A method for realizing original information tracing based on data coding to suppress LLM illusion
By introducing data source identifiers and identity codes into large language models through data lineage technology, the problem of untraceable training data is solved, the verifiability of generated content is realized, and the credibility of the model in knowledge-intensive fields is improved.
Patent Information
- Authority / Receiving Office
- CN · China
- Patent Type
- Applications(China)
- Current Assignee / Owner
- JIANGSU UNIV OF SCI & TECH
- Filing Date
- 2026-02-27
- Publication Date
- 2026-06-02
AI Technical Summary
Existing large language models lack a source verification mechanism when generating responses, making it difficult to guarantee the authenticity of the output information. In particular, the lack of traceability of training data in knowledge-intensive tasks makes it even more difficult to detect and intervene in hallucination phenomena.
By adopting the data lineage technology approach, the traceability of training data is achieved by designing a data source identification system and introducing identity coding. This is integrated into the model training process and supports the retrospective verification of the training data corresponding to the generated content during the inference stage. The decoder structure is extended to introduce sample label coding, and the traceability and verifiability of the content are achieved by combining the data traceability triggering mechanism.
Without affecting the model's inference efficiency, we can achieve an accurate correspondence and traceability between the generated content and the original training information, thereby enhancing the credibility of the model in knowledge-sensitive fields.
Smart Images
Figure CN122132837A_ABST
Abstract
Description
Technical Field
[0001] This application relates to the field of Large Language Model (LLM) training technology, specifically to LLM illusion suppression technology in the knowledge-intensive field, and particularly to a method for suppressing LLM illusion by tracing back original information based on data encoding. Background Technology
[0002] In recent years, Large Language Models (LLMs) have demonstrated outstanding performance in numerous application scenarios due to their powerful language understanding and generation capabilities. They have been widely applied in fields such as intelligent question answering, content generation, assisted programming, and academic research, accelerating the integration of artificial intelligence with practical business and greatly improving work efficiency and innovation experience.
[0003] However, with the widespread application of LLM across various industries, especially in knowledge-intensive fields such as healthcare, finance, law, and scientific research, problems such as "deviation from facts, lack of evidence, and logical inconsistencies" in the generated content have become increasingly prominent, leading to the widely discussed phenomenon of "LLM illusion." This illusion can mislead users, reduce information credibility, and potentially cause serious consequences, such as security risks and legal liabilities, becoming a significant obstacle to the high-credibility deployment of LLM. Current methods for suppressing illusions mainly focus on cue word design, model architecture improvement, or external knowledge injection, but there are relatively few technical solutions for training data processing, and they generally lack effective mechanisms for deep collaboration with the model.
[0004] Data lineage is a technology used to track the various changes and flows of data from its source to its final use, ensuring data traceability, interpretability, and compliance. Therefore, data lineage technology is a possible approach to addressing the LLM illusion and can be significantly correlated with each stage of LLM construction, thereby helping to improve the effectiveness of LLM illusion suppression. However, currently, there is no LLM illusion suppression method that integrates similar technologies. Summary of the Invention
[0005] This invention aims to address the problem that existing large language models lack verifiable source mechanisms when generating responses, making it difficult to guarantee the authenticity of output information. Especially in knowledge-intensive tasks, the lack of traceability of training data makes LLM (Low Language Modeling) illusions even more difficult to detect and intervene in, necessitating the construction of a traceable data-level oversight mechanism. Therefore, this invention proposes a method based on data encoding to achieve traceability of original information and suppress LLM illusions, drawing inspiration from data lineage technology. This method can be integrated into the model training process and supports traceable verification of the generated content corresponding to the training data during the inference phase, providing a complete technical path to improve the verifiability and credibility of model output content.
[0006] To achieve the above objectives, the present invention adopts the following technical solution:
[0007] The present invention provides a method for suppressing LLM hallucinations by tracing original information based on data encoding, comprising the following steps:
[0008] Step 1: Conduct multi-source raw information collection;
[0009] Step 2: Introduce the original information identity encoding operation; design a data source identification system to generate a unique identity code for each piece of original information, and design a placeholder retention mechanism to ensure the retention of identity codes during the semantic fusion of training texts. Map the original information identity codes that constitute the same training sample to sample label codes to realize the correspondence between the training text and the original information that constitutes it and the identity identification code.
[0010] Step 3: Use semantic fusion technology to process the multi-source raw information into LLM training text format;
[0011] Step 4: Feed the LLM training text into the constructed LLM training model to carry out LLM training;
[0012] Step 5: During LLM training, extend the existing LLM decoder structure by introducing sample label encoding as supervision labels to guide the model to learn the one-to-one correspondence between sample semantic information and its label encoding, so that the trained model can output its corresponding label encoding while generating text content.
[0013] Step Six: When the trained LLM is used, output the generated LLM content;
[0014] Step 7: Data traceability. The model's working architecture is divided into default mode and verification mode. In verification mode, after LLM generates content, a data traceability trigger mechanism is designed to decode the implicitly carried sample label encoding in the generated text when traceability is triggered, thereby restoring the original information corresponding to the training data and realizing the traceability and verifiability of the content.
[0015] As a further preferred option, in step two, the data source identification system needs to have a hierarchical and structured design, using appropriate granularity to avoid coarse-grained ambiguity and fine-grained redundancy. The data source identification system can be composed of triplet identifiers to realize the source and type positioning of the original information.
[0016] As a further preferred option, the mathematical expression of the triplet identifier system is: ,in , , These represent "source document identifier", "section identifier", and "data type identifier", respectively. A hierarchical encoding structure of "[Document Type]-[Version Number]-[Hash Fingerprint]" is adopted. Based on structured markup specification coding, An extended classification code is defined, and these three elements work together to accurately identify the source document and other information. At the same time, each part has an appropriate code length, improving the efficiency of original information tracing while ensuring the code's collision resistance.
[0017] As a further preferred option, in step two, the placeholder retention mechanism refers to setting up a corresponding identity encoding information channel that is independent of the semantic information while retaining the semantic information of the original information, so as to avoid the identity encoding of the original information affecting the semantic information during the semantic fusion process.
[0018] Semantic information and identity encoding information are in one-to-one correspondence, and the placeholder length of the placeholder retention mechanism is determined by the original information identity encoding algorithm and the collision resistance requirements.
[0019] As a further preferred option, in step two, the encoding algorithm for the original information identity encoding is the Base64 encoding algorithm. The selected encoding algorithm needs to have strict encoding and decoding rules, and needs to ensure the collision resistance and reversibility of the original information identity encoding process.
[0020] The encoding algorithm for sample label encoding is as follows: The encryption algorithm must have strict encoding and decoding rules, and ensure the collision resistance and reversibility of the sample label encoding process.
[0021] As a further preferred option, in step five, extending the existing LLM decoder structure means extracting the state of the training process and converting it into "identity features"; at the same time, designing a corresponding auxiliary loss function to align the identity features with the sample label encoding of the training data without affecting the normal training of the LLM, thereby realizing supervised modeling of identity encoding and ensuring that the model has the ability to synchronously output sample label encoding.
[0022] As a further preferred option, in step seven, the default mode in the model's working architecture refers to the model interacting normally with the user, while the verification mode refers to the model tracing the original information of the LLM-generated content and verifying the accuracy of the answer.
[0023] As a further preferred option, in step seven, the data tracing triggering mechanism includes active model triggering and passive model triggering. Active triggering depends on the confidence score of the LLM-generated content and the knowledge conflict detection results, while passive triggering depends on the user's needs.
[0024] The confidence score of LLM-generated content refers to the LLM's self-assessment of the correctness or reliability of its generated content, which can help users or the model itself assess the credibility of the model's output; the knowledge conflict detection result of LLM-generated content refers to the LLM's knowledge consistency check of its generated content, which can avoid knowledge conflicts and thus ensure the consistency of the model's output.
[0025] User needs refer to users' questions about the content generated by LLM and their requests for content verification.
[0026] As a further preferred option, step seven, the process of encoding and decoding the sample labels includes three steps: decoding mapping, reverse indexing, and restoration of the original information.
[0027] As a further preferred option, decoding mapping refers to mapping the identity feature vectors during LLM training to sample label encodings corresponding to the training statements;
[0028] Inverse indexing refers to decoding all the original information triplet identifiers corresponding to the training sample based on the decoding algorithm corresponding to the sample label encoding and the original information identity encoding;
[0029] Original information restoration refers to locating the original information that constitutes the training text based on the original information triplet identifier information, and the original information is fed back to the verification requester in the storage format used (such as JSON format storage).
[0030] The beneficial effects of this technical solution are as follows: Based on the identity encoding reverse decoding mechanism, the generated content and the original training information are accurately corresponded and traceable without affecting the reasoning efficiency of the large language model, thereby enhancing the credibility of the model in knowledge-sensitive fields. Attached Figure Description
[0031] Figure 1 This is a schematic diagram illustrating how the original information tracing mechanism proposed in this invention supplements the conventional LLM training process.
[0032] Figure 2 Embed a flowchart for the original information identity encoding;
[0033] Figure 3 An integrated diagram illustrating the identity encoding-assisted training mechanism;
[0034] Figure 4 This is a flowchart illustrating the data tracing trigger mechanism and the process of restoring original information based on identity coding. Detailed Implementation
[0035] The technical solutions of the present invention will be clearly and completely described below with reference to the accompanying drawings in the embodiments of the present invention. Obviously, the described embodiments are only some embodiments of the present invention, and not all embodiments.
[0036] like Figure 1 As shown, this application provides an LLM illusion suppression method based on an original information tracing mechanism, particularly an LLM illusion suppression method in knowledge-intensive fields. This method supplements the existing LLM training process.
[0037] The typical LLM training process can be divided into the following steps: First, collect multi-source raw information 1; second, use semantic fusion technology to process the multi-source raw information 1 into LLM training text 2; then, feed the LLM training text 2 into the constructed LLM training model to carry out LLM training 3; finally, when the trained LLM is used, output the LLM generated content 4.
[0038] The original information tracing mechanism proposed in this application consists of three parts and can be used in conjunction with existing LLM training processes to suppress LLM illusions: ① In the stage of generating LLM training text 2 from multi-source original information 1, the operation of original information identification encoding S1 is introduced; ② In the training stage from LLM training 3 to LLM generated content 4, the operation of label embedding regression S2 is introduced; ③ After LLM training is completed, the data tracing trigger mechanism S3 is set for LLM generated content 4. S1~S3 together constitute the original information tracing of LLM training data, which can significantly improve the quality of training data in knowledge-intensive fields.
[0039] The original information identity encoding S1 operation process proposed in this application is as follows: Figure 2 As shown.
[0040] The original information identity encoding provides a data source identification system design method based on data encoding to realize the original information embedding encoding in the LLM training data construction stage.
[0041] First, a Hierarchical Triple Identifier (HTI) system is designed to encode the original information, i.e., to construct a data tracing framework based on the Hierarchical Triple Identifier (HTI), the mathematical expression of which is:
[0042]
[0043] In the formula, (S101) is the "source document identifier"; (S102) is the "chapter identifier"; (S103) is a "data type identifier".
[0044] (S101): "Source Document Identifier" uses a hierarchical encoding structure to encode the source document. To ensure efficient tracing of original information during actual operation, an appropriate encoding length must be set, while also considering collision resistance. For example, a hierarchical encoding structure like [Document Name]-[Document Type]-[Version Number]-[Hash Fingerprint] can be designed. Here, "Document Type" follows the ISO 25964-1 standard classification, "Version Number" uses semantic versioning, and the hash fingerprint is encoded using information such as the source document name and author. The hash fingerprint is the first 4 bytes of SHA-256, used for uniqueness verification. For example, "xxx-JiShu-1.2-3a5f" represents "xxx-Technical Manual-v1.2.pdf". To ensure efficient tracing of original information during actual operation, an appropriate hash fingerprint length must be set to balance collision resistance and subsequent decoding efficiency (the length of the hash fingerprint is affected by collision resistance requirements).
[0045] (S102): “Chapter Identifier”, based on the RFC 8949 Structured Markup Specification, such as “SEC-10.1.3” which represents “Chapter 10, Section 1, Subsection 3”. It is extensible when dealing with unstructured data (such as free text reports). To further achieve sentence segmentation and positioning of specific content;
[0046] (S103): “Data type identifier”, defines extended classification codes: 0: Image (can be further subdivided, such as 0.1=X-ray image, 0.2=Magnetic resonance imaging image, etc.); 1: Text (can be further subdivided, such as 1.1=Natural paragraph, 1.2=Table description); 2: Table (can be further subdivided, such as 2.1=Structured CSV, 2.2=Unstructured HTML table).
[0047] This triplet data source identification system possesses an appropriate granularity to adapt to the data structure characteristics of knowledge-intensive domains: it can both... (S101) (S102) (S103) and other multi-level coding methods achieve precise location of the source, chapter position, and content type of training samples (avoiding the ambiguity of tracing the source caused by coarse-grained coding), and avoid the storage and retrieval resource consumption caused by excessive subdivision through hierarchical coding design (avoiding the redundant costs caused by fine-grained coding). For example, in the medical field, through (S101) (S102) (S103) can determine the source of the work and the position of the chapter, as well as the content type (text description or image content) of a certain training sample of LLM.
[0048] Secondly, based on the defined HTI data tracing framework, the original information identity code value is generated: An encoding strategy with strict encoding and decoding rules is employed to transform the HTI into a globally unique database location code value. Specifically, a collision-resistant and reversible encoding algorithm is used to transform the complex HTI triplet encoding of the original information identity code into a globally unique database location code value—the original information identity code S104. For example, Base64 encoding is a collision-resistant and reversible encoding algorithm; in this case, Base64 encoding can be used to transform the complex HTI triplet encoding into the original information identity code. (S104):
[0049]
[0050] In the medical field, for example, a particular viewpoint or treatment method mentioned in a book can be coded into a corresponding... .
[0051] Due to the reversibility of the encoding algorithm, through this The HTI triplet information of the original information can be decoded:
[0052]
[0053] To restore the "original information" of LLM training data;
[0054] In the medical field, for example, based on a certain viewpoint / treatment method This allows us to trace the specific source of the viewpoint / method.
[0055] Then, a placeholder retention mechanism S105 is designed: Generally, multiple pieces of original information from multiple sources 1 need to undergo semantic fusion processing S106 and other operations to form LLM training text 2. Therefore, in the semantic fusion stage, a dual-channel embedding architecture is adopted, where the semantic channel is the semantic information vector of the data, and the position channel is the position vector. The position vector has the protection of position information isolation and reversibility, avoiding the identity encoding of the original information. The occurrence of (S104) affects semantic deviations and other issues that may occur in the semantic fusion processing S106 operation.
[0056] The raw information from different sources can only form the training dataset for LLM after semantic fusion and other operations. Therefore, in the cross-modal semantic fusion stage, a dual-channel embedding architecture is adopted, in which the semantic channel is the semantic information vector of the data and the positional channel is the positional vector. The positional vector has the guarantee of positional information isolation and reversibility.
[0057] Finally, after semantic fusion processing (S106) and other operations, the multiple pieces of original information will form an LLM training text 2. In terms of content composition, this training text will contain the original information text content (S109) and its corresponding identity code. (S107) After semantic fusion processing S106, the text content will be transformed into training statement S110, which corresponds to multiple identity codes. (S107) Transform the code into the sample label code corresponding to the training statement S110 using a collision-resistant encoding algorithm. (S108).
[0058] For example, a block of natural language text after semantic fusion Depend on The original information consists of a series of elements, and its corresponding elements are as follows: indivual gather Using coding methods that are collision-resistant and reversible (such as...) Encryption algorithm generation Globally unique sample label encoding ;
[0059] A natural language text block of a training statement S110 after semantic fusion processing S106. Depend on The original information consists of a series of elements, and its corresponding elements are as follows: A related gather ,and It is a collision-resistant and reversible encoding algorithm, so at this time, natural language text blocks Sample label encoding (S108) is:
[0060]
[0061] In the formula, Arranging by lexicographical order ensures deterministic generation.
[0062] Due to the reversibility of the encoding algorithm, through this The corresponding original information location can be decoded.
[0063] gather:
[0064]
[0065] In the medical field, for example, the different emphases on a particular condition in multiple books (corresponding to text content S109) can be unified into a single expression (corresponding to training statement S110), and the positions of the emphasized content in each book can be combined (corresponding to the combination of the positions of each explanatory method itself). The unified encoding is the sample label encoding corresponding to this expression method. (S108).
[0066] The original information identity coding constructs a coding system for multi-source original data, which can realize the accurate mapping between LLM training data and its corresponding original information, and provides a verifiable coding traceability framework for ensuring the quality of training data.
[0067] The label embedding regression operation S2 proposed in this application is as follows: Figure 3 As shown.
[0068] Tag embedding regression provides an LLM fine-tuning optimization method based on the original information tracing mechanism to achieve tag embedding regression training during the model training phase. While ensuring the quality of LLM-generated content, it enhances the modeling ability of the correspondence between training data and identity tags, providing support for subsequent data tracing.
[0069] First, the existing LLM decoder structure is extended. Based on the final layer of the Transformer decoder S21, the original text output path is retained, while a new latent variable output branch is added. This branch, through a linear mapping module, maps the hidden states to "identity feature variables" S22, representing the identity feature information corresponding to each training text during training, and is encoded with the sample label of the training text. (S108) One-to-one correspondence.
[0070] Then, for the newly added branch, an auxiliary supervision mechanism of "label embedding regression" is introduced. On the basis of ensuring the normal generation of LLM text content, an auxiliary loss function is constructed to measure the relationship between the identity feature variable S22 and the sample label encoding. The degree of correspondence (S108). During training, the results of the latent variable output branch only participate in the auxiliary loss calculation and do not directly pass through the standard decoder for text output, thus avoiding impact on the conventional LLM training process 5.
[0071] Finally, during the model inference phase, the model generates output text normally without explicitly exposing the latent variable results. Only when data verification or detection of model illusions is needed does the model call the internal decoding module or external tools to read the latent variable information in order to accurately reconstruct the corresponding sample label encoding. (S108) and its corresponding identity code (S107) enables the backtracking and verification of training data.
[0072] Label embedding regression, through its mechanism, ensures that the identity code of each training sample can participate in the model training process without affecting the quality of text generation, thereby effectively supporting the realization of the original information tracing mechanism.
[0073] The data tracing triggering operation S3 proposed in this application is as follows: Figure 4 As shown.
[0074] The data tracing triggering method is based on the original information tracing mechanism to achieve the original information tracing of LLM output content during the model inference stage.
[0075] First, the model's working architecture is divided into two modes: default mode S301 and verification mode S302. Default mode S301 is used for normal model usage and does not enable the tracing function. Verification mode S302, on the other hand, activates the tracing module when a controversial answer is detected or a user's verification request is received, mapping the "identity feature vector" S22 to the corresponding sample label code. (S108) Further restore the original information content corresponding to the answer.
[0076] Secondly, the triggering methods for verification mode S302 include two types of mechanisms: model-active triggering (S303) and model-passive triggering (S304). Model-active triggering (S303) refers to the LLM's internal integration of the answer confidence assessment mechanism (S305) and the knowledge conflict detection module (S306). When the confidence of the generated answer is lower than a preset threshold or there is a knowledge inconsistency, the system automatically enters verification mode S302, prompting the user to conduct data tracing (checking training data) and perform content verification. Model-passive triggering (S304) refers to the user explicitly requesting data tracing of the current output content during the interaction process, i.e., raising questions about the current output content (S307) or issuing an instruction (S308). After receiving the instruction, the system activates the corresponding module and executes the data tracing process.
[0077] Finally, after entering the data tracing process in verification mode, the original information reverse indexing mechanism S309 based on the identity feature vector S22 is adopted. The specific process is as follows:
[0078] ① Decoding Mapping S310: Utilizing the mapping relationship established during training, accurately convert the identity feature vector S22 into a unique sample label code. (S108), identity feature vector S22 and (S108) Design of auxiliary loss function in correspondence-dependent label embedding regression operation S2;
[0079] ② Reverse Index S311: Encoding using sample labels The encoding method and the corresponding decoding rules (such as...) (Decryption algorithm), based on (S108) The query constitutes its gather And based on the decoding rules corresponding to the original information identity encoding (such as...) The decoding algorithm will Restore to the initial HTI triplet;
[0080] ③Original Information Restoration S312: Based on the restored HTI information (S101, S102, S103), accurately locate the corresponding multi-dimensional original information 1 in the storage database, and then conduct quality verification between the LLM output content and the corresponding original information, and make corrections if necessary.
[0081] The knowledge correction process involves inviting domain experts to correct the controversial LLM-generated content 4 and the multi-source original information 1 that generated the content, and optimizing the corresponding LLM training text 2, and then performing training text optimization S313.
[0082] In the medical field, for example, when a user questions the LLM's solution for a certain symptom (corresponding to LLM generated content 4), the original information tracing mechanism will trigger verification mode S302, and based on LLM generated content 4 and its corresponding sample tag encoding... (S108) Restore the multi-source original information 1 content corresponding to the LLM generated content 4 (i.e., which works contain methods or principles for treating the symptoms), and users can make subsequent judgments or decisions based on the more authoritative multi-source original information 1 content.
[0083] The above description is only a preferred embodiment of the present invention, but the scope of protection of the present invention is not limited thereto. Any equivalent substitutions or modifications made by those skilled in the art within the scope of the technology disclosed in the present invention, based on the technical solution and inventive concept of the present invention, should be covered within the scope of protection of the present invention.
Claims
1. A method for suppressing LLM illusion by tracing original information based on data encoding, characterized in that, Includes the following steps: Step 1: Conduct multi-source raw information collection; Step 2: Input the original information identity encoding operation; Design a data source identification system to generate a unique identity code for each piece of original information, and design a placeholder retention mechanism to ensure the retention of the identity code during the semantic fusion of training text. Map the identity codes of the original information that constitute the same training sample to the sample label code to realize the correspondence between the training text and the original information that constitutes it and the identity identification code. Step 3: Use semantic fusion technology to process the multi-source raw information into LLM training text format; Step 4: Feed the LLM training text into the constructed LLM training model to carry out LLM training; Step 5: During LLM training, extend the existing LLM decoder structure by introducing sample label encoding as supervision labels to guide the model to learn the one-to-one correspondence between sample semantic information and its label encoding, so that the trained model can output its corresponding label encoding while generating text content. Step Six: When the trained LLM is used, output the generated LLM content; Step 7: Data traceability. The model's working architecture is divided into default mode and verification mode. In verification mode, after LLM generates content, a data traceability trigger mechanism is designed to decode the implicitly carried sample label encoding in the generated text when traceability is triggered, thereby restoring the original information corresponding to the training data and realizing the traceability and verifiability of the content.
2. The method for suppressing LLM hallucination by tracing original information based on data encoding according to claim 1, characterized in that: In step two, the data source identification system consists of triplet identifiers, which enables the identification of the source and type of the original information.
3. The method for suppressing LLM hallucinations by tracing original information based on data encoding according to claim 2, characterized in that: The mathematical expression of the triplet identifier system is as follows: ,in , , These represent "source document identifier", "section identifier", and "data type identifier", respectively. A hierarchical encoding structure of "[Document Type]-[Version Number]-[Hash Fingerprint]" is adopted. Based on structured markup specification coding, Define extended classification codes.
4. The method for suppressing LLM hallucinations by tracing original information based on data encoding according to claim 1, characterized in that: In step two, the placeholder retention mechanism refers to setting up a corresponding identity encoding information channel that is independent of the semantic information while retaining the semantic information of the original information, so as to avoid the identity encoding of the original information affecting the semantic information during the semantic fusion process. Semantic information and identity encoding information are in one-to-one correspondence, and the placeholder length of the placeholder retention mechanism is determined by the original information identity encoding algorithm and the collision resistance requirements.
5. The method for suppressing LLM hallucination by tracing original information based on data encoding according to claim 1, characterized in that: In step two, the encoding algorithm for the original information identity is Base64 encoding; the encoding algorithm for the sample tag is... Encryption algorithm.
6. The method for suppressing LLM hallucination by tracing original information based on data encoding according to claim 1, characterized in that: In step five, extending the existing LLM decoder structure means extracting the state of the training process and converting it into "identity features"; at the same time, designing a corresponding auxiliary loss function to align the identity features with the sample label encoding of the training data without affecting the normal training of the LLM, thereby realizing supervised modeling of identity encoding and ensuring that the model has the ability to synchronously output sample label encoding.
7. The method for suppressing LLM illusion by tracing original information based on data encoding according to claim 1, characterized in that: In step seven, the default mode in the model's working architecture refers to the model interacting normally with the user, while the verification mode refers to the model tracing the original information of the LLM-generated content and verifying the accuracy of the answer.
8. The method for suppressing LLM hallucination by tracing original information based on data encoding according to claim 1, characterized in that: In step seven, the data tracing triggering mechanism includes active model triggering and passive model triggering. Active triggering depends on the confidence score of the LLM-generated content and the knowledge conflict detection results, while passive triggering depends on the user's needs. The confidence score of LLM-generated content refers to the LLM's self-assessment of the correctness or reliability of its generated content, which can help users or the model itself assess the credibility of the model's output; the knowledge conflict detection result of LLM-generated content refers to the LLM's knowledge consistency check of its generated content, which can avoid knowledge conflicts and thus ensure the consistency of the model's output. User needs refer to users' questions about the content generated by LLM and their requests for content verification.
9. The method for suppressing LLM hallucination by tracing original information based on data encoding according to claim 1, characterized in that: In step seven, the decoding process of the sample label encoding includes three steps: decoding mapping, reverse indexing, and original information restoration.
10. A method for suppressing LLM hallucinations by tracing original information based on data encoding according to claim 9, characterized in that: Decoding mapping refers to mapping the identity feature vectors during LLM training to sample label encodings corresponding to the training statements; Inverse indexing refers to decoding all the original information triplet identifiers corresponding to the training sample based on the decoding algorithm corresponding to the sample label encoding and the original information identity encoding; Original information restoration refers to locating the original information that constitutes the training text based on the original information triplet identifier information, and the original information is fed back to the verification requester in the storage format used.