present invention.

[0143] Referring to FIG. 7, the content creation control device (900) may include a processor (910) and a memory (920). The content creation control device (900) may correspond to an example in which the content creation control device (100) of FIG. 1 is implemented in hardware.

[0144] The processor (910) can control the operation of the content creation control device (900) by executing instructions stored in the memory (920). In one embodiment, the processor (910) can perform at least one of the operations of the prompt configuration unit (110), content acquisition unit (120), and verification unit (130) described with reference to FIG. 1, and the operations of the publishing gate unit (140) and history storage unit (150) by executing instructions stored in the memory (920). The processor (910) may be composed of one or more units and may be implemented as a central processing unit, a graphics processing unit, or other computational processing means.

[0145] The memory (920) can store instructions executed by the processor (910). Additionally, the memory (920) can store data necessary for the operation of the content creation control device (900). In one embodiment, the memory (920) can store at least one of target content, previous creation content, policy, and creation prompt. The memory (920) may include at least one of volatile memory and non-volatile memory.

[0146] The processor (910) and the memory (920) can be connected to transmit and receive data to and from each other. In one embodiment, the processor (910) and the memory (920) can be connected via a bus.

[0147] In one embodiment, the history storage unit (150) of FIG. 1 may be implemented by a memory (920) and may also be implemented by an external storage device linked with a content creation control device (900).

[0148] Meanwhile, FIG. 7 illustrates an embodiment in which a content creation control device (900) includes a processor (910) and a memory (920), but this is exemplary. The content creation control device (900) may further include at least one of an input / output interface, a communication module, and a storage device. Additionally, the operation of the content creation control device (900) may be implemented by a computer-readable recording medium storing instructions that cause the operation to be performed when executed by one or more processors.

[0149] As described above, according to one embodiment of the present invention, whether the generated content violates a preset policy and whether the generated content has a repeating pattern with respect to previous generated content can be verified together within a single generation control flow. Accordingly, it is possible to simultaneously suppress repetitive content that is not filtered by a method that checks only the policy compliance of the content itself, and policy-violating content that is not filtered by a method that checks only for repetition. Furthermore, by including an instruction in the generation prompt to avoid repetition with previous generated content, the generation of repetitive content can be suppressed in advance at the generation stage, while repeating patterns in the generated content can be detected afterward.

[0150] According to one embodiment of the present invention, when verification fails, the likelihood of the same type of verification failure occurring repeatedly can be reduced by configuring a regeneration prompt that includes the reason for the failure and regenerating candidate content. In particular, when configuring a regeneration prompt using a failure reason vector that combines the reason for failure based on policy verification and the reason for failure based on repeat pattern detection, heterogeneous reasons for failure can be integrated and reflected into a single regeneration instruction. Furthermore, by identifying the factor that increased the repetition score and including instructions to avoid it in the regeneration prompt, closed-loop control can be implemented that feeds the cause of the repetition back to subsequent generation.

[0151] According to one embodiment of the present invention, the style and form of the generated content can be diversified by configuring a generation prompt based on a generation mode selected from a plurality of content generation modes, and cyclic control can be implemented by feeding back the verification results to the mode selection by adjusting the selection weight for each generation mode based on the verification results of the previously generated content. In addition, the possibility of generating general content unrelated to the target content can be reduced by configuring a generation prompt that instructs a response to specific clues included in the target content, and the processing path of the candidate content can be differentiated according to the degree of repetition by classifying the candidate content as a target for regeneration or a target for human approval based on a plurality of thresholds for repetition scores.

[0152] Meanwhile, the method according to the present invention is written as a program executable on a computer and can be implemented on various recording media such as magnetic storage media, optical reading media, and digital storage media.

[0153] Implementations of the various technologies described herein may be implemented in digital electronic circuits, computer hardware, firmware, software, or a combination thereof. Such implementations may be implemented as computer programs stored on a computer-readable storage medium to control the operation of a data processing device. Such implementations may be implemented as computer program products, i.e., information carriers, such as computer programs stored on a machine-readable storage device or a computer-readable non-transient storage medium, for processing by the operation of a data processing device, e.g., a programmable processor, a computer, or a plurality of computers, or for controlling such operation. Computer programs such as the computer program(s) described above may be written in any form of programming language including compiled or interpreted languages, and may be deployed in any form including as a standalone program or as a module, component, subroutine, or other unit suitable for use in a computing environment. Computer programs may be deployed to be processed on one computer or a plurality of computers at a single site, or distributed across a plurality of sites and interconnected by a communication network.

[0154] Processors suitable for processing computer programs include, for example, both general-purpose and special-purpose microprocessors, and any one or more processors of any type of digital computer. Generally, the processor will receive instructions and data from read-only memory or random access memory, or both. The elements of the computer may include at least one processor that executes instructions and one or more memory devices that store instructions and data. Generally, the computer may include one or more mass storage devices that store data, for example, magnetic, magneto-optical disks, or optical disks, or may be combined to receive data from these, transmit data to these, or both. Information carriers suitable for embodying computer program instructions and data include, for example, semiconductor memory devices, magnetic media such as hard disks, floppy disks, and magnetic tapes; optical media such as CD-ROMs (Compact Disk Read Only Memory) and DVDs (Digital Video Disks); magneto-optical media such as floptical disks; ROMs (Read Only Memory); RAMs (Random Access Memory); flash memory; EPROMs (Erasable Programmable ROM); EEPROMs (Electrically Erasable Programmable ROM); etc. Processors and memory may be supplemented by or included in special-purpose logic circuit organizations.

[0155] Although this specification contains details of a number of specific embodiments, they should not be understood as limiting the scope of any invention or claimables, but rather as descriptions of features that may be characteristic of a specific embodiment of a specific invention. Specific features described in this specification in the context of individual embodiments may be implemented in combination in a single embodiment. Conversely, various features described in the context of a single embodiment may also be implemented in multiple embodiments individually or in any appropriate sub-combination. Furthermore, while features may operate in a specific combination and be described as initially claimed, one or more features from the claimed combination may be excluded from the combination in some cases, and the claimed combination may be changed to a sub-combination or a variation of the sub-combination.

[0156] Likewise, although operations are depicted in the drawings in a specific order, this should not be understood as requiring that such operations be performed in that specific or sequential order depicted to obtain a desirable result, or that all depicted operations must be performed. In certain cases, multitasking and parallel processing may be advantageous. Furthermore, the separation of the various device components of the above-described embodiments should not be understood as requiring such separation in all embodiments, and it should be understood that the described program components and devices can generally be integrated together into a single software product or packaged into multiple software products.

[0157] Meanwhile, the embodiments of the present invention disclosed in this specification and drawings are merely specific examples provided to aid understanding and are not intended to limit the scope of the present invention. It is obvious to those skilled in the art that other variations based on the technical concept of the present invention are possible in addition to the embodiments disclosed herein.

[0158] [Explanation of the symbol]

[0159] 100, 900: Content creation control device

[0160] 110: Prompt configuration

[0161] 120: Content Acquisition Section

[0162] 130: Verification Department

[0163] 140: Posting Gate Department

[0164] 150: History Storage

[0165] 200: Generative AI Model

[0166] 910: Processor

[0167] 920: Memory

Claims

Claim 1 A content generation control device that controls content generation using generative artificial intelligence, comprising: a prompt configuration unit that configures a generation prompt including instructions to avoid repetition with said previous generation content based on a target content and said previous generation content; a content acquisition unit that acquires candidate content from a generative artificial intelligence model into which the generation prompt is input; and a verification unit that verifies whether said candidate content satisfies a preset policy and whether said candidate content has a repeating pattern with respect to said previous generation content, wherein the prompt configuration unit configures a regeneration prompt including a reason for failure when the verification by the verification unit fails, and the content acquisition unit acquires regenerated candidate content from the generative artificial intelligence model into which the regeneration prompt is input. Claim 2 A content generation control device according to claim 1, wherein the verification unit calculates a repetition score by combining two or more of partial segment similarity, n-gram overlap, embedding similarity, frequent vocabulary inclusion, and sentence structure similarity between the candidate content and the previously generated content, and determines whether the repetition pattern is present based on the repetition score. Claim 3 In paragraph 2, the prompt configuration unit identifies a factor that increased the repetition score and includes instructions to avoid the identified factor in the regeneration prompt, a content generation control device. Claim 4 A content creation control device according to paragraph 2, wherein the verification unit classifies the candidate content as a regeneration target if the repetition score is greater than or equal to a first threshold, and classifies the candidate content as a human approval target if the repetition score is less than the first threshold and greater than or equal to a second threshold. Claim 5 A content creation control device according to claim 1, wherein the prompt configuration unit comprises a failure reason vector formed by combining a failure reason resulting from policy verification failure of the verification unit and a failure reason resulting from repeated pattern detection, and a regeneration prompt formed using the failure reason vector. Claim 6 A content generation control device according to claim 1, wherein the content acquisition unit acquires a skip signal from the generative artificial intelligence model and transmits the skip signal to the verification unit, and when the skip signal is acquired, the verification by the verification unit for the target content is not performed. Claim 7 In claim 1, the prompt configuration unit is a content creation control device that configures the creation prompt based on one creation mode selected from a plurality of content creation modes. Claim 8 A content creation control device according to claim 7, wherein the creation mode is selected in a performance-based manner, and the performance-based manner adjusts the selection weight for each creation mode based on the verification result of the previously created content. Claim 9 In claim 1, the prompt configuration unit comprises a content generation control device that configures a generation prompt instructing to respond to one or more specific clues among numbers, names, claims, events, and terms included in the target content. Claim 10 A content creation control device according to claim 1, wherein the previously created content is stored in association with one or more of the content body, the creation mode used for creation, the reason for failure, the number of regenerations, whether it is published, and the target content identifier. Claim 11 A content creation control method comprising a content creation control device including a processor and memory, wherein the content creation control device controls content creation using a generative artificial intelligence, the method comprising: a step of configuring a creation prompt including instructions to avoid repetition with the previously created content based on the target content and the previously created content; a step of the content creation control device obtaining candidate content from a generative artificial intelligence model into which the creation prompt is input; a step of the content creation control device verifying whether the candidate content satisfies a preset policy and whether the candidate content has a repeating pattern with respect to the previously created content; a step of the content creation control device configuring a regeneration prompt including a reason for failure if the verification fails; and a step of the content creation control device obtaining regenerated candidate content from the generative artificial intelligence model into which the regeneration prompt is input. Claim 12 A content generation control method according to claim 11, wherein the verification step comprises calculating a repetition score by combining two or more of partial interval similarity, n-gram overlap, embedding similarity, frequent vocabulary inclusion, and sentence structure similarity between the candidate content and the previously generated content, and determining whether the repetition pattern is present based on the repetition score. Claim 13 A content generation control method according to claim 12, wherein the step of configuring the regeneration prompt comprises identifying a factor that increased the repetition score and including instructions to avoid the identified factor in the regeneration prompt. Claim 14 A content creation control method according to claim 12, wherein the verification step comprises classifying the candidate content as a regeneration target if the repetition score is greater than or equal to a first threshold, and classifying the candidate content as a human approval target if the repetition score is less than the first threshold and greater than or equal to a second threshold. Claim 15 A content creation control method according to claim 11, wherein the step of configuring the regeneration prompt comprises configuring a failure reason vector by combining a failure reason due to a failure in verification of the policy and a failure reason due to the detection of a repeating pattern, and configuring the regeneration prompt using the failure reason vector. Claim 16 In claim 11, the content generation control method further comprises: a step in which the content generation control device obtains a skip signal from the generative artificial intelligence model; and a step of terminating the content generation processing for the target content when the skip signal is obtained. Claim 17 A content creation control method according to claim 11, wherein the step of configuring the creation prompt comprises the step of configuring the creation prompt based on one creation mode selected from a plurality of content creation modes. Claim 18 A content generation control method according to claim 17, wherein the generation mode is selected in a performance-based manner, and the performance-based method adjusts the selection weight for each generation mode based on the verification result of the previously generated content. Claim 19 A content generation control method according to claim 11, wherein the step of configuring the generation prompt comprises configuring the generation prompt to instruct to respond to one or more specific clues among numbers, names, claims, events, and terms included in the target content. Claim 20 A content creation control method according to claim 11, wherein the previously created content is stored in association with one or more of the content body, the creation mode used for creation, the reason for failure, the number of regenerations, whether it is published, and the target content identifier.

Citation Information

Patent Citations

  • Generative AI-based automated question and answer system and how to implement it

    KR1020250113091A

  • Question answering system using generative model and method thereof

    KR1020250128787A

  • System and Method for Correcting Artificial Intelligence Inference Errors Using Accumulated Rotational-Avoidance Phase Control

    KR1020260093231A

  • Detecting evasive prompts for generative artificial intelligence systems

    US20250384132A1