REAL-TIME NEURO-SYMBOLIC VERIFICATION SYSTEM

Abstract

This invention relates to a system (1) developed to provide a neuro-symbolic validation architecture to prevent factual hallucinations that occur in large language model-based production processes during production and token selection phase instead of correcting them after the response is generated.

Description

1 TARIFF REAL-TIME NEURO-SYMBOLIC VERIFICATION SYSTEM Technical Area This discovery reveals factual changes in production processes based on large language models. hallucinations are corrected during production and with the token instead of after the response is generated. providing a neuro-symbolic validation architecture aimed at preventing interference during the selection process. It is related to a system. Previous Technique One of the common problems in validation systems is the errors or inaccuracies produced by the model. mechanisms that attempt to correct fabricated information are not effective enough. The absence of post-hoc hallucination correction is the problem. This situation is incorrect. 15 Allowing the content to be produced in the initial stage and then having subsequent checks fully detect the error. This leads to its inability to resolve the issue. In addition, sentence-end validation delay, because the verification process only kicks in after the output is complete This leads to the inability to prevent erroneous information during production. Another problem. The uniform intervention intensity varies depending on the different risk levels of the verification mechanism. It reacts to errors with the same intensity and thus provides context-sensitive corrections. The inability to do so. Finally, the problem of over-intervention in the case of insufficient QC (Information Graphics), When there is insufficient data in the information graph of the verification system, this is called uncertainty. instead of evaluating it as such, it unnecessarily discards content that might be correct. This leads to suppression or rejection. When these problems are considered together, 25 Verification systems are important in terms of both accuracy and flexibility of use. It has been observed that performance losses can occur. 2 Therefore, the phenomena that emerge in large language model-based production processes hallucinations are corrected during production and with the token instead of after the response is generated. providing a neuro-symbolic validation architecture aimed at preventing interference during the selection process. It is clear that a system is needed. United States patent number US2025259075, which falls under the prior art. The document discusses optimizing artificial intelligence systems, including large language models, and An advanced model management platform is mentioned to ensure security. This invention is relevant to generative artificial intelligence, such as large language models and diffusion models. an advanced 10 developed to optimize and secure systems It describes the model management platform. The platform uses existing generative artificial intelligence. Their systems produce hallucinations, lack verification, have security vulnerabilities, and are inadequate. It combines various techniques to overcome its limitations, such as model management. The system introduces reinforcement learning algorithms for model optimization. Yield-assisted production (RAG) for hallucination reduction, field 15 versus expert knowledge. Specifically, verification, model distillation and similarity scoring for security purposes, neurosymbolic with adventarial training and improved management to increase resilience AI models attention mechanisms for routine combinations. It uses blending. The integrated use of these techniques. Thanks to this, the platform automatically combines the strongest aspects of symbolic and relational techniques. by combining planning and modeling with simulation, in a wide variety of tasks and fields. The producer significantly improves the performance, reliability, and safety of the artificial intelligence. It increases. Brief Description of the Invention 25 The aim of this invention is to investigate the factual issues that arise in large language model-based production processes. hallucinations are corrected during production and with the token instead of after the response is generated. 3 To provide a neuro-symbolic validation architecture for prevention during the selection process. The goal is to implement a system developed for this purpose. Detailed Description of the Invention The “Real-Time Neuro-Symbolic” technology was developed to achieve the purpose of this invention. The "Verification System" is shown in the attached figure; Figure 1 shows a schematic view of the system that is the subject of the invention. The parts shown in the figure are individually numbered, and the corresponding numbers correspond to these numbers. It is given below. 1. System 2. Application 15 3. Server Phenomenal hallucinations that arise in production processes based on the grand language model Instead of correcting the response after it has been generated, correct it during the production and token selection phase. 20 The developed invention subject system (1); -to enable the user to enter input and select field profiles. at least one application configured (2), -semantic sequences of partial tokens generated by the large language model at fixed intervals the transformation of these propositions into verified information, business rules, and 25 Comparison with a temporal accuracy graph consisting of domain ontologies, identification The level of trust in the identified contradictions will have a real impact on the next token selection. 4 at least one configured to enable timed logit intervention. Includes server (3). The application (2) in the system (1) which is the subject of the invention does not require any communication protocol. 5 to communicate with the server (3) and exchange data using It is being structured. The server (3) in the system (1) which is the subject of the invention, uses any communication protocol to communicate with application (2) and exchange data using is configured. The server (3) is configured with the field profile of the request entered by the user and 10 defining the validation policy, the field type upon user request, validation frequency, time context, trust policy, information graph sub-category to be used the cluster's evaluation of soft intervention and hard intervention thresholds, output the domain validation profile to be obtained and which ontology to use with that profile whether it will be used, whether temporal validity control is mandatory, uncertainty 15 buffer window size, logit interference coefficients and escape route The mechanism is structured to enable the determination of the confidence threshold. Server (3) triggers the model token generation in partial proposition inference. monitoring, not waiting for the completion of the relevant sentence, the flow of generated tokens the model should be evaluated in terms of sliding windows of a specific size, sentence 20 determining whether it is moving towards a possible factual claim before it ends, partial subject, partial predicate, beginning of existence and relationship, and progressing towards subject, action, and object. Searching for early patterns in structural types, in real time. The verification is initiated within the token stream and a partial proposition is triggered as output. The event is 25, which enables the generation of the window ID and location marker. The server (3) is configured semantically for the triggered window. the decomposition and conversion of a token into a partial or complete propositional structure bringing the sequence into a machine-verifiable logical form, entity the removal of candidates, normalization, identification of relationship candidates, If necessary, apply an ontological equivalence analysis and provide a triple as output. the production of representation, triple representation, full triple, partial triple, incomplete object relationship and the possibility of an ambiguous relationship being in the form of a candidate proposition and the semantic representation created 5 to ensure that it is brought into a format that can be questioned in the accuracy graph is structured. The server (3) provides verified factual information for the proposition derived. ontological relationship rules, institutional business logic, field-specific constraints, and on a triple basis. temporal accuracy, which includes parameters in the form of temporal validity information. The graph should be worked on, and the question is whether the triple graph in question is included in the graph. whether the relationship is ontologically valid, whether the object type is appropriate, 10 whether the relationship was correct during the relevant time period, or whether it was contrary to area policy. whether it involves matching and whether it contradicts corporate rules Parameter verification, information assets with a validity range of triples. it should be considered as such, and the truth mapping, the ontological outcome, and It is structured to enable the production of temporal validity results. 15 Server (3) uses the results returned from the graph and ontology query to identify a conflict. the calculation of the score, the direct contradiction of the score, the severity of the ontological violation, temporal invalidity level, asset matching confidence, and the model's current token confidence signal. deriving it from components in this form, using a single contradiction in a single window Avoid aggressive intervention, certain final windows have conflict scores of 20 keeping uncertainty within a buffer, the buffer mechanism preventing false positives. reduction, filtering out early false alarms that may arise from missing triples, QA inquiry the functions of softening the effect of the delay and making the abdomen more stable. fulfilling the sustainability control of the contradiction in successive windows and the conflict confidence threshold value should be above 25 for a specified consecutive window. to ensure that actual intervention is initiated in such cases The server (3) is configured with the intervention trigger signal and conflict confidence level. According to the model, the next token selection will be influenced by soft printing intervention. 6 and its implementation in the form of a hard barrier, the possibility of steel in soft pressure it is possible but not certain, that the logit value of the relevant token candidates will be a certain lowering the coefficient and the model's natural flow to the correct probability its orientation and the contradiction in the rigid obstacle is highly secure and ontological or temporal. In cases where a breach is confirmed, the relevant token or group of tokens will be directly blocked. This involves preventing the model from generating the incorrect statement, and then moving on to the next most suitable option. switching to an alternative, not only by not running the ban but also Supporting an alternative token path prevents the model from becoming pointless. passing through and being directed in the right direction and the hallucination token selection It is configured to prevent its occurrence during this phase. Server 10 (3), managing situations where the infographic is incomplete or has low reliability, if there is no direct match in the graph, if triple confidence is low, if ontological support is insufficient, and If the temporal context is unclear, the relevant matter is automatically subject to punitive intervention without any prior warning. the proposal is added to a candidate update queue, and the queue is external to the RAG-based system. Source verification, reliable corporate databases, human endorsement, and domain expert 15 to ensure that it is processed through mechanisms such as verification is being configured. Server (3), the model's corrected or released token producing the final text through the flow, not just leaving the text but during the production process compilation of verification events, final response, confidence score, which the windows were interfered with, which ontological conflicts were identified, which 20 tokens that are suppressed or blocked and which propositions are evaded the production of outputs in this form, generating responses to the user, and simultaneously providing speech. The issue is to provide a confidence trail on how the answer passed through the accuracy filter. It is configured to provide the server (3), which is generated during the model execution. learning from all events, triples in the candidate update queue, successful and 25 Examples of failed interventions, false positive or false negative statistics, which the performance of the uncertainty buffer and the areas where the hard barrier is used too much or too little, and Data in the form of resource usage and delay records from the relevant sources. 7 the collection and updating of the graph using that data, field adjusting thresholds, optimizing soft pressure coefficients, hard barrier updating its boundaries and reweighting ontology priorities It is configured to provide the server (3), the model from past interventions. Neuro-symbolic control management that learns and manages future production more accurately 5 It is structured to enable it to be accomplished. Industrial Applicability Thanks to the system (1) that is the subject of the invention, 10 can be produced in large language model based production processes. instead of correcting the resulting factual hallucinations after a response is generated, the production of these hallucinations... neuro-symbolic validation aimed at preventing problems during and after token selection. Its architecture is provided. Around these fundamental concepts, the subject of the invention is “Real-Time Neuro-Symbolic 15 It is possible to develop a wide variety of applications related to the Verification System (1), The invention cannot be limited to the examples described here, but is primarily defined in the claims. It is like that.

Claims

8 REQUESTS 1. Factual phenomena arising in production processes based on large language models. hallucinations are corrected during production rather than after the response is generated, and Neuro-symbolic validation to prevent token selection during the selection phase 5 providing the architecture; -to enable the user to enter input and select field profiles. containing at least one configured application (2) and -partial token sequences generated at fixed intervals by the large language model the transformation of semantic propositions into verified 10 propositions with a temporal accuracy graph consisting of information, business rule, and field ontologies comparison, based on the confidence level of the identified discrepancies, to the next stage. Real-time logit intervention applied to token selection a server characterized by having at least one server (3) configured to provide system (1). 15 2. To communicate with the server (3) using any communication protocol and The application is characterized by (2) which is structured to perform data exchange. A system like the one in Request 1 (1).

3. Communicate with the application (2) using any communication protocol. and characterized by the server (3) configured to carry out data exchange. A system like the one in Request 1 or 2 (1).

4. The domain profile and validation policy of the request entered by the user. 25 determining the field type, validation frequency, and time as requested by the user. the context, the trust policy, the subset of infographics to be used, Evaluation of soft intervention and hard intervention thresholds, output 9 obtaining the domain validation profile and which profile is used with that profile. whether the ontology will be used and whether temporal validity checks are mandatory that it is not, the uncertainty buffer window size, logit intervention determining the coefficients and the confidence threshold of the escape route mechanism The above 5 is characterized by the server (3) configured to provide a system like any of the requests (1).

5. In partial proposition inference triggering, the model monitors token generation. not waiting for the completion of the relevant sentence, the flow of generated tokens the model should be evaluated in terms of sliding windows of a specific size, sentence 10 determining whether it is moving towards a possible factual claim before it ends, partial subject, partial predicate, origin of existence and relationship, and subject, action, and object. the search for early patterns in types with a correctly progressing structural form, real timely verification is initiated within the token stream and output The partial proposition trigger event, window ID, and position marker are 15. characterized by the server (3) configured to enable its production a system like any of the above requests (1).

6. Semantic parsing of the triggered window and its partial or complete analysis. the conversion of the token sequence into a propositional structure by the machine 20 bringing the candidate entities into a verifiable logical form removal, normalization, identification of relationship candidates, If necessary, apply an ontological equivalence analysis and provide an output. generating triple representations, triple representations with full triple, partial triple, and incomplete object representations. It can be in the form of a nominee proposition with a relationship and an uncertain relationship, and the 25 created bringing the semantic representation into a format that can be questioned on the truth graph the above characterized by the server (3) configured to provide a system like any of the requests (1).

7. The derived proposition is supported by verified factual information and ontological relationship rules. corporate business logic, field-specific constraints, and temporal aspects based on triples. a temporal accuracy graph that includes parameters in the form of validity information It needs to be studied, is the question being investigated regarding the triple graph in question? Whether the relationship is ontologically valid, whether the object type is appropriate, 5 whether it is true or not, whether the relationship was true during the relevant time period, area whether it involves a match that violates the policy and corporate rules Checking the parameters such as whether they contradict each other, triples They should be treated as information assets with a validity period and as output. truth mapping, ontology result and temporal validity result 10 characterized by the server (3) configured to enable its production a system like any of the above requests (1).

8. A conflict score is calculated using the results returned from the graph and ontology query. calculation, the direct contradiction in the score, the severity of the ontological violation, 15 temporal invalidity level, entity matching confidence, and the model's current state deriving the token from components in the form of a trust signal, all in a single window. avoiding aggressive intervention by using singular contradictions, specific outcomes keeping the conflict scores of the windows within an uncertainty buffer, buffer mechanism to reduce false positives, missing triple 20 Filtering out potential early false alarms, reducing QA query delay. It performs the functions of softening the impact and making the abdomen more stable. bringing about, sustainability control of the contradiction in successive windows and the conflict confidence threshold value to be determined over a specified consecutive window. 25 from the above requests characterized by the server (3) configured for a system like any other (1). 11 9. The next step in the model is based on the intervention trigger signal and the conflict confidence level. Influencing token selection, intervention through soft pressure and hard barrage the possibility of steel being used in soft printing, to be implemented in this way However, this is not certain, as the logit value of the relevant token candidates is a certain... the coefficient should be reduced and the model should reach the correct probability of 5 within the natural flow. its guidance and the contradiction in the rigid obstacle, high confidence and ontological or In cases where a temporal violation is certain, the relevant token or group of tokens directly blocking the model's path to generating the false statement. closure, switching to the next most suitable alternative, only the ban not being activated also means the alternative token path is 10 to ensure that the model is supported, and to prevent it from becoming pointless, and being directed in the right direction and the hallucination during the token selection phase with the server (3) configured to prevent its occurrence a system like any of the above characterized demands (1).

10. Managing situations where the infographic is incomplete or has low reliability. If there is no direct match in the graph, and triple confidence is low, then ontological support... If insufficient and the temporal context is unclear, it is automatically punitive. The relevant proposal is added to a candidate update queue without any intervention. retrieval and queue RAG-based external validation, reliable 20 corporate databases, human endorsement, and domain expert verification. server (3) configured to enable processing with mechanisms a system like any of the above characterized demands (1).

11. The final text of the model via the revised or released token stream 25 producing it, not just leaving the text, but also the verification that takes place during the production process. compilation of events, final answer, confidence score, which windows interventions were made, which ontological conflicts were identified, which tokens are suppressed or blocked, and which propositions are available as escape routes. 12 generating outputs in the form of received data, generating responses to the user, and also an indication of how that answer passed through the truth filter. Characterized by the server (3) configured to provide a trust trail. a system like any of the above-mentioned requests (1).

12. The candidate ensures that the model learns from all events that occur during execution. Triples in the update queue, examples of successful and unsuccessful interventions, False positive or false negative statistics, in which areas are there more severe barriers? or is used sparingly, the performance of the uncertainty buffer and resource utilization and the collection of data in the form of delay records from relevant sources and the word 10 updating the graph using the relevant data, setting field thresholds adjustment, optimization of soft pressure coefficients, hard barrier updating its boundaries and redefining ontology priorities characterized by the server (3) configured to enable weighting. a system like any of the above requests (1). 15 13. The model learns from past interventions and makes future production more accurate. to enable it to perform a neuro-symbolic control management that governs any of the above requests characterized by the configured server (3) a system like one of them (1). 20

Images