Information processing method and information processing system

Abstract

This invention provides an information processing method, etc., that can suitably prove the validity of the information used by LLM. [Solution] The information processing method involves a first computer 1 executing a process to acquire user information, generate processed data by processing the user information according to a predetermined logic, calculate a zero-knowledge proof value indicating the validity of the processed data, and output the processed data and the zero-knowledge proof value. A second computer 2 receives a verification request from a language model system that generates a response using the processed data, which includes the processed data and the zero-knowledge proof value. The second computer 2 executes a process to verify the validity of the processed data based on the zero-knowledge proof value and output the verification result to the language model system.

Description

Technical Field

[0001] The present invention relates to an information processing method and an information processing system.

Background Art

[0002] As represented by the RAG (Retrieval-Augmented Generation) technique, a mechanism for controlling the output of a large language model (Large Language Model, hereinafter simply referred to as "LLM") by connecting the LLM to specific knowledge is called grounding (see Non-Patent Document 1). By applying the grounding mechanism, there is a possibility that an LLM service can collect user information (for example, the user's age, address, financial situation, family composition, etc.) and provide a service that presents some personalized answers (for example, life planning, investment advice, etc.).

Prior Art Documents

Non-Patent Documents

[0003]

Non-Patent Document 1

Summary of the Invention

Problems to be Solved by the Invention

[0004] In one aspect, an object is to provide an information processing method or the like that can suitably prove the legitimacy of information used by an LLM.

Means for Solving the Problems

[0005] The information processing method involves a first computer performing the following processes: acquiring user information, generating processed data by processing the user information according to a predetermined logic, calculating a zero-knowledge proof value indicating the validity of the processed data, and outputting the processed data and the zero-knowledge proof value; receiving a verification request from a language model system that generates a response using the processed data, which includes the processed data and the zero-knowledge proof value; verifying the validity of the processed data based on the zero-knowledge proof value; and outputting the verification result to the language model system. [Effects of the Invention]

[0006] In one respect, it can effectively demonstrate the validity of the information used by LLMs. [Brief explanation of the drawing]

[0007] [Figure 1] This is a diagram showing an example of the configuration of an information processing system. [Figure 2] This block shows an example configuration for a certification server. [Figure 3] This is a block diagram showing an example configuration for a verification server. [Figure 4] This is a diagram illustrating the overview of Embodiment 1. [Figure 5] This is an explanatory diagram regarding processing when using zk-SNARK. [Figure 6] This is an explanatory diagram regarding the processing when using zk-STARK. [Figure 7] This is an explanatory diagram regarding the response generation process using LLM. [Figure 8] This figure shows an example of LLM input / output. [Figure 9] This flowchart shows the procedure for calculating the Proof value. [Figure 10] This flowchart shows the procedure for verifying processed data. [Figure 11] This is a flowchart showing the steps involved in the response generation process. [Figure 12]This is an explanatory diagram regarding the response generation process according to Embodiment 2. [Figure 13] This is a flowchart showing the procedure for the response generation process according to Embodiment 2. [Modes for carrying out the invention]

[0008] The present invention will be described in detail below with reference to the drawings illustrating its embodiments. (Embodiment 1) Figure 1 shows an example of the configuration of an information processing system. In this embodiment, we describe an information processing system that proves the validity of information provided to an LLM using zero-knowledge proof. The information processing system includes a proof server 1 (first computer), a verification server 2 (second computer), a user terminal 3, a provider system 4, and an LLM system 5 (language model system). Each device is connected via a network N such as the Internet.

[0009] LLM System 5 is a system (server computer) that provides a chat service using LLM. LLM System 5 accepts arbitrary queries (instructions that tell the system to generate an answer) from the user, inputs the query into LLM, generates an answer, and presents it to the user.

[0010] User terminal 3 is a terminal device used by users of this system, such as a personal computer, smartphone, or tablet. User terminal 3 receives query input from the user, outputs it to the LLM system 5, and retrieves and displays the answer from the LLM system 5.

[0011] As described above, the LLM system 5 provides a chat service using LLM. In this embodiment, in order to personalize the response to the user, user information about the user (e.g., age, asset status, deposit / withdrawal history, investment history, etc.) is used to generate the response.

[0012] The provider system 4 is a system (such as a server computer) that stores user information and provides the user information used for generating answers by the LLM. The provider of the user information (data provider) is, for example, a financial institution, an insurance company, a communication company, a credit card company, etc., but is not particularly limited. By giving the user information to the LLM, personalized answers can be generated.

[0013] When providing user information to the LLM system 5 as described above, there is a problem of how to balance privacy and the reliability of the answers. For example, if all the user information (personal information) stored in the provider system 4 is provided to the LLM system 5, the reliability of the answers will increase, but the user has to disclose more privacy than necessary. On the other hand, if privacy is prioritized and only fragmented information is provided, there is a risk of generating contradictory answers (for example, answers including mistakes in the total of various expenses, mistakes between the take-home pay and annual income, etc.), which is related to the quality of the LLM service.

[0014] Therefore, in this embodiment, the user information given to the LLM is processed, and the legitimacy of the processed data is proved using zero-knowledge proof. The processed data whose legitimacy is proved is used as information for grounding.

[0015] The proof server 1 is a server computer capable of various information processing and information transmission and reception. Note that the computer corresponding to the proof server 1 does not have to be a server computer. The proof server 1 acquires user information from the provider system 4 and generates processed data obtained by processing the user information. Then, the proof server 1 calculates a zero-knowledge proof value (hereinafter referred to as "Proof value") indicating the legitimacy of the processed data using the zero-knowledge proof algorithm, and provides it to the user together with the processed data.

[0016] Verification Server 2 is a server computer capable of various information processing and information transmission / reception. Note that the computer corresponding to Verification Server 2 does not necessarily have to be a server computer. Verification Server 2 verifies the validity of the processed data based on the Proof value calculated by Proof Server 1 and outputs the verification result to the LLM system 5. If the verification is successful, the processed data is used to generate the answer by LLM.

[0017] Figure 2 is a block diagram showing an example configuration of the certification server 1. The certification server 1 comprises a control unit 11, a main memory unit 12, a communication unit 13, and an auxiliary memory unit 14. The control unit 11 has one or more processors such as a CPU (Central Processing Unit), MPU (Micro-Processing Unit), and GPU (Graphics Processing Unit), and performs various information processing by reading and executing the program P1 stored in the auxiliary storage unit 14. The main memory unit 12 is a temporary storage area such as SRAM (Static Random Access Memory) or DRAM (Dynamic Random Access Memory), and temporarily stores the data necessary for the control unit 11 to perform calculation processing. The communication unit 13 is a communication module for performing communication-related processing, and sends and receives information with the outside. The auxiliary storage unit 14 is a non-volatile storage area such as a hard disk or large-capacity memory, and stores the program P1 (program product) and other data necessary for the control unit 11 to perform processing.

[0018] The certification server 1 may be a multi-computer system consisting of multiple computers, or it may be a virtual machine created virtually by software.

[0019] Furthermore, in this embodiment, the certification server 1 is not limited to the above configuration and may include, for example, an input unit for receiving operation input, a display unit for displaying images, etc. Also, the certification server 1 may be equipped with a reading unit for reading portable storage media 1a such as CD (Compact Disk)-ROM or DVD (Digital Versatile Disc)-ROM, and may read and execute program P1 from the portable storage media 1a.

[0020] Figure 3 is a block diagram showing an example configuration of the verification server 2. The verification server 2 comprises a control unit 21, a main memory unit 22, a communication unit 23, and an auxiliary memory unit 24. The control unit 21 has one or more processors such as CPUs and performs various information processing by reading and executing the program P2 stored in the auxiliary storage unit 24. The main memory unit 22 is a temporary storage area such as RAM and temporarily stores data necessary for the control unit 21 to perform calculation processing. The communication unit 23 is a communication module for performing communication-related processing and sends and receives information with the outside. The auxiliary storage unit 24 is a non-volatile storage area such as a hard disk and stores the program P2 (program product) and other data necessary for the control unit 21 to perform processing.

[0021] The verification server 2 may be a multi-computer system consisting of multiple computers, or it may be a virtual machine created virtually by software.

[0022] Furthermore, the verification server 2 may include, for example, an input unit for receiving operation inputs, a display unit for displaying images, etc. The verification server 2 may also be equipped with a reading unit for reading the portable storage medium 2a, and may read and execute the program P2 from the portable storage medium 2a.

[0023] Figure 4 is a diagram illustrating the overview of Embodiment 1. The overview of this embodiment will be described based on Figure 4.

[0024] First, user terminal 3 requests provider system 4 to output processed user information in order to provide processed user information data to LLM system 5. Upon receiving the request for processed data output, provider system 4 outputs the user information (raw data) to certification server 1.

[0025] The certification server 1 generates processed data from user information obtained from the provider system 4 according to a predetermined logic (program). For example, the certification server 1 generates processed data by anonymizing the user information (for example, changing age from "31 years old" to "30s" or changing assets from "6.2 million yen" to "5 to 7 million yen") or by performing predetermined inferences (for example, analyzing and labeling the user's consumption trends).

[0026] Proof Server 1 calculates a Proof value indicating the validity of the processed data according to a zero-knowledge proof algorithm. Examples of zero-knowledge proof algorithms that can be used include zk-SNARK (Zero Knowledge Succinct Non-interactive Argument of Knowledge) and zk-STARK (Zero Knowledge Scalable Transparent Argument of Knowledge).

[0027] The certification server 1 outputs the processed data and Proof values ​​to the provider system 4. The certification server 1 also shares the logic (program) used to generate the processed data with the verification server 2 in advance. The provider system 4 outputs the processed data and Proof values ​​output from the certification server 1 to the user terminal 3.

[0028] User terminal 3 outputs the processed data obtained from provider system 4 to LLM system 5 in order to use the processed data for generating answers by LLM. In this case, user terminal 3 also outputs the Proof value. LLM system 5 outputs a verification request for the processed data, including the Proof value, to verification server 2 to verify the validity of the processed data.

[0029] Verification server 2 outputs the verification results to LLM system 5. Only if verification results indicating successful verification of the validity of the processed data are obtained from verification server 2, LLM system 5 stores (saves) the processed data in user DB 501 and uses it to generate the answer.

[0030] Figure 5 is an explanatory diagram of the process when using zk-SNARK. Based on Figure 5, the proof and verification process of the validity of processed data when using zk-SNARK as the zero-knowledge proof algorithm will be explained.

[0031] zk-SNARK is a non-interactive zero-knowledge proof algorithm that allows a prover to prove a proposition to be true without revealing any secrets (user information in this embodiment) to the verifier. zk-SNARK has advantages such as being able to simplify the size of the proof.

[0032] First, let's explain the certification step for processed data. Certification Server 1, the certifier, obtains a certification key generated based on the logic used to generate the processed data (the logic to be certified) and predetermined public parameters. The public parameters are parameters created through a process called trusted setup, and in this embodiment, they are fixed throughout the entire system. Generally, certification keys are generated by a trusted third party (not shown).

[0033] The proof server 1 generates (calculates) processed data (output of the logic) and calculates a Proof value based on the logic to be proven, user information to be kept confidential (input of the logic), and the proof key. As this is publicly known technology, a detailed explanation will be omitted, but the proof server 1 obtains a scalar value by multiplying the QAP (Quadratic Arithmetic Programs) obtained from the proof key and logic, the user information, and predetermined public information. The proof server 1 then encrypts the scalar value using the proof key and calculates the Proof value based on the logic by summing the encrypted values.

[0034] Next, we will describe the verification step for the processed data. Verification server 2, which is the verifier, obtains a verification key generated based on the logic used to generate the processed data and the public parameters. Similar to the proof key mentioned above, the verification key is generally generated by a trusted third party.

[0035] Verification Server 2 verifies the validity of the processed data based on the Proof value, the processed data (logic output), and the verification key. While a detailed explanation is omitted, Verification Server 2 determines whether the paired (cryptographically processed) Proof value matches the result of pairing it with the verification key and public information.

[0036] Figure 6 is an explanatory diagram of the process when using zk-STARK. Based on Figure 6, the proof and verification process of the validity of processed data when using zk-STARK as the zero-knowledge proof algorithm will be explained.

[0037] zk-STARK, like zk-SNARK, is a zero-knowledge proof algorithm that proves a proposition is true without revealing any secrets. Unlike zk-SNARK, zk-STARK has advantages such as not requiring a trusted setup and being quantum resistant.

[0038] First, let's explain the proof step for processed data. Proof Server 1, the prover, generates (calculates) processed data (output of the logic) based on the logic used to generate the processed data (the logic to be proven) and the user information to be kept secret (input of the logic), and also obtains an execution trace of the logic used to generate the processed data. The execution trace refers to the intermediate states necessary to prove the statement, and refers to each term when the logic is transformed into a polynomial problem.

[0039] Furthermore, the proof server 1 obtains (calculates) verification parameters to identify the logic to be proven. The reason for providing verification parameters is that, as a challenge with zk-STARK, it is not possible to determine from the zk-STARK protocol alone whether the proof (Proof) provided to the verifier is truly generated from the intended logic. In this embodiment, by providing verification parameters, the verifier can verify that the logic provided by the prover is the intended logic.

[0040] For example, proof server 1 obtains the hash value (code hash) of the logic as a verification parameter. In addition, proof server 1 may also obtain the hash value of the initial conditions of the logic as a verification parameter.

[0041] Proof Server 1 calculates the Proof value based on the execution trace and verification parameters. A detailed explanation is omitted as it is a publicly known technique, but for example, Proof Server 1 calculates the Proof value from datasets such as the polynomial commitment converted from the execution trace (evaluation results and Merkle paths at randomly selected evaluation points), the polynomial commitment evaluated using the FRI (Fast Reed-Solomon Interactive Oracle Proofs of Proximity) protocol (evaluation results and Merkle paths), the random oracle challenge (evaluation points), and initial and boundary condition data (specific conditions at the beginning and in the middle of the execution trace).

[0042] In this case, proof server 1 configures the Proof value to include the above verification parameters. For example, proof server 1 embeds a code hash in part of the Merkle tree.

[0043] In this embodiment, verification parameters are included as part of the Proof value, but the verification parameters may be output separately from the Proof value.

[0044] Next, we will explain the verification step for the processed data. Verification server 2, which is the verifier, obtains (calculates) verification parameters from the logic used to generate the processed data. Verification server 2 then confirms that the Proof value contains these verification parameters. This allows verification server 2 to determine whether the logic used to generate the processed data is correct.

[0045] If the Proof value is found to contain the validation parameters, the validation server 2 verifies the validity of the processed data based on the Proof value, the processed data (logic output), and the validation parameters. While a detailed explanation is omitted, the validation server 2 performs AIR (Arithmetic Intermediate Representation) validation, FRI validation, and other similar verifications based on the Proof value.

[0046] The certification server 1 and verification server 2 perform proof and verification of the validity of the processed data using the algorithms described in Figures 5 and 6. If the verification server 2 outputs a verification result indicating that the verification was successful, the LLM system 5 stores the processed data in the user DB 501 and uses it to generate the answer.

[0047] Figure 7 is an explanatory diagram of the response generation process using LLM. Based on Figure 7, the process of how the LLM system 5 generates responses using LLM will be explained.

[0048] The LLM system 5 receives a query (instruction statement) from the user terminal 3 that commands the generation of an answer. When a query is received, the LLM system 5 searches the user DB 501 for processed data to be used to generate the answer. For example, the LLM system 5 uses LLM to convert the query into a vector value and then searches the user DB 501 for processed data based on the converted vector value.

[0049] The LLM system 5 creates a prompt that includes processed data retrieved from user DB 501 as context, in addition to the query entered by the user. The LLM system 5 generates an answer by inputting the created prompt into LLM. The LLM system 5 outputs the generated answer to user terminal 3.

[0050] Figure 8 shows an example of LLM input and output. In the example in Figure 8, the system responds to the query "What is a recommended way to build wealth?" by recommending regular investment. In addition to the query entered by the user, the LLM system 5 can generate responses that depend on user information by including anonymized results of the user's age and asset status, as well as the results of inferring the user's type, as context in the prompt.

[0051] In this case, if processed data is used to generate the answer, it is preferable to add a mark to the answer indicating that processed data was used. For example, as shown in Figure 8, the LLM system 5 adds a mark "[P]" to indicate that processed data was used. This makes it possible to show the user that the answer was generated based on user information.

[0052] Based on the above, according to this embodiment, privacy can be protected by processing user information, while the reliability of the information (processed data) provided to the LLM can be ensured by zero-knowledge proofs.

[0053] Figure 9 is a flowchart showing the procedure for calculating the Proof value. Based on Figure 9, the process by which the proof server 1 generates processed data and calculates the Proof value will be explained. Note that this flowchart assumes that the process follows the zk-SNARK algorithm.

[0054] The certification server 1 obtains user information from the provider system 4 in response to a request from the user (step S11). The certification server 1 obtains a certification key generated based on the logic used to generate processed data and predetermined public parameters (step S12).

[0055] The certification server 1 generates processed data by processing user information according to the logic (step S13). The certification server 1 calculates a Proof value indicating the validity of the processed data based on the acquired certification key (step S14). The certification server 1 outputs the processed data and the Proof value to the user terminal 3 via the provider system 4 (step S15), and the series of processes ends.

[0056] Figure 10 is a flowchart showing the procedure for verifying the processed data. Based on Figure 10, the processing steps taken by the verification server 2 to verify the validity of the processed data will be explained. Note that this flowchart assumes that the processing is carried out according to the zk-SNARK algorithm.

[0057] The verification server 2 receives a verification request from the LLM system 5, which generates a response using LLM with the processed data, including the processed data and the Proof value (step S31). The verification server 2 obtains a verification key generated based on the logic used to generate the processed data and the public parameters (step S32). The verification server 2 verifies the validity of the processed data based on the obtained verification key and the Proof value (step S33). The verification server 2 outputs the verification result to the LLM system 5 (step S34). If the verification is successful, the LLM system 5 stores the processed data in the user DB 501. The verification server 2 then terminates the series of processes.

[0058] Figure 11 is a flowchart showing the steps of the response generation process. Based on Figure 11, the process by which the LLM system 5 generates responses using processed data will be explained.

[0059] The LLM system 5 receives a response generation request from the user terminal 3, which includes a query instructing the generation of a response (step S51). Based on this query, the LLM system 5 searches the user DB 501 for processed data to be used to generate the response (step S52).

[0060] The LLM system 5 creates a prompt that includes a query instructing the generation of an answer and a context containing the processed data (step S53). The LLM system 5 generates an answer by inputting the created prompt into the LLM (step S54). The LLM system 5 adds a mark to the answer indicating that processed data was used (step S55). The LLM system 5 outputs the answer to the user terminal 3 (step S56) and completes the series of processes.

[0061] In the above, the user terminal 3 downloads the processed data (and Proof values) from the provider system 4 and provides it to the LLM system 5. However, this embodiment is not limited to this. For example, the certification server 1 may output the processed data directly to the LLM system 5 via the provider system 4, and the verification server 2 may receive a verification request from the LLM system 5 when the LLM system 5 has acquired the processed data. In other words, it is not necessary for the user to download the processed data once; the LLM system 5 may download the processed data directly.

[0062] Based on the above, according to this embodiment 1, the validity of the information used by the LLM can be suitably proven.

[0063] (Embodiment 2) In this embodiment, in order to make the response from LLM more reliant on user information, we describe a configuration in which processed user information data is preferentially added to the prompt to repeatedly generate (correct) the response. Note that the same reference numerals are used for parts that overlap with Embodiment 1, and their explanation is omitted.

[0064] Figure 12 is an explanatory diagram relating to the response generation process according to Embodiment 2. Based on Figure 12, the outline of this embodiment will be explained.

[0065] In this embodiment, the LLM system 5 creates a prompt that includes data other than processed data with a Proof value in the context. Data other than processed data is, for example, text data such as web articles, and is stored in knowledge DB 502, which is different from user DB 501. When the LLM system 5 receives a query input from the user, it searches for text data in knowledge DB 502 based on the query. The search for text data is performed, for example, by vector search.

[0066] The LLM system 5 creates a prompt (first prompt) that includes the query entered by the user, the processed data, and a context containing data other than the processed data. The LLM system 5 generates an answer (first answer) by inputting this prompt into the LLM.

[0067] In this embodiment, the LLM system 5 modifies the response to make it more personalized. Specifically, the LLM system 5 sequentially removes data other than the processed data from the context and causes the LLM to regenerate the response.

[0068] In the example in Figure 12, the context includes processed data with Proof values, such as "User is in their 30s, with assets of 5-7 million yen" and "User has a balanced savings / investment portfolio." Simultaneously, the context includes data without Proof values, such as "Articles on speculative techniques" and "Articles on life planning." The LLM system 5 removes the "Articles on life planning," which are data other than processed data, from this context.

[0069] The LLM system 5 creates a prompt (second prompt) that includes the previously generated answer (first answer), the query, and the context with data other than processed data removed. By inputting this prompt into the LLM, the LLM system 5 generates a new answer (second answer) that corrects the previous answer.

[0070] The LLM system 5 repeats this process, correcting the responses until no data other than processed data remains in the context. For example, in the example in Figure 12, the "speculative techniques article" is deleted next. By repeating this process, processed data with Proof values ​​is preferentially used in generating responses, making the responses more reliant on user information.

[0071] Figure 13 is a flowchart showing the procedure for the response generation process according to Embodiment 2. After searching for processed data from the user DB 501 (step S52), the LLM system 5 executes the following process. The LLM system 5 searches for data other than processed data from the knowledge database 502 based on the query entered by the user (step S201). For example, the knowledge database 502 stores text data such as web articles, and the LLM system 5 searches for text data related to the query entered by the user.

[0072] The LLM system 5 creates a prompt (first prompt) that includes a query instructing the generation of an answer, a context that includes the processed data retrieved in step S52 and data other than the processed data retrieved in step S201 (step S202). The LLM system 5 generates an answer (first answer) by inputting the created prompt into the LLM (step S203).

[0073] The LLM system 5 determines whether or not to correct the generated response (step S204). For example, the LLM system 5 determines whether or not the context of the prompt entered into the LLM contained data other than processed data.

[0074] If it is determined that the answer should be corrected (S204: YES), the LLM system 5 returns to step S202. In this case, the LLM system 5 creates a prompt (second prompt) again that includes the answer generated immediately before, the query, and the context with data other than the processed data removed (step S202). The LLM system 5 inputs this prompt into the LLM to generate a new answer (second answer) that corrects the previous answer (step S203).

[0075] If it is determined that no correction is needed to the answer (S204: NO), the LLM system 5 proceeds to step S55.

[0076] Based on the above, according to this embodiment 2, the answers generated using LLM can be made to be more reliant on user information.

[0077] (modified version) Embodiment 2 described a method for generating (correcting) responses by prioritizing the use of processed data with Proof values, by removing data other than processed data with Proof values ​​from the context. On the other hand, according to the inventor's research, it has been found that in some use cases, it is possible to generate more appropriate responses by removing processed data with Proof values ​​from the context, rather than using data without Proof values. For example, if processed data with Proof values ​​is objective data based on user information, such as "aggressive investment style," and data without Proof values ​​is a subjective opinion of the user, such as "I want to operate safely and securely," then prioritizing the use of data without Proof values ​​can generate responses that are more in line with the user's subjective opinion.

[0078] Therefore, the LLM system 5 may generate a corrected answer (second answer) by inputting a prompt (second prompt) into the LLM that includes the answer output from the LLM (first answer), the query, and a context from which the processed data has been removed (not data other than processed data). In this way, the data to be input into the LLM may be selected depending on the use case.

[0079] The embodiments disclosed herein should be considered in all respects to be illustrative and not restrictive. The scope of the present invention is indicated by the claims, not in the sense described above, and all modifications within the sense and scope equivalent to the claims are intended.

[0080] The matters described in each embodiment can be combined with each other. Furthermore, the independent and dependent claims described in the claims can be combined with each other in any combination, regardless of the form of reference. In addition, the claims use a form in which claims referencing two or more other claims (multi-claim form), but are not limited to this. A form in which multi-claims referencing at least one multi-claim (multi-multi-claim) may also be used. [Explanation of symbols]

[0081] 1. Certification Server (First Computer) 11 Control Unit 12 Main memory 13 Communications Department 14 Auxiliary storage P1 Program 2. Verification Server (Second Computer) 21 Control Unit 22 Main memory 23 Communications Department 24 Auxiliary storage P2 Program 3. User terminals 4. Provider System 5. LLM System (Language Model System)

Claims

1. Obtain user information, Processed data is generated by processing the user information according to a predetermined logic. A zero-knowledge proof value is calculated to demonstrate the validity of the processed data. Output the processed data and zero-knowledge proof values. The first computer executes the process, The language model system that generates responses using the processed data receives a verification request including the processed data and zero-knowledge proof values. Based on the aforementioned zero-knowledge proof values, the validity of the processed data is verified, The verification results are output to the language model system. An information processing method in which a second computer performs the processing.

2. The first computer is, Obtain a proof key generated based on the aforementioned logic and predetermined public parameters. Based on the aforementioned proof key, the zero-knowledge proof value is calculated, The second computer is, Obtain the verification key generated based on the aforementioned logic and the aforementioned public parameters. The processed data is verified based on the zero-knowledge proof value and the verification key. The information processing method according to claim 1.

3. The first computer is, Calculate verification parameters to identify the aforementioned logic, Based on the execution trace of the logic used to generate the processed data, the zero-knowledge proof value is calculated. The processed data, zero-knowledge proof values, and verification parameters are output. The second computer is, The language model system receives the verification request, which includes the processed data, zero-knowledge proof values, and verification parameters. Based on the aforementioned verification parameters, it is determined whether the logic used to generate the processed data is correct. If it is determined that the logic is correct, the validity of the processed data is verified based on the zero-knowledge proof value. The information processing method according to claim 1.

4. The first computer outputs the processed data and zero-knowledge proof values ​​to the user. The second computer receives the verification request from the language model system, which includes the processed data and the zero-knowledge proof, when the language model system obtains the processed data and the zero-knowledge proof from the user. The information processing method according to claim 1.

5. The first computer outputs the processed data and zero-knowledge proof values ​​to the language model system. The second computer receives the verification request from the language model system, which includes the processed data and zero-knowledge proof values, when the language model system obtains the processed data and zero-knowledge proof values ​​from the first computer. The information processing method according to claim 1.

6. The aforementioned language model system is A first response is generated by inputting a first prompt to the language model, which includes a query instructing the generation of a response, the processed data, and a context containing data other than the processed data. By inputting a second prompt to the language model, which includes the generated first response, the query, and a context that includes data other than the processed data or the processed data with the processed data removed, a second response is generated that corrects the first response. The information processing method according to claim 1.

7. When the language model system uses the processed data to generate a response, it adds a mark to the response indicating that the processed data was used. The information processing method according to claim 1.

8. An information processing system having a first computer and a second computer, The first computer, Obtain user information, Processed data is generated by processing the user information according to a predetermined logic. A zero-knowledge proof value is calculated to demonstrate the validity of the processed data. Output the processed data and zero-knowledge proof values. The second computer, The language model system that generates responses using the processed data receives a verification request including the processed data and zero-knowledge proof values. Based on the aforementioned zero-knowledge proof values, the validity of the processed data is verified, The verification results are output to the language model system. Information processing system.

Images