An autonomous agent rumor verification system based on dynamic causal argumentation graph

Abstract

The application provides a rumor verification system for autonomous intelligent agents based on dynamic causal proof graph, which comprises a planning module for analyzing the topological structure of the dynamic causal proof graph read from a memory module and generating action instructions; an action module for executing the action instructions issued by the planning module, converting unstructured evidence information into structured operations on the dynamic causal proof graph, finding the best proof path and generating a final report; a memory module responding to reading requests from the planning module and writing requests from the action module; an environment module responding to query API calls from the action module and returning original unstructured evidence information; and a source authority management module responding to query requests from the action module and providing evidence information with information source authority rating. The application can realize deep, accurate and interpretable real-time identification of complex network rumors by constructing an intelligent agent with a dynamic causal proof graph as the core memory and reasoning center.

Description

Technical Field

[0001] This invention belongs to the fields of network content security and artificial intelligence technology, specifically relating to an autonomous intelligent agent rumor verification system based on dynamic causal verification graphs. Background Technology

[0002] Currently, the governance of online rumors mainly relies on methods such as "keyword retrieval, regular scanning, and manual review." These methods have shown significant limitations in addressing the massive, high-speed, and ever-changing nature of information dissemination on modern social networks. While existing AI-assisted verification technologies have improved efficiency to some extent, they still suffer from the following fundamental paradigmatic limitations:

[0003] (1) Limitations in evidence processing and reasoning capabilities: Existing methods treat the collected evidence as an unstructured "list of facts," ignoring the complex supporting, contradictory, premise-based, and inferential relationships between different pieces of evidence. This flattened approach makes it difficult for the model to perform true multi-step, non-linear reasoning, and it is easily misled by surface information. At the same time, due to the lack of modeling of the evidence structure, the planning capabilities of existing systems are linear and heuristic, unable to perceive "where the entire chain of evidence breaks" from a global perspective, and unable to proactively identify and resolve "the inherent contradictions between two pieces of evidence," making them inefficient and fragile when dealing with complex rumors.

[0004] (2) Limitations of dynamic response and real-time performance: Traditional methods are mostly one-time matching and classification, which cannot dynamically collect the latest evidence from authoritative sources during the verification process. They are difficult to deal with sudden events with rapidly changing information, resulting in poor real-time performance and a high rate of missed identification. This static verification mode makes it difficult for the system to adapt to the rapid evolution of online rumors, especially during major public events, and it is unable to capture and respond to new forms of rumors in a timely manner.

[0005] (3) Limitations in learning and adaptation: Existing models treat each verification as an isolated task, failing to record and learn from its successes and failures in complex verification processes. This can lead to the system repeatedly making invalid API calls or losing key intermediate conclusions in multi-step reasoning, thereby reducing efficiency and robustness and making it difficult to optimize verification strategies and improve governance effectiveness through continuous learning.

[0006] (4) Limitations in transparency and interpretability: Some methods that use large models for reasoning have a reasoning process that is disconnected from real-time external information and relies entirely on the unstable internal knowledge of the model, which can easily create illusions that contradict the facts. At the same time, the reasoning logic of these models is hidden in the parameters of the neural network or the "thinking" text generated by natural language. This black-box decision-making process is not conducive to the machine to perform strict logical consistency verification, nor is it easy to provide human regulators with a clear, credible, and visualized decision-making path, thus limiting its application depth in human-machine collaborative governance. Summary of the Invention

[0007] To address the problems of static verification, flattened evidence processing, implicit reasoning, and lack of adaptive learning in existing online rumor identification systems, this invention provides an autonomous agent rumor verification system based on a dynamic causal verification graph. The core of this system lies in an innovative autonomous fact-checking agent centered on a dynamic causal verification graph (DCVG). Through modular design, this agent system transforms the fact-checking process from list-based verification to a process of dynamically constructing and reasoning about the DCVG. The agent no longer passively records facts but actively integrates each new piece of evidence as a node or edge of the graph into a global, dynamically evolving knowledge graph. This achieves a complete closed loop from evidence collection and structured memorization to strategic planning and interpretable reasoning. The planning module analyzes the topology of the DCVG to determine the next action, thus systematically solving the aforementioned technical challenges and providing technological empowerment for efficient, accurate, and reliable rumor governance.

[0008] To solve the above-mentioned technical problems, the present invention adopts the following technical solution:

[0009] The autonomous intelligent agent rumor verification system based on dynamic causal verification graphs includes a planning module, an action module, a memory module, an environment module, and a source authority management module; among which,

[0010] The planning module is used to read the complete dynamic causal verification graph from the memory module and the failure query log in the auxiliary memory, analyze the topology of the dynamic causal verification graph, and generate specific action instructions based on the analysis results and avoid the failed attempts recorded in the failure query log, and issue them to the action module.

[0011] The action module is used to execute action instructions from the planning module, and its functions include:

[0012] (a) Interact with the environment module to perform a query and obtain unstructured evidence information;

[0013] (b) Update the query log list in the auxiliary memory within the memory module in real time based on the results of the query action;

[0014] (c) Upon successful query, a fact extraction operation is performed, which interacts with the source authority management module to obtain an authority rating and transforms unstructured evidence information into structured knowledge to update the main memory in the memory module;

[0015] (d) When the termination condition is met, execute the graph traversal algorithm on the main memory to find the best verification path and generate the final report;

[0016] The memory module includes a main memory as a long-term structured knowledge base and an auxiliary memory as a long-term contextual memory log. The main memory is a weighted dynamic causal verification graph used to explicitly store all facts, entities and their logical relationships. The auxiliary memory is a query log list used to record the agent's action history to improve the efficiency of subsequent decision-making and provide traceability.

[0017] The environment module is used to respond to query API calls from the action module and return raw, unstructured evidence information most relevant to the query;

[0018] The source authority management module is used to respond to query requests from the action module, provide an authority rating for the source of the evidence information, and allow the action module to set the initial confidence level of the fact nodes.

[0019] Furthermore, the planning module includes:

[0020] State reading unit: At the beginning of each decision cycle, the planning module first reads the complete dynamic cause-effect verification graph structure from the memory module, as well as the failure query log in the auxiliary memory;

[0021] The Graph Analysis Unit is used to call the built-in graph analysis algorithm to perform a comprehensive topological evaluation and analysis on the dynamic causal verification graph. The priority of the analysis is contradiction detection, breakpoint location, and path evaluation.

[0022] Decision-making unit: Based on the topology evaluation analysis results of the analysis landscape unit and to avoid failed attempts recorded in the failure query log, it generates specific action instructions and issues them to the action module.

[0023] Furthermore, the action module includes:

[0024] Instruction receiving unit: used to receive a specific action instruction from the planning module;

[0025] The search action unit is used to interact with the environment module to execute queries and perform subsequent operations based on the query results: if the query successfully obtains valid evidence information, the query is recorded in the query history list and the fact extraction unit is invoked; if the query fails to obtain valid evidence information, the query is recorded in the failed query log list.

[0026] The execution of the fact extraction unit is used to interact with the source authority management module to query the source authority of each piece of evidence information to obtain the authority rating, and to extract atomic facts and key entities from the evidence information, create corresponding nodes in the subject memory and establish logical edges;

[0027] The synthesis and completion unit is used to perform a graph traversal algorithm on the dynamic causal proof graph of the memory module, find the optimal proof path, and generate a final report.

[0028] Furthermore, the action module performs node creation and edge establishment operations in the main memory through its execution fact extraction unit. The node creation includes the creation of entity nodes, fact nodes, and hypothesis nodes, and the edge establishment operation includes the establishment of relational edges and logical edges.

[0029] Furthermore, the auxiliary memory includes a query history list and a failed query log list.

[0030] Furthermore, the source authority management module includes:

[0031] Request receiving unit: Used to receive query requests from the action module, the request containing a source URL;

[0032] Query and rating unit: Used to match in the authority database maintained within the source authority management module to find the rating corresponding to the URL;

[0033] Return to rating unit: Used to return the queried authority rating to the action module, so that the action module can set the initial confidence level of the fact node.

[0034] Compared with existing technologies, the autonomous intelligent agent rumor verification system based on dynamic causal verification graphs provided by this invention has the following advantages:

[0035] 1. Achieve deep logical reasoning and improve recognition accuracy: By constructing a dynamic causal verification graph, the way evidence is processed is fundamentally changed, from simple "fact listing" to "logical construction". This enables the agent to explicitly model and analyze the support, contradiction and causal relationship between evidence, thereby accurately identifying complex rumors that rely on deep logical fallacies and effectively overcoming the "fact illusion" problem of traditional methods.

[0036] 2. Achieve strategic dynamic verification and effectively respond to emergencies: The planning module of this invention makes decisions based on topological analysis of the global evidence graph, possessing a strategic and global perspective. It can locate the most critical breakpoints and contradictions in the evidence chain in real time and accurately, and proactively plan the most efficient action path. This dynamic "research" process improves the response speed and adaptability in scenarios where information evolves rapidly, such as emergencies.

[0037] 3. Possesses adaptive learning and conflict resolution capabilities, enhancing system robustness: Structured dynamic memory, especially the recording of failed queries and contradictory evidence, enables the agent to have self-evolution capabilities. It can not only learn from past failed explorations to improve efficiency, but also proactively identify conflicting evidence and trigger specialized resolution mechanisms (such as seeking higher authoritative sources). This makes the system highly robust when facing noisy and contradictory information on the network.

[0038] 4. Providing ultimate explainability and empowering trustworthy human-machine collaborative governance: The final conclusion of this invention is presented as a traceable and visualized verification path, clearly demonstrating the source of evidence and logical chain for each judgment. This "glass box" transparency provides a solid and intuitive technical foundation for manual review, rumor tracing, and governance decision-making, and is a key link in building an efficient and trustworthy human-machine collaborative governance system. Attached Figure Description

[0039] Figure 1 This is a diagram of the autonomous intelligent agent rumor verification system based on dynamic causal verification graph provided by the present invention.

[0040] Figure 2 This is the decision-making flowchart of the planning module provided by the present invention.

[0041] Figure 3 This is a flowchart of the action module execution and DCVG update provided by the present invention.

[0042] Figure 4 This is a diagram of the DCVG construction and structure provided by the present invention.

[0043] Figure 5 This is a working structure diagram of the dynamic memory module provided by the present invention.

[0044] Figure 6 This invention provides a step-by-step update trajectory diagram for the fact-checking case memory module. Detailed Implementation

[0045] To make the technical means, creative features, objectives and effects of this invention easier to understand, the invention will be further described below with reference to specific illustrations.

[0046] Please refer to Figure 1 As shown, this invention provides an autonomous intelligent agent rumor verification system based on a dynamic causal verification graph, including a planning module, an action module, a memory module, an environment module, and a source authority management module; wherein,

[0047] The planning module, as the decision-making core (brain) of the agent, is used to read the complete dynamic causal verification graph from the memory module and the failure query log in the auxiliary memory, analyze the topological structure of the dynamic causal verification graph, generate specific action instructions based on the analysis results and avoid the failed attempts recorded in the failure query log, and issue them to the action module to formulate the globally optimal next action strategy. It is responsible for locating evidence gaps, identifying logical contradictions and planning solutions.

[0048] The Action Module, acting as the agent's executors (both hands), executes action instructions from the Planning Module. Its functions include: (a) interacting with the Environment Module to perform queries and obtain unstructured evidence information; (b) updating the query log list in the auxiliary memory of the Memory Module in real time based on the query action results; (c) performing a fact extraction operation upon successful query, which interacts with the Source Authority Management Module to obtain the source authority rating of the evidence information and transforms the unstructured evidence information into structured knowledge to update the main memory of the Memory Module; and (d) executing a graph traversal algorithm on the main memory when the verification meets the termination condition, finding the optimal verification path and generating a final report.

[0049] The memory module, serving as the core memory and reasoning center of the agent (notebook and whiteboard), includes a main memory as a long-term structured knowledge base and an auxiliary memory as a long-term contextual memory log. The main memory is a weighted dynamic causal verification graph (DCVG) used to explicitly store all facts, entities, and their logical relationships. The auxiliary memory is a query log list used to record the agent's action history to improve subsequent decision-making efficiency and provide traceability.

[0050] The Environment Module refers to the external world through which the intelligent agent interacts, mainly including external information sources such as search engines and professional databases. The Environment Module is used to respond to query API calls from the Action Module and return the raw, unstructured evidence information most relevant to the query.

[0051] The Source Authority Module, as a key auxiliary module, is used to respond to query requests from the Action Module and provide an authority rating for the source of the evidence information. In other words, it provides the system with prior knowledge about the reliability of the information source, allowing the Action Module to set the initial confidence level of the fact nodes. This is the foundation for ensuring the quality of graph construction and realizing automated conflict resolution.

[0052] The five core modules of this intelligent agent system work together through a clearly defined interaction protocol to form a closed loop from "planning" to "action" and then to "memory update". The core of its workflow is to read, analyze and write dynamic causal verification graphs.

[0053] For a specific embodiment, please refer to Figure 2 As shown, the core responsibility of the planning module is to act as the "brain" of the intelligent agent. By analyzing the DCVG state in the current memory module, it determines the next action to be taken, including a state reading unit, a scenario analysis unit, and a decision-making unit. Specifically,

[0054] The read state unit is used at the beginning of each decision cycle. The planning module first reads the complete dynamic causal verification graph structure from the memory module, as well as the failure query log (M_failed_queries) in the auxiliary memory.

[0055] The Graph Analysis Unit: This unit invokes built-in graph analysis algorithms to perform a comprehensive topological evaluation and analysis of the dynamic causal verification graph. The analysis priority is conflict detection, breakpoint location, and path evaluation, with the specific analysis priorities as follows:

[0056] Contradiction detection: First, check if there are any high-confidence contradictory edges in the graph. This is a strong signal that there is a fundamental conflict in the chain of evidence.

[0057] Breakpoint location: Next, analyze the connectivity of the graph to find isolated entity nodes or evidence subgraphs that cannot be connected to the core fact cluster.

[0058] Path evaluation: Finally, evaluate whether the verification path from the high-confidence fact node to the final hypothesis node has been formed, is complete, and is credible.

[0059] Decision-making unit: Based on the topology evaluation analysis results of the analysis scenario unit and to avoid failed attempts recorded in the failure query log, it generates specific action instructions and issues them to the action module. An example of the decision logic is as follows:

[0060] If a contradiction is detected → Decision: Plan a high-priority search action, instructing it to find a higher-authority source to resolve the contradiction.

[0061] If a break in the chain of evidence is found → Decision: Plan a search action to find intermediate facts that can bridge the break.

[0062] If an isolated entity is found → Decision: Plan a search operation to collect background information about the entity.

[0063] If the verification path is complete → Decision: Plan the synthesize_and_finish action, terminate the verification and generate a report.

[0064] If no clear path is found → Decision: Continue planning exploratory search actions to expand the evidence map.

[0065] For a specific embodiment, please refer to Figure 3 As shown, the core responsibility of the action module is to act as the "hands and feet" of the intelligent agent, accurately executing instructions from the planning module, interacting with the external environment module and other modules, and finally writing the results back to the memory module. It includes an instruction receiving unit, a search action unit, an extract fact unit, and a synthesize and finish unit. Specifically,

[0066] The instruction receiving unit is used to receive a specific action instruction from the planning module, such as search[query] or synthesize_and_finish.

[0067] The execution search action unit is used to interact with the environment module to execute queries and perform subsequent operations based on the query results: if the query is successful and obtains valid evidence information, the query is recorded in M_queries (query history list), and the execution fact extraction unit is invoked; if the query fails and fails to obtain valid evidence information, the query is recorded in M_failed_queries (failed query log list).

[0068] The Fact Extraction Unit (Core Graph Operation) is invoked only after a successful search. It interacts with the Source Authority Management Module to query the authority of each piece of evidence, obtain an authority rating, and extract atomic facts and key entities from the evidence. Corresponding nodes are then created and logical edges are established in the DCVG of the memory module. This is the second step in writing to memory: internalizing the action results. A specific example is as follows:

[0069] Query source authority: Send the source_url of the evidence to the source authority management module to obtain the authority rating of the source, which is the basis for all subsequent operations.

[0070] Extracting and creating nodes: The instruction LLM (Large Language Model) extracts atomic facts and key entities from the evidence text and creates corresponding fact nodes and entity nodes in the DCVG of the memory module. The initial confidence score of the fact node is determined by the authority rating obtained in the previous step.

[0071] Establishing logical edges (two-step method): To ensure accuracy, this step is strictly constrained. The specific action module submits new fact nodes to the LLM in pairs with existing related nodes in the graph, uses a multiple-choice template ("support", "contradictory", or "irrelevant") to determine the relationship, and then creates corresponding weighted logical edges in the DCVG.

[0072] The synthesis and completion unit is used to perform a graph traversal algorithm on the dynamic causal proof graph of the memory module, find the optimal proof path, and generate a final report.

[0073] In a specific embodiment, the core innovation of the memory module lies in using a Dynamic Causal Verification Graph (DCVG) as the main memory, replacing the traditional list-based memory. The memory module consists of two parts: main memory and auxiliary memory. The main memory is a dynamic, weighted graph used to explicitly store entities, facts, hypotheses, and their logical relationships. The auxiliary memory includes a query history list (M_queries) and a failed query log list (M_failed_queries), used to record action history to support decision-making. The DCVG, as the main memory, is the core of the system's persistent state, but it only passively responds to read and write requests and does not contain active logic.

[0074] For a specific embodiment, please refer to Figure 4 As shown, the main memory, as a dynamic, weighted graph, explicitly represents a network of "entity-fact-logical relationships." The confidence score of a node reflects the reliability of the evidence source, while the confidence score of an edge reflects the reliability of the logical judgment. The action module, through its fact extraction unit, performs node creation and edge establishment operations in the main memory. Node creation includes the creation of entity nodes, fact nodes, and hypothesis nodes, while edge establishment includes the establishment of relational edges and logical edges. In other words, the structure of the main memory includes nodes and edges, where nodes carry entities, facts, and hypotheses, and edges represent the relationships between them. A specific example is as follows:

[0075] Entity Node: Key entities (people, events, things, places, concepts) involved in rumors, such as "N95 masks" and "CDC".

[0076] Fact Node: An atomic fact with independent semantics extracted from external evidence. For example: "N95 masks are more efficient at filtering aerosols than surgical masks." Each fact node contains metadata: {fact_id, source_url, confidence_score, timestamp, status} (status such as: verified, pending verification, outdated, disputed). The initial value of the confidence_score is directly determined by the rating of its source in the source authority management module.

[0077] Hypothesis Node: The original rumor itself to be verified.

[0078] Relational Edge: Connects entity nodes and fact nodes, representing the role an entity plays in a fact (e.g., subject, object).

[0079] Logical Edge: Connects two fact nodes, representing the logical relationship between them, which is crucial for deep reasoning. Edge types include: supports and contradictions, and each edge is assigned a confidence_score, representing the reliability of the logical relationship judgment.

[0080] In a specific embodiment, the auxiliary memory includes a query history list (M_queries) and a failed query log list (M_failed_queries). The query history list stores all executed search actions and their query parameters to avoid duplicate queries and trace the source of DCVG fact nodes. The failed query log list stores queries that did not return valid evidence or led to errors, serving as a negative experience base to guide the planning module in optimizing its exploration strategy.

[0081] Figure 5The flowchart illustrates the workflow of the dynamic memory module. Starting with a declaration, the process triggers a retrieval after declaration verification, pushing the declaration into the query history list to form a pool of pending issues. The existence of verified fact nodes (M_verified_facts logical view) in the main memory is checked to determine whether to directly return a conclusion or initiate a query process. If the query is successful, the DCVG fact node status is updated to verified and added to the M_verified_facts logical view. If the query fails, it is written to the failed query log list. After a retrieval count (threshold < 2), the query is marked as permanently failed. Finally, a traceable conclusion containing evidence sources, logical chains, and confidence scores is generated using the DCVG graph traversal algorithm, achieving a declaration-driven dynamic verification closed loop and ensuring the system's high efficiency and robustness in noisy environments.

[0082] In a specific embodiment, the environment module is a passive responder. It receives searchAPI calls from the action module and returns the original web page content or data fragments that it can find that are most relevant to the query. The breadth (how many information sources can be accessed) and quality (the authenticity of the information sources) of the environment directly determine the upper limit of the verification capability of this intelligent agent system. The design of this system, especially the information source authority management module and the contradiction detection mechanism, is precisely to work as robustly as possible in an environment where the quality is uncontrollable.

[0083] In a specific embodiment, the core responsibility of the source authority management module is to act as an independent, configurable "information filter" or "fact preprocessor," assigning a reliability label to every piece of external information entering the system. Its function is to maintain a white / black / gray list of sources and its authority rating database. The rating can be pre-set manually or automatically assigned an initial weight based on the domain name type (e.g., .gov, .edu, .org). For example:

[0084] First-tier sources (high weight): government official websites, authoritative research institutions, and national-level news organizations.

[0085] Secondary sources (medium weight): mainstream commercial media, local governments and news organizations.

[0086] Level 3 information sources (low weight): User-generated content from self-media, forums, and social networks.

[0087] The source authority management module includes a request receiving unit, a query and rating unit, and a rating return unit. Specifically,

[0088] Request Receiver: Used to receive query requests from the action module (when extract_fact is executed), and the request contains a source URL.

[0089] Query and rating unit: Used to match the URL in the authority database maintained within the source authority management module to find the corresponding rating (such as level 1, level 2, level 3).

[0090] Return rating unit: Used to return the queried authority rating (or a numerical value representing its weight) to the action module, so that the action module can set the initial confidence level of the fact node.

[0091] The authority rating output by the source authority management module will serve as a core parameter to determine the initial confidence level of fact nodes and as a key basis in subsequent conflict resolution. At the same time, this module is a key bridge connecting the external environment and the internal memory module, ensuring that the problem of "garbage in, garbage out" is initially controlled at the source.

[0092] To better understand the autonomous intelligent agent rumor verification system based on dynamic causal verification graph provided by this invention, the following will provide a specific intelligent agent workflow case analysis to construct and analyze DCVG as the driving force.

[0093] Case 1: Product quality rumors ("The bottled water of a well-known brand XX is bottled with tap water and the pH value does not meet the standards")

[0094] ------------------------------------------------------------------------------------------------------

[0095] t=0: Initialization

[0096] Input: Rumor to be verified: "The bottled water of XX well-known brand is bottled with tap water and the pH value does not meet the standards."

[0097] Memory Module: The system initializes a DCVG in memory.

[0098] Assume node (H1): {"label": "XX brand water is bottled from tap water and its pH value does not meet the standard", "status":"Unverified"}

[0099] Entity nodes (E1, E2, E3): {"label": "XX brand bottled water"}, {"label": "tap water"}, {"label": "pH value does not meet standard"}

[0100] Edge: H1 is connected to E1, E2, and E3, indicating that these entities are assumed to be included.

[0101] ------------------------------------------------------------------------------------------------------

[0102] t=1: First Iteration - Task Decomposition and Core Entity Verification

[0103] (1) Planning Module

[0104] Read status: Read the initial DCVG from the memory module.

[0105] Analysis of the scenario: The analysis reveals that hypothesis H1 can be decomposed into two independent verification paths (water source, pH value). It is determined that the "water source" issue is a more fundamental and serious accusation, and should be the highest priority verification target.

[0106] Decision Making & Prompt:

[0107]

[0108] (2) Action Module

[0109] Receive command: search[Official water source information for XX brand bottled water].

[0110] Perform a search: The module interacts with the environment and executes `search[XX brand bottled water official water source information]`. If the search is successful, results from the official website are obtained. For example, a result from the brand's official website might be returned: `source_url: "http: / / xxbrand.com / our-water"`, with the content: "...Every drop of our water originates from the deep mineral springs of the XX National Nature Reserve...".

[0111] Update auxiliary memory: At this point, the search action unit immediately records the successful query into M_queries.

[0112] Execute the extract_fact subroutine:

[0113] a. Query source authority: Invoke the source authority management module and enter source_url. This module queries the database and returns the rating: "Level 2 source (commercial official website), initial confidence level 0.8".

[0114] b. Extracting and Creating Nodes: The LLM instruction extracts facts. A new fact node F1 is created in the DCVG of the memory module: {"fact_id": 1, "label": "XX brand water officially states its water source is deep mineral spring", "source_url":"...", "confidence_score": 0.8, "status": "Verified"}.

[0115] c. Establish logical edges: The action module submits new facts F1 and hypothesis H1 in pairs to the LLM for judgment.

[0116] * Prompt (Multiple Choice Question Template):

[0117] Task: Determine the relationship between fact F1 and the "water source is tap water" part of hypothesis H1. - Hypothesis H1: "XX brand water is bottled tap water" - Fact F1: "XX brand water officially claims its water source is deep mineral spring" # Options: A. Supports B. Contradicts C. Irrelevant--- B

[0118] * Create a proof edge from F1 to H1 in DCVG, of type contradicts, with confidence_score provided by LLM (e.g., 0.95).

[0119] (3) Memory Module (DCVG Update)

[0120] The DCVG adds a new fact node F1, which is connected to H1 through a high-confidence contradicts edge, making the graph structure richer.

[0121] ------------------------------------------------------------------------------------------------------

[0122] t=2: Second Iteration - Bridging the Information Gap and Applying Authority

[0123] (1) Planning Module

[0124] Read status: Reading the updated DCVG.

[0125] Analysis of the scenario reveals preliminary evidence refuting the "water source" path (F1). However, the "pH value not meeting standards" (E3) remains a break in the chain of evidence. Furthermore, the authority of source F1 is only level two, requiring the planning module to find more authoritative third-party evidence for cross-verification.

[0126] Decision Making & Prompt:

[0127]

[0128] (2) Action Module & Environment Observation

[0129] The search returned a sampling inspection notice from the official website of a municipal market supervision administration bureau (source_url: "http: / / gov.cn / xxx"). The notice stated: "...XX brand bottled water (batch number...) passed the sampling inspection, with a measured pH value of 7.3... The national standard is 6.5-8.5..."

[0130] (3) Memory Module (DCVG Update)

[0131] extract_fact subroutine:

[0132] a. Query source authority: The source authority management module returns: "Level 1 source (government official website), initial confidence level 0.98".

[0133] b. Extracting and Creating Nodes: In DCVG, create fact node F2: {"fact_id": 2, "label": "The Market Supervision Bureau's announcement shows that the pH value of XX brand water is 7.3", "source_url": "...", "confidence_score":0.98, "status": "Verified"} and fact node F3: {"fact_id": 3, "label": "The national standard defines the pH value range of bottled water as 6.5-8.5", "source_url": "...", "confidence_score": 1.0, "status": "Verified"}.

[0134] c. Establish logical edges:

[0135] First, LLM is instructed to derive a new, directly verifiable fact from fact F2 (measured pH value is 7.3) and fact F3 (national standard is 6.5-8.5). LLM generates the fact statement: "The pH value of XX brand water meets the national standard."

[0136] * The system creates a new fact node F4 based on this: {"label": "XX brand water pH value meets national standards", "source_url": "inferred_from_F2_F3", "confidence_score": 0.97, "status":"Verified"}. (Note: This is a new fact node, its source can be internally marked as derived, and its confidence level can be calculated based on the confidence levels of F2 and F3, for example, taking the lower one or a weighted average).

[0137] Next, the system establishes supporting edges from F2 to F4 and from F3 to F4, indicating that F2 and F3 together support the conclusion of F4.

[0138] Finally, the system determines that the new fact node F4 directly contradicts the "pH value not meeting the standard" part of the original hypothesis H1, and therefore establishes a contradictory edge from F4 to H1.

[0139] At this point, both subpropositions of hypothesis H1 in DCVG are refuted by fact nodes with high confidence and authoritative sources.

[0140] ------------------------------------------------------------------------------------------------------

[0141] t=3: Third iteration - Conclusion summary

[0142] (1) Planning Module

[0143] Analysis of the graph: It is found that all sub-propositions of H1 have been refuted by the fact nodes with high confidence (0.8, 0.98, 1.0) in the graph through the contradicts edge. The proof path is complete.

[0144] (2) Action Module

[0145] Execute synthesize_and_finish to traverse the contradicts paths from nodes such as F1 and F4 to H1 in the DCVG and generate a report.

[0146] Final output:

[0147] Reasoning process:

[0148] ■ Regarding the water source: Rumors claim it is tap water. However, verification shows that the brand's official label states the water source is deep mineral spring (Source: Brand website, Source Level: Level 2).

[0149] ■ Regarding pH value: Rumors claim it does not meet standards. However, a verification report from a municipal market supervision administration bureau shows a pH value of 7.3 (Source: Government website, Source Level: Level 1).

[0150] ■ Regarding standards: According to verification, the national standard for bottled drinking water in my country has a pH range of 6.5-8.5 (Source: Government website, Source Level: Level 1).

[0151] ■ Overall assessment: The product's pH value (7.3) is within the national standard range (6.5-8.5), therefore the accusation of "pH value not meeting the standard" is inconsistent with the facts. Both core arguments of the rumor have been refuted by highly authoritative sources.

[0152] Final verdict: Rumor

[0153] ------------------------------------------------------------------------------------------------------

[0154] Case 2: Common Science Rumors ("Radiation from cell phone base stations can cause insomnia and hair loss in nearby residents")

[0155] This case study will demonstrate how an agent can weigh and make decisions amidst conflicting information by using a source authority module.

[0156] ------------------------------------------------------------------------------------------------------

[0157] t=1-2: Verification of authoritative sources and processing of low-quality information

[0158] (1) Initialization & Initial Planning: DCVG initialization, including entity nodes such as "mobile phone base station", "radiation", and "insomnia". The planning module recognizes that this is a health science issue, and the primary action to make the decision is to consult the conclusions of top authoritative institutions.

[0159] (2) Action & Authoritative Observation: Search [World Health Organization Base Station Radiation Health] to obtain the official WHO conclusions from the Environment module. The source authority module is rated as "Level 1 source, confidence level 1.0".

[0160] (3) Memory update: Add a new fact node F1 in DCVG: {"label": "WHO says there is no evidence that base station radiation has adverse health effects", "source_url": "who.int", "confidence_score": 1.0, ...}, and establish a contrast edge with hypothesis H1.

[0161] (4) Secondary planning and discovery of contradictory information: In order to verify completeness, the planning module executes search[cases of hair loss caused by mobile phone base station radiation]. The environment module returns a post from a health forum source_url: "health-forum.net / ...", in which a user claims that he started losing hair after moving near a base station.

[0162] (5) Actions & Handling of Low Authority:

[0163] extract_fact subroutine:

[0164] a. Query source authority: The source authority management module returns: "Level 3 source (user-generated content), initial confidence level 0.2".

[0165] b. Extracting and creating nodes: Create fact node F2: {"label": "A forum user claims that a base station caused his hair loss", "source_url": "...", "confidence_score": 0.2, "status": "Verified_Anecdotal"}.

[0166] c. Establish logical edges: F2 establishes a support edge with H1. At this point, a contradiction appears in the graph: F1 (confidence_score: 1.0) and F2 (confidence_score: 0.2) make opposite judgments about H1.

[0167] ------------------------------------------------------------------------------------------------------

[0168] t=3: Third iteration - Weight-based conflict resolution

[0169] (1) Planning Module

[0170] Analysis of the scenario: A high-priority conflict was detected: the logical relationship between F1 and F2 and H1 is conflicted. The core task of the planning module is to resolve this conflict. It compares the confidence_score of the two nodes.

[0171] Decision Making & Prompt:

[0172]

[0173] (2) Action Module & Final Output

[0174] The `synthesize_and_finish` command will generate a report that clearly points out conflicting information and explains why the conclusions of authoritative sources were adopted.

[0175] Reasoning process:

[0176] ① Rumors claim that radiation from cell phone base stations can cause insomnia and hair loss.

[0177] ② Upon investigation, the official conclusion of the World Health Organization (WHO) is that there is no conclusive scientific evidence that radiation from base stations that meet the standards will have adverse health effects (Source: WHO official website).

[0178] ③Comprehensive Analysis: As the world's top health authority, the World Health Organization (WHO) draws its official conclusions based on extensive scientific evidence, and its authority and strength of evidence far exceed those of anonymous personal statements on online forums.

[0179] ④ Conclusion: The core claim of this rumor contradicts the current consensus of the mainstream scientific community.

[0180] Final verdict: Rumor

[0181] Therefore, the fact-checking reasoning process provided by this system is as follows:

[0182] Step 1: Read the state (from the memory module to the planning module)

[0183] ● At the start of each decision cycle, the planning module sends a read request to the memory module.

[0184] ● It obtains a complete topological snapshot of the current Dynamic Causal Verification Graph (DCVG), as well as the failure query log (M_failed_queries) in auxiliary memory.

[0185] ● This is the starting point for planning; all decisions are based on the present moment and the complete state of memory.

[0186] Step 2: Make a decision (within the planning module, output to the action module)

[0187] ● The planning module performs graph topology analysis on the read DCVG to identify key issues such as broken links in the chain of evidence, isolated entities, or contradictory evidence.

[0188] ● Based on the failure experience in M_failed_queries, it determines the optimal course of action for the current situation.

[0189] ● Ultimately, it generates a specific, standardized action instruction (such as search[XX brand water third-party testing report]) and issues it to the action module.

[0190] Step 3: Execution and Observation (Interact with the environment in the Actions module)

[0191] ● After receiving the instruction, the action module begins execution.

[0192] ● After receiving instructions, the action module interacts with the external environment (such as calling the search engine API) and obtains the "observation" results.

[0193] ● (Write order 1) As soon as the search operation ends, the operation module immediately sends a write request to the auxiliary memory to update the M_queries or M_failed_queries list, depending on whether it was successful or not.

[0194] Step 4: Update memory (from action module to memory module)

[0195] ● If step 3 is successful, the action module processes the "observation" results and transforms unstructured information into structured knowledge through subroutines such as extract_fact.

[0196] ● (Write order 2) Then, it sends a write request to the main memory to create new nodes and edges.

[0197] ● At this point, the state of the memory module has been updated, containing richer information than in the previous round. The agent has completed a full learning iteration and is ready to enter the next, higher-level planning cycle.

[0198] Figure 6 The trajectory diagram of a specific fact-checking case details how an agent starts from an initial state and, through multiple iterations and planning, including proposing queries, extracting facts, handling conflicts, and summarizing conclusions, gradually updates its memory module until it reaches the final state, demonstrating the evolution of the memory module's content.

[0199] Compared with existing technologies, the autonomous intelligent agent rumor verification system based on dynamic causal verification graphs provided by this invention has the following advantages:

[0200] 1. Achieve deep logical reasoning and improve recognition accuracy: By constructing a dynamic causal verification graph, the way evidence is processed is fundamentally changed, from simple "fact listing" to "logical construction". This enables the agent to explicitly model and analyze the support, contradiction and causal relationship between evidence, thereby accurately identifying complex rumors that rely on deep logical fallacies and effectively overcoming the "fact illusion" problem of traditional methods.

[0201] 2. Achieve strategic dynamic verification and effectively respond to emergencies: The planning module of this invention makes decisions based on topological analysis of the global evidence graph, possessing a strategic and global perspective. It can locate the most critical breakpoints and contradictions in the evidence chain in real time and accurately, and proactively plan the most efficient action path. This dynamic "research" process improves the response speed and adaptability in scenarios where information evolves rapidly, such as emergencies.

[0202] 3. Possesses adaptive learning and conflict resolution capabilities, enhancing system robustness: Structured dynamic memory, especially the recording of failed queries and contradictory evidence, enables the agent to have self-evolution capabilities. It can not only learn from past failed explorations to improve efficiency, but also proactively identify conflicting evidence and trigger specialized resolution mechanisms (such as seeking higher authoritative sources). This makes the system highly robust when facing noisy and contradictory information on the network.

[0203] 4. Providing ultimate explainability and empowering trustworthy human-machine collaborative governance: The final conclusion of this invention is presented as a traceable and visualized verification path, clearly demonstrating the source of evidence and logical chain for each judgment. This "glass box" transparency provides a solid and intuitive technical foundation for manual review, rumor tracing, and governance decision-making, and is a key link in building an efficient and trustworthy human-machine collaborative governance system.

[0204] Finally, it should be noted that the above embodiments are only used to illustrate the technical solutions of the present invention and are not intended to limit it. Although the present invention has been described in detail with reference to preferred embodiments, those skilled in the art should understand that modifications or equivalent substitutions can be made to the technical solutions of the present invention without departing from the spirit and scope of the technical solutions of the present invention, and all such modifications or substitutions should be covered within the scope of the claims of the present invention.

Claims

1. A rumor verification system for autonomous intelligent agents based on dynamic causal verification graphs, characterized in that, It includes a planning module, an action module, a memory module, an environment module, and a source authority management module; among which, The planning module is used to read the complete dynamic causal verification graph from the memory module and the failure query log in the auxiliary memory, and analyze the topology of the dynamic causal verification graph. The priority of the analysis is contradiction detection, breakpoint location, and path evaluation. Based on the analysis results and avoiding the failed attempts recorded in the failure query log, specific action instructions are generated and issued to the action module. The action module is used to execute action instructions from the planning module, and its functions include: (a) Interact with the environment module to perform a query and obtain unstructured evidence information; (b) Update the query log list in the auxiliary memory within the memory module in real time based on the results of the query action; (c) Upon successful query, a fact extraction operation is performed. This operation interacts with the source authority management module to obtain an authority rating. The source authority rating directly determines the initial confidence score of the fact node and transforms unstructured evidence information into structured knowledge to update the main memory in the memory module. (d) When the termination condition is met, execute the graph traversal algorithm on the main memory to find the best verification path and generate the final report; The memory module includes a main memory serving as a long-term structured knowledge base and an auxiliary memory serving as a long-term contextual memory log. The main memory is a weighted dynamic causal verification graph used to explicitly store all facts, entities, hypotheses, and their logical relationships. The structure of the main memory includes nodes and edges. The nodes include entity nodes, fact nodes, and hypothesis nodes, and the edges include relational edges and logical edges. The auxiliary memory is a query log list used to record the agent's action history to improve subsequent decision-making efficiency and provide traceability. The auxiliary memory includes a query history list and a failed query log list. The query history list stores all executed search actions and their query parameters to avoid duplicate queries and trace the source of DCVG fact nodes. The failed query log list stores queries that did not return valid evidence or led to errors as a negative experience base to guide the planning module in optimizing the exploration strategy. The environment module is used to respond to query API calls from the action module and return raw, unstructured evidence information most relevant to the query; The source authority management module is used to respond to query requests from the action module, provide an authority rating for the source of the evidence information, and allow the action module to set the initial confidence level of the fact nodes.

2. The dynamic causal provenance graph based autonomous agent rumor verification system according to claim 1, wherein, The planning module includes: State reading unit: At the beginning of each decision cycle, the planning module first reads the complete dynamic cause-effect verification graph structure from the memory module, as well as the failure query log in the auxiliary memory; The Graph Analysis Unit is used to call the built-in graph analysis algorithm to perform a comprehensive topological evaluation and analysis on the dynamic causal verification graph. The priority of the analysis is contradiction detection, breakpoint location, and path evaluation. Decision-making unit: Based on the topology evaluation analysis results of the analysis landscape unit and to avoid failed attempts recorded in the failure query log, it generates specific action instructions and issues them to the action module.

3. The dynamic causal evidential graph based autonomous agent rumor verification system of claim 1, wherein, The action module includes: Instruction receiving unit: used to receive a specific action instruction from the planning module; The search action unit is used to interact with the environment module to execute queries and perform subsequent operations based on the query results: if the query successfully obtains valid evidence information, the query is recorded in the query history list and the fact extraction unit is invoked; if the query fails to obtain valid evidence information, the query is recorded in the failed query log list. The execution of the fact extraction unit is used to interact with the source authority management module to query the source authority of each piece of evidence information to obtain the authority rating, and to extract atomic facts and key entities from the evidence information, create corresponding nodes in the subject memory and establish logical edges; The synthesis and completion unit is used to perform a graph traversal algorithm on the dynamic causal proof graph of the memory module, find the optimal proof path, and generate a final report.

4. The dynamic causal evidential graph-based autonomous agent rumor verification system of claim 1, wherein, The action module performs node creation and edge establishment operations in the main memory through its execution fact extraction unit. The node creation includes the creation of entity nodes, fact nodes and hypothesis nodes, and the edge establishment operation includes the establishment of relational edges and logical edges.

5. The dynamic causal evidential graph-based autonomous agent rumor verification system of claim 1, wherein, The source authority management module includes: Request receiving unit: Used to receive query requests from the action module, the request containing a source URL; Query and rating unit: Used to match in the authority database maintained within the source authority management module to find the rating corresponding to the URL; Return to rating unit: Used to return the queried authority rating to the action module, so that the action module can set the initial confidence level of the fact node.

Images