Association and reasoning dual-mode large model system and method
By using a dual-mode system of association and reasoning, resources are dynamically allocated to collaboratively process tasks, solving the problem of insufficient creativity and reasoning ability of large language models in complex tasks, and improving the overall performance and interpretability of the model.
Patent Information
- Authority / Receiving Office
- CN · China
- Patent Type
- Applications(China)
- Current Assignee / Owner
- CHONGSHANG KESAI (HANGZHOU) TECHNOLOGY CO LTD
- Filing Date
- 2026-01-15
- Publication Date
- 2026-04-17
AI Technical Summary
Existing large-scale language models suffer from insufficient creativity and reasoning ability, inefficient resource allocation, and inability to dynamically adjust cognitive patterns according to task characteristics when dealing with complex tasks.
It adopts a dual-mode system of association and inference, and dynamically allocates computing resources through a mode routing controller, a dual-mode processing engine and a coordinating integrator to collaboratively process divergent and convergent tasks.
It enhances the model's creativity in open-domain tasks, improves its accuracy in reasoning tasks, and increases the transparency and interpretability of the decision-making process.
Smart Images
Figure CN121882276A_ABST
Abstract
Description
Technical Field
[0001] This invention relates to the fields of artificial intelligence and large-scale language model technology, and in particular to a large-scale language model system architecture and its control method that simulates the dual-track cognitive process of humans, integrates divergent association and convergent reasoning, and can dynamically allocate computing resources according to task characteristics. Background Technology
[0002] Current large language models (LMs) primarily process information based on a single, general-purpose parameterized knowledge base pre-trained on massive amounts of data. While these models perform well in many natural language processing tasks, they suffer from the following inherent limitations when handling complex tasks requiring different cognitive modes: Limitations of creative tasks: The content generated by the model is heavily dependent on the statistical distribution of the training data. It is essentially a recombination and interpolation of existing patterns, making it difficult to make truly divergent associations and original innovations that break through the data distribution. It is easy to fall into formulaic and templated outputs. Limitations of complex reasoning tasks: When faced with tasks that require multi-step logical deduction, strict causal inference or symbolic operation, the model lacks a reliable internal reasoning mechanism, which can easily lead to factual errors, logical breaks or so-called hallucinations, and has insufficient convergent reasoning ability. Rigid cognitive mode switching: Existing single models cannot dynamically adjust their internal processing mechanisms according to the characteristics of the problem. They handle the creation of a poem and the solution of a math problem with the same or similar "thinking methods", making it difficult to achieve adaptive switching between the two fundamentally different cognitive modes of divergence and convergence; Inefficient resource allocation: For simple queries, the model may activate all parameters, resulting in wasted computing power; for complex problems, a fixed computation graph may not provide sufficient depth of thought, leading to performance bottlenecks.
[0003] Therefore, there is an urgent need in this field for a new intelligent computing architecture that can simulate the dual-mode cognitive process of humans, so as to synergistically leverage the respective advantages of association and reasoning, thereby comprehensively improving the comprehensiveness, adaptability and reliability of large models when dealing with complex, open-domain problems. Summary of the Invention
[0004] (a) Technical problems to be solved This invention aims to overcome the aforementioned shortcomings of existing technologies and solve the problem that current large models cannot dynamically coordinate creative association and rigorous logical reasoning according to task requirements. This invention provides a system and method capable of adaptively allocating computing resources and coordinating dual-mode computation, with the aim of significantly improving the creativity of models in open-domain tasks, their accuracy in reasoning tasks, and enhancing the transparency and interpretability of their decision-making process.
[0005] (II) Technical Solution To achieve the above objectives, the present invention adopts the following technical solution: In a first aspect, the present invention provides a large model system with both association and reasoning modes, characterized in that it includes: The pattern routing controller is used to analyze input information in real time and generate routing control signals; A dual-mode processing engine, connected to the mode routing controller, includes an associative processing module and an inference processing module. The associative processing module is configured to perform divergent information processing for open-domain problems, and the inference processing module is configured to perform convergent information processing for closed-domain problems. The co-integrator, connected to the mode routing controller and the dual-mode processing engine, is configured to integrate the outputs of the associative processing module and the inference processing module based on the strategy indicated by the routing control signal to generate the final output.
[0006] Preferably, the routing control signal is used to dynamically allocate computing resources between the associative processing module and the inference processing module or to indicate the processing flow. Further, the pattern routing controller analyzes input information to generate the routing control signal through at least one of the following methods: semantic intent recognition, syntactic complexity analysis, keyword matching, and task type classification.
[0007] Preferably, the association processing module and the inference processing module are implemented using any of the following architectures: a) Functional specialization is achieved using different fine-tuning strategies based on the same pre-trained large model; b) As two relatively independent neural network models; c) Within the same neural network model, it consists of a set of functionally specialized subnetworks or expert networks.
[0008] Preferably, the integration strategy employed by the collaborative integrator includes any one or more combinations of the following: Weighted fusion strategy: The outputs of the association processing module and the reasoning processing module are weighted at the vocabulary probability distribution level or the feature representation level; Hypothesis-verification loop strategy: The associative processing module generates multiple candidate outputs as hypotheses, and the reasoning processing module performs logical verification and scoring on each hypothesis, and selects or optimizes the final output based on the scoring results; Generate-Verify-Repair Strategy: The associative processing module generates an initial result, which is then verified by the inference processing module, which provides error feedback to guide the associative processing module in performing repair iterations.
[0009] Secondly, the present invention provides a large model method with both association and reasoning modes, characterized by comprising the following steps: S100: Analyzes input information through the mode routing controller and generates routing control signals for dynamic control processing modes; S200: Based on the routing control signal, the input information is allocated to the association processing module and the inference processing module for parallel or sequential processing; S300: Receive the intermediate output results generated by the association processing module and the inference processing module respectively; S400: Through the collaborative integrator, based on the integration strategy corresponding to the routing control signal, the intermediate output results are collaboratively processed to generate the final output.
[0010] Preferably, the collaborative processing includes a multi-round iterative process, in which the output of the associative processing module is provided to the inference processing module as a creative hypothesis for verification, and / or the output of the inference processing module is fed back to the associative processing module as a logical constraint to guide a new round of generation.
[0011] Preferably, the system or method is further configured to output decision information related to the routing control signal, and / or intermediate processing results of the association processing module and the inference processing module, to enhance the interpretability of the model decision-making process.
[0012] (III) Beneficial Effects Compared with the prior art, the technical solution provided by the present invention has the following significant advantages: Dynamic adaptation and precise control capabilities: By introducing a pattern routing controller, the system achieves real-time perception of the nature of the problem and on-demand allocation of cognitive patterns, enabling the model to "adapt to the specific problem," solving the rigidity problem of single-pattern processing and improving energy efficiency.
[0013] Breakthrough improvement in overall performance: Through the synergy and balance of association and reasoning modes, it can stimulate broader divergent thinking in creative tasks and ensure the rigor of logic and the accuracy of conclusions in reasoning tasks, thus achieving a joint enhancement of the model's capabilities in both "breadth" and "depth".
[0014] Flexible and scalable architecture: The architecture proposed in this invention does not depend on a specific model implementation, is compatible with various implementation methods such as shared parameters, independent modules and deep integration (such as MoE), and allows for flexible expansion of new integration strategies, with good versatility and foresight.
[0015] Transparent decision-making process and enhanced interpretability: By explicitly outputting routing decisions and intermediate results of dual modes, the model's "thinking process" becomes visible to users, greatly improving the model's reliability and debuggability, and providing a technical foundation for responsible AI. Attached Figure Description
[0016] To more clearly illustrate the technical solutions in the embodiments of the present invention or the prior art, the drawings used in the description of the embodiments or the prior art will be briefly introduced below. Obviously, the drawings described below are some embodiments of the present invention. For those skilled in the art, other drawings can be obtained based on these drawings without creative effort.
[0017] Figure 1 This is a system architecture block diagram provided in an embodiment of the present invention; Figure 2 This is a flowchart of a method provided in an embodiment of the present invention; Figure 3 This is a schematic diagram illustrating the workflow of a mode routing controller according to an embodiment of the present invention. Figure 4 This is a diagram illustrating the application of a strategy integrator provided in an embodiment of the present invention. Figure 5 A schematic diagram of the internal architecture of a system based on a hybrid expert model (MoE) provided for another embodiment of the present invention. Detailed Implementation
[0018] To make the objectives, technical solutions, and advantages of this invention clearer, the invention will be described in detail below with reference to the accompanying drawings and specific embodiments.
[0019] Example 1: Dual-mode system based on weighted fusion In this embodiment, the system is constructed as follows: Base model: The open-source model LLaMA-2-7B is used as the base. The dual-mode processing engine adopts the implementation approach of architecture a). The associative processing module fine-tunes the base model on datasets such as creative writing, poetry, and story generation, and encourages the use of diverse strategies such as Top-p sampling; the inference processing module fine-tunes the model using chain thinking (CoT) on mathematical problems and logical reasoning datasets, encouraging the model to generate progressive reasoning steps. Both share most of the model parameters but are functionally specialized through different fine-tuning strategies. Pattern routing controller: Employs a BERT-based classifier fine-tuned on a task classification dataset. This classifier categorizes the input question into "creativity," "reasoning," or "mixed" types and outputs a two-dimensional routing weight vector [α, β] as the routing control signal, where α represents the weight assigned to the associative pattern, β represents the weight assigned to the reasoning pattern, and α + β = 1. Collaborative Integrator: Employs a weighted fusion strategy. The final output is generated by weighted fusion of the outputs of the two modules at the probability distribution level, i.e., Output = α . P_a + β . P_r (where P_a is the vocabulary probability distribution output by the association module and P_r is the vocabulary probability distribution output by the inference module), and then the final text is generated through decoding strategies (such as greedy decoding and beam search).
[0020] Application Example 1: The input problem is "Write a five-character poem about autumn and parting." The routing controller analyzes it as a highly creative task and outputs a weight of [0.95, 0.05]. The system mainly relies on the associative processing module to generate poetic and imagery-rich phrases, while the reasoning module plays only a minor constraining role (such as ensuring conformity to the five-character format).
[0021] The input problem is "A pool has an inlet pipe and an outlet pipe... How many hours will it take to fill the pool if both pipes are open simultaneously?" The routing controller determines this as a mathematical reasoning task and outputs a weight of [0.1, 0.9]. The system primarily relies on the reasoning processing module for step-by-step mathematical calculations, with the associative module contributing very little.
[0022] Example 2: A dual-mode system based on a hypothesis-verification loop In this embodiment, the collaborative integrator adopts a "hypothesis-verification loop strategy," which is suitable for tasks that require innovative solutions but also need to meet strict constraints.
[0023] Step 1 (Association-Hypothesis): The routing controller identifies the task as "hybrid" and first activates the association processing module, which generates N (e.g., 3-5) possible solutions or creative hypotheses {H1, H2, ..., Hn}. Step 2 (Reasoning-Verification): The reasoning processing module is activated to perform multi-dimensional verification on each hypothesis Hi, including: logical consistency, factual compliance, and constraint satisfaction, and to give a comprehensive score S_i; Step 3 (Collaboration-Decision): The collaborative integrator sorts the hypotheses according to the score S_i and selects the highest-scoring hypothesis as the final output. Alternatively, it can guide the associative module to conduct a new round of refinement and divergence based on the highest-scoring hypothesis, starting the next "hypothesis-verification" cycle until predetermined conditions are met (such as the score exceeding a threshold or reaching the maximum number of iterations). Application Example: Input the command "Design a water-saving planting scheme suitable for urban balconies". The association module proposes multiple creative hypotheses, the reasoning module verifies their feasibility, and the collaborative integrator selects the optimal solution that balances feasibility, water-saving effect, and cost.
[0024] Example 3: A dual-mode system based on expert hybrid model (MoE) and gating network This embodiment demonstrates the implementation of architecture c), which deeply integrates the two modes into a single model.
[0025] System Architecture: A Transformer-based MoE model is used as the foundation, which contains a large number of functionally specialized "expert" feedforward neural networks (FFNs). Through training, these expert networks spontaneously differentiate into "association experts" who excel at association and creative expression, and "reasoning experts" who excel at logic and symbolic operations. Pattern Routing Controller: The gating network within the model acts as a fine-grained routing controller. This gating network computes a set of routing weights for each input token or input sequence. These weights precisely control which "association experts" and which "reasoning experts" are activated in the current computation, and the strength of their contributions. This is a concrete implementation of "routing control signals" within the model. Co-integrator: The integration process is completed automatically and implicitly within the model. During the forward propagation process of the MoE model, the outputs of the selected expert networks are dynamically weighted and summed according to the weights calculated by the gating network, thereby achieving seamless fusion and coordination of associative and inference information flows in each layer of the neural network; Workflow Example: Input sequence "Please design an environmentally friendly express delivery package". The model's gating network dynamically assigns weights. When generating "biodegradable mushroom mycelium material", high weights are assigned to the association expert to generate innovative concepts; when generating "its compressive strength needs to reach 50 kPa", high weights are assigned to the inference expert to introduce specific engineering parameters. The entire process is completed efficiently within a single model.
[0026] Example 4: Multi-round dynamic interaction for complex decision-making tasks This embodiment demonstrates the "multi-round iteration" method flow protected by claim 7, which is applicable to complex decision-making scenarios.
[0027] Application scenario: The user requests the model to "develop a marketing strategy for the company's next smartwatch".
[0028] System workflow: Round 1 (Divergent Dominance): Routing control weights are biased towards the association module [0.8, 0.2]. The association module generates multiple creative directions, and the reasoning module performs a rapid feasibility screening. The collaborative integrator outputs a list of filtered creative ideas and suggests proceeding to detailed evaluation; Round Two (Convergence Dominance): Routing control weights shift to the inference module [0.3, 0.7]. The inference module performs SWOT analysis and cost-benefit estimation on the optimal solution. The association module supplements the conceptual details within the inference framework.
[0029] The collaborative integrator integrates a comprehensive, data-driven, and logically sound promotion plan document. This process simulates the human "divergent-convergent" cyclical thinking process.
[0030] Example 5: A closed loop for code generation, verification, and repair. This example demonstrates the application of the "generate-verify-repair strategy" in a specific field, showcasing the advantages of multiple iterations.
[0031] System Construction: Associative processing module: trained to directly generate code snippets; Inference processing module: This is embodied as a code analyzer / compiler, used for static analysis, logic verification, and unit testing.
[0032] Workflow: The user entered: "Write a function to calculate the mode of a list." The Lenovo module generates candidate code A; The reasoning module (code analyzer) analyzes code A and finds that multiple modes have not been processed, and returns an error message; The co-integrator combines error messages with the original instructions and feeds them back to the Lenovo module; Lenovo module generates repaired code B; The inference module passed verification, and the co-integrator output the final code. This closed-loop process significantly improves the reliability of the output.
[0033] Example Comparison and Explainability Output The core comparisons of the various embodiments are as follows: Example 1: The core is weighted fusion, with the routing signal as the weight vector, making it simple and intuitive to implement; Example 2: The core is the hypothesis-verification cycle, which is suitable for innovative solution design; Example 3: The core is the internal integration of MoE, and the routing signal is the gated network weight to achieve fine-grained and highly efficient deep fusion; Example 4: The core is multi-round dynamic interaction, which demonstrates collaboration in the time dimension; Example 5: The core is generation-verification-repair, which is suitable for domains with clear rule constraints.
[0034] In all embodiments, the system can be configured to output interpretability information, such as: classification results or weight vectors ([0.8, 0.2]) output by the pattern routing controller, the strategy name adopted by the co-integrator, multiple candidate hypotheses generated by the association module, and validation scores or error reports given by the inference module. This information is presented in the form of logs or additional outputs, greatly enhancing the user's understanding and trust in the model's decision-making process.
[0035] The above description is merely a preferred embodiment of the present invention and is not intended to limit the scope of protection of the present invention. Any modifications, equivalent substitutions, improvements, etc., made within the spirit and principles of the present invention should be included within the scope of protection of the present invention.
Claims
1. A large-scale model system with both association and reasoning modes, characterized in that, include: The pattern routing controller is used to analyze input information in real time and generate routing control signals; A dual-mode processing engine, connected to the mode routing controller, includes an associative processing module and an inference processing module. The associative processing module is configured to perform divergent information processing for open-domain problems, and the inference processing module is configured to perform convergent information processing for closed-domain problems. The co-integrator, connected to the mode routing controller and the dual-mode processing engine, is configured to integrate the outputs of the associative processing module and the inference processing module based on the strategy indicated by the routing control signal to generate the final output.
2. The system according to claim 1, characterized in that, The routing control signal is used to dynamically allocate computing resources between the associative processing module and the inference processing module or to indicate the processing flow.
3. The system according to claim 1 or 2, characterized in that, The pattern routing controller analyzes input information to generate routing control signals through at least one of the following methods: semantic intent recognition, syntactic complexity analysis, keyword matching, and task type classification.
4. The system according to claim 1, characterized in that, The association processing module and the inference processing module are implemented using any of the following architectures: a) Functional specialization is achieved using different fine-tuning strategies based on the same pre-trained large model; b) As two relatively independent neural network models; c) Within the same neural network model, it consists of a set of functionally specialized subnetworks or expert networks.
5. The system according to claim 1, characterized in that, The integration strategy employed by the collaborative integrator includes any one or more combinations of the following: Weighted fusion strategy: The outputs of the association processing module and the reasoning processing module are weighted at the vocabulary probability distribution level or the feature representation level; Hypothesis-verification loop strategy: The associative processing module generates multiple candidate outputs as hypotheses, and the reasoning processing module performs logical verification and scoring on each hypothesis, and selects or optimizes the final output based on the scoring results; Generate-Verify-Repair Strategy: The associative processing module generates an initial result, which is then verified by the inference processing module, which provides error feedback to guide the associative processing module in performing repair iterations.
6. A large-scale model method with both association and reasoning modes, characterized in that, Includes the following steps: S100: Analyzes input information through the mode routing controller and generates routing control signals for dynamic control processing modes; S200: Based on the routing control signal, the input information is allocated to the association processing module and the inference processing module for parallel or sequential processing; S300: Receive the intermediate output results generated by the association processing module and the inference processing module respectively; S400: Through the collaborative integrator, based on the integration strategy corresponding to the routing control signal, the intermediate output results are collaboratively processed to generate the final output.
7. The method according to claim 6, characterized in that, The collaborative processing includes a multi-round iterative process. In the iteration, the output of the association processing module is provided to the reasoning processing module as a creative hypothesis for verification, and / or the output of the reasoning processing module is fed back to the association processing module as a logical constraint to guide the next round of generation.
8. The system according to claim 1 or the method according to claim 6, characterized in that, The system or method is further configured to output decision information related to the routing control signal, and / or intermediate processing results of the association processing module and the inference processing module.