Method for reducing cognitive load of user based on complex intention understanding

Abstract

The invention discloses a method for reducing cognitive load of a user based on complex intention understanding. The method comprises the following steps: receiving a user instruction, retrieving or example a construction mode from a data set, decomposing a complex intention into hierarchical elements, and ensuring that a prerequisite dependency relationship is reserved; organizing and presenting clarified questions in a hierarchical structure; grouping independent questions into a single-round interactive table, and putting forward dependency questions in sequence to keep logic consistency; user-large language model clarification interaction is simulated through Monte Carlo sampling, and lexical element level intention importance and generation confidence are fused to evaluate and screen a clarification track; and finally, through a fine-tuning iterative optimization model on the high-quality reward data, each iteration generates higher-quality data for a subsequent optimization round. According to the invention, the system firstly identifies the core challenge of complex intention understanding in a large model, and provides an innovative solution; and the cognitive load of the user is reduced through logic coherent intention clarification.

Description

Technical Field

[0001] This invention belongs to the field of text recognition technology, and in particular relates to a method for reducing the cognitive load on users based on understanding complex intent. Background Technology

[0002] The internet has become the primary medium for human-computer interaction, and large language model assistants are increasingly embedded in social platforms to aid in user information discovery, creation, and management. Modern large language models excel at generating high-quality single-turn responses with explicit input; however, in online interactions, users often provide only rough descriptions due to the ambiguity of their intent, forcing them to repeatedly clarify their needs. The efficiency of this clarification process directly determines user experience and task success rate, highlighting the importance of intent understanding.

[0003] Recent research has explored interactive mechanisms that enable large language models to proactively clarify user intent. Mistral-Interact uses sequential question-answer pairs to progressively obtain missing information, while ITIU has developed a table-based interactive mechanism that generates multiple clarification questions in parallel within a single round. While these pioneering methods improve efficiency, they implicitly assume that the clarification questions are independent and are only applicable to simple intent scenarios.

[0004] Real-world tasks often involve complex intent scenarios, where clarifying questions involves logical dependencies. The model should suggest activities only after confirming the "destination." Ignoring these dependencies can lead to logically confusing suggestions (e.g., "diving in a certain location in December" when it's freezing cold), increasing the cost for users to obtain useful information—that is, cognitive load. The performance degradation of existing methods when migrating from simple to complex scenarios confirms the complexity of this task. Therefore, understanding the logical dependencies in clarifying questions is the core challenge of understanding complex intents. Understanding complex user intents is crucial for developing truly collaborative large language models. While existing methods improve efficiency and interactivity in simple intent scenarios, when clarifying questions involves logical dependencies, users are forced to manage these relationships themselves, leading to increased cognitive load. Summary of the Invention

[0005] To systematically address this challenge, this application draws upon Cognitive Load Theory (CLT), which academically categorizes cognitive load into two types: intrinsic load stemming from the inherent complexity of the task, and extrinsic load caused by poor task presentation or interaction design. This application aims to reduce the inherent complexity of user intent (intrinsic load) while minimizing unnecessary reasoning burden (extrinsic load). This application treats complex user intent as "white light" and logic-driven patterns as a "prism" to refract it into hierarchical elements, forming a "color spectrum." These hierarchical elements guide the order (parallel or sequential) of clarifying questions, thereby achieving a low-friction and logically coherent clarification process. This application first determines the "destination," then suggests relevant "activities" in subsequent stages, ultimately generating practical recommendations such as "whale watching boat tour."

[0006] Specifically, this application first constructs a retrieval dataset (CID) to provide a framework for decomposing complex intents across 429 intents in 27 domains. This allows fuzzy, high-dimensional intents to be decomposed into clearly structured sub-elements, reducing task complexity. Second, this application ensures that clarification questions are structured according to logical premise dependencies. By hierarchically ordering clarification questions, users are prevented from reasoning about unresolved elements. Third, intent-aware rewards and Monte Carlo sampling are introduced to simulate large-scale user-model interactions. Finally, to ensure practical and user-centered responses, this application employs a self-evolutionary intent tuning method to fine-tune the large model and optimize logical clarification. Prism is evaluated through extensive experiments across three dimensions: clarification interaction, intent execution, and cognitive load. Most notably, Prism achieves state-of-the-art logical clarification, reducing logical conflicts in clarification questions to 11.5%. Furthermore, compared to Mistral-Interact, Prism improves interaction scores by 14.4% and reduces average task time by 34.8%, significantly improving user experience.

[0007] To achieve the above objectives, the method for reducing user cognitive load based on complex intent understanding disclosed in this application includes the following steps: Receive user instructions, retrieve or build patterns from datasets or examples, break down complex intents into hierarchical elements, and preserve prerequisite dependencies; Organize and present clarification questions in a hierarchical structure: group independent questions into single-round interactive tables, and present dependent questions in sequence to maintain logical consistency; The user-large language model clarification interaction was simulated by Monte Carlo sampling, and the filtering clarification trajectory was evaluated by fusing word-level intent importance and generation confidence. By fine-tuning and iteratively optimizing the model based on reward data, each iteration generates higher-quality data for subsequent optimization rounds.

[0008] Furthermore, the complex intent It is decomposed into multiple elements; there are prerequisite dependencies between the elements. Given initial instructions Then, generate the final output. Before, proceed Round of clarification interaction; in the first round Based on the time of the turn And previous clarification of history Generate clarification content The user then responded. ;No. The round of clarification is The complete interaction process is represented as ,in This represents the complete clarification trajectory. Indicates the first The round of clarification.

[0009] Furthermore, the quality of the clarified trajectory and the final output is determined by the intent-aware reward function. The evaluation, in which the intent-aware reward function quantifies the clarification process, affects the output. The effect of improved alignment with user intent; formally expressed as: , The above formula represents maximizing the alignment between the model output and the user's intent.

[0010] Furthermore, step S1 also includes constructing a complex intent decomposition dataset. The construction process is divided into two stages: multi-source integration and dependency recognition. Both stages are marked with large language models and then manually reviewed. The multi-source data integration utilizes existing datasets of related tasks as data sources. These datasets contain annotation information of domain, intent, and elements. A semantic clustering method based on a large language model is used to standardize synonyms in each dataset. Each cluster is assigned a canonical label corresponding to the most frequent or most standardized expression. The dependency identification utilizes a large language model to assist manual annotators in annotating the logical dependencies between elements. The specific steps are as follows: First, guide GPT-4o to predict the prerequisite elements of each target element. The annotator then verifies these predictions through counterexamples. Based on the identified prerequisites, the elements are organized into a hierarchical structure under two constraints: (1) elements in the same layer are independent of each other, and (2) the dependency relationship strictly flows from the early layer to the late layer.

[0011] Further, in step S1, an initial user instruction is given. The goal of the decomposition is to deduce its hierarchical elements. , It is the first Hierarchical elements; adopting a dataset-based retrieval construction paradigm: if the user intent matches an existing intent in the dataset, reuse its labeled hierarchical elements; otherwise, construct new hierarchical elements by performing few-sample generalization from the most similar intent in the dataset; The specific steps in S1 include: First, predict the initial user command. Most likely area - intention pair From a formal perspective, the large language model estimates are as follows: in This represents the probability distribution provided by the large language model. and It uses predefined domains and intent sets in the dataset to ensure naming consistency. These represent the domain and complex intent, respectively; if the dataset contains... If the specified level is true, then its labeled hierarchical elements are directly reused; otherwise, a small sample is generated. Secondly, a few-shot construction is performed: when no exact match is found, the semantically most similar match is retrieved from the dataset. One intent is used as an example; these retrieved examples and their labeled hierarchical structure elements will be used as few-sample examples. Provided to the model; the model then uses these examples as a basis for... Generate new hierarchical structure elements: It is a few sample examples; The above process enables large language models to migrate decomposition logic from existing intents in the dataset to unseen intents, while maintaining logical consistency between dependencies.

[0012] Furthermore, in step S2, according to element hierarchy Organize a sequence of questions to facilitate logically coherent interactions with users; for each element level The large language model generates an interactive table containing corresponding clarification questions and recommended options: in This indicates that the previous rounds of clarification of history, This represents an interactive table containing corresponding clarification questions and recommended options; same level The clarification questions within are independent of each other, allowing for parallel interaction within a single table; user responses are... and clarifying history Combine; Subsequently based on Generate the next level The clarification ensures that the clarification path follows the logical dependencies defined by the element hierarchy.

[0013] Furthermore, in step S3, Monte Carlo sampling is used to generate large-scale, high-quality fine-tuning training data. During this process, intent-aware rewards are used to evaluate the quality of each clarified trajectory; Clarification of the cycle The corresponding intention-aware reward IR is: in Indicates as of the date The history of dialogue between wheels Representing the The subsequent dialogue after the round, distributed Modeling potential future dialogues based on historical premises; Use word-level rewards Together, we can obtain the importance of each word in expressing the user's intent, as well as the confidence level of the model in generating that word; Given a clarified trajectory obtained through forward sampling and the final output of the simulation ,calculate: in, calculate The intention importance score vector, calculate middle Generate a confidence score vector for each word; each vector dimension corresponds to... One of the lexical elements reflects its criticality in expressing the user's intent and the model's grasp of the lexical element's generation. Use intention importance score Measure each word pair The contribution of the overall intention expression in the evaluation The score is calculated based on sensitivity to semantic contradictions; specifically, a loss function is defined. Quantification through natural language reasoning models The possibility of "contradiction" with itself; reverse the label and calculate the "contradiction" loss relative to each lexical embedding. gradient The formula for calculating the importance score of intent is: in, This indicates element-wise multiplication; The Dimensions reflect the word Sensitivity to disturbances in the overall intent representation Indicates the number of lexical units. It is the lexical sequence number; Used to measure the final output generated by a large language model Confidence level at time; for each lexical unit Large language models are based on the preceding context. Generate conditional probability distribution: in The Dimensions correspond to the large language model for lexical units The predicted probability; Through integration and The lexical-level reward function highlights lexical units that are both semantically important and generated with high confidence. from Mid-sampling; use intent reward to estimate the intent-aware reward score for each round.

[0014] Further, step S7 includes: Intent tuning of a large language model is performed in a self-evolutionary manner. This self-evolutionary intent tuning process involves multiple iterations, with data evolution and model evolution mutually reinforcing each other. In the first round, the GPT-4o model is used as the policy model to generate a large-scale intent-clarifying dialogue record, and high-quality data is selected based on IR (Intent Reduction) to fine-tune the open-source large language model. In subsequent rounds, the fine-tuned large language model serves as the new policy model, generating higher-quality training data for the next iteration. Specifically, the self-evolutionary intent tuning includes the following stages: Clarification Trajectory Collection: Monte Carlo sampling is used to generate intent clarification trajectories. Each trajectory is evaluated and filtered by IR score for supervised fine-tuning and direct preference optimization. For supervised fine-tuning, the top-ranked trajectories are retained to form the clarification dialogue training data. For direct preference optimization, the clarification trajectories are ranked according to IR score to construct "preferred" and "rejected" clarification pairs. Self-evolutionary intent tuning: A large language model is selected as the base model; in each iteration, Monte Carlo sampling is used to generate intent clarification trajectories, high IR samples are fine-tuned, and the resulting model is deployed as a new strategy for subsequent rounds.

[0015] The beneficial effects of this application are as follows: We were the first to systematically identify the core challenges of understanding complex intentions in large models and propose innovative solutions; Introducing CLT theory into this task reduces the user's cognitive load through logically coherent intent clarification. Attached Figure Description

[0016] Figure 1 This is a method framework diagram of this application; Figure 2 These are experimental results. Figure 2 (a) is the user's time spent. Figure 2 (b) is the number of dialogue markers. Participants' responses to... Figure 2 (c) Overall interaction experience score, and in Figure 2 (d) An additional evaluation is conducted every two rounds; Figure 3 This represents the trend of absolute power spectral density (PSD) across the entire frequency band during the cognitive load experiment. Detailed Implementation

[0017] To further illustrate the technical means and effects of the present invention in achieving its intended purpose, the following detailed description of the specific implementation methods, structures, features, and effects of the present invention, in conjunction with the accompanying drawings and preferred embodiments, is provided below.

[0018] Artificial intelligence (AI) is the theory, methods, technology, and application systems that use digital computers or machines controlled by digital computers to simulate, extend, and expand human intelligence, perceive the environment, acquire knowledge, and use that knowledge to achieve optimal results. In other words, AI is a comprehensive technology within computer science that attempts to understand the essence of intelligence and produce a new kind of intelligent machine that can react in a way similar to human intelligence. AI studies the design principles and implementation methods of various intelligent machines, enabling them to possess the functions of perception, reasoning, and decision-making.

[0019] Natural Language Processing (NLP) is an important field within computer science and artificial intelligence. It studies the theories and methods for enabling effective communication between humans and computers using natural language. NLP is a science that integrates linguistics, computer science, and mathematics. Therefore, research in this field involves natural language—the language people use in daily life—and thus it has a close relationship with linguistic research. NLP techniques typically include text processing, semantic understanding, machine translation, question answering, and knowledge graphs.

[0020] The technical solutions provided in this application involve technologies such as machine learning and natural language processing in artificial intelligence, which are specifically described and illustrated through the following embodiments.

[0021] Before introducing the embodiments of this application, some terms involved in this application will be explained.

[0022] 1. Non-interactive User Intent Understanding: User intent understanding refers to the process by which a large model infers a user's true intention, typically achieved through non-interactive or interactive methods. Non-interactive intent understanding focuses on predicting user intent without direct feedback or dialogue. Existing methods can be divided into two categories: human preference alignment and cue word optimization. Human preference alignment adjusts the large model through supervised fine-tuning (SFT) and human feedback reinforcement learning (RLHF) to align its output with human intent. Cue word optimization refines user input to improve intent prediction without modifying the underlying model.

[0023] 2. Interactive User Intent Understanding: Interactive intent understanding aims to capture personalized user intent through effective human-computer interaction. Existing methods neglect the crucial dimension of clarifying the logical ordering of questions. Due to the lack of modeling dependencies between questions, they struggle to handle the complex intents commonly found in real-world scenarios.

[0024] 3. Cognitive Load in Human-Computer Interaction: Cognitive Load Theory (CLT), originating from cognitive psychology, emphasizes the limitations of human working memory in processing information. This theory distinguishes between two types of cognitive load: intrinsic load, stemming from the complexity of the task itself; and extrinsic load, caused by poor task presentation or interaction design. Effective interaction design aims to manage intrinsic load while minimizing extrinsic load, thereby improving task performance and user experience. Existing human-computer interaction designs primarily focus on efficiency (e.g., parallel question clarification) or personalization (e.g., reward modeling). While these strategies improve surface outcomes, they often overlook the implicit cognitive costs arising from improper question sequence structure—particularly detrimental in complex intent scenarios with precondition dependencies. Based on CLT, this application's framework reduces extrinsic load through dependency-aware clarification sequences and manages intrinsic load through structured intent decomposition. This perspective connects cognitive psychology with practical large language model interaction design, promoting cognitively efficient human-computer collaboration.

[0025] This application will have complex intentions. Defined as: (1) decomposable into multiple elements; (2) having prerequisite dependencies among these elements. For example, the intention to "plan a trip" includes elements such as "destination," "travel date," and "activity arrangements." Logically, determining the "destination" should precede selecting the "activity arrangements," which allows the model to generate relevant activity suggestions based on the selected location.

[0026] Multiple rounds of clarification: given initial instructions Then, the model generates the final output. Before, proceed The first round of clarification interaction. During the round, the model is based on And previous clarification of history Generate clarification content (e.g., an interactive form), and the user responds. . No. The clarification record is as follows The complete interaction process is represented as ,in This represents a complete clarification trajectory.

[0027] refer to Figure 1The framework of this application is as follows: Given a user instruction, the policy model first performs a hierarchical decomposition of complex intents by retrieving the CID dataset, then organizes clarification questions and implements them through an interactive table; a logical interaction is then performed. Intent-aware rewards (IRs) are then estimated through Monte Carlo sampling. Finally, self-evolutionary intent tuning iteratively improves the quality of training data, enhancing the large language model's ability to understand complex intents.

[0028] The objective of this application is to generate a logically coherent and clear trajectory. Ultimately, high-quality output is achieved. The quality of the clarified trajectory and the final output is determined by the intent-aware reward function. The evaluation function quantifies the effect of the clarification process on the output. The effect of improved alignment with user intent. Formalized as: , This means that the optimization goal of this application is to maximize the alignment between the model output and the user's intent.

[0029] To address the identified challenges, this application designs four customized modules based on Cognitive Load Theory (CLT). The overall framework of this application's system is as follows: Figure 1 As shown, upon receiving a user instruction, the model first decomposes complex intents into hierarchical elements by retrieving patterns from CIDs or constructing patterns from a small number of examples, ensuring that prerequisite dependencies are preserved. Then, clarification questions are organized and presented in a hierarchical structure: independent questions are grouped into single-round interaction tables, while dependent questions are presented sequentially to maintain logical consistency. Subsequently, this application simulates user-large language model clarification interactions through Monte Carlo sampling, employing an intent-aware reward mechanism—integrating lexical-level intent importance and generation confidence—to evaluate the selection of clarification trajectories. Finally, the self-evolutionary intent tuning module iteratively optimizes the model through fine-tuning on high-quality reward data (SFT / DPO), with each iteration generating higher-quality data for subsequent optimization rounds.

[0030] S1 Complex Intent Decomposition Complex Intent Decomposition (CID) aims to break down user intents into smaller, more clearly structured elements, thereby reducing task complexity. To support this process, this application constructs a Complex Intent Decomposition dataset (CID), which provides labeled element dependencies for 429 common intents across 27 domains.

[0031] S11 CID Retrieval Dataset Construction: This application follows two key principles to construct the CID retrieval dataset: (1) adopting multi-source integration and unified naming to reduce data heterogeneity; (2) explicitly labeling prerequisites to capture logical dependencies between elements. The construction process is divided into two stages: multi-source integration and dependency identification, both of which adopt LLM-assisted labeling followed by manual review.

[0032] Multi-source data integration: This application utilizes existing datasets related to the task as data sources. These datasets contain annotation information for key components such as domain, intent, and features. Given their heterogeneity, this application employs a semantic clustering method based on a large language model to standardize synonymous expressions in each dataset, thereby ensuring consistency. Each cluster is assigned a canonical label, typically corresponding to the most frequent or most standardized expression. For example, expressions such as "Plan a trip" and "Travelplanning" are uniformly grouped under the canonical intent "make a travel plan," and their associated elements are also merged accordingly.

[0033] Dependency Identification: This application utilizes a Large Language Model (LLM) to assist human annotators in identifying logical dependencies between elements. First, GPT-4o is guided to predict the prerequisite elements for each target element. Annotators then verify these predictions using counterexamples (e.g., checking whether clarifying "activity arrangements" can be achieved without first specifying the "destination"). Based on the identified prerequisites, elements are organized into a hierarchical structure under two constraints: (1) elements within the same level are independent of each other, and (2) dependencies strictly flow from earlier to later levels. This hierarchical structure optimizes the clarification process—independent clarification questions can be raised in parallel within a single round of dialogue, while dependency questions are deferred to subsequent rounds.

[0034] Through these two stages, this application constructs a CID retrieval dataset. A significant feature of CID is its explicit encoding of the preconditions between intent elements, which provides a reference framework for the decomposition of complex intents.

[0035] S12 Intent Decomposition Process: Given the initial user instruction The goal of the decomposition is to deduce its hierarchical elements. This application adopts a retrieval-construction paradigm based on the CID dataset: if the user intent matches an existing intent in the CID, the hierarchical elements labeled therein are reused; otherwise, new hierarchical elements are constructed by performing few-sample generalization from the most similar intent in the CID.

[0036] Domain and Intent Recognition: This application first predicts the initial user instruction. Most likely area - intention pair From a formal perspective, the Large Language Model (LLM) estimate is as follows: in This represents the probability distribution provided by the Large Language Model (LLM). and This refers to the predefined domains and intent sets within the CID to ensure naming consistency. If the CID contains... If the specified level is true, then the labeled level elements are reused directly; otherwise, a small sample is generated.

[0037] Few-shot construction: When no exact match is found, this application retrieves the semantically most similar item from the CIDs. One intent is used as an example. These retrieved examples and their labeled hierarchical structure elements will serve as few-sample examples. These examples are provided to the model. The model then uses these examples to... Generate new hierarchical structure elements: This process enables LLM to migrate decomposition logic from existing intents in CID to unseen intents, while maintaining logical consistency between dependencies.

[0038] S2 Logic Clarification Generation This module is organized by element hierarchy. Organize a sequence of clarifying questions to facilitate logically coherent interactions with the user. This applies to each element level. LLM will generate an interactive table containing corresponding clarification questions and recommended options: in This refers to the clarification of the previous rounds of history.

[0039] same level The clarification questions within are independent of each other, so they are presented in a single table to allow for parallel interaction. User responses are denoted as... and clarifying history (i.e.) ) combined. Then based on Generate the next level The clarification ensures that the clarification path follows the logical dependencies defined by the element hierarchy.

[0040] S3 Intent-Aware Reward To enhance the logical clarification capability of LLM, this application employs Monte Carlo sampling to generate large-scale, high-quality fine-tuning training data. During this process, this application designs an intent-aware reward to evaluate the quality of each clarification trajectory. Clarification of the cycle The corresponding intention-aware reward (IR) is: in Indicates as of the date The history of dialogue between wheels Representing the The subsequent dialogue after the round. Distribution We modeled potential future dialogues based on historical premises.

[0041] However, calculating Equation 4 remains challenging, as it requires: (1) a method for evaluating each clarifying trajectory. lexical-level reward function (2) Obtain the distribution of subsequent dialogues The sampling strategy.

[0042] S31 Lexical-level reward function: This application defines a lexical-level reward. This is used to collectively capture the importance of each lexical in expressing user intent, as well as the confidence level of the model in generating that lexical. Given a clarification trajectory obtained through forward sampling... and the final output of the simulation ,calculate: in, calculate The intention importance score vector, calculate middle The generated confidence score vector for each term. Each vector dimension corresponds to... One of the lexical terms reflects its criticality in expressing the user's intent and the model's grasp of the lexical term in generating it.

[0043] For example, the morpheme "a certain place" is crucial for expressing the intent of "destination," thus earning a high intent importance score. If the model also generates this morpheme with high confidence, the trajectory is clarified. The intent of "destination" is effectively clarified. Conversely, low generated confidence for such keyword elements indicates that the clarification process is still incomplete or inconsistent.

[0044] Intention Importance Score: Used to measure each word pair The contribution of the overall intent expressed in the application. This application passed the evaluation. The score is calculated based on sensitivity to semantic contradictions. Specifically, this application defines a loss function. Quantification through Natural Language Inference (NLI) models The possibility of "contradiction" with itself. Although models typically predict "implication" relationships with high probability, this application reverses the label and calculates the "contradiction" loss relative to each lexical embedding. The gradient. The formula for calculating the importance score of intent is: in This indicates element-wise multiplication. The Dimensions reflect the word Sensitivity to perturbations in the overall intent representation. A higher value indicates that the term is more critical to the intent expression. The loss function L can be the mean squared error (MSE) function, which is sensitive to outliers.

[0045] Generate confidence scores: Used to measure the final output generated by a Large Language Model (LLM). The confidence level at that time. For each terminology. LLM will be based on the context above. Generate conditional probability distribution: in The Dimensions correspond to LLM lexical units The predicted probability. A higher probability value indicates a stronger confidence level, suggesting a clearer trajectory. For reliable generation It provided sufficient information.

[0046] Through integration and The lexical-level reward function highlights lexical units that are both semantically important and generated with high confidence. This design ensures that the reward accurately reflects whether the clarification trajectory resolves key intent elements with logical consistency. The lexical-level reward function can be either cross-entropy loss or negative log-likelihood loss.

[0047] S32 Forward Sampling: As per Formula 4, this application requires... Sampling is performed on a conditional forward dialogue distribution based on dialogue history. A simple approach is to use Monte Carlo sampling, extending the dialogue round by round until completion. Unlike standard Monte Carlo sampling, this application does not rely on an external reward model, but instead uses intent reward (IR) to estimate the intent-aware reward score for each round. This method reduces computational overhead while more accurately measuring the final output. The ability to perceive intent.

[0048] While real-world dialogues can be collected from human participants, sampling multiple forward dialogues during training is costly and inefficient. To achieve scalable training, this application introduces a user simulator. This drives the LLM to simulate user behavior. The model is required to mimic the user's previous language style and inject typical user behaviors, thereby achieving efficient and realistic forward sampling.

[0049] S4 self-evolutionary intention optimization By leveraging a lexical-level reward function and a forward sampling strategy, this application can compute the Intent Resolver (IR) of any clarifying trajectory without relying on an independent reward model. This enables the application to perform intent tuning of the LLM in a self-evolutionary manner. The self-evolutionary intent tuning process involves multiple iterations, with data evolution and model evolution mutually reinforcing each other. In the first round, a powerful closed-source model (GPT-4o) is used as the policy model to generate a large-scale intent-clarifying dialogue record, and high-quality data is selected based on the IR to fine-tune the smaller open-source LLM (LLaMA-3.1-8B-Intruct). In subsequent rounds, the fine-tuned LLM serves as the new policy model, generating higher-quality training data for the next iteration. This application collects... User instructions serve as the data source. Specifically, self-evolutionary intent tuning includes the following stages.

[0050] Clarification Trajectory Collection: This application employs Monte Carlo sampling to generate intent clarification trajectories. Each trajectory is evaluated and filtered using an Indicator Reduction (IR) score for Supervised Fine-Tuning (SFT) and Direct Preference Optimization (DPO). For SFT, the top-ranked trajectories are retained to form the clarification dialogue training data; for DPO, the clarification trajectories are ranked according to their IR scores to construct "preferred" and "rejected" clarification pairs.

[0051] Self-evolutionary intent tuning: LLaMA-3.1-8B-Instruct and Mistral-7B-Instruct-v0.3 were selected as the base models for this application. The number of self-evolutionary rounds was determined based on empirical experience and previous research results. Three rounds of self-evolutionary intent tuning were conducted, with each round generating data of higher quality than the previous one. In each iteration, this application uses Monte Carlo sampling to generate intent clarification trajectories, fine-tunes high-IR samples, and deploys the resulting model as a new strategy for subsequent rounds. By continuously optimizing for higher IR values, the model learns to generate responses that improve the clarification of interactions and coherence. This self-evolutionary process enables Prism to automatically generate scalable, high-quality, complex intent understanding data without manual annotation, thereby supporting robust generalization capabilities across tasks.

[0052] The ability of a large language model to understand user intent can be evaluated from three complementary dimensions: clarifying interaction, intent execution, and cognitive load. Clarifying interaction assesses the model's effectiveness in understanding and optimizing user intent during the clarification process; intent execution reflects the model's ability to accurately and efficiently complete tasks; and cognitive load quantifies the mental exertion of the user throughout the interaction. Accordingly, this application organizes experiments around these three dimensions.

[0053] Cognitive load can be assessed using three types of indicators: behavioral indicators (user actions), subjective indicators (user self-rating scores), and physiological indicators (biometric data). To evaluate the model's effectiveness from the user's perspective, we conducted a cognitive load study with 20 participants as the core.

[0054] Each participant was randomly assigned four tasks from complex intent scenarios of TIN, IN3, or ABP, interacting sequentially with this application, ColLABLLM, Mistral-Interact, and ITIU (the model order was not disclosed beforehand). Interaction scores were given based on real-time experience after every two rounds of dialogue. After a complete dialogue, the quality of the model's final output and the overall interaction experience were evaluated (subjective metrics), with scores ranging from 1 to 10. Total interaction time and total number of ephemerals were also recorded to assess behavioral efficiency (behavioral metrics). Furthermore, participants' electroencephalogram (EEG) data were recorded using a multimodal physiological signal acquisition system, and cognitive load (physiological metrics) was measured using power spectral density (PSD) analysis.

[0055] Figure 2 and Figure 3 The results of the cognitive load assessment were presented. Prism performed as follows: (1) More efficient user behavior. Figure 2 (a) shows that Prism reduces task time by 34.8% compared to Mistral-Interact and by 22.3% compared to ColLABLLM. Figure 2 (b) Prism controls the total number of morphemes in the dialogue to below 1000, indicating that users need to receive and output less information, which confirms the important role of logical clarification in improving interaction efficiency. (2) Better user experience. Figure 2 (c) Prism consistently leads in user ratings, with an average interaction score of 8.3 and 88.6% of participants rating it as "good" (8-9 points). Rated as "Excellent" (8-10 points), while Mistral-Interact is respectively and . Figure 2 (d) It shows that the scores of ITIU and COLLABLLM decreased after 4-6 rounds, indicating that user satisfaction decreased in long interactions; while the Prism score increased over time, proving that its logically coherent interaction design can continuously improve the dialogue experience. (3) Lower user cognitive load. Figure 3 The data shows that PSD values ​​gradually increased as participants interacted with Prism, COLLABLLM, Mistral-Interact, and ITIU in sequence. Prism consistently maintained the lowest PSD value, indicating that users exerted the least cognitive effort when interacting with it, which is consistent with its user-centered design goals.

[0056] This application decomposes ambiguous intentions into well-structured small elements and serializes the clarification process in a logically coherent manner, achieving smooth and low-friction interaction. Extensive experiments and practical evaluations show that this application achieves state-of-the-art performance in logical clarification, improving user satisfaction and shortening task completion time. It can drive large language models towards becoming empathetic and proactive collaborators, capable of understanding and guiding users to complete complex and dynamic goals.

[0057] The above description is merely a preferred embodiment of the present invention and is not intended to limit the present invention in any way. Although the present invention has been disclosed above with reference to preferred embodiments, it is not intended to limit the present invention. Any person skilled in the art can make some modifications or alterations to the above-disclosed technical content to create equivalent embodiments without departing from the scope of the present invention. Any simple modifications, equivalent changes and alterations made to the above embodiments based on the technical essence of the present invention without departing from the scope of the present invention shall still fall within the scope of the present invention.

Claims

1. A method for reducing cognitive load of a user based on complex intent understanding, characterized in that, The method comprises the following steps: Receiving user instructions, retrieving or constructing patterns from data sets, decomposing complex intentions into hierarchical elements, and retaining premise dependencies; Organizing and presenting clarified questions in a hierarchical structure: grouping independent questions into single-round interaction tables and presenting dependent questions in sequence to maintain logical consistency; Simulating user-large language model clarification interactions through Monte Carlo sampling, combining word-level intent importance and generation confidence to evaluate and filter clarification trajectories; Iteratively optimizing the model through fine-tuning on reward data, generating higher-quality data for subsequent optimization rounds. 2.The method of reducing cognitive load of a user based on complex intention understanding according to claim 1, wherein, decomposing the complex intent into a plurality of elements; the elements have prerequisite dependency relationships; Given initial instruction After generating final output Before, perform Round clarification interaction; in the Round, generate clarification content based on and previous clarification history , the user then responds ; the Round clarification is , the complete interaction process is represented as , where represents the complete clarification trajectory, represents the Round clarification. 3.The method of reducing cognitive load of a user based on complex intention understanding according to claim 2, characterized in that, The quality of the clarification trajectory and the final output is rewarded by an intent-aware reward function The intent-aware reward function quantifies how well the clarification process aligns with the user intent The effect of the promotion of the alignment with the user intent; formally expressed as: ， The above formula represents the maximization of the alignment between the model output and the user's intention. 4.The method of reducing cognitive load of a user based on complex intention understanding according to claim 3, wherein, In step S1, a complex intention decomposition dataset is also constructed, and the construction process is divided into two stages: multi-source integration and dependency relationship identification, both of which use large language models to assist in manual review; The multi-source data integration uses existing datasets related to the task as data sources, which contain domain, intent, and element annotation information. A semantic clustering method based on large language models is used to standardize the synonymous expressions in each dataset. Each cluster is assigned a standard label corresponding to the most frequent or most standardized expression form. The dependency relationship identification uses a large language model to assist human annotators in labeling the logical dependencies between elements. The specific steps are as follows: First, GPT-4o is used to predict the prerequisite elements for each target element, and then the annotators verify these predictions through counterexamples. Based on the identified prerequisites, the elements are organized into a hierarchical structure under two constraints: (1) elements at the same level are independent of each other, and (2) dependency relationships strictly flow from early layers to late layers. 5.The method of reducing cognitive load of a user based on complex intention understanding according to claim 4, characterized in that, In step S1, an initial user instruction is given , the goal of intent decomposition is to derive its hierarchical elements , is the first layer element; adopt the retrieval-based construction paradigm: if the user intent matches an existing intent in the dataset, reuse its annotated hierarchical elements; otherwise, construct new hierarchical elements by few-shot generalization from the most similar intent in the dataset; The specific steps of S1 include: First predict initial user instruction Most likely domain-intent pair From a formal perspective, the large language model estimates as follows: wherein denotes the probability distribution provided by the large language model, and is a predefined set of domains and intents in the dataset to ensure naming consistency, represent the domain and complex intent, respectively; if there is in the dataset, directly reuse the hierarchical elements annotated by it; otherwise, small sample construction is performed to generate; Secondly, a few-shot construction is performed: when no exact match is found, the semantically most similar match is retrieved from the dataset. One intent is used as an example; these retrieved examples and their labeled hierarchical structure elements will be used as few-sample examples. Provided to the model; the model then uses these examples as a basis for... Generate new hierarchical structure elements: is a few-shot example; The above process enables the large language model to transfer the decomposition logic from existing intentions in the dataset to unseen intentions while maintaining logical consistency between dependencies.

6. The method for reducing cognitive load of a user based on complex intention understanding according to claim 5, characterized in that, In step S2, the sequence of clarification questions is organized to facilitate logically coherent interactions with the user; for each element hierarchy the sequence of clarification questions is organized to facilitate logically coherent interactions with the user; for each element hierarchy a large language model generates an interaction table containing corresponding clarification questions and recommended options: wherein represents a history of previous rounds of clarification, represents an interactive table containing corresponding clarification questions and recommended options; The same level The clarification issues within are independent of each other, implemented in a single table for parallel interaction; user responses are combined with the clarification history ​ Subsequently based on Generating the next level of clarification, ensuring that the clarification trajectory follows the logical dependency relationships defined by the element hierarchy.

7. The method for reducing cognitive load of a user based on complex intention understanding according to claim 6, characterized in that, In step S3, large-scale high-quality fine-tuning training data is generated using Monte Carlo sampling, and in this process, the quality of each clarification trajectory is evaluated using the intention-aware reward; the first corresponding intention-aware reward IR is: corresponding intention-aware reward IR is: wherein represents the number of turns up to the dialogue history, represents the number of turns after the subsequent dialogue, distributed models potential future dialogues based on historical premises; Using token-level rewards to jointly capture the importance of each token in expressing the user intent, as well as the model's confidence in generating that token; Given the clarified trajectory obtained by forward sampling and the simulated final output , compute: wherein, computing an intent importance score vector, computing in a generation confidence score vector for each of the word pieces; each vector dimension corresponds to a word piece in the intent, reflecting its keyness in expressing the user intent and the model's confidence in generating the word piece; Use intention importance score Measure each word pair The contribution of the overall intention expression in the evaluation The score is calculated based on sensitivity to semantic contradictions; specifically, a loss function is defined. Quantification through natural language reasoning models The possibility of "contradiction" with itself; reverse the label and calculate the "contradiction" loss relative to each lexical embedding. gradient The formula for calculating the importance score of intent is: wherein, denotes element-wise multiplication; the first The dimension reflects the word sensitivity to the overall intent representation perturbation, denotes the number of tokens, is the token index; Measuring confidence levels of large language models in generating final outputs at the time of generation; for each token A large language model generates a conditional probability distribution based on the preceding context A large language model generates a conditional probability distribution based on the preceding context wherein the first dimension corresponds to a predicted probability of a token by the large language model; By integrating and , the token-level reward function highlights tokens that are both semantically important and high-confidence generated; From sampling; using the intent reward to estimate the intent-aware reward score for each round. 8.The method of reducing cognitive load of a user based on complex intention understanding according to claim 7, wherein, Step S7 includes: Self-evolutionary intent tuning of the large language model, which undergoes multiple iterations of data evolution and model evolution, with the first round using model GPT-4o as the strategy model to generate large-scale intent clarification dialogue records and fine-tune the open-source large language model based on IR to filter high-quality data; in subsequent rounds, the fine-tuned large language model serves as the new strategy model to generate higher-quality training data for the next iteration; specifically, the self-evolutionary intent tuning includes the following stages: Clarification trajectory collection: Monte Carlo sampling is used to generate intent clarification trajectories, each trajectory is evaluated and filtered by IR score for supervised fine-tuning and direct preference optimization; for supervised fine-tuning, the top-ranked trajectories are retained to form clarification dialogue training data; for direct preference optimization, the clarification trajectories are sorted according to the IR score to construct "preferred" and "eliminated" clarification pairs; Self-evolutionary intent tuning: Select a large language model as the base model; in each iteration, use Monte Carlo sampling to generate intent clarification trajectories, fine-tune high-IR samples, and deploy the resulting model as the new strategy for the next round.

Images