Hybrid intelligent system architecture combining negative rule driving and probabilistic reasoning

Abstract

The invention discloses a hybrid intelligent system architecture combining rule driving and probabilistic reasoning and an operation method thereof. The system comprises a rule processing and debugging module, a probabilistic reasoning and candidate generation module, a hierarchical decision and fusion module and a semantic data packet management and circulation module. The system adopts a processing mechanism of'rule first and probability second ', logical reasoning and error elimination are carried out through a rule module, and then candidate results are generated and sorted in a limited space through a probability module. The semantic data packet is used for bearing information of each module. Model behaviors are constrained by using a digital natural normal form, so that the system improves the long text understanding ability, reduces the language model illusion occurrence rate, and optimizes the resource use efficiency.

Description

Technical Field

[0001] This invention belongs to the field of computer technology, specifically, to artificial intelligence (AI) systems, particularly a system architecture and information processing methods related to hybrid intelligence (Hybrid Intelligence) or neuro-symbolic AI. More specifically, the invention relates to a novel intelligent system architecture that integrates deterministic logical reasoning with uncertain probabilistic reasoning capabilities, coordinates the operation of various components within the system through a structured, semantically rich information exchange mechanism (i.e., "semantic data packets"), and methods for implementing intelligent information processing, decision support, and knowledge management based on this architecture. Background Art

[0002] Large language models (LLMs), based on the Transformer architecture, have demonstrated revolutionary capabilities and significant technological advancements in many areas of natural language processing (NLP), including text understanding, content generation, machine translation, dialogue systems, and intelligent question-answering (Analysis of Mainstream Large Language Model Architectures). These models are pre-trained on text data containing hundreds of billions or even trillions of tokens (A Review of Research on the Hallucination Problem in Large Language Models), thereby learning and internalizing rich linguistic patterns, syntactic structures, and world knowledge. However, despite the tremendous success and continued evolution of LLMs, they still face a series of inherent challenges and limitations in practical deployment and application. These issues, to varying degrees, hinder their reliability and practicality in critical, high-precision, high-risk, or resource-sensitive applications in finance, healthcare, and law.

[0003] Hallucination: LLMs are prone to hallucinations when generating responses, especially when responding to queries outside their knowledge boundaries, requiring complex multi-step logical reasoning, or when training data contains biased or contradictory information. These can manifest as fabricated facts, content inconsistent with input information or established facts, or output that appears grammatically and fluently reasonable but lacks logical coherence, is inconsistent with the context, or is completely meaningless. The root causes of this phenomenon are complex, primarily including inherent biases in the training data, internal model mechanisms (such as the limitations of the attention mechanism on long-range dependencies or a statistical fit to probability distributions rather than a true logical understanding), and an overreliance on generating the most probable, rather than the most accurate, sequences during reasoning. (References note that the causes of hallucination in large models can be categorized into data, training, and reasoning; combining knowledge graphs with large models can effectively mitigate hallucinations; and that models may generate seemingly reasonable content that contradicts common sense or facts.) Hallucinations seriously undermine the reliability of LLM output and user trust.

[0004] Context Window Limitation: Current mainstream LLMs have a fixed upper limit on the length of input text (i.e., context) that they can effectively process. This limit is typically measured in "tokens" and ranges from a few thousand to several hundred thousand. When processing tasks that require integrating information from long documents, conducting complex multi-turn conversational interactions, or capturing long-range dependencies, if the input exceeds the model's context window size, early, potentially critical information may be truncated, forgotten, or diluted by the attention mechanism. This can lead to a biased and incomplete understanding of the overall context, resulting in a lack of coherence in the generated text, or the loss of important decision-making basis. (The reference mentions that limiting the context length is a challenge; simply submitting all relevant documents to the LLM at once may exceed the context window limit.)

[0005] High Resource Consumption: Both LLM training and inference processes place extremely high demands on computing and storage resources. Training a large LLM typically requires massive (terabytes or even petabytes) of high-quality datasets, supported by hundreds or even thousands of high-performance computing units (such as GPUs or TPUs), and training cycles can take weeks to months. Even during inference (deployment), achieving acceptable response speeds requires powerful hardware support. This significant resource consumption not only leads to high model development and maintenance costs, but also limits its widespread deployment and adoption in resource-constrained environments (such as edge devices) or cost-sensitive applications. (References note that the datasets required for large model training can reach hundreds of billions of datasets, and increasing model size requires even more resources.)

[0006] Heavy reliance on corpora: An LLM's knowledge base, reasoning capabilities, and even potential biases are largely derived directly from the vast corpora used for pre-training. If the corpus lacks coverage in certain knowledge domains, contains outdated information, exhibits systematic biases, or fails to adequately represent diverse perspectives, the model's performance in these areas will significantly decline, and even lead to erroneous perceptions. For low-resource languages, highly specialized knowledge domains, or the nuanced semantics of specific cultural contexts, acquiring and constructing high-quality, large-scale, and unbiased corpora is extremely difficult and costly. This limits the LLM's adaptability in these scenarios, often requiring additional, costly, domain-specific fine-tuning to achieve a usable level. (The reference mentions: Model knowledge primarily comes from pre-training data, and data biases lead to knowledge errors; insufficient training data leads to incorrect predictions about unseen situations.)

[0007] Difficulty in Semantic Capture: Although LLM performs well in surface semantic understanding (such as lexical meaning and basic syntactic structure), it still faces challenges in dealing with deeper and more complex semantic phenomena. This includes accurately understanding the logical relationships in complex syntactic structures, identifying and resolving the polysemy of words, capturing the deep implicit meaning in the text (such as irony, metaphor, humor), and understanding the subtle semantic differences and implications in specific contexts. These difficulties limit the effectiveness and reliability of LLM in applications that require highly accurate semantic understanding and deep logical reasoning (such as legal text analysis, complex contract review, and scientific literature understanding). (The reference mentions: Neural symbolic AI can refine semantic processing and enhance contextual understanding and logical reasoning capabilities).

[0008] Existing technical approaches to address these issues primarily include: improving performance by continuously expanding model parameter size and training datasets; and employing techniques such as Retrieval-Augmented Generation (RAG) to incorporate external knowledge bases for generation assistance (large models + RAG, a comprehensive introduction!). However, simply increasing model size further increases resource consumption and training difficulty. While methods like RAG can mitigate hallucinations and knowledge limitations to a certain extent, their effectiveness is limited by the accuracy and efficiency of the retrieval module, potentially introducing new noise, and failing to fundamentally address the model's inherent logical reasoning flaws and limitations in context processing. In recent years, neuro-symbolic AI (Neuro-Symbolic AI - The Alan Turing Institute), a paradigm that combines the learning capabilities of neural networks with the reasoning capabilities of symbolic systems, has garnered widespread attention. It offers a new approach to addressing the fundamental flaws of LLM. However, designing an efficient, deeply integrated hybrid intelligence architecture that systematically addresses these multiple pain points remains a key and challenging area of ​​current AI research. Summary of the Invention

[0009] The present invention aims to provide an innovative hybrid intelligent system architecture and its operation method, with the core goal of overcoming a series of key technical bottlenecks commonly found in existing large language models in practical applications. Specifically, the present invention is committed to systematically solving the following problems: 1. Alleviating or even eliminating the illusion phenomenon in the model output, and improving the factual accuracy and logical consistency of the generated content. 2. Effectively expanding the model's ability to understand and process long-range contexts, breaking through the limitations of existing fixed context windows, and ensuring the coherence and integrity of information in long-sequence processing. 3. Significantly optimizing the system's computing and storage resource consumption, reducing the training and inference costs of LLMs, and improving their deployment feasibility in various environments. 4. Alleviating the system's over-reliance on large-scale, high-quality, unbiased pre-training corpora, and enhancing its adaptability and robustness in scenarios with few samples, professional fields, or low resources. 5. Improving the system's accuracy and depth in capturing deep semantics, subtle semantics, and implicit meanings in complex scenarios, and enhancing its ability to perform complex logical reasoning and precise understanding. The present invention achieves the above goals through a novel mechanism that deeply combines rule-driven deterministic reasoning with uncertainty exploration of probabilistic models, thereby providing a more reliable, efficient, economical and intelligent AI solution.

[0010] To address the aforementioned technical issues, the present invention discloses a hybrid intelligent system architecture that combines rule-driven and probabilistic reasoning. The core design philosophy of this architecture draws on the efficient processing mechanism used in human cognitive learning to quickly eliminate incorrect options and focus on the feasible solution space, as well as the concept of precise information flow directed by routing tables in network dynamic routing protocols. This approach enables the system to achieve the powerful guidance and constraints of deterministic rules, organically combining and complementing the advantages of probabilistic models in recognizing uncertainty and complex patterns. The proposed system primarily consists of the following core modules.

[0011] Rule Processing and Error Exclusion Module: This module serves as the system's deterministic gatekeeper. Based on a predefined, dynamically updateable rule base, it performs rapid, deterministic logical reasoning, pattern matching, constraint checking, and obvious error elimination on input information or intermediate processing results circulating within the system. Its primary responsibility is to significantly narrow the solution space for subsequent processing, filtering out paths and options that are clearly unreasonable, illogical, or violate established knowledge. This provides a purified and preliminarily screened subset of high-quality, highly relevant information for the subsequent probabilistic reasoning module. The design of this module draws inspiration from the brain's "lazy system" and inhibitory mechanism hypothesis, discussed in the "Large Model Rule-Driven Design Approach" document. This module prioritizes and rapidly identifies and eliminates "erroneous" options to improve overall processing efficiency. Human learning is essentially a process of communication with nature. Nature is a vast generator of reactions, constantly informing us of what will happen under certain circumstances. The brain rapidly develops rules and experiences to facilitate avoidance, reasoning, and prediction. This invention aims to return the right to define right and wrong to nature, allowing the system to adapt to learning and thus solve the directional problems of existing large models.

[0012] Probabilistic Inference and Candidate Generation Module: This module serves as the system's "uncertainty explorer." It typically embeds one or more optimized and adapted deep learning models (e.g., a domain-specific fine-tuned variant of a large language model). Within the "remaining compliance space" filtered and constrained by the rule processing module, this module performs in-depth uncertainty analysis, sophisticated pattern recognition, and detailed semantic understanding. Based on this information, it probabilistically generates one or more candidate answers, action plans, or text sequences, and assigns corresponding confidence or probability scores to these candidate solutions.

[0013] Hierarchical Decision and Fusion Module: This module is the system's "decision-making hub." It strictly adheres to the hierarchical decision-making mechanism of "rules first, probabilities second." Based on pre-defined fusion strategies (e.g., rule-based priorities, probability confidence thresholds, dynamic weighting of information entropy, or complex gating mechanisms), it carefully integrates the deterministic outputs / strong constraints from the rule processing module with the probabilistic candidate solutions from the probabilistic reasoning module. The core mission of this module is to ensure that the system, while pursuing the probabilistically optimal solution, never violates core rules, domain knowledge, and common sense, thereby generating a final, reliable decision or output.

[0014] Semantic Data Packet Management and Flow Module: This module serves as the "transportation system" for the efficient and orderly flow of information within the system. It defines a standardized, structured data unit—semantic data packets—and constructs, parses, validates, and efficiently transmits these data packets between system modules. These semantic data packets, the core vehicle for information exchange within the system, encapsulate critical information such as the contextual information required for task processing, precise semantic labels, identified entities and their relationships, current module processing status, and inter-module control instructions. They are the foundation for modular decoupling and complex process management.

[0015] Knowledge and Rule Dynamic Update Module: This module empowers the system to continuously learn and evolve. Based on feedback collected from various modules during operation (such as decision success, user feedback, and external real-world verification results), newly acquired external knowledge input (such as updated databases and new industry standards), and new knowledge connections discovered through the internal "collision algorithm," it dynamically adjusts, optimizes, and expands the content of the rule and knowledge bases, and even fine-tunes the parameters of the probability model.

[0016] The technical principle of combining rule-driven and probabilistic selection: The technical principle of combining rule-driven and probabilistic selection proposed in this invention aims to build a hybrid intelligent system that leverages the precision of symbolic logic and the pattern recognition capabilities of neural networks. Its core operating mechanism is as follows: 1. Preemptive Rule-based Filtering and Space Pruning: When the system receives input information (or intermediate output from the previous round during iterative processing), it is first fed into the rule processing and troubleshooting module. This module applies the rules in its rule base. These rules can be formal logical rules (such as first-order logic and description logic), domain-specific knowledge constraints (such as industry regulations and physical laws), heuristic rules (such as best practices derived from experience), or common sense judgments. Through an efficient rule matching and inference engine, this module quickly eliminates a large number of obviously incorrect, irrelevant, or unreasonable possibilities. This process can be understood as first subtracting a set of "clearly incorrect" options from the vast potential solution (entropy) space. This mechanism draws on the principle of dynamic routing protocols in network communications (such as OSPF and BGP), where the routing table can quickly determine which paths should not be used for packet forwarding (a large-model rule-driven design approach). It also mimics the human brain's high sensitivity to erroneous information and its rapid suppression mechanism when faced with a problem, thereby avoiding wasting cognitive resources on invalid paths. 2. Refined Probabilistic Model Selection within the Residual Space: After preliminary screening and strong constraints at the rule layer, the solution space for the original problem is significantly narrowed and purified. At this point, the focus shifts to the "residual space" that is more likely to contain the correct answer. The probabilistic reasoning and candidate generation module operates within this space, which is clearly defined by the rule system and has a higher probability of containing the correct answer. This module leverages the powerful capabilities of its deep learning models (such as LLM) to perform more detailed probability calculations, pattern recognition, and candidate answer generation on the remaining options or information. This step primarily handles complex patterns, subtle semantic nuances, and inherently uncertain aspects of the problem that are difficult for rules to accurately cover. By prioritizing rules, the probabilistic model can focus more on valuable computations and reduce inefficient exploration of vast but irrelevant spaces. 3. Hierarchical Decision-making and Iterative Optimization: The final decision is made by the hierarchical decision-making and fusion module.This module comprehensively considers the outputs from the rule module (such as the directly provided deterministic solution, strong constraints on the solution, and the confidence or priority of the rule itself) and the outputs from the probability module (such as the probability distribution of multiple candidate answers and their respective confidence scores). The decision-making logic follows the principle of "rules first, probabilities second": If the rule layer can provide a highly deterministic solution that fully meets the task requirements, the system may directly adopt it. If the rule layer only provides constraints or a partial solution, the output of the probability model is evaluated within these constraints, and the solution with the highest probability or the best overall score is selected. Furthermore, as described in "Large Model Rule-Driven Design Principles," the system can incorporate a "looping and debugging" mechanism: When faced with multiple plausible possible paths, the system does not stop at a single choice. Instead, it can randomly select one or select a path based on a specific strategy (trial and error), then re-initiate the debugging mechanism (rule verification and probability evaluation). Through this iterative cycle, it gradually approaches the global optimal or most reliable solution. The "Collision and Update" feedback learning loop (Feedback Loop of "Collision and Update") To achieve dynamic knowledge growth and continuous improvement in reasoning capabilities, this system introduces a bidirectional reasoning mechanism, similar to "from concrete to abstract, and back again," that seeks convergence points. This mechanism, known as a "collision algorithm," originates from the principle of large-scale model rule-driven design. Specifically, the system may simultaneously maintain two reasoning paths: one starting from specific input information or low-level features, abstracting and generalizing upward, while eliminating obviously unreasonable combinations and inferences in the process (bottom-up); the other starting from high-level abstract rules, concepts, or hypotheses, concretizing and instantiating downward, attempting to match and explain specific details with existing knowledge systems (top-down). When these two reasoning paths "collide" at an intermediate knowledge level or representation level—that is, when a consistent, mutually corroborative connection point or explanation is found—this typically indicates effective knowledge discovery, verification, or the formation of a deeper understanding. The results of this "collision" (such as new associations, confirmed rules, or disproven hypotheses) are fed back to the system through the knowledge and rule dynamic update module, used to update the existing knowledge base and rule base, and even adjust the weights of the probability model, thus forming a closed-loop learning and evolutionary process.Theoretical Underpinnings: This technology draws heavily on the core principles of Neuro-Symbolic AI, which aims to combine the powerful pattern recognition and learning capabilities of neural networks with the precise logical reasoning, knowledge representation, and interpretability of symbolic AI systems (e.g., logic- and rule-based systems). (Neuro-Symbolic Artificial Intelligence: Integrating Learning and Reasoning; Neurosymbolic AI: Bridging Neural Networks and Symbolic...) The goal is to create AI systems that can both learn from large amounts of data and perform rigorous and transparent reasoning. Furthermore, the design philosophy draws heavily on the user-contributed "Large Model Rule-Driven Design Approach" document, which discusses dynamic routing protocols and assumptions about how the brain works (particularly the "lazy system," inhibition mechanisms, focusing on correct answers by eliminating errors, and the metaphor of entropy increase and excess space). National scientific research plans also point out the importance of such hybrid methods. For example, the "2024 Project Guidelines for the Major Research Program on Interpretable and Universal Next-Generation Artificial Intelligence Methods" issued by the National Natural Science Foundation of China clearly proposes to "develop artificial intelligence methods that combine logical reasoning, knowledge and rules, and establish an artificial intelligence theoretical framework that is interpretable and universal."

[0017] System Architecture: The hybrid intelligent system architecture proposed in this invention aims to achieve a deep integration of rule-driven and probabilistic reasoning through a modular design. Its typical components and the interaction processes between them are as follows: · Input Interface: Serving as a bridge between the system and the external environment, it is responsible for receiving various inputs, such as user natural language queries, structured or unstructured data streams, and API calls from other systems. · Semantic Data Packet Construction Unit: Receives raw information from the input interface, performs preliminary parsing and standardization, and encapsulates it into a "semantic data packet" that circulates uniformly within the system. This unit is responsible for initializing the key fields of the data packet. · Rule Processing and Error Exclusion Module: · Rule Base: Stores a large amount of deterministic knowledge, including formal logic rules (such as if-then statements), business constraints, domain ontology knowledge, heuristic strategies, common sense libraries, and error pattern libraries. Rules can be predefined by experts or dynamically generated and optimized by the system through learning. The Rule Engine is the core of rule execution, responsible for retrieving applicable rules from the rule library, matching, reasoning, validating, and executing pre-defined error-correction actions on input semantic data packets. The Action Vector Library (derived from the large-model rule-driven design approach) works with the rule engine in specific implementations to guide agents or processing units to select the next inference path or execute specific actions based on the current rule matching results and the state of the semantic data packet. The Probabilistic Inference and Candidate Generation Module (LICM) is a pre-trained Optimized Large Model Core (LLM Core), typically consisting of one or more large language models (or other deep learning models). These models are pre-trained and may be fine-tuned and optimized for specific tasks or domains to suit the needs of the system (for example, to enhance their ability to follow instructions or reduce specific types of hallucinations). The Context Manager extracts and manages valid contextual information from incoming semantic data packets, providing it to the LLM core in a format that the probabilistic model can understand and utilize (e.g., prompt engineering or embedding vector sequences). It needs to process the constraints passed by the rule module.Hierarchical Decision and Fusion Module: The Decision Logic Unit (DLU) contains one or more decision algorithms, implementing a "rules first, probabilities second" decision-making process. It evaluates the outputs from the rule and probability modules and synthesizes the final decision or output based on fusion strategies (such as voting, weighting, prioritization, and confidence thresholds). Semantic Data Packet Management and Flow Module: The Semantic Data Packet Bus (SDPB) provides a standardized communication mechanism to ensure efficient, orderly, and reliable delivery and routing of semantic data packets between modules within the system. Global Link State Pool (optional, derived from a large-model rule-driven design): In distributed or multi-agent implementations, this component can serve as a shared, real-time database, storing the current status of key components of the entire agent network or system (such as load, availability, processing queue length, and knowledge update versions). Each module can query this information to assist in its own decision-making (such as task allocation and path selection). Knowledge and Rule Dynamic Update Module: The Learning and Feedback Unit collects performance data, user feedback, external validation results, and valid knowledge connections generated by the collision algorithm during system operation. Based on this information, this unit executes learning algorithms (such as reinforcement learning, online learning, and rule induction) to adjust the content of the rule base, rule priorities, or fine-tune parameters of the probabilistic model. Agent Operation Mechanism (optional, used for complex distributed implementations, derived from the large-model rule-driven design approach): In some architectures, system functions may be carried out collaboratively by multiple agents. Each agent can play one or more roles and load the corresponding action and task libraries based on its semantic labels. Agents operate on received semantic data packets according to rules or update their own configuration (such as the action and task libraries) based on processing results and feedback. Output Interface: Responsible for delivering the system's final decision results, generated text, executed action status, or other processing results to users or external systems in an appropriate format. Data Flow and Interaction Logic: The typical processing flow begins with the input interface receiving an external request. The request is encapsulated into an initial semantic data packet by the semantic data packet building unit.This packet is first routed to the Rule Processing and Debugging module, which analyzes the packet content according to the rule base, performs debugging operations, and may preliminarily tag semantic information, add constraints, and then update the packet. Next, the semantic packet carrying the rule processing results (such as constraints) is passed to the Probabilistic Reasoning and Candidate Generation module. Within its defined "compliant space," this module utilizes the LLM core to perform deep semantic understanding and generate candidate answers, populating these probabilistic results back into the packet. The packet then enters the Hierarchical Decision and Fusion module. Based on pre-defined decision logic and fusion strategies, this module integrates the deterministic judgments of the rule module and the candidate recommendations of the probabilistic module to make a final decision or generate the final output, and further update the packet. Finally, the processed semantic packet is presented to users or downstream systems through an output interface. Throughout this process, the Knowledge and Rules Dynamic Update module continuously monitors system performance and external feedback. Based on the results of mechanisms such as the "collision algorithm," it iteratively optimizes the rule base and model parameters, forming a closed learning loop. Architecture Design Rationale: This system architecture adheres to key software engineering principles such as modular design, hierarchical processing, and separation of concerns. Its core components and interaction logic are primarily inspired by the functional modules proposed in the user-provided "Large Model Rule-Driven Design Ideas" document. It also draws on various architectural patterns already established in the field of neuro-symbolic AI (e.g., serial processing—pre-processing by the symbolic system followed by the neural model, or verifying the neural model output by the symbolic system; parallel processing; and more tightly coupled fusion models based on attention mechanisms or shared internal representations) (Unlocking the Potential of Generative AI through Neuro-Symbolic…). This architecture is unique in its emphasis on pre-deterministic rule debugging, hierarchical decision-making mechanisms, and structured semantic data packets as the core information carrier, all of which are designed to specifically address the inherent shortcomings of LLM. The decision-making mechanism in this invention is a multi-layered, dynamically adjusted process designed to ensure the accuracy, robustness, and compliance of the system's output. It operates as follows: 1. Triggering Conditions and Inputs: The decision process is triggered by a semantic data packet containing the task to be processed. The data packet should at least include: task type (such as factual question answering, text generation, risk assessment), current context information (obtained from the ContextPayload field), the user's original input (in the PrimaryContent^ field), and preliminary constraints or labels that may be imposed by the previous module (in fields such as RuleConstraints or SemanticLabel).2. Decision-making Process: The decision-making mechanism strictly adheres to the hierarchical principle of "rules first, probabilities later": a. Rule-First Evaluation: The semantic data packet is first submitted to the rule engine of the rule processing and troubleshooting module for evaluation. If high-priority, highly deterministic rules are matched, and these rules directly lead to a clear and complete final decision (for example, answering a factual question that can be directly retrieved from the knowledge base, or executing a predefined, unambiguous instruction), the decision and its usually high confidence (or certainty mark) are recorded and populated in the corresponding field of the semantic data packet. At this point, if the rule certainty is sufficient to meet the task requirements, the subsequent probabilistic reasoning steps may be skipped or used only as auxiliary verification. If only some rules are matched, or the rules are restrictive rather than prescriptive in nature (for example, eliminating certain inappropriate options, specifying certain conditions that the final answer must meet, or setting the target direction or style of content generated by the probabilistic model), the rule module clearly encodes these constraints in the RuleConstraints field of the semantic data packet and passes the packet to the next stage. b. Probabilistic Enhancement (if Needed): If the rules don't directly address the problem, or the task requires richer semantic understanding, creative generation, or the ability to handle uncertainty (for example, writing a summary report, performing sentiment analysis, or predicting future trends), the semantic data packet, processed by the rule module and containing explicit constraints, is passed to the Probabilistic Reasoning and Candidate Generation module. The probabilistic model (the core of the LLM) performs reasoning and content generation within the framework of the rules (i.e., the constraints specified in `RuleConstraints`). For example, if a rule requires that generated content must contain specific keywords and not exceed a certain length, these conditions will guide the LLM's generation process. The probabilistic model outputs one or more candidate solutions (e.g., different versions of text, multiple possible answer options), each accompanied by its internally calculated confidence score or probability value. This information is organized and populated into the `ProbabilisticOutput` field of the semantic data packet. c. Fusion and Final Decision: The Hierarchical Decision and Fusion module receives a semantic data packet containing the deterministic outputs / constraints from the rule module and the candidate solutions and their confidence scores from the probabilistic module. Consistency Check: First, the module rigorously checks that all candidate solutions generated by the probability model are fully compatible with the RuleConstraints imposed by the rule module. Any incompatible candidate solutions are typically excluded or have their confidence significantly reduced.If all candidate solutions are incompatible, the system may trigger a re-inference of the probabilistic model (possibly adjusting prompts or parameters) or flag the conflict, indicating a potential underlying inconsistency between the rules and the model's understanding. Selection and Integration: From the candidate solutions that pass consistency checks, the decision module selects the final decision based on a pre-defined fusion strategy. This strategy can be diverse: If the rule module produces a deterministic output that is consistent with a highly confident candidate solution from the probabilistic model, that solution is favored. If there is a certain degree of conflict between the two (but within the permitted limits of the rules), a dynamic weighted average or voting selection can be performed based on the priority and confidence of the rules and the confidence of the probabilistic model. If the rules impose very strong constraints (e.g., "the answer must be one of A, B, or C"), the system selects only the most confident candidate probabilistic model that meets these constraints. In some cases, rules may be used to perform post-processing corrections to the probabilistic model output (e.g., format normalization, sensitive word filtering). Iterative Refinement through Error Exclusion: When handling particularly complex tasks or when confidence in an initial decision is low, the system can initiate an iterative cycle (see "Large Model Rule-Driven Design"). It may select one of multiple seemingly plausible paths (candidate decisions) through random selection or a specific strategy (such as an exploratory strategy) for a trial run or further in-depth simulation / evaluation. The results of this trial run are then fed back to the rule processing module and probabilistic reasoning module for another round of error correction and evaluation. Through this "try-verify-correct" cycle, suboptimal solutions are gradually eliminated, and a more reliable final solution is approached. 3. Key Decision Parameters: The effective operation of the decision-making mechanism depends on the appropriate setting of a series of key parameters. These parameters may include: the priority of different rule types (e.g., compliance rules take precedence over advisory rules), the confidence threshold for rule triggering, the minimum confidence threshold for adopting the probabilistic model output, the fusion weight coefficient when rules conflict with probabilities, and the termination condition for the iterative loop (e.g., reaching the maximum number of iterations or exceeding a certain high threshold for the decision confidence). These parameters can be set offline through an expert system, optimized through machine learning based on historical performance data, or adaptively adjusted online during system runtime based on the characteristics of the current task. Theoretical Basis: This decision-making mechanism draws on the negotiation and collaborative decision-making mechanisms in multi-agent systems (MAS), specifically how agents with different capabilities (the rule module and the probability module can be considered two different types of agents) reach a common goal through information exchange and certain agreements. Furthermore, the constraints imposed by rules on the probabilistic model are conceptually similar to the satisfaction of constraints in mixed-integer programming problems in operations research.The process of iterative error correction and approaching the optimal solution shares similarities with the policy evaluation and improvement cycle (Policy Iteration / Value Iteration) and trial-and-error learning in reinforcement learning. As emphasized in the "2024 Project Guidelines for the Major Research Program on Universal Next-Generation Artificial Intelligence Methods," developing "artificial intelligence methods that combine logical reasoning, knowledge, and rules" is a key direction for building next-generation AI. This decision-making mechanism embodies this philosophy.

[0018] The semantic data packet: The "Semantic Data Packet" (SDP) is the core mechanism for information flow and inter-module collaboration in the hybrid intelligent system architecture proposed in this paper. It is a standardized, self-contained, structured data unit, carefully designed to efficiently, accurately, and unambiguously transmit contextual information, semantic representations, processing instructions, and real-time status information between modules within the system. The introduction of the SDP aims to achieve loose coupling between system components, ensuring the orderly flow and consistent understanding of information within complex processing flows. It is a key support for the advantages of this architecture. Definition: A semantic data packet is more than just a simple data container; it is more like a work unit carrying rich "metacognitive" information. It encapsulates all relevant information required for the current task at a specific processing stage and records its progress and results within the system. Its design draws on the initial concept of the semantic data packet in the "Large Model Rule-Driven Design Approach" and has been expanded based on the completeness of the patent application and actual application requirements. Core Structure: A typical semantic data packet may contain the following fields (which may be added, deleted, or adjusted according to specific implementation requirements): Function and value, especially the shortcomings of LLM: Overcoming context window limitations: Through a carefully designed ContextPayload field, semantic data packets can dynamically encapsulate and deliver the most important context summaries for the task at hand, embedding vectors for key information points, or pointers to external long-term memory, rather than simply delivering the entire original long text. This enables the system to "indirectly memorize" and effectively utilize critical historical information that far exceeds the physical context window length of the LLM core, resulting in improved performance in long document analysis, complex multi-turn dialogues, and long-term dependency reasoning tasks.

[0019] Improved semantic consistency and accuracy: Structured fields such as Entities, Relations, and SemanticLabel ensure high consistency and clarity when core semantic elements are transferred across system modules. This reduces distortion, ambiguity, or loss of information during transmission and conversion, helping the probabilistic model (LLM) more accurately capture deeper semantics and reduce misunderstandings.

[0020] Promotes module collaboration and standardized information flow: The standardized data packet format greatly simplifies the interface design between modules within the system and reduces the coupling between them. This makes information flow more efficient, controllable, and easy to trace. Fields such as SourceModule, TargetModule, and ProcessingState facilitate monitoring processing flow, debugging system behavior, and implementing complex routing logic.

[0021] Support for precise transfer and enforcement of deterministic constraints: The RuleConstraints field is the key channel for the rule module to effectively transfer its deterministic knowledge and hard constraints to the probabilistic module. This allows the generation of probabilistic models to be clearly guided and constrained, significantly reducing the generation of hallucinations and avoiding the generation of irrelevant or non-compliant outputs.

[0022] Optimizing computing resource utilization: Pre-screening input information by the rule processing module and precisely targeting information delivery using semantic data packages prevents the input of large amounts of irrelevant, redundant, or low-value information to the computationally intensive probabilistic model (LLM core). This directly reduces unnecessary computational load on the LLM core, helping to lower overall system inference latency and operating costs.

[0023] Theoretical Support and Similar Concepts: The design concept of semantic data packets is conceptually similar to data packets in computer network protocol stacks (such as TCP / IP packets). These packets encapsulate routing instructions, sequence numbers, checksums, and other information through standardized header information, ensuring accurate and efficient data transmission across the network. Furthermore, its structured representation of semantic elements such as entities and relationships draws on the organization of nodes and edges in knowledge graphs and the description of resources, their attributes, and relationships in semantic web technologies (such as RDF and OWL), aiming to make information easier for machines to understand and process. The design of its core fields draws on the initial conception of "semantic data packets" in the user-provided "Large Model Rule-Driven Design Approach," which included "semantic labels," "content," "sender," and "receiver," and has been significantly enriched and refined.

[0024] Beneficial Effects: Compared to existing technologies (primarily traditional large language models that rely solely on probabilistic models and their simple enhancement methods, such as RAG), the hybrid intelligent system architecture and operational method proposed in this invention, combining rule-driven and probabilistic reasoning, is expected to achieve one or more of the following significant benefits: 1. Significant Reduction in Hallucinations: The rule processing and error correction module verifies input information and potential LLM outputs against factual knowledge, logical consistency, and common sense judgment, proactively filtering or correcting a large amount of content that is inconsistent with facts or contains obvious logical contradictions. Furthermore, the hierarchical decision-making and fusion module performs rule-based cross-validation and screening on multiple candidate solutions generated by the probabilistic model. This "double insurance" mechanism effectively suppresses the model's tendency to generate inaccurate, irrelevant, or meaningless text, significantly reducing the incidence of hallucinations and improving the credibility of the output content. (See "Enhancing Large Language Models through Neuro-Symbolic..." for examples of how neuro-symbolic methods enhance the consistency and reliability of LLMs.) 2. Effective Extension of Long-Context Processing Capability: Traditional LLMs are limited by their fixed physical context window size. The semantic data packet in this invention, through its structured ContextPayload field, can dynamically manage and deliver key context summaries most relevant to the current task, embedded representations of historical information, or pointers to external knowledge sources. This means the system no longer needs to feed the entire, lengthy raw sequence to the LLM at once. Instead, through a method of "intelligent compression" and "on-demand extraction," the LLM can indirectly "perceive" and utilize effective information far beyond its physical window. This is crucial for understanding long documents, conducting deep multi-turn conversations, and capturing long-range dependencies. (Some ideas are borrowed from RAG, but further optimized through rules and structured context. For example, RAG is one solution to address the context window limitation, and this invention provides a deeper integration.) 3. Optimization of Computational Resource Consumption and Inference Efficiency: Rule processing modules are typically based on efficient logical matching and symbolic operations, resulting in computational costs far lower than the complex neural network operations of large probabilistic models. Through rule pre-screening, error correction and constraint imposition, the amount of data and processing complexity entering the computationally intensive probabilistic reasoning and candidate generation module (LLM core) can be greatly reduced.The precise information transfer of semantic data packets also avoids inputting large amounts of irrelevant or redundant information into the LLM module. These mechanisms work together to significantly reduce the frequency of calls to expensive LLM cores and the inference load of each call, while maintaining or even improving task performance. This optimizes overall system computing resource consumption and reduces inference latency. (This draws on the "lazy system" concept from the "Large Model Rule-Driven Design" approach to avoid unnecessary expensive computation.) 4. Reduced Dependency on Large-Scale Perfect Corpora: Traditional LLMs rely heavily on the quality and coverage of their pre-trained corpora for knowledge and capabilities. By introducing a rule base, this invention enables the system to directly encode, store, and utilize precise domain knowledge, business logic, and deterministic causal relationships. This, to a certain extent, compensates for potential knowledge gaps, biases, or outdated information in pre-trained corpora. Consequently, this system demonstrates greater adaptability and robustness in few-shot learning scenarios, highly specialized domains (which often lack large-scale annotated data), and low-resource language environments, reducing its absolute reliance on massive, high-quality, and completely unbiased training data. 5. Enhanced Capability in Capturing Complex and Subtle Semantics: Semantic data packets, through their refined structures (e.g., entities, relations, and semantic labels), provide a clear semantic framework for information. The rule-based module can leverage these structures to perform precise logical judgments and relational reasoning. Assisted by this structured information and guided by rules, the probabilistic model can more accurately parse complex syntactic structures, identify underlying implicit intent, effectively resolve polysemy, and be more sensitive to subtle semantic variations in context (e.g., emotional tone and intensity). The synergy between rules and probabilistic models enables the overall system to surpass the depth and accuracy of single-model semantic understanding. 6. Improved System Interpretability and Controllability: Because the rule-based processing module operates based on a clear, human-readable (or at least understandable) set of rules, some of the system's decision paths and behavioral logic become more transparent and explainable. When the system output is wrong or not as expected, it is easier to trace back to which rule or rules caused the problem, or the possible conflict points between the rules and the probabilistic model.At the same time, by directly modifying, adding, or deleting rules, maintainers can more directly and effectively control and intervene in the system's behavior to meet specific business needs, security constraints, or ethical considerations, without having to retrain or complexly fine-tune the large and elusive LLM. 7. Enhanced Modularity, Maintainability, and Scalability: The system architecture's clear modular division (rule processing, probabilistic reasoning, decision fusion, etc.) and standardized semantic data packet interface design naturally support a high degree of modularity. This means that different modules in the system can be independently developed, tested, optimized, and upgraded with minimal impact on other modules. For example, the LLM core can be easily replaced or upgraded, or the rule base can be expanded for specific domains without reconfiguring the entire system. This modularity also facilitates long-term maintenance and functional expansion of the system. The feasibility and effectiveness of the technical solution provided by this invention are not only theoretically supported by the core concepts of the "Large Model Rule-Driven Design Approach" document, but also highly consistent with current research trends in the field of artificial intelligence, particularly in cutting-edge areas such as neural symbolic computing, knowledge graph applications, and hybrid intelligent systems. For example, the National Natural Science Foundation of China's "Major Research Program for Explainable and Interpretable Next-Generation Artificial Intelligence Methods" explicitly emphasizes the importance of developing AI methods that combine logical reasoning, knowledge, and rules to overcome the current limitations of deep learning. Preliminary prototype system validation demonstrates that the proposed architecture demonstrates statistically significant performance improvements and defect reductions compared to traditional LLM approaches in terms of output accuracy, resource efficiency, and the ability to follow complex instructions when handling specific tasks.

Claims

1. A hybrid intelligent system architecture combining rule-driven and probabilistic reasoning, characterized by: include: A rule-driven processing module for deterministic logical reasoning, fact checking, preliminary decision-making, and task planning based on a formal rule base; The probabilistic reasoning module is used for reasoning, pattern recognition, and natural language understanding and generation in uncertain environments based on statistical learning models. The hybrid coordination and decision-making module is used to receive, integrate, and evaluate the outputs from the rule-driven module and the probabilistic reasoning module, and make final decisions based on preset strategies or dynamic evaluations. The semantic data packet interaction bus serves as a unified carrier and communication protocol standard for information exchange between modules within the system. The knowledge and rule management module is used for persistent storage, dynamic updating, version control, efficient retrieval, and management of various types of knowledge and rules required for system operation.

2. The hybrid intelligent system architecture according to claim 1, characterized in that: The rule-driven processing module includes a rule engine and a rule base interface. The rule engine adopts an optimization algorithm based on Rete network or a modern rule engine to achieve efficient rule matching and execution.

3. The hybrid intelligent system architecture according to claim 1, characterized in that: The probabilistic reasoning module uses a pre-trained large language model as its basis and is fine-tuned or combined with other types of neural networks according to the needs of specific application areas.

4. The hybrid intelligent system architecture according to claim 1, characterized in that: The hybrid coordination and decision module adopts a hierarchical decision-making mechanism of "rules first, probability later", that is, the output of the rules has a higher priority or veto power. When the rules have no clear conclusion or are within the allowed range, the high-confidence output of the probability module is adopted.

5. The hybrid intelligent system architecture according to claim 1, characterized in that: The semantic data packet interaction bus includes a semantic data packet manager, which is responsible for implementing the registration, routing, distribution, monitoring and necessary conversion services of the semantic data packets.

6. The hybrid intelligent system architecture according to claim 1, characterized in that: The semantic data packet has a predefined structure, including one or more of the following fields: data packet unique identifier, timestamp, source agent identifier, target agent identifier, session identifier, semantic type, priority, confidence score, payload format, payload content, context reference identifier, processing instructions, verification status and metadata.

7. The hybrid intelligent system architecture according to claim 6, characterized in that: The semantic type of the semantic data packet defines the core semantic category carried by the data packet, adopts a hierarchical structure design, and supports multiple types such as query, assertion, response, and control.

8. The hybrid intelligent system architecture according to claim 1, characterized in that: The knowledge and rule management module includes a structured knowledge base, a rule base, a dynamic knowledge / state pool and a knowledge acquisition and learning unit, and is used to manage the formalized rule base, the structured knowledge base and the dynamic knowledge.

9. The hybrid intelligent system architecture according to claim 8, characterized in that: The dynamic knowledge / state pool draws on the idea of ​​the global link state pool to store information such as the capability description, current task status and context of each agent within the system for query and use by other modules.

10. An information processing method based on the hybrid intelligent system architecture according to any one of claims 1 to 9, characterized in that: The method comprises the following steps: receiving input information and converting it into a semantic data packet; routing the semantic data packet to a rule-driven processing module for preliminary analysis; The rule-driven processing module analyzes the rule base to identify and eliminate paths that are non-compliant, conflict with known facts, or have logical errors; Within the compliance scope defined by the rule-driven module, the processed semantic data packets are forwarded to the probabilistic reasoning module for further processing; the probabilistic reasoning module generates probabilistic conclusions or content with confidence scores; the hybrid coordination and decision-making module receives outputs from the rule-driven module and the probabilistic reasoning module, resolves potential conflicts and makes the final decision; the final decision or processing result is converted into output form and presented to the user.

11. The information processing method according to claim 10, wherein: It also includes a reasoning process that implements a "bidirectional cycle from concrete to abstract and back to concrete," namely: refining specific input information into an abstract representation through a probabilistic model; using a rule-driven module to perform logical reasoning and constraints on the abstract representation; and then converting the constrained abstract representation into a specific, detailed output through a probabilistic model.

12. The information processing method according to claim 10, wherein: A "high-speed error correction mechanism" is implemented in the rule-driven processing module to quickly identify and eliminate obviously erroneous or unreasonable assumptions, inferences, or generated content through predefined heuristic rules and fact bases.

13. The information processing method according to claim 10, wherein: For complex tasks, task decomposition and collaborative processing based on semantic data packets are achieved through hybrid coordination and decision-making modules. The original task is decomposed into multiple subtasks and assigned to appropriate processing modules, and then the processing results of each subtask are integrated.

14. A knowledge enhancement system using the hybrid intelligent system architecture according to any one of claims 1 to 9, characterized in that: By combining rule-driven deterministic reasoning with the statistical reasoning capabilities of probabilistic models, it can understand, analyze, enhance and generate input information. It is suitable for application scenarios such as intelligent question answering, automatic summarization, content creation, knowledge reasoning, decision support, and educational guidance.

15. The knowledge enhancement system according to claim 14, characterized in that: The system explicitly links related historical interaction information through the context reference ID field in the semantic data packet structure, and combines it with the dynamic context knowledge graph constructed by the knowledge and rule management module to effectively capture and utilize long-range dependency information, breaking through the limitations of fixed context windows.