Search

CN-121980006-A - Large model anti-illusion method and system based on multi-mode fact verification and dynamic constraint

CN121980006ACN 121980006 ACN121980006 ACN 121980006ACN-121980006-A

Abstract

The invention relates to the technical field of artificial intelligence large models, and discloses a large model anti-illusion method based on multi-mode fact verification and dynamic constraint. The method carries out full-link multi-dimensional management and control on user inquiry by constructing a multi-source multi-mode fact retrieval and depth fusion, cross-mode depth fact verification, dynamic intelligent prompt constraint generation and self-adaptive optimization, external rule bottom and iterative optimization four-level anti-illusion system. The system comprises a multi-source multi-mode searching and fusing module, a cross-mode depth checking module, a dynamic intelligent prompt generating and optimizing module, an external rule and cross-checking module and a closed loop iteration optimizing module. The method solves the technical problems of incomplete management, weak search fusion and stiff prompt constraint of the conventional large model multi-type illusion, realizes accurate management of multi-scene illusions, remarkably improves the credibility and suitability of generated contents, and is suitable for the field of high credibility and creative scenes.

Inventors

  • LI QIANG

Assignees

  • 上海才匠智能科技有限公司

Dates

Publication Date
20260505
Application Date
20251226

Claims (10)

  1. 1. The large model anti-illusion method based on multi-mode fact verification and dynamic constraint is characterized by comprising the following steps: (1) The multi-source multi-mode fact retrieval and depth fusion is carried out by dynamically adapting the retrieval source end according to the query type and the domain attribute after receiving the user query, adopting the differentiated retrieval strategy to execute the retrieval task in parallel, and carrying out de-duplication, entity linking and timeliness on the retrieval result Preprocessing filtering, namely generating a structured multi-modal fact library through a three-level depth fusion mechanism of semantic association, feature weighting and credibility quantification; (2) Performing cross-mode depth fact verification, namely performing full-dimension verification by adopting a three-section architecture of 'pre-verification-in-generation verification-post-generation verification', and performing correction or refusal when the comprehensive reliability is calculated and is lower than a threshold value through four types of verification units of real type, logic type, format type and cross-mode consistency; (3) Dynamic intelligent prompt constraint generation and self-adaptive optimization, namely carrying out multi-dimensional analysis of intention recognition, field classification and risk level evaluation on user inquiry, constructing a high-dimensional input feature vector containing fact library features, and adopting basic constraint layer and field adaptation The four-layer architecture of the layer, the risk reinforcement layer and the dynamic adjustment layer generates personalized prompt words, adapts different constraint strategies aiming at different illusions, and dynamically adjusts constraint parameters based on reinforcement learning; (4) Constructing a structured domain rule base for secondary spam verification, adopting a 'main model + auxiliary model' architecture for multi-model cross verification, recording full-link traceable logs, collecting user feedback and illusion cases, and passing through Reinforcement learning and supervised fine tuning continuously optimizes the correlation model and knowledge base.
  2. 2. The method of claim 1, wherein the multi-source multi-modal fact retrieval in step (1) supports dynamic expansion of a retrieval source, and an optimal retrieval source combination is intelligently selected according to query modality types and implicit field attributes, wherein a retrieval recall rate is greater than or equal to 92%, a cross-modality pair qi accuracy rate is greater than or equal to 93%, and a conflict resolution accuracy rate is greater than or equal to 95%.
  3. 3. The method of claim 1, wherein the cross-modal depth fact verification in step (2) uses a knowledge-graph-based "entity-relationship-attribute" three-dimensional network for fact authenticity verification, and an improved NLI model logic consistency determination criterion The determination rate is more than or equal to 95%, the checking error of complex mathematical calculation is less than or equal to 0.5%, and the overall checking accuracy is improved by 40% compared with the traditional single-stage checking.
  4. 4. The method of claim 1, wherein the dynamic intelligent hint constraint generation in step (3) supports a hallucination type differentiated constraint policy for realistic type hallucination forcing fact references and source labels for logical type hallucination embedding thoughts Chain examples, checklists are explicitly formatted for instruction type hallucinations, consistency check requirements are increased for cross-modal hallucinations, and constraint parameter tuning periods are less than or equal to 24 hours.
  5. 5. The method of claim 1, wherein the external rule spam and iteratively optimized domain rule base in step (4) supports visual low code configuration, multi-model cross-validation consistency thresholds are customizable, closed loop iterative optimization default perimeter The period is 1 time per week, and the anti-illusion overall effect is improved by more than or equal to 10 percent after each iteration.
  6. 6. A large model anti-illusion system implementing the method of any one of claims 1 to 5, comprising: The multi-source multi-mode searching and fusing module supports dynamic adaptation of a searching source end, parallel execution of a differential searching strategy and three-level deep fusion processing, generates a structured multi-mode fact library, and has a searching recall rate of more than or equal to 92% and an updating delay of less than or equal to 10 minutes; the cross-mode depth verification module comprises four types of verification units, supports three-section verification and reliability quantification scoring, and has an overall verification accuracy rate of more than or equal to 95% and a verification response delay of less than or equal to 500ms; The dynamic intelligent Prompt generation and optimization module is used for generating four-layer architecture personalized Prompt words based on multidimensional input analysis, supporting accurate adaptation of illusion types and dynamic optimization of constraint parameters, wherein the response time of the Prompt generation is less than or equal to 300ms, and the constraint hit rate is more than or equal to 92%; The system comprises an external rule and cross verification module, a closed loop iteration optimization module and an automatic iteration update module, wherein the external rule and cross verification module comprises a structured field rule base, a multi-model cross verification unit and a full-link log tracing component, the rule matching accuracy rate is more than or equal to 98%, and the closed loop iteration optimization module is used for collecting full-link log data and user feedback, continuously optimizing a related model and a knowledge base through reinforcement learning and supervision fine tuning, and supporting automatic iteration update.
  7. 7. The system of claim 6, wherein the multi-source multi-mode search and fusion module supports an asynchronous parallel search execution and timeout fusing mechanism, and wherein a core domain fact library update delay is less than or equal to 5 minutes, and supports incremental synchronization and version management.
  8. 8. The system of claim 6, wherein the dynamic intelligent hint generation and optimization module supports creation, categorization, and multiplexing of personalized constraint templates, and wherein a template library supports continuous evolution based on practical effects.
  9. 9. The system of claim 6, wherein the system provides a visual configuration interface supporting rule base configuration, parameter adjustment, effect monitoring and report generation, and opens an API interface for third party systems to integrate.
  10. 10. The system of claim 6, wherein the system supports multi-modal input and output, adapts to cross-modal interaction scenarios of text, images, speech, video, and has multi-language processing capabilities.

Description

Large model anti-illusion method and system based on multi-mode fact verification and dynamic constraint Technical Field The invention relates to the technical field of artificial intelligence large models, in particular to a large model anti-illusion method and a large model anti-illusion system based on multi-mode fact verification and dynamic constraint, which are suitable for high-reliability scenes such as finance, medical treatment, government affairs, law and the like, and different illusion tolerance scenes such as creative writing, intelligent customer service, cross-mode interaction and the like. Background The illusion of the large model is expressed as a plurality of types such as inconsistent generated content and facts, logic contradiction, disordered format, cross-modal information dislocation and the like, and the application of the large model in the field of high credibility is seriously restricted. The existing anti-illusion scheme has the following technical defects that 1, retrieval fusion capability is weak, single source end or simple multi-source splicing is adopted, cross-source semantic association and deep fusion mechanisms are lacked, 2, a checking mechanism is rough, single-stage and single-mode checking is adopted, deep checking capability on complex reasoning chains and cross-mode association information is lacked, 3, prompt constraint is fixed and stiff, a generalized prompt word template is adopted, multi-dimensional input characteristics, field characteristics and fact library state dynamic adjustment are not combined, 4, multi-mode adaptation is lacked, illusion problems in cross-mode interaction such as text-image and text-voice are difficult to deal with, 5, an optimization mechanism is closed, full-link log tracing and user feedback closed loops are not constructed, and continuous iteration of anti-illusion effects cannot be realized. Thus, there is a need for a full link solution that can systematically handle multi-type, multi-modal illusions. Disclosure of Invention Aiming at the technical defects of weak retrieval fusion, extensive verification mechanism, stiff prompt constraint, insufficient multi-mode adaptation and lack of continuous optimization of the existing anti-illusion scheme, the invention provides a large model anti-illusion method based on multi-mode fact verification and dynamic constraint, which realizes multi-source information deep fusion and multi-type illusion full-link refinement treatment. The technical scheme of the invention is to construct a four-level anti-illusion system: 1. Multisource multi-mode fact retrieval and depth fusion are performed, and a high-quality fact library is generated through intelligent source end adaptation, differential retrieval strategies and a three-level depth fusion mechanism; 2. the cross-mode depth fact verification is realized by adopting a three-section architecture and four types of verification units; 3. Dynamic intelligent prompt constraint generation and self-adaptive optimization, wherein personalized prompt words are generated based on multidimensional analysis and a four-layer architecture, and dynamic parameter tuning is supported; 4. And (3) external rule spam and iterative optimization, providing a rigid spam through a structural rule base and multi-model cross verification, and establishing a full-link data closed loop to realize continuous optimization. The beneficial effects of the invention include: 1. the multisource information fusion capability is obviously improved, the retrieval recall rate is more than or equal to 92%, and the conflict resolution accuracy rate is more than or equal to 95%; 2. The multi-type illusion is fully covered, and the treatment range is improved by 60% compared with the traditional scheme; 3. The verification precision and reliability are outstanding, the fact accuracy is more than or equal to 95%, and the logic consistency is more than or equal to 95%; 4. The scene suitability and flexibility are extremely strong, the constraint hit rate is more than or equal to 92%, and the customization of the personalized template is supported; 5. the method realizes the treatment of the cross-mode phantom system for the first time and supports the reliable application of the multi-mode large model; 6. The continuous optimization capability is outstanding, and the effect of each iteration is improved by more than or equal to 10%; 7. The deployment cost is low, the compatibility is strong, and the modularized integration is supported. Drawings FIG. 1 is a schematic diagram of a four-level anti-hallucination architecture provided in accordance with one embodiment of the present invention. FIG. 2 is a schematic diagram of a multi-source multi-modal fact retrieval and depth fusion process according to an embodiment of the present invention. FIG. 3 is a schematic diagram of dynamic intelligent hint constraint generation and adaptive optimization logic provided by an embodiment of the