Mathematical inference task-oriented large model interpretable inference registration system and method

Abstract

The invention discloses a mathematical inference task-oriented large-model interpretable inference registration system and method, which are used for responding to a mathematical inference task in an input natural language form, and carrying out symbolic extraction and formalized concept alignment based on mathematical domain knowledge to obtain a structured concept set. A formalized language is used as an intermediate interpretation layer, a natural language reasoning process is mapped into a formalized reasoning sequence, and then a reasoning path is realized by symbolic representation. And verifying the logic consistency between the intermediate hypothesis generated by the model and the constraint relationship through a symbol solver, and continuously correcting a model reasoning path under an iterative feedback mechanism to realize transparency, verifiability and knowledge consistency of the reasoning process. The method can improve the stability of a large model reasoning result in a complex reasoning scene, and can be widely applied to practical application scenes containing complex mathematical reasoning requirements, including but not limited to reasoning enhancement and interpretability analysis of large models in the vertical field, such as mathematical reasoning, symbolization calculation and financial reasoning.

Description

Technical Field

[0001] This invention belongs to the fields of artificial intelligence, machine learning, and large model interpretability, and particularly relates to a large model interpretability inference registration system and method for mathematical reasoning tasks. Background Technology

[0002] With the rapid development of large-scale pre-trained language models (LLMs), their reasoning abilities have shown significant progress in tasks such as natural language processing, scientific computing, and question answering. Particularly in mathematical reasoning tasks, large models can generate complex problem-solving processes through step-by-step reasoning or chain-of-thought (CoT), marking a significant leap in artificial intelligence from "language generation" to "logical reasoning." However, despite the high accuracy of these models on well-defined, standardized datasets, their capabilities still show considerable limitations in reasoning tasks geared towards real-world applications.

[0003] In real-world scenarios such as engineering design, scientific computing, automated decision-making, and complex system analysis, reasoning results not only need to "provide answers," but also require traceable processes, verifiable logic, and auditable constraints. However, even when existing large-scale language models achieve high accuracy on standardized, well-defined datasets, the reasoning processes they generate are often difficult to formally verify, and the relationship between reasoning paths and domain rules is opaque, failing to meet the reliability and consistency requirements of real-world scenarios. Current large-scale language models demonstrate strong computational and linguistic capabilities in mathematical reasoning tasks, but two prominent problems remain in complex reasoning scenarios: first, the reasoning process is unverifiable and lacks interpretability, making it difficult to cope with diverse and complex real-world reasoning scenarios; second, the logical chains generated by the model are easily affected by surface semantic biases, leading to inconsistencies between reasoning paths and mathematical constraints, making it difficult to achieve sufficient modeling in one go and requiring continuous iterative optimization. Traditional cue-based fine-tuning or reinforcement learning methods often focus on result optimization but fail to achieve deep registration between reasoning paths and domain knowledge. Therefore, a technical solution is urgently needed that can make the internal logical structure of the model transparent and align it with formal mathematical knowledge systems.

[0004] Existing methods often attempt to enhance the "logical consistency" of models at the reasoning level, such as through process supervision, CoT refinement, or by introducing external verifiers to improve the quality of inference paths. While these methods improve the output stability of models to some extent, they still remain at the linear generation mode of natural language and lack a mechanism to verify the consistency between the reasoning process and the result, making the reasoning structure difficult to track and interpret.

[0005] Meanwhile, the development of formal reasoning techniques has provided new approaches to solving the problem of reasoning consistency. By explicitly introducing variables, constraints, and logical relationships into the reasoning chain, the intermediate reasoning states of the model can be structurally modeled and their consistency verified. Some studies have attempted to integrate symbolic methods with neuro-symbolic methods to enhance interpretability, but several challenges remain: First, linguistic reasoning and symbolic modeling exist in different semantic spaces, and direct splicing can easily lead to semantic drift; second, the correspondence between variable definitions, constraints, and the problem-solving process lacks automatic mapping and calibration; finally, there is a lack of a unified "reasoning registration framework" to achieve logical consistency tracking and dynamic alignment in multi-step reasoning.

[0006] Therefore, how to construct a systematic framework that can be oriented towards real-world application scenarios and achieve dynamic registration of logical consistency, verifiable reasoning process, and interpretable results in mathematical reasoning has become a key issue for large model reasoning capabilities to move from "usable" to "reliable and deployable". It is also the core gap that urgently needs to be bridged between current research on reasoning interpretability and practical implementation. Summary of the Invention

[0007] The summary section of this application is intended to provide a brief overview of the concepts, which will be described in detail in the detailed description section below. This summary section is not intended to identify key or essential features of the claimed technical solutions, nor is it intended to limit the scope of the claimed technical solutions.

[0008] To address the problems and shortcomings of existing technologies, this invention aims to provide a large-scale interpretable reasoning registration system and method for mathematical reasoning tasks. By introducing formal language as an intermediate interpretation layer, it achieves dynamic alignment between model-generated semantics and mathematical symbol constraints, thereby ensuring the accuracy of reasoning while making the reasoning process logically verifiable and interpretable. This solves the problems mentioned in the background technology.

[0009] To achieve the above objectives, the present invention provides the following technical solution: As a first aspect of this application, the present invention discloses a large-model interpretability reasoning registration method for mathematical reasoning tasks, comprising the following steps: Step 1, responding to a mathematical reasoning task in the form of natural language input; Step 2: Based on the characteristics of the mathematical reasoning task, a structured concept set is obtained by symbolic extraction and formal concept alignment based on mathematical domain knowledge. Step 3: Introduce a formal language into the structured concept set, map the natural language reasoning process into a formal reasoning sequence, and implement the reasoning path using symbolic representation; Step 4: Use a symbolic solver or theorem prover to perform formal verification of the logical consistency and mathematical validity of the reasoning path; Step 5: Based on the formal verification results and the prediction results of the large language model, the inference path is iteratively corrected through a search strategy; Step 6: Output the structured mathematical reasoning results and their interpretable structure after formal verification and iterative correction.

[0010] Preferably, in step 2, based on the characteristics of the mathematical reasoning task, mathematical entities, relations, and operators involved are first extracted from the domain knowledge base and the question corpus to construct a formal concept representation; then, through symbolic extraction and semantic alignment, the alignment mapping between the natural language description and the formal concept is formed to create the structured concept set.

[0011] Preferably, in step 3, the formal reasoning sequence can be modeled as a constraint system, which includes partial equality constraints and inequality constraints. The partial equality constraints involve various assignment operations, while the inequality constraints consist of various constraints containing real-world knowledge.

[0012] Preferably, in step 4, the correctness of the algebraic equations, logical conditions and numerical calculations involved in the reasoning are first verified based on the symbolic solver or theorem prover, and then the global constraint consistency check is performed on the reasoning path. If a logical contradiction or calculation error is detected, a correction signal is generated and fed back to perform iterative correction.

[0013] Preferably, in step 5, a heuristic search or Monte Carlo search strategy is used to simulate and analyze the distribution of variables and combination of constraints in the symbolic space, and the reasoning path and model generation strategy are dynamically corrected based on the results of the formal verification, so as to achieve the gradual alignment of the neural network output and the symbolic system logic.

[0014] Preferably, in step 6, by integrating the intermediate results from the structured concept set and the reasoning path after formal verification, and utilizing variable-constraint mapping tables, reasoning graph structures, and visual representations, the reasoning process is made transparent and logically traceable. This not only outputs the final numerical solution but also the symbolic logic basis and verification certificate of the reasoning.

[0015] Preferably, in step 5, the heuristic search employs an iterative optimization mechanism based on variable queues and constraint pools in the symbolic space, and further includes the following steps: Step 5.1: Obtain the variable to be processed from the variable queue and retrieve the set of constraints associated with it from the constraint pool; Step 5.2: Evaluate the current state using a large language model and introduce a candidate set of real-world knowledge generation constraints; Step 5.3: Filter the candidate constraint set based on confidence level or consistency score to obtain a new constraint set, and replace the constraint system obtained in step 3 with the new constraint set to form a new constraint system; Step 5.4: Introduce a large language model to evaluate the original constraint system and the new constraint system, and verify them using the results of the symbolic verifier. Step 5.5: Select the better-performing constraint system as the new initial state and the starting input for the next iteration.

[0016] As a second aspect of this application, the present invention also discloses a large-model interpretable reasoning registration system for mathematical reasoning tasks, comprising: The concept registration module, including a symbol extraction unit and a semantic alignment unit, is used to extract symbols and align formal concepts based on mathematical knowledge to obtain a structured concept set. The symbolic reasoning registration module, including the hypothesis generation unit and the variable constraint search unit, is used to introduce formal language to map the natural language reasoning process into a formal reasoning sequence, and then realize it as a reasoning path using symbolic representation; The formal verification module includes a symbolic solver unit and a consistency determination unit, which are used to perform formal verification of the logical consistency and mathematical validity of the reasoning path using a symbolic solver or a theorem prover. The iterative update control module includes a hypothesis screening unit, a conflict detection unit, and a feedback optimization unit, which is used to iteratively correct the inference path based on the formal verification results and the prediction results of the large language model through a search strategy. The interpretable reasoning output terminal module is used to output the verified and corrected reasoning results and their interpretation structure.

[0017] As a third aspect of this application, the present invention also discloses an electronic device comprising: At least one processor, and a memory communicatively connected to said at least one processor; The memory stores instructions that can be executed by the at least one processor, which enable the at least one processor to perform the steps of the above-described large model interpretability reasoning registration method for mathematical reasoning tasks.

[0018] As a fourth aspect of this application, the present invention also discloses a computer storage medium having a computer program stored thereon, characterized in that, when the computer program is executed by a processor, it implements the steps of the above-described large model interpretability reasoning registration method for mathematical reasoning tasks.

[0019] Compared with the prior art, the beneficial effects of the present invention are as follows: This invention discloses a large-scale interpretable reasoning registration system and method for mathematical reasoning tasks. Responding to a mathematical reasoning task input in natural language form, and based on the characteristics of the mathematical reasoning task, it obtains a structured concept set through symbolic extraction and formal concept alignment based on mathematical domain knowledge. Formal language is introduced into the structured concept set, mapping the natural language reasoning process into a formal reasoning sequence, which is then symbolically represented as a reasoning path. A symbolic solver or theorem prover is used to formally verify the logical consistency and mathematical validity of the reasoning path. Based on the formal verification results and the prediction results of the large language model, the reasoning path is iteratively corrected through a search strategy. The output is the structured mathematical reasoning result after formal verification and iterative correction, along with its interpretable structure. This invention effectively improves the logical consistency and interpretability of the large-scale model's generated reasoning path by introducing symbolic concept alignment and formal verification mechanisms into mathematical reasoning tasks. Through the combination of variable-constraint registration and symbolic solving, a structured mapping from natural language to formal reasoning is achieved. Meanwhile, by leveraging a closed-loop iterative optimization mechanism, the system can automatically identify and correct inference biases, thereby improving the stability, reliability, and application feasibility of large models in complex mathematical reasoning scenarios. Attached Figure Description

[0020] The accompanying drawings, which form part of this application, are used to provide a further understanding of the application and to make other features, objects, and advantages of the application more apparent. The illustrative embodiments and descriptions of this application are used to explain the application and do not constitute an undue limitation of the application.

[0021] In the attached diagram: Figure 1 This is a flowchart illustrating the main steps of the large model interpretability reasoning registration method for mathematical reasoning tasks in this embodiment of the invention. Figure 2This is a flowchart illustrating the search strategy steps of the large-model interpretability reasoning registration method for mathematical reasoning tasks in an embodiment of the present invention. Figure 3 This is a connection block diagram of a large model interpretable reasoning registration system for mathematical reasoning tasks in an embodiment of the present invention. Detailed Implementation

[0022] Embodiments of this disclosure will now be described in more detail with reference to the accompanying drawings. While some embodiments of this disclosure are shown in the drawings, it should be understood that this disclosure can be implemented in various forms and should not be construed as limited to the embodiments set forth herein. Rather, these embodiments are provided to provide a more thorough and complete understanding of this disclosure. It should be understood that the accompanying drawings and embodiments of this disclosure are for illustrative purposes only and are not intended to limit the scope of protection of this disclosure.

[0023] It should also be noted that, for ease of description, only the parts relevant to the invention are shown in the accompanying drawings. Unless otherwise specified, the embodiments and features described in this disclosure can be combined with each other.

[0024] Example In this embodiment of the invention, a large-model interpretable reasoning registration system and method for mathematical reasoning tasks are disclosed. The disclosure will be described in detail below with reference to the accompanying drawings and embodiments. (Refer to...) Figure 1 As shown, the present invention mainly includes the following steps: Step 1, responding to a mathematical reasoning task in the form of natural language input; Step 2: Based on the characteristics of mathematical reasoning tasks, a set of structured concepts is obtained by symbolic extraction and formal concept alignment based on mathematical domain knowledge. Step 3: Introduce formal language into the structured concept set, map the natural language reasoning process into a formal reasoning sequence, and implement the reasoning path with symbolic representation; Step 4: Use a symbolic solver or theorem prover to perform formal verification of the logical consistency and mathematical validity of the reasoning path; Step 5: Based on the formal verification results and the prediction results of the large language model, the inference path is iteratively corrected through a search strategy; Step 6: Output the structured mathematical reasoning results and their interpretable structure after formal verification and iterative correction.

[0025] Specifically, in response to mathematical reasoning tasks input in natural language form, and based on the characteristics of mathematical reasoning tasks, a structured concept set is obtained through symbolic extraction and formal concept alignment using mathematical domain knowledge. Specifically, firstly, based on the characteristics of mathematical reasoning tasks, a structured concept set is obtained from the domain knowledge base... With the corpus of questions Extracting the mathematical entities, relations, and operators involved, a formal conceptual representation is constructed as follows: }.in, Represents a set of mathematical entities (such as variables and constants). Represents a set of mathematical relations. This represents a set of operators. Then, through symbolic extraction and semantic alignment, a mapping is achieved between natural language problems and formalized concepts, forming a structured concept representation, i.e., a structured concept set. Here, we will utilize a large language model for mathematical expression recognition, processing the input text... Perform structured parsing to generate a symbolic representation, as follows: ;in, Represented as a set of variables, Represented as a set of constants, This is represented as operations or constraint functions (such as addition, subtraction, multiplication, division, equations, logical predicates, etc.). Aligning and mapping natural language descriptions with mathematical concepts in the mathematical domain provides a solid foundation for subsequent steps through this logical representation structure. Therefore, this step ensures that the model has a unified and verifiable semantic foundation at the input stage by achieving a structured mapping between natural language issues and formalized concepts in mathematical reasoning tasks.

[0026] After obtaining the structured concept set, we introduce a formal language as an intermediate interpretation layer to make the reasoning path explicitly visible and logically verifiable. Step 3 introduces a formal language into the structured concept set, mapping the natural language reasoning process into a formal reasoning sequence, which is then symbolically represented as the reasoning path. This transforms the natural language reasoning process generated by the large language model into symbolic and operable logical expressions, thereby establishing a precise structured mapping between inputs, variables, and constraints, achieving interpretable registration of "natural language—formal language". Specifically, a formal language is introduced for the structured concept set obtained in Step 1. The formal language can be a satisfaction-based language (SMT) or a domain-specific language (DSL). This will establish a natural language inference chain. Mapped to a formal reasoning sequence Each of them This represents logical constraints, conditions, or intermediate conclusions. This formalized reasoning sequence is the reasoning path, achieving the goal of symbolically representing the reasoning path. This formalized reasoning sequence can be further modeled as a constraint system, and this constraint system can be represented as... .in, V The set of variables involved in the constraint system. CThis represents all constraints involved in the constraint system. Constraints include partial equality constraints and inequality constraints. The former involves various assignment operations, while the latter consists of various constraints containing knowledge of the real world (such as the number of items needing to be greater than 0).

[0027] Next, the reasoning path is checked for logical consistency and mathematical validity to ensure the correctness and verifiability of the reasoning process. Specifically, step 4 involves using a symbolic solver or theorem prover to formally verify the logical consistency and mathematical validity of the reasoning path. By introducing a symbolic solver, the reasoning chain generated by the large language model is verified and constrained at the formal logic level, ensuring that the model's reasoning output meets interpretability and consistency requirements in semantics, structure, and numerical values. The symbolic solver can use formal tools such as Z3 and SMT solvers to verify the correctness of the algebraic equations, logical conditions, and numerical calculations involved in the reasoning. Formal verification is then performed on the formal reasoning sequence (i.e., the constraint system) generated in step 3. This symbolic solver supports global constraint consistency checks on the entire reasoning chain, represented as... If a logical contradiction is detected (such as mutually exclusive conditions, circular dependencies, or computational errors), the system will generate a correction signal and feed it back for iterative correction. This is used to locally correct or regenerate the inference path, thus forming a self-correcting closed loop. This step not only provides feedback for inference path correction but also provides quantitative support for the overall interpretability evaluation of the system, enabling the inference results to be traceable, verifiable, and optimizable.

[0028] Step 5 involves iteratively refining the inference path based on the verification results and prediction results of the large language model, using a search strategy. Specifically, heuristic search or Monte Carlo search strategies are employed to simulate and analyze variable distribution and constraint combinations within the symbolic space. The inference path and model generation strategy are dynamically adjusted based on the verification results, achieving gradual alignment between the neural network output and the symbolic system logic. Since directly generating complete variable-constraint mappings in a high-dimensional constraint space often leads to an excessively large search space or unstable solutions, this invention introduces an iterative optimization mechanism based on variable queues and constraint pools into the symbolic space. This mechanism allows for a phased and controllable completion of the inference registration process, addressing the challenge of simultaneously ensuring variable-constraint consistency and inference correctness in a single inference registration process for large language models. (Refer to...) Figure 2 As shown. During system operation, two core data structures are maintained: a variable queue and a variable queue. and constraint pool The variable queue stores target variables to be verified or optimized, while the constraint pool stores logical constraints, equations, or conditional expressions related to the current inference task. The process also includes the following steps: Step 5.1: Obtain the variable to be processed from the variable queue and retrieve the set of constraints associated with it from the constraint pool; Step 5.2: Evaluate the current state using a large language model and introduce a candidate set of real-world knowledge generation constraints; Step 5.3: Select a new constraint set by filtering the candidate constraint set based on confidence or consistency score, and replace the constraint system obtained in step 3 with the new constraint set to form a new constraint system; Step 5.4 introduces a large language model to evaluate the original constraint system and the new constraint system, and verifies them using the results of the symbolic validator. Step 5.5: Select the better-performing constraint system as the new initial state and the starting input for the next iteration.

[0029] Specifically, the above steps can be divided into three phases: preparation, exploration, and verification. During the preparation phase, the variables are retrieved from the queue... A variable to be processed pops up at the head of the queue. And retrieve the set of constraints associated with it in the constraint pool, denoted as .in, Representing constraints The set of variables involved. This stage completes the extraction and initialization of variable-constraint pairs, providing input for subsequent optimization. During the exploration phase, a large language model is used. For the current state Conduct an assessment and incorporate real-world knowledge. Generate a new set of candidate constraints And filter based on confidence level or consistency score, The new constraint set Replace the original constraint system to form a new constraint system. Listed as an alternative This will be verified in the next section. During the verification phase, a large language model is introduced to evaluate the original and new constraint systems, and the results from the symbolic verifier are used for verification. The constraint system with the better performance is selected as the new initial state and the starting input for the next iteration. Furthermore, the system first calls the symbolic verifier to perform logical consistency and satisfiability checks on both sets of constraint systems to obtain the verification result set, represented as follows: ; Subsequently, the evaluation module of the large language model calculates a comprehensive evaluation score based on the symbolic verification results and the inference objective function. Finally, the system selects the better-performing constraint system as the new initial state based on the comprehensive score and uses it as the starting input for the next iteration, thereby achieving dynamic optimization and steady-state convergence of the inference registration process. This design enables the system to form a closed loop between symbolic verification and the large language model evaluation, ensuring that each iteration evolves towards logical consistency and optimal results.

[0030] Step 6 outputs the structured mathematical reasoning results and their interpretable structure after formal verification and iterative correction. Specifically, after logical consistency checks and numerical validity verification, structured mathematical reasoning results and their interpretable reasoning chain representations are generated and output. The variable-constraint mapping relationships, reasoning steps, and intermediate conclusions are displayed in a structured form, facilitating user understanding of the reasoning logic and verification process, and achieving traceability and transparency of the reasoning results. By integrating intermediate results from the structured concept set and the verified reasoning path, the reasoning process is presented transparently and logically traceable through variable-constraint mapping tables, reasoning graph structures, and visual representations. This step not only outputs the final numerical solution (such as the problem-solving results)... It also outputs the symbolic logic basis of the reasoning and the verification certificate (such as the satisfiability proof of the SMT solver), thereby achieving the unity of formal verifiability and interpretability of the reasoning results.

[0031] This invention aims to address the problems of black-box, unverifiable reasoning, lack of interpretable reasoning mechanisms, and instability in open environments in current large language models (LLMs) for mathematical reasoning. It seeks to solve the key technical challenges in transforming the reasoning capabilities of large models from "usable" to "reliable and deployable." This method proposes a variable-constraint registration search framework, using formal language as an intermediate interpretation layer. It achieves dual alignment between model semantics and mathematical knowledge through concept registration and reasoning registration modules. The system verifies the logical consistency of intermediate hypotheses and constraint relationships generated by the model through a symbolic solver and continuously corrects the model's reasoning path through an iterative feedback mechanism, thereby achieving transparency, verifiability, and knowledge consistency in the reasoning process. This invention can improve the stability and reliability of large model reasoning results in complex reasoning scenarios and can be widely applied to real-world applications with complex mathematical reasoning requirements, including but not limited to reasoning enhancement and interpretability analysis of large models in vertical fields such as mathematical reasoning, symbolic computation, and financial reasoning.

[0032] To achieve the above embodiments, this application also proposes a large-model interpretable reasoning registration system for mathematical reasoning tasks, referring to... Figure 3As shown, the system includes a concept registration module, a symbolic reasoning registration module, a formal verification module, an iterative update control module, and an interpretable reasoning output terminal module. The concept registration module is used to extract symbolic meaning and align formal concepts based on mathematical knowledge to obtain a structured concept set. The symbolic reasoning registration module introduces formal language to map the natural language reasoning process into a formal reasoning sequence, which is then represented symbolically as a reasoning path. The formal verification module uses a symbolic solver or theorem prover to perform formal verification of the logical consistency and mathematical validity of the reasoning path. The iterative update control module uses a search strategy to iteratively correct the reasoning path based on the formal verification results and the prediction results of a large language model. The interpretable reasoning output terminal module outputs the verified and corrected reasoning results and their interpretive structure.

[0033] Furthermore, the concept registration module includes a symbolic extraction unit and a semantic alignment unit, used to map mathematical entities, relations, and operators appearing in natural language problems into formal expressions, and to achieve precise alignment with domain concepts through a large language model and structural constraints. The symbolic reasoning registration module includes a hypothesis generation unit and a variable constraint search unit. The hypothesis generation unit extracts candidate reasoning hypotheses from intermediate expressions generated by the large model, while the variable-constraint search unit constructs an interpretable reasoning path structure by traversing variable combinations and constraints. The formal verification module includes a symbolic solution unit and a consistency determination unit, used to verify the mathematical validity and constraint consistency of the reasoning path based on a domain symbolic system (such as an SMT solver). If a logical contradiction is detected, a correction signal is fed back to the iterative update module. The iterative update control module includes a hypothesis screening unit, a conflict detection unit, and a feedback optimization unit. It uses heuristic search or Monte Carlo methods to simulate and analyze variable distribution and constraint combinations in the symbolic space, dynamically correcting the reasoning path and model generation strategy based on the verification results, achieving gradual alignment between the neural network output and the symbolic system logic.

[0034] To implement the above embodiments, this application also discloses an electronic device. The electronic device may include a processing unit (e.g., a central processing unit, a graphics processing unit, etc.) that can perform various appropriate actions and processes according to a program stored in a read-only memory (ROM) or a program loaded from a storage device into a random access memory (RAM). Various programs and data required for the operation of the electronic device are also stored in the RAM. The processing unit, ROM, and RAM are interconnected via a bus. An input / output (I / O) interface is also connected to the bus. Typically, the following devices can be connected to the I / O interface: input devices including, for example, a touchscreen, touchpad, keyboard, mouse, camera, microphone, accelerometer, gyroscope, etc.; output devices including, for example, a liquid crystal display (LCD), speaker, vibrator, etc.; storage devices including, for example, magnetic tape, hard disk, etc.; and communication devices. The communication device allows the electronic device to communicate wirelessly or wiredly with other devices to exchange data. Although electronic devices with various devices are shown, it should be understood that it is not required to implement or possess all of the shown devices. More or fewer devices may be implemented or possessed alternatively.

[0035] In particular, according to some embodiments of this disclosure, the processes described above with reference to the flowcharts can be implemented as computer software programs. For example, some embodiments of this disclosure include a computer program product comprising a computer program carried on a computer storage medium, the computer program containing program code for performing the methods shown in the flowcharts. In such embodiments, the computer program can be downloaded and installed from a network via a communication device, or installed from a storage device, or installed from a ROM. When the computer program is executed by a processing device, it performs the functions defined above in the methods of some embodiments of this disclosure.

[0036] It should be noted that the computer storage medium in some embodiments of this disclosure may be a computer-readable signal medium or a computer-readable storage medium, or any combination thereof. A computer-readable storage medium may be, for example, but not limited to, an electrical, magnetic, optical, electromagnetic, infrared, or semiconductor system, apparatus, or device, or any combination thereof. More specific examples of a computer-readable storage medium may include, but are not limited to: an electrical connection having one or more wires, a portable computer disk, a hard disk, random access memory (RAM), read-only memory (ROM), erasable programmable read-only memory (EPROM or flash memory), optical fiber, portable compact disk read-only memory (CD-ROM), optical storage device, magnetic storage device, or any suitable combination thereof.

[0037] In some embodiments of this disclosure, a computer-readable storage medium can be any tangible medium that contains or stores a program that can be used by or in connection with an instruction execution system, apparatus, or device. In some embodiments of this disclosure, a computer-readable signal medium can include a data signal propagated in baseband or as part of a carrier wave, carrying computer-readable program code. Such propagated data signals can take various forms, including but not limited to electromagnetic signals, optical signals, or any suitable combination thereof. A computer-readable signal medium can also be any computer storage medium other than a computer-readable storage medium, which can send, propagate, or transmit a program for use by or in connection with an instruction execution system, apparatus, or device. The program code contained on the computer storage medium can be transmitted using any suitable medium, including but not limited to: wires, optical fibers, RF (radio frequency), etc., or any suitable combination thereof.

[0038] In other implementations, clients and servers may communicate using any currently known or future-developed network protocol such as HTTP (Hypertext Transfer Protocol), and may interconnect with digital data communication (e.g., communication networks) of any form or medium. Examples of communication networks include local area networks (“LANs”), wide area networks (“WANs”), the Internet (e.g., the Internet of Things), and end-to-end networks (e.g., ad hoc end-to-end networks), as well as any currently known or future-developed networks.

[0039] The aforementioned computer storage medium may be included in the aforementioned electronic device, or it may exist independently and not assembled into the electronic device. The aforementioned computer storage medium carries one or more programs that, when executed by the electronic device, enable the electronic device to implement a large-model interpretable reasoning registration method for mathematical reasoning tasks.

[0040] Computer program code for performing operations of some embodiments of this disclosure can be written in one or more programming languages ​​or a combination thereof, including object-oriented programming languages ​​such as Java, Smalltalk, and C++, as well as conventional procedural programming languages ​​such as the "C" language or similar programming languages. The program code can be executed entirely on the user's computer, partially on the user's computer, as a standalone software package, partially on the user's computer and partially on a remote computer, or entirely on a remote computer or server. In cases involving remote computers, the remote computer can be connected to the user's computer via any type of network—including a local area network (LAN) or a wide area network (WAN)—or can be connected to an external computer (e.g., via the Internet using an Internet service provider).

[0041] The flowcharts and block diagrams in the accompanying drawings illustrate the architecture, functionality, and operation of possible implementations of systems, methods, and computer program products according to various embodiments of the present disclosure. In this regard, each block in a flowchart or block diagram may represent a module, segment, or portion of code containing one or more executable instructions for implementing a specified logical function. It should also be noted that in some alternative implementations, the functions indicated in the blocks may occur in a different order than those indicated in the drawings. Each block in the block diagrams and / or flowcharts, and combinations of blocks in the block diagrams and / or flowcharts, may be implemented using a dedicated hardware-based system that performs the specified function or operation, or using a combination of dedicated hardware and computer instructions. Units described in some embodiments of the present disclosure may be implemented in software or hardware. The described units may also be located in a processor, and the names of these units do not necessarily constitute a limitation on the unit itself.

[0042] The functions described above in this document can be performed, at least in part, by one or more hardware logic components. For example, exemplary types of hardware logic components that can be used, without limitation, include: Field Programmable Gate Arrays (FPGAs), Application-Specific Integrated Circuits (ASICs), Application Standard Products (ASSPs), System-on-Chip (SoCs), Complex Programmable Logic Devices (CPLDs), and so on.

[0043] All technologies not described in detail in this invention are existing technologies. The above descriptions are merely some preferred embodiments of this disclosure and explanations of the technical principles employed. Those skilled in the art should understand that the scope of the invention involved in the embodiments of this disclosure is not limited to technical solutions formed by specific combinations of the above-described technical features, but should also cover other technical solutions formed by arbitrary combinations of the above-described technical features or their equivalent features without departing from the above-described inventive concept. For example, technical solutions formed by substituting the above-described features with (but not limited to) technical features with similar functions disclosed in the embodiments of this disclosure.

Claims

1. A method for interpretable reasoning registration of large models for mathematical reasoning tasks, characterized in that, Includes the following steps: Step 1, responding to a mathematical reasoning task in the form of natural language input; Step 2: Based on the characteristics of the mathematical reasoning task, a structured concept set is obtained by symbolic extraction and formal concept alignment based on mathematical domain knowledge. Step 3: Introduce a formal language into the structured concept set, map the natural language reasoning process into a formal reasoning sequence, and implement the reasoning path using symbolic representation; Step 4: Use a symbolic solver or theorem prover to perform formal verification of the logical consistency and mathematical validity of the reasoning path; Step 5: Based on the formal verification results and the prediction results of the large language model, the inference path is iteratively corrected through a search strategy; Step 6: Output the structured mathematical reasoning results and their interpretable structure after formal verification and iterative correction.

2. The large-model interpretability reasoning registration method for mathematical reasoning tasks according to claim 1, characterized in that: In step 2, based on the characteristics of the mathematical reasoning task, mathematical entities, relations and operators involved are extracted from the domain knowledge base and the question corpus to construct formal concept representations. Then, through symbolic extraction and semantic alignment, the alignment mapping between natural language descriptions and formal concepts is formed to create the structured concept set.

3. The large-model interpretability reasoning registration method for mathematical reasoning tasks according to claim 2, characterized in that: In step 3, the formal reasoning sequence can be modeled as a constraint system. The constraints involved in the constraint system include partial equality constraints and inequality constraints. The partial equality constraints involve various assignment operations, while the inequality constraints consist of various constraints containing knowledge of the real world.

4. The large-model interpretability reasoning registration method for mathematical reasoning tasks according to claim 3, characterized in that: In step 4, the correctness of the algebraic equations, logical conditions and numerical calculations involved in the reasoning are first verified based on the symbolic solver or theorem prover. Then, the global constraint consistency check is performed on the reasoning path. If a logical contradiction or calculation error is detected, a correction signal is generated and fed back for iterative correction.

5. The large-model interpretability reasoning registration method for mathematical reasoning tasks according to claim 4, characterized in that: In step 5, heuristic search or Monte Carlo search strategies are used to simulate and analyze the distribution of variables and combination of constraints in the symbolic space. Based on the results of the formal verification, the inference path and model generation strategy are dynamically corrected to achieve gradual alignment between the neural network output and the symbolic system logic.

6. The large-model interpretability reasoning registration method for mathematical reasoning tasks according to claim 5, characterized in that: In step 6, by integrating the intermediate results from the structured concept set and the reasoning path after formal verification, and utilizing variable-constraint mapping tables, reasoning graph structures, and visual representations, the reasoning process is made transparent and logically traceable. Not only is the final numerical solution output, but also the symbolic logic basis of the reasoning and the verification certificate are also output.

7. The large-model interpretability reasoning registration method for mathematical reasoning tasks according to claim 5, characterized in that: In step 5, the heuristic search employs an iterative optimization mechanism based on variable queues and constraint pools in the symbolic space, and also includes the following steps: Step 5.1: Obtain the variable to be processed from the variable queue and retrieve the set of constraints associated with it from the constraint pool; Step 5.2: Evaluate the current state using a large language model and introduce a candidate set of real-world knowledge generation constraints; Step 5.3: Filter the candidate constraint set based on confidence level or consistency score to obtain a new constraint set, and replace the constraint system obtained in step 3 with the new constraint set to form a new constraint system; Step 5.4: Introduce a large language model to evaluate the original constraint system and the new constraint system, and verify them using the results of the symbolic verifier. Step 5.5: Select the better-performing constraint system as the new initial state and the starting input for the next iteration.

8. A large-scale model interpretability reasoning registration system for mathematical reasoning tasks, characterized in that, include: The concept registration module, including a symbol extraction unit and a semantic alignment unit, is used to extract symbols and align formal concepts based on mathematical knowledge to obtain a structured concept set. The symbolic reasoning registration module, including the hypothesis generation unit and the variable constraint search unit, is used to introduce formal language to map the natural language reasoning process into a formal reasoning sequence, and then realize it as a reasoning path using symbolic representation; The formal verification module includes a symbolic solver unit and a consistency determination unit, which are used to perform formal verification of the logical consistency and mathematical validity of the reasoning path using a symbolic solver or a theorem prover. The iterative update control module includes a hypothesis screening unit, a conflict detection unit, and a feedback optimization unit, which is used to iteratively correct the inference path based on the formal verification results and the prediction results of the large language model through a search strategy. The interpretable reasoning output terminal module is used to output the verified and corrected reasoning results and their interpretation structure.

9. An electronic device, characterized in that, include: At least one processor, and a memory communicatively connected to said at least one processor; The memory stores instructions that can be executed by the at least one processor to enable the at least one processor to perform the steps of the method according to any one of claims 1 to 7.

10. A computer storage medium storing a computer program thereon, characterized in that: When the computer program is executed by the processor, it performs the steps as described in any one of claims 1 to 7.

Images