Paper influence evolution trajectory intelligent agent prediction method and system
By combining large language models with external academic information interfaces and reinforcement learning, an interpretable paper impact prediction method is constructed. This method solves the problems of uninterpretability and weak supervision in existing technologies, achieves highly reliable prediction results, and is applicable to scientific research evaluation and resource allocation.
Patent Information
- Authority / Receiving Office
- CN · China
- Patent Type
- Applications(China)
- Current Assignee / Owner
- ZHEJIANG UNIV CITY COLLEGE
- Filing Date
- 2026-01-30
- Publication Date
- 2026-05-08
AI Technical Summary
Existing methods for predicting paper impact suffer from problems such as unexplainable reasoning processes, limited learning capacity under weak supervision, and a lack of feedback-based dynamic reasoning strategy optimization mechanisms.
By employing a large language model and an interface with external academic information, and through multi-round reasoning and reward signal construction, reinforcement learning is used to optimize the model's reasoning path, thus forming an interpretable prediction process.
It achieves highly reliable and interpretable prediction of paper impact under weak supervision, improving the credibility and generalization ability of the prediction results, and is suitable for high-reliability scenarios such as scientific research evaluation and resource allocation.
Smart Images
Figure CN121997050A_ABST
Abstract
Description
Technical Field
[0001] This application relates to the field of artificial intelligence technology, specifically a method and system for predicting the evolution trajectory of paper influence in intelligent agents. Background Technology
[0002] Paper impact prediction is of great significance for scientific research evaluation, academic recommendation, and resource allocation. With the increasing complexity of academic networks, research methods in this field have undergone the following evolution and have corresponding limitations: 1. Limitations of traditional statistical and deep learning methods Early methods based on statistical models (such as regression and time series) had simple structures but struggled to characterize the complex nonlinear relationships within academic networks. Subsequent methods based on deep learning, particularly graph neural networks (such as GNN and R-GCN), improved prediction performance by vectorizing papers and related entities through end-to-end learning. However, their decision-making process relied on high-dimensional implicit vectors, resulting in completely uninterpretable prediction logic, untraceable judgment criteria, and difficulty in applying them to high-reliability scenarios.
[0003] 2. Existing advanced methods and their inherent problems Recent research has proposed more complex graph neural network models to improve prediction accuracy, such as: The HINTS method predicts future citations of papers by embedding dynamic heterogeneous information networks with RNNs, but its inference depends on the complete network embedding trajectory.
[0004] The H2CGL method constructs a hierarchical heterogeneous graph and employs contrastive learning to improve the sensitivity of the representation to references.
[0005] However, these methods generally suffer from three core flaws: Lack of interpretability: The decision-making process is a black box, unable to provide traceable and auditable reasoning.
[0006] Poor adaptability under weak supervision: It relies heavily on complete, long-term time-series labeled data (such as annual citations), while newly published papers usually only have short-term or single-point cumulative citation data, and the model's learning ability is limited under weak supervision.
[0007] Static and fixed inference strategies: Once the model is trained, the strategy is fixed and cannot be dynamically adjusted and optimized based on the feedback of the prediction results. It lacks iterative cognition and self-improvement capabilities.
[0008] 3. Opportunities and Challenges of Introducing Large Language Models Large language models have shown great potential in natural language understanding and reasoning, and some studies have attempted to use them for predicting the impact of academic papers. Although they enhance the utilization of textual semantics, their reasoning process remains implicit within the model, lacking explicit structured constraints and external evidence annotations, resulting in unstable reasoning paths and insufficient interpretability and reliability.
[0009] In summary, existing technologies have not yet systematically solved the three key problems of constructing interpretable reasoning chains, effective learning under weak supervision, and feedback-driven dynamic policy optimization. Therefore, there is an urgent need for a new paradigm that can integrate the reasoning capabilities of large language models with reinforcement learning optimization mechanisms to achieve a highly transparent, reliable, and continuously evolving method for predicting paper impact. Summary of the Invention
[0010] This invention addresses the problems of unexplainable reasoning processes, limited learning capacity under weak supervision, and lack of feedback-based dynamic reasoning strategy optimization mechanisms in paper influence prediction tasks.
[0011] To address the aforementioned issues, this invention provides a method, system, electronic device, and storage medium for predicting the evolution trajectory of a paper's influence.
[0012] In a first aspect, the present invention provides a method for predicting the evolution trajectory of paper influence using an intelligent agent, comprising the following steps: S1. Construct a training dataset containing the target papers and their true influence labels; S2. For the target paper in the training dataset, construct a large language model inference input and configure an external academic information interface set. The large language model performs multiple rounds of inference and external information retrieval process. When the termination condition is met, the predicted influence value of the target paper is output. S3. Construct a reward signal based on the deviation between the true influence value and the predicted influence value of the target paper; S4. With the goal of maximizing the expected reward on the training dataset, the reasoning strategy of the large language model in the multi-round reasoning and information retrieval process is optimized and updated using reinforcement learning methods, and the trained model parameters are output.
[0013] Optionally, constructing the training dataset in step S1 specifically includes: S11. Construct an academic database based on academic network data in the target research field, wherein the academic database contains at least paper meta-information and relational data; S12. Set the length of the time window for impact statistics. For each target paper, calculate the cumulative number of citations within the time window after publication, and use it as the true impact value of the target paper. S13. Divide the labeled sample set into training set, validation set and test set, and construct the training dataset representation.
[0014] Optionally, the set of external academic information interfaces mentioned in step S2 includes at least: The paper metadata query interface is used to query the title, publication year, author list, and publication medium information of a paper based on its identifier. The citation relationship query interface is used to query the reference set or citation relationship of a paper based on the paper identifier; The author or publication carrier statistics interface is used to query the corresponding list of historical papers or historical citation statistics based on the author identifier or publication carrier identifier. The similar paper retrieval interface is used to retrieve and return a collection of similar papers based on paper identifiers or paper text descriptions.
[0015] Optionally, the multi-round reasoning and external information retrieval process described in step S2 specifically includes: Under the constraint of the maximum number of inference rounds T, the large language model generates new external information call requests or intermediate inference conclusions round by round based on the initial input and the external information call results of the historical rounds. When the large language model outputs a termination command or reaches the maximum number of inference rounds T, the inference process ends, and a predicted influence value of the target paper is generated based on the current inference context.
[0016] Optionally, the construction of the reward signal based on the deviation in step S3 specifically includes: Based on predicted influence values Compared with the true influence label value Construct the reward function ,in Used to measure how close the predicted result is to the true value; The reward signal is subjected to numerical constraints or normalization to ensure it meets a preset value range. .
[0017] Optionally, the reward function is defined as: .
[0018] Optionally, the optimization update using reinforcement learning methods described in step S4 has the following optimization objective: ,in The parameters of the large language model are represented. Represents the training dataset. This indicates that samples were randomly taken from the training set during training. The corresponding reward signal Expected value Indicates the target paper, This indicates the corresponding influence rating.
[0019] Optionally, the termination condition includes: reaching a preset maximum number of training rounds, or the average prediction error or expected reward on the validation dataset no longer improving over multiple consecutive training rounds.
[0020] Secondly, embodiments of the present invention provide an intelligent agent prediction system for the evolution trajectory of paper influence, comprising: The data processing module is used to construct a training dataset containing the target papers and their true impact labels. The reasoning and interaction module, including a large language model and a set of external academic information interfaces, is used to perform multiple rounds of reasoning and external information retrieval for the target paper and output the predicted influence value. The reward calculation module is used to construct a reward signal based on the deviation between the actual influence value and the predicted influence value of the target paper. The strategy optimization module is used to optimize and update the reasoning strategy of the large language model using reinforcement learning methods with the goal of maximizing expected reward.
[0021] Thirdly, embodiments of the present invention provide an electronic device, including a processor, a communication interface, a memory, and a bus, wherein the processor, the communication interface, and the memory communicate with each other through the bus, and the processor can call logical instructions in the memory to execute the steps of the method provided in the first aspect.
[0022] Fourthly, embodiments of the present invention provide a non-transitory computer-readable storage medium having a computer program stored thereon, which, when executed by a processor, implements the steps of the paper influence evolution trajectory intelligent agent prediction method as described in the first aspect.
[0023] The beneficial effects of the paper influence evolution trajectory intelligent agent prediction method of the present invention are as follows: This method addresses the shortcomings of existing technologies in predicting paper impact, such as uninterpretable reasoning processes, limited learning capabilities under weak supervision, and a lack of feedback-based dynamic reasoning strategy optimization mechanisms. It models impact prediction as a traceable and interventionizable multi-round reasoning process. By constructing an interaction mechanism between a large language model agent and external academic information interfaces, and introducing reward signals based on the consistency between predicted results and actual annotations, reinforcement learning is used to optimize the model's reasoning path and information retrieval strategy end-to-end. This method can guide the model to gradually form reasonable, stable, and interpretable reasoning logic under weak supervision conditions that rely solely on single-point or finite-time-window impact annotations. This effectively improves the credibility, generalization ability, and decision transparency of prediction results, thus better serving high-reliability application scenarios such as scientific research evaluation, academic recommendation, and resource allocation. Attached Figure Description
[0024] Figure 1 This is a flowchart of the paper influence evolution trajectory intelligent agent prediction method in an embodiment of the present invention; Figure 2 This is a schematic diagram of the paper influence evolution trajectory intelligent agent prediction method in an embodiment of the present invention; Figure 3 This is a flowchart illustrating the construction of the training dataset in an embodiment of the present invention; Figure 4 This is a structural block diagram of the intelligent agent prediction system for the evolution trajectory of paper influence in an embodiment of the present invention; Figure 5 This is a structural block diagram of the electronic device in an embodiment of the present invention. Detailed Implementation
[0025] To better understand the purpose, technical solution, and advantages of this application, the application is described and explained below in conjunction with the accompanying drawings and embodiments.
[0026] Unless otherwise defined, the technical or scientific terms used in this application shall have the general meaning understood by one of ordinary skill in the art to which this application pertains. Words such as “a,” “an,” “an,” “the,” “the,” and “these” used in this application do not indicate quantitative limitation and may be singular or plural. The terms “comprising,” “including,” “having,” and any variations thereof used in this application are intended to cover non-exclusive inclusion; for example, a process, method, system, product, or device that comprises a series of steps or modules (units) is not limited to the listed steps or modules (units) but may include steps or modules (units) not listed, or may include other steps or modules (units) inherent to these processes, methods, products, or devices. Words such as “connected,” “linked,” and “coupled” used in this application are not limited to physical or mechanical connections but may include electrical connections, whether direct or indirect. “Multiple” used in this application refers to two or more. “And / or” describes the relationship between related objects, indicating that three relationships may exist; for example, “A and / or B” can represent: A alone, A and B simultaneously, and B alone. Normally, the character " / " indicates that the objects before and after it are in an "or" relationship. The terms "first," "second," "third," etc., used in this application are merely to distinguish similar objects and do not represent a specific order of objects.
[0027] This method addresses the shortcomings of existing technologies in predicting paper impact, such as uninterpretable reasoning processes, limited learning capabilities under weak supervision, and a lack of feedback-based dynamic reasoning strategy optimization mechanisms. It models impact prediction as a traceable and interventionizable multi-round reasoning process. By constructing an interaction mechanism between a large language model agent and external academic information interfaces, and introducing a reward signal based on the consistency between predicted results and actual annotations, reinforcement learning is used to optimize the model's reasoning path and information retrieval strategy end-to-end. This method can guide the model to gradually form reasonable, stable, and interpretable reasoning logic under weak supervision conditions that rely solely on single-point or finite-time-window impact annotations. This effectively improves the credibility, generalization ability, and decision transparency of prediction results, thus better serving high-reliability application scenarios such as scientific research evaluation, academic recommendation, and resource allocation.
[0028] To address the problems of uninterpretable reasoning processes, limited learning capabilities in weakly supervised scenarios, and the lack of feedback-driven dynamic reasoning strategy optimization mechanisms in existing paper influence prediction methods, this application proposes an intelligent agent prediction method for the evolution trajectory of paper influence. The core technical solution is as follows: 1. Explainable multi-turn reasoning and structured external information retrieval mechanism: Influence prediction is modeled as a multi-turn natural language reasoning process supported by a large language model agent and predefined academic information interfaces (such as paper metadata, citation relationships, author / conference statistics, and similar paper retrieval). Each step of the reasoning explicitly generates query intent or intermediate conclusions, forming a traceable and auditable reasoning chain, significantly improving the transparency and credibility of the prediction results.
[0029] 2. Reward signal construction and normalization method for single-point weak supervision annotation: Relying only on the cumulative citations of the target paper within a fixed time window (such as 5 years after publication) as the true influence annotation, a sparse reward function based on prediction error is constructed, and it is normalized to a uniform interval through numerical constraints to ensure that the model can still be stably and fairly guided to learn under weak supervision conditions where there is a lack of complete evolutionary trajectory.
[0030] 3. A Dynamic Reasoning Strategy Optimization Framework Based on Reinforcement Learning: Aiming to maximize expected cumulative reward, this framework utilizes reinforcement learning to optimize the information retrieval order, evidence integration method, and termination decision of a large language model in multi-round reasoning. This enables the model to continuously improve its reasoning strategy based on the deviation between predicted results and actual feedback, achieving human-like expert-level iterative cognition and adaptive decision-making capabilities. The specific methods are detailed below.
[0031] like Figure 1 and Figure 2 As shown in the figure, the method for predicting the evolution trajectory of paper influence provided in this embodiment of the invention includes the following steps: S1. Construct a training dataset containing the target papers and their true influence labels; In this embodiment, as Figure 3 As shown, step S1 involves constructing the training dataset, including: S11. Construct an academic database based on academic network data in the target research field. The academic database shall contain at least paper meta-information and relational data. S12. Set the length of the time window for impact statistics. For each target paper, calculate the cumulative number of citations within the time window after publication, and use it as the true impact value of the target paper. S13. Divide the labeled sample set into training set, validation set and test set, and construct the training dataset representation.
[0032] Specifically, an academic database is constructed based on academic network data in the target research field, and the true influence label value of the target papers is calculated within a preset time window based on paper citation relationships. After labeling, the dataset is divided and a training sample representation is constructed to obtain a training dataset containing the target papers and their influence label values. The specific process is as follows: Data source identification: An academic database is constructed based on academic network data in a specific research field. The paper metadata in the academic database includes: paper identifier, title, publication year, publication medium identifier, author identifier, and abstract. Relational data in the academic database includes: "Author-written papers," "Papers cited by the paper," and "Papers published in the publication medium."
[0033] Influence Labeling Construction: Based on the requirements of the paper impact prediction task, a 5-year time window is set for impact statistics. Starting from the publication date of the target paper, the cumulative number of citations it receives within 5 years of publication is counted as the true impact value of the target paper. Specifically, based on paper citation relationship data, for each target paper... The number of citations generated for the paper within the stated time window is counted to obtain the corresponding influence rating. In the absence of precise citation timestamps, the publication year of the cited paper is used as an approximation of the citation time, and citation records that meet the time window constraints are included in the influence label value.
[0034] Dataset partitioning and training sample construction: After constructing the impact rating values for all target papers, 500 samples were randomly selected from the sample set as the training set, 500 samples as the validation set, and 500 samples as the test set.
[0035] After completing the dataset partitioning, construct the training sample representation. Let the training dataset be: ; in, Indicates the first The target paper can be represented by a unique identifier or a textual description of the paper. This represents the influence rating value corresponding to the target paper.
[0036] S2. For the target paper in the training dataset, construct a large language model for inference input and configure a set of external academic information interfaces. The large language model performs multiple rounds of inference and external information retrieval process. When the termination condition is met, the predicted influence value of the target paper is output. Specifically, for the target papers in the training dataset, a large language model is constructed as the inference input, and a set of external academic information interfaces is configured. During the training phase, the large language model performs multiple rounds of inference and external information retrieval. When the termination condition is met, the predicted influence value of the target paper is output as the basis for subsequent reward signal calculation and optimization updates. The specific process is as follows: Inference Input Construction and External Academic Information Interface Configuration For each target paper Input for constructing a large language model The inputs include at least: The target paper's paper identifier or text description; the task instruction for directing the model to perform the paper impact prediction task (requiring the model to output the predicted paper impact value); and a description of an external academic information interface that can be invoked, wherein the interface description includes the interface name, input parameter types, and return field descriptions.
[0037] To ensure the reproducibility of the reasoning process, a set of external academic information interfaces is predefined, and each interface has defined input and output fields. The set of external academic information interfaces includes at least: The paper metadata query interface is used to query the title, publication year, author list, and publication medium information of a paper based on the paper identifier (i.e., input the paper identifier and output the paper title, publication year, author list, and publication medium information).
[0038] The citation relationship query interface is used to query the reference set or citation relationship of a paper based on the paper identifier (i.e., input the paper identifier and output the reference set or citation relationship edge set of the paper).
[0039] The author or publication carrier statistics interface is used to query the corresponding list of historical papers or historical citation statistics based on the author identifier or publication carrier identifier (i.e., input the author identifier or publication carrier identifier, and output the corresponding list of historical papers or historical citation statistics).
[0040] The similar paper retrieval interface is used to retrieve and return a set of similar papers based on paper identifiers or paper text descriptions (i.e., input a paper identifier or paper text description, output a set of papers similar to the target paper, and the set of similar papers may contain similarity information).
[0041] Multi-round inference execution and prediction result generation During the training phase, based on the constructed input For each target paper, multiple rounds of reasoning and external information retrieval are performed. Specifically: Under the constraint of the maximum number of inference rounds T, the large language model generates new external information call requests or intermediate inference conclusions round by round based on the initial input and the external information call results from the previous rounds. When the large language model outputs a termination instruction or reaches the maximum number of inference rounds T, the inference process ends, and a predicted influence value for the target paper is generated based on the current inference context.
[0042] For example, the maximum number of rounds of multi-round inference is T, where T=10. In the first round of inference, the large language model bases its inference on the input... The first external information retrieval request or intermediate inference conclusion is generated. In the t-th round of inference, the large language model continues to generate new retrieval requests or determines whether to terminate the inference process based on the return results of external information retrieval in previous rounds. The multi-round inference process ends when the model outputs a termination command or reaches the maximum number of inference rounds T.
[0043] After the inference terminates, the large language model generates a predicted impact value for the target paper based on the current inference context. This information is then used for subsequent reward signal calculation and reinforcement learning optimization of model parameters.
[0044] In some implementations, the same inference process can be performed multiple times on the same target paper under the same input conditions to generate multiple sets of predicted influence values, thereby enhancing the stability of the reinforcement learning training process.
[0045] S3. Construct a reward signal based on the deviation between the true influence value and the predicted influence value of the target paper; A reward function is constructed based on the deviation between the actual impact value of the target paper and the model's predicted impact value. The corresponding reward signal is then obtained, and numerical constraints or normalization are applied to the reward signal to ensure comparability and training stability among different samples. The specific process is as follows: Reward signals are constructed based on true influence values: Based on the target paper's true influence label value obtained in step S1 and the predicted influence results obtained in step S2. A reward signal based on the true value is constructed. Specifically, the reward function is calculated based on the deviation between the predicted influence result and the actual influence label value: ; in, This is a reward function based on the true value, used to measure the closeness between the predicted result and the true influence label value.
[0046] In this embodiment, the reward function is defined as: .
[0047] The reward function uses the absolute value of the references predicted by the large language model agent and the true value as a signal to guide the optimization of the large language model agent. If the prediction is inaccurate, then... Low, if the prediction is accurate. high, The highest value is 0.
[0048] The closer the prediction results are to the true influence label value, the higher the reward will be.
[0049] To ensure the comparability and numerical stability of reward signals among different training samples, the reward signals are subjected to numerical constraints and normalization to ensure that the reward values meet a preset range. ; in, and This is a preset constant. By fixing the range of reward values, instability in the training process due to differences in the scale of the actual influence labels is avoided.
[0050] S4. With the goal of maximizing the expected reward on the training dataset, use reinforcement learning to optimize and update the reasoning strategy of the large language model in the process of multi-round reasoning and information retrieval, and output the trained model parameters.
[0051] With the goal of maximizing the expected reward on the training dataset, reinforcement learning is used to optimize and update the inference strategy of a large language model during multi-round inference and information retrieval processes. The trained model parameters are then output when a preset termination condition is met. The specific process is as follows: Optimization of inference strategies based on true value rewards: Based on the reward signal obtained in step S3 The reasoning and information retrieval behaviors generated by the large language model in step S2 are optimized and updated to maximize the expected reward of the model on the training dataset. The optimization objective is expressed as: ; in The parameters of the large language model are represented. Represents the training dataset. This indicates that samples were randomly taken from the training set during training. The corresponding reward signal Expected value Indicates the target paper, This represents the corresponding influence label value. Through this optimization process, the model is guided to gradually generate prediction results with smaller prediction errors and closer to the true influence label values.
[0052] Training termination and model output: When the training process meets the preset termination conditions, the model parameter updates are stopped and the parameters of the trained large language model are output. The termination conditions include reaching the maximum number of training rounds, or verifying that the average prediction error or expected reward on the dataset no longer improves over several consecutive training rounds. Table 1
[0053] As shown in Table 1, compared with existing technologies, the paper influence evolution trajectory prediction agent method based on large language model reasoning proposed in this application effectively solves the key problems of traditional methods, such as the lack of interpretability in the reasoning process, weak learning ability under weak supervision, and lack of feedback-driven dynamic strategy optimization, in scientific research evaluation scenarios with weak supervision and high interpretability requirements. It has the following significant technical advantages and effects: 1. Significantly improves prediction accuracy, especially under weakly supervised conditions: This method was experimentally evaluated on the standard academic dataset DBLP, using Mean Absolute Error (MAE) as the evaluation metric and the cumulative citation count within 5 years of the target paper's publication as the true impact label. Under the setting of using only single-point cumulative citations as the supervision signal and not relying on the complete citation evolution trajectory, our method (Ours) achieved an MAE of 0.98, significantly outperforming the current representative graph neural network methods HINTS (1.54) and H2CGL (1.87). This result demonstrates that through a reinforcement learning-guided multi-round inference mechanism, this method can more effectively mine discriminative signals from limited labels, achieving higher accuracy in impact prediction.
[0054] 2. Achieve a traceable and interventionable highly reliable reasoning process: Unlike existing end-to-end black-box models, this method explicitly models the prediction process as a multi-round interaction between a large language model agent and a structured external academic interface. Each step of reasoning generates intermediate judgments or information query requests in natural language form. This mechanism not only allows the prediction basis to be reviewed and verified by human experts, but also supports manual intervention or correction of the reasoning path in practical applications, significantly improving the credibility and usability of the system in high-risk scenarios such as scientific research evaluation and talent selection.
[0055] 3. Possesses the ability to dynamically evolve reasoning strategies for optimization: By introducing a reward signal based on the prediction-to-reality bias and employing reinforcement learning for end-to-end optimization of the reasoning strategy, this method can continuously adjust the information retrieval order, evidence integration method, and termination decision logic based on feedback. This "reasoning, learning, and optimizing simultaneously" mechanism enables the model to adaptively form reasonable judgment strategies when facing new domains, newly published papers, or data-sparse situations, overcoming the limitations of insufficient generalization ability of traditional static models.
[0056] This invention also provides a paper influence evolution trajectory intelligent agent prediction system, which is used to implement the above-described method embodiments; details already described will not be repeated. The terms "module," "unit," and "subunit," etc., used below refer to combinations of software and / or hardware that achieve a predetermined function. Although the system described in the following embodiments is preferably implemented in software, hardware implementation or a combination of software and hardware is also possible and contemplated.
[0057] like Figure 4 As shown, Figure 4 This is a structural block diagram of the intelligent agent prediction system for the evolution trajectory of paper influence in this invention. The system includes: Data processing module 101 is used to construct a training dataset containing the target paper and its true influence annotation values; The reasoning and interaction module 102 includes a large language model and a set of external academic information interfaces, which are used to perform multiple rounds of reasoning and external information retrieval for the target paper and output the predicted influence value. The reward calculation module 103 is used to construct a reward signal based on the deviation between the true influence label value and the predicted influence value of the target paper. The strategy optimization module 104 is used to optimize and update the reasoning strategy of the large language model using reinforcement learning methods with the goal of maximizing expected reward.
[0058] In this embodiment, the system is used to implement the above method embodiment. Its working principle and technical effect are basically the same as those of the method embodiment, so they will not be described again here.
[0059] like Figure 5As shown in the figure, an electronic device provided by an embodiment of the present invention includes: a processor 610, a communication interface 620, a memory 630, and a communication bus 640, wherein the processor 610, the communication interface 620, and the memory 630 communicate with each other through the communication bus 640. The processor 610 can call logical instructions in the memory 630 to execute the following method: S1. Construct a training dataset containing the target papers and their true influence labels; S2. For the target paper in the training dataset, construct a large language model inference input and configure an external academic information interface set. The large language model performs multiple rounds of inference and external information retrieval process. When the termination condition is met, the predicted influence value of the target paper is output. S3. Construct a reward signal based on the deviation between the true influence value and the predicted influence value of the target paper; S4. With the goal of maximizing the expected reward on the training dataset, the reasoning strategy of the large language model in the multi-round reasoning and information retrieval process is optimized and updated using reinforcement learning methods, and the trained model parameters are output.
[0060] Furthermore, the logical instructions in the aforementioned memory 630 can be implemented as software functional units and, when sold or used as independent products, can be stored in a computer-readable storage medium. Based on this understanding, the technical solution of the present invention, or the part that contributes to the prior art, or a part of the technical solution, can be embodied in the form of a software product. This computer software product is stored in a storage medium and includes several instructions to cause a computer device (which may be a personal computer, server, or network device, etc.) to execute all or part of the steps of the methods described in the various embodiments of the present invention. The aforementioned storage medium includes various media capable of storing program code, such as USB flash drives, portable hard drives, read-only memory (ROM), random access memory (RAM), magnetic disks, or optical disks.
[0061] This invention also provides a non-transitory computer-readable storage medium storing a computer program thereon, which, when executed by a processor, is implemented to perform the methods provided in the above embodiments.
[0062] Through the above description of the embodiments, those skilled in the art can clearly understand that each embodiment can be implemented by means of software plus necessary general-purpose hardware platforms, and of course, it can also be implemented by hardware. Based on this understanding, the above technical solutions, in essence or the part that contributes to the prior art, can be embodied in the form of a software product. This computer software product can be stored in a computer-readable storage medium, such as ROM / RAM, magnetic disk, optical disk, etc., including several instructions to cause a computer device (which may be a personal computer, server, or network device, etc.) to execute the methods of various embodiments or some parts of embodiments.
[0063] Finally, it should be noted that the above embodiments are only used to illustrate the technical solutions of the present invention, and not to limit them; although the present invention has been described in detail with reference to the foregoing embodiments, those skilled in the art should understand that modifications can still be made to the technical solutions described in the foregoing embodiments, or equivalent substitutions can be made to some of the technical features; and these modifications or substitutions do not cause the essence of the corresponding technical solutions to deviate from the spirit and scope of the technical solutions of the embodiments of the present invention.
Claims
1. A method for predicting the evolution trajectory of a paper's influence using an intelligent agent, characterized in that, Includes the following steps: S1. Construct a training dataset containing the target papers and their true influence labels; S2. For the target paper in the training dataset, construct a large language model inference input and configure an external academic information interface set. The large language model performs multiple rounds of inference and external information retrieval process. When the termination condition is met, the predicted influence value of the target paper is output. S3. Construct a reward signal based on the deviation between the true influence value and the predicted influence value of the target paper; S4. With the goal of maximizing the expected reward on the training dataset, the reasoning strategy of the large language model in the multi-round reasoning and information retrieval process is optimized and updated using reinforcement learning methods, and the trained model parameters are output.
2. The paper influence evolution trajectory prediction method of the intelligent agent according to claim 1, characterized in that, The construction of the training dataset in step S1 specifically includes: S11. Construct an academic database based on academic network data in the target research field, wherein the academic database contains at least paper meta-information and relational data; S12. Set the length of the time window for impact statistics. For each target paper, calculate the cumulative number of citations within the time window after publication, and use it as the true impact value of the target paper. S13. Divide the labeled sample set into training set, validation set and test set, and construct the training dataset representation.
3. The paper influence evolution trajectory prediction method of the intelligent agent according to claim 1, characterized in that, The set of external academic information interfaces mentioned in step S2 includes at least: The paper metadata query interface is used to query the title, publication year, author list, and publication medium information of a paper based on its identifier. The citation relationship query interface is used to query the reference set or citation relationship of a paper based on the paper identifier; The author or publication carrier statistics interface is used to query the corresponding list of historical papers or historical citation statistics based on the author identifier or publication carrier identifier. The similar paper retrieval interface is used to retrieve and return a collection of similar papers based on paper identifiers or paper text descriptions.
4. The paper influence evolution trajectory prediction method of the intelligent agent according to claim 1, characterized in that, The multi-round reasoning and external information retrieval process described in step S2 is as follows: Under the constraint of the maximum number of inference rounds T, the large language model generates new external information call requests or intermediate inference conclusions round by round based on the initial input and the external information call results of the historical rounds. When the large language model outputs a termination command or reaches the maximum number of inference rounds T, the inference process ends, and a predicted influence value of the target paper is generated based on the current inference context.
5. The paper influence evolution trajectory prediction method of the intelligent agent according to claim 1, characterized in that, The construction of the reward signal based on the bias in step S3 specifically involves: Based on predicted influence values Compared with the true influence label value Construct the reward function ,in Used to measure how close the predicted result is to the true value; The reward signal is subjected to numerical constraints or normalization to ensure it meets a preset value range. ,in and All of these are preset constants.
6. The paper influence evolution trajectory prediction method for intelligent agents according to claim 5, characterized in that, The reward function is defined as follows: .
7. The paper influence evolution trajectory prediction method for intelligent agents according to claim 1, characterized in that, The optimization and update using reinforcement learning methods described in step S4 has the following optimization objective: ,in The parameters of the large language model are represented. Represents the training dataset. This indicates that samples were randomly taken from the training set during training. The corresponding reward signal Expected value Indicates the target paper, This indicates the corresponding influence rating.
8. The paper influence evolution trajectory prediction method for intelligent agents according to claim 1, characterized in that, The termination conditions include: reaching a preset maximum number of training rounds, or the average prediction error or expected reward on the validation dataset no longer improving over multiple consecutive training rounds.
9. A predictive system for the evolution trajectory of a paper's influence using an intelligent agent, characterized in that, include: The data processing module is used to construct a training dataset containing the target papers and their true impact labels. The reasoning and interaction module, including a large language model and a set of external academic information interfaces, is used to perform multiple rounds of reasoning and external information retrieval for the target paper and output the predicted influence value. The reward calculation module is used to construct a reward signal based on the deviation between the actual influence value and the predicted influence value of the target paper. The strategy optimization module is used to optimize and update the reasoning strategy of the large language model using reinforcement learning methods with the goal of maximizing expected reward.
10. An electronic device, characterized in that, Including memory and processor; The memory is used to store computer programs; The processor is configured to, when executing the computer program, implement the paper influence evolution trajectory intelligent agent prediction method as described in any one of claims 1 to 8.