A multimodal knowledge post-editing model reasoning ability strengthening method and system

Abstract

The application discloses a multimodal knowledge editing post-model reasoning ability strengthening method and system. The method obtains a multimodal knowledge editing sample, extracts and normalizes objective detectable rules, and generates a multi-hop stress case of a labeled target rule subset; inputs the multi-hop stress case into a knowledge editing post-multimodal large language model configured with a hybrid expert routing strategy, generates a candidate response, and performs rule-by-rule compliance judgment; calculates case-level accuracy, rule-level accuracy, target rule accuracy, and hierarchical retention score based on rule satisfaction results; constructs editing performance rewards, expert co-activation rewards, and sequence consistency rewards according to the rule satisfaction results and the hierarchical retention score, and optimizes and updates the hybrid expert routing strategy or the expert parameters through a group relative strategy. Through the method, the retention ability of editing knowledge under multi-hop, restatement and counter-pressure can be improved, and the overall deployment reliability and interpretability of the model can be improved.

Description

Technical Field

[0001] This application relates to the field of multimodal large-scale language model technology, and in particular to a method and system for enhancing the reasoning ability of a multimodal knowledge-edited model. Background Technology

[0002] Multimodal large-scale language models can simultaneously process information from different modalities such as text and images. Through continuous updates to external knowledge, factual relationships, or security boundaries, they adapt to application scenarios such as enterprise-level knowledge base maintenance, intelligent agent deployment, strategic key decision support, and multimodal question answering. In these applications, knowledge editing is typically used to inject new facts, corrected relationships, or constraint rules into the model, enabling the model to follow the edited knowledge during subsequent reasoning and generation. For practical applications, whether the model can reproduce the edited knowledge in a single round of question answering is not the only objective. More importantly, it must maintain the priority and consistency of the edited knowledge under multi-hop reasoning, cross-modal associations, complex user-side prompts, and potentially conflicting instructions.

[0003] Existing knowledge editing evaluation schemes typically focus on metrics such as editing reliability, generalization, locality, or cross-modal consistency. Some schemes further introduce multi-hop question answering or visual language knowledge editing benchmarks to observe whether edited knowledge can be transferred to related questions. However, most of these schemes primarily evaluate fact recall or the correctness of pre-defined question answers, failing to adequately characterize the hierarchical degradation phenomena commonly seen in real-world deployments. For example, when user prompts conflict with the edited knowledge, the model may exhibit single-round drift, multi-round drift, inference-to-answer drift, role boundary expansion, partial compliance, indirect execution, or nested injection issues. These problems do not necessarily manifest as simple factual errors but rather as the model's failure to consistently maintain the edited rules in multi-hop inference chains. Existing schemes struggle to uniformly categorize these failure types and to objectively assess model responses at the rule-by-rule level, resulting in insufficient interpretability and reproducibility of evaluation results.

[0004] Furthermore, while existing reinforcement learning-based large-scale model inference optimization methods can improve model output through reward functions, they typically lack rule extraction, stress case generation, and hierarchy-preserving diagnostic mechanisms adapted to multimodal inference scenarios after knowledge editing. Using only multi-hop question-answering results or overall accuracy as reward signals can easily mask the location of specific rule violations and makes it difficult to distinguish between insufficient basic instruction following ability of the model and the degradation of the inference chain after knowledge editing.

[0005] Therefore, there is an urgent need for a technical solution that can interpretably evaluate and further enhance the reasoning ability of multimodal knowledge-edited models by combining objective and detectable rules, multi-hop stress cases, rule-by-rule compliance judgments, and reinforcement optimization mechanisms. Summary of the Invention

[0006] This application provides a method and system for enhancing the reasoning ability of a multimodal knowledge-edited model, in order to at least solve the problems that existing multimodal knowledge-edited models are prone to hierarchical degradation under multi-hop reasoning pressure scenarios, the rule preservation ability is difficult to objectively quantify, and the enhancement optimization lacks fine-grained rule feedback.

[0007] This application provides a method for enhancing the reasoning ability of a multimodal knowledge-edited model in a first aspect, comprising: acquiring multimodal knowledge-edited samples, the samples including edited knowledge, source hints, and at least two types of modal information associated with the edited knowledge, the at least two types of modal information including text modal information and visual modal information; extracting and normalizing objectively detectable rules from the edited knowledge to obtain a rule set; generating case-conditionalized multi-hop stress cases based on a preset hierarchical failure type, the source hints, and the rule set, and labeling the multi-hop stress cases with a subset of target rules; inputting the multi-hop stress cases into a knowledge-edited multimodal large-scale language model configured with a hybrid expert routing strategy to generate a set of candidate responses; performing rule-by-rule compliance judgment on each candidate response and the rule set to obtain a rule satisfaction result; calculating case-level accuracy, rule-level accuracy, target rule accuracy, and hierarchical preservation score based on the rule satisfaction result; constructing an editing performance reward, an expert co-activation reward, and a sequence consistency reward based on the rule satisfaction result and the hierarchical preservation score, and optimizing and updating the hybrid expert routing strategy or expert parameters through a group relative strategy to enhance the multi-hop reasoning ability of the model.

[0008] Furthermore, the step of extracting and standardizing objectively detectable rules from the edited knowledge to obtain a rule set includes: extracting candidate rules from the knowledge triples corresponding to the edited knowledge; performing detectability screening on the candidate rules, retaining candidate rules that include at least one of precise numerical restrictions, fixed output formats, required fields, prohibited fields, and structured serialization constraints; excluding subjective rules, purely descriptive rules, and rules that cannot be verified through structured judgment; and unifying the fields, constraints, and target objects of the retained candidate rules to obtain the rule set.

[0009] Furthermore, the preset hierarchical failure types include at least two of the following: forgetting type, misunderstanding type, and attack type; the forgetting type includes at least one of single-round drift, multi-round drift, and reasoning-to-answer drift; the misunderstanding type includes at least one of completion bias, role boundary expansion, and partial compliance fallacy; the attack type includes at least one of explicit overriding, role reallocation, task restructuring, indirect execution, nested injection, reference injection, constraint trade-off, progressive inducement, and authority overriding.

[0010] Further, the step of generating conditional multi-hop stress cases based on preset hierarchical failure types, the source hints, and the rule set includes: determining a multi-hop inference chain based on the source hints and the rule set; determining multiple sequentially associated knowledge triples in the multi-hop inference chain, each knowledge triple including a subject, a relation, and an object; determining a target knowledge triple hit by fact editing from the multiple sequentially associated knowledge triples, and replacing the object in the target knowledge triple with the edited object; updating the inference object of subsequent jumps in the multi-hop inference chain according to the edited object; and rewriting the updated multi-hop inference chain according to the preset hierarchical failure types to obtain the conditional multi-hop stress cases.

[0011] Furthermore, the step of inputting the multi-hop stress case into a knowledge-edited multimodal large language model configured with a hybrid expert routing strategy to generate a set of candidate responses includes: obtaining the edited hidden state formed by the knowledge-edited multimodal large language model when processing the multi-hop stress case; generating multiple expert parameters based on the edited hidden state using an expert generator; generating gating probabilities corresponding to the multiple expert parameters using a gating network; determining active experts from the multiple experts according to the gating probabilities; and fusing the outputs of the active experts to generate the set of candidate responses.

[0012] Furthermore, before updating the hybrid expert routing strategy or expert parameters through the group relative strategy, the method further includes: performing random singular value decomposition on the expert parameter matrix to obtain the decomposed matrix components and singular values; injecting normalized noise into the singular values ​​to obtain perturbed singular values; reconstructing candidate expert parameters based on the decomposed matrix components and the perturbed singular values; generating candidate experts based on the candidate expert parameters; constructing deterministic rule checking conditions according to the rule set; and verifying the responses generated by the candidate experts based on the deterministic rule checking conditions to retain candidate experts that meet the rule verification conditions.

[0013] Further, the calculation of case-level accuracy, rule-level accuracy, target rule accuracy, and hierarchy preservation score based on the rule satisfaction results includes: for each multi-hop stress case, determining the case judgment result of the multi-hop stress case based on the rule satisfaction results of a set of candidate responses corresponding to the multi-hop stress case, and judging the multi-hop stress case as correct when the case judgment result indicates that all rules are satisfied; determining the case-level accuracy based on the proportion of correct multi-hop stress cases in all multi-hop stress cases; determining the rule-level accuracy based on the proportion of satisfied rules in all rules; determining the target rule accuracy based on the proportion of satisfied target rules in all target rules; weighting and summing the case-level accuracy, rule-level accuracy, and target rule accuracy according to preset weights to obtain the hierarchy preservation score; the method further includes: retaining the multimodal knowledge editing samples that satisfy the rule set in the first round of responses to form a first-round compliance subset, and recalculating the hierarchy preservation score on the first-round compliance subset to distinguish between baseline instruction following defects and hierarchy degradation in the multi-hop inference process.

[0014] This application also provides a system for enhancing the reasoning ability of a multimodal knowledge editing model in a second aspect, comprising: a sample acquisition module for acquiring multimodal knowledge editing samples, the samples including edited knowledge, source hints, and at least two types of modal information associated with the edited knowledge, the at least two types of modal information including text modal information and visual modal information; a rule normalization module for extracting and normalizing objective and detectable rules from the edited knowledge to obtain a rule set; a stress case generation module for generating case-conditionalized multi-hop stress cases based on a preset level of failure type, the source hints, and the rule set, and labeling the multi-hop stress cases with a subset of target rules; and a response generation module for... The multi-hop stress cases are input into a knowledge-edited multimodal large-scale language model configured with a hybrid expert routing strategy to generate a set of candidate responses. A rule-by-rule judgment module is used to perform rule-by-rule compliance judgment on each candidate response and the rule set to obtain the rule satisfaction result. An index calculation module is used to calculate case-level accuracy, rule-level accuracy, target rule accuracy, and hierarchy preservation score based on the rule satisfaction result. A strategy optimization module is used to construct editing performance rewards, expert co-activation rewards, and sequence consistency rewards based on the rule satisfaction result and hierarchy preservation score, and to optimize and update the hybrid expert routing strategy or expert parameters through a group relative strategy to enhance the multi-hop inference capability of the model.

[0015] Compared with existing technologies, the above technical solution has at least the following beneficial effects: By extracting objective and detectable rules from multimodal knowledge editing samples and standardizing them, the edited knowledge is transformed from a natural language description that is difficult to judge directly into a set of rules that can be verified one by one. Furthermore, by combining preset hierarchical failure types, source hints, and rule sets to generate conditional multi-hop stress cases, the test scenarios can cover common hierarchical degradation scenarios in real deployments, such as single-round drift, multi-round drift, reasoning-to-answer drift, misunderstanding-based failure, and adversarial failure. Therefore, the model response no longer relies solely on the overall correctness of the question-and-answer response, but can locate the satisfaction and violation of specific rules through rule-by-rule compliance judgment, thereby improving the objectivity, interpretability, and reproducibility of the evaluation of the model's reasoning ability after multimodal knowledge editing.

[0016] The aforementioned technical solutions also offer the following advantages: By constructing editing performance rewards, expert co-activation rewards, and sequence consistency rewards based on rule satisfaction results and hierarchy preservation scores, and combining these with group-relative strategies to optimize and update hybrid expert routing strategies or expert parameters, the reinforcement process can revolve around the ability of edited knowledge to be continuously preserved in multi-hop inference chains, rather than simply optimizing the surface accuracy of a single output. Simultaneously, case-level accuracy, rule-level accuracy, target rule accuracy, and hierarchy preservation scores can reflect the model's hierarchy preservation capability at different granularities. Recalculation on the first-round compliant subset can also distinguish between baseline instruction following defects and hierarchy degradation during multi-hop inference. Therefore, these solutions can provide a fine-grained, diagnosable, and feedback-enabled inference capability enhancement path for multimodal knowledge-edited models without relying on full parameter fine-tuning. This makes them suitable for improving model reliability in scenarios such as model security auditing, intelligent agent constraint verification, continuous enterprise knowledge updates, and policy-critical scenarios. Attached Figure Description

[0017] The present invention will be further described below with reference to the accompanying drawings and embodiments: Figure 1 A flowchart illustrating the method for enhancing the reasoning ability of a multimodal knowledge-edited model provided in this application; Figure 2 This is a schematic diagram of the multi-hop inference chain cascading update provided in this application; Figure 3 A schematic diagram of the hybrid expert response generation structure provided in this application; Figure 4 A schematic diagram of the two-stage random SVD sampling and expert verification process provided for this application; Figure 5 A schematic diagram illustrating the multi-hop portability comparison results provided in this application; Figure 6 A schematic diagram illustrating the experimental results of the impact of chain-like reasoning on multi-hop accuracy provided in this application; Figure 7 A schematic diagram of the structure of the multimodal knowledge-edited model reasoning ability enhancement system provided in this application; Figure 8 A schematic diagram of the structure of the computer device provided in this application. Detailed Implementation

[0018] To make the technical solution, implementation method, and beneficial effects of this application clearer, the following description, in conjunction with the accompanying drawings, further illustrates this application. It should be understood that the following embodiments are only for explaining this application and should not be construed as limiting the scope of protection of this application. For those skilled in the art, adaptive adjustments can be made to the step sequence, module deployment method, model type, parameter form, and data organization method in the following embodiments without departing from the technical concept of this application.

[0019] In the description of this application, the terms "first," "second," etc., are used only to distinguish different objects and do not indicate any hierarchy of importance, execution order, or quantity among the objects. The term "multiple" can mean two or more. The terms "including," "comprising," "having," etc., should be understood as open-ended expressions, meaning that in addition to the listed steps, modules, or objects, other unlisted steps, modules, or objects may also be included. The term "and / or" indicates that the relevant objects can exist individually or in any combination.

[0020] In this application, a multimodal large-scale language model can be a large-scale language model capable of processing at least two types of modal information, such as text and images, or a visual language model capable of encoding, reasoning, and generating the relationships between textual and visual information. Knowledge editing refers to the process of injecting facts, correcting relationships, or updating boundaries into existing knowledge, external knowledge, or applied rules within the model. Edited knowledge can be represented as knowledge triples, rule text, cross-modal entity relationships, or security constraints. Source hints can be original user queries, system hints, edit sample hints, or basic question-and-answer hints corresponding to the edited knowledge. A knowledge triple refers to a structured knowledge representation consisting of a subject, a relation, and an object. Objectively detectable rules refer to rules whose satisfaction can be verified through structured judgment methods, such as precise numerical constraints, fixed output formats, required fields, prohibited fields, and structured serialization constraints. A target rule subset refers to a subset of rules that are primarily examined in a multi-hop stress case. A multi-hop stress case refers to an input case that requires passing through two or more reasoning stages and applies test or reinforcement pressure to the ability to maintain edited knowledge.

[0021] In this application, MoE is an abbreviation for Mixture of Experts, representing a hybrid expert structure; GRPO is an abbreviation for Group Relative Policy Optimization; MoE-GRPO represents a reinforcement module that combines a hybrid expert structure with group relative policy optimization; SVD is an abbreviation for Singular Value Decomposition; CoT is an abbreviation for Chain-of-Thought, representing chain-of-thought reasoning; HPS is an abbreviation for Hierarchical Preservation Score; IF-Controlled represents the first-round compliance filtering diagnostic, used to distinguish between basic instruction following defects and hierarchical degradation in multi-hop reasoning processes; MLLM represents a multimodal large-scale language model; QA represents a question-answering task; and LiveEditRL represents an overall framework for organizing the reinforcement of model reasoning capabilities after multimodal knowledge editing.

[0022] Figure 1 A flowchart illustrating the method for enhancing the reasoning ability of a model after multimodal knowledge editing, as provided in this application. Figure 1 As shown, the method for enhancing the reasoning ability of a multimodal knowledge-edited model provided in this application may include the following steps.

[0023] In step S1, multimodal knowledge editing samples are obtained. These samples may include edited knowledge, source hints, and associated textual and visual modal information. Edited knowledge may originate from enterprise knowledge base updates, factual relationship corrections, security policy updates, multimodal question-and-answer revisions, or intelligent agent rule maintenance scenarios. Textual modal information may include natural language facts, structured fields, system hints, user queries, or rule descriptions; visual modal information may include images, entity regions within images, visual entity annotations, or visual evidence corresponding to text entities. These multimodal knowledge editing samples can be used to describe the knowledge content that the model should update, and the constraint boundaries that this knowledge needs to maintain in subsequent multi-hop inference.

[0024] In step S2, objectively detectable rules are extracted and standardized to obtain a rule set. Specifically, candidate rules can be identified from the editing knowledge, and these rules are screened for detectability. Candidate rules that can be verified through structured judgment are retained, while subjective rules, purely descriptive rules, or rules that cannot be verified through deterministic judgment are excluded. The retained candidate rules may include at least one of precise numerical restrictions, fixed output formats, required fields, prohibited fields, and structured serialization constraints. Subsequently, the retained candidate rules can be standardized in terms of fields, constraint expressions, target objects, and triggering conditions to form a rule set. This rule set may include information such as rule identifier, applicable object, triggering condition, constraint type, constraint content, and whether it belongs to a subset of target rules.

[0025] In step S3, multi-hop stress cases are generated and a subset of target rules is labeled. Specifically, conditional multi-hop stress cases can be generated based on the rule set, source hints, and preset hierarchical failure types. Preset hierarchical failure types can include at least two of the following: forgetting types, misunderstanding types, and attack types. Forgetting types can include single-round drift, multi-round drift, and reasoning-to-answer drift; misunderstanding types can include completion bias, role boundary expansion, and partial compliance fallacies; attack types can include explicit coverage, role reassignment, task restructuring, indirect execution, nested injection, reference injection, constraint trade-offs, progressive inducement, and authority coverage. Through these hierarchical failure types, the generated multi-hop stress cases can cover conflict hints, restatement hints, multi-round interaction hints, and adversarial hints that the model may encounter in real-world deployments.

[0026] When generating multi-hop stress cases, a multi-hop inference chain can be determined based on source hints and a set of rules. In one embodiment, the multi-hop inference chain can consist of multiple sequentially linked knowledge triples, each knowledge triple including a subject, a relation, and an object. If a target knowledge triple is matched by fact editing, the object in the target knowledge triple is replaced with the edited object, and the inference object of subsequent jump steps is updated based on the edited object, so that subsequent jump steps no longer infer based on the object before editing, but form a new multi-hop inference chain based on the edited object. When generating multi-hop stress cases, a subset of target rules can also be labeled for each multi-hop stress case, enabling subsequent judgments to distinguish between cases where all rules are satisfied and cases where the target rules are maintained.

[0027] Figure 2 This is a schematic diagram illustrating the multi-hop inference chain cascading update provided in this application. For example... Figure 2As shown, the pre-edit multi-hop inference chain can include knowledge triple 1, knowledge triple 2, and knowledge triple 3. Knowledge triple 1 includes subject 1, relation 1, and object 1; knowledge triple 2 includes subject 2, relation 2, and object 2; and knowledge triple 3 includes subject 3, relation 3, and object 3. Adjacent knowledge triples can be associated through the previous hop object, the previous hop inference result, or an object determined by the previous hop inference result. When the target knowledge triple is edited in the k-th hop, the object in the target knowledge triple is replaced with the edited object, and the inference object in subsequent hops is updated accordingly, forming an updated multi-hop inference chain. Therefore, the model not only needs to maintain the edited knowledge at the target knowledge triple but also needs to continuously perform cascading inference based on the edited object in subsequent hops.

[0028] In one embodiment, a multi-hop inference chain can be represented as:

[0029] in, This represents a multi-hop reasoning chain consisting of multiple knowledge triples. , This indicates the main body in different jump steps. , This indicates the relationship between different jump steps. , Represents the objects in different jump steps. For the first... The single-fact editing process for the target knowledge triplet that is jumped to can be represented as follows:

[0030] in, Indicates the first The main body of the jump Indicates the first The relationship of jumping, Indicates the object before editing. This represents the edited object. After the target knowledge triple is edited, the inference objects of subsequent jumps are updated in a cascading fashion along with the edited object. Figure 2 In the illustrated embodiment, the edited subsequent reasoning object can be represented as:

[0031] The object of subsequent reasoning before editing can be represented as:

[0032] in, This represents the mapping relationship where the object of the previous jump determines the object of the next jump inference. Therefore, when... Replaced with After that, the The objects of reasoning for jumps and subsequent jump steps also need to be updated accordingly, so that the model can perform cascading reasoning based on the edited knowledge in multi-hop stress cases, instead of continuing to use the old reasoning chain corresponding to the object before editing.

[0033] In step S4, the multi-hop stress case is input into the knowledge-edited MLLM and a set of candidate responses is generated. This knowledge-edited MLLM can be configured with a hybrid expert routing strategy. For the same multi-hop stress case, the model can generate multiple candidate responses, which can be used for subsequent rule-by-rule compliance judgment and group relative strategy optimization. By generating a set of candidate responses, the degree of rule satisfaction among different responses can be compared under the same input conditions, thereby forming a relative advantage signal and avoiding coarse-grained optimization based solely on a single response.

[0034] Figure 3 This is a schematic diagram of the hybrid expert response generation structure provided in this application. Figure 3 As shown, after knowledge editing, MLLM forms a hidden state when handling multi-hop stress cases. Hide after editing You can input the expert generator separately. and gated networks Expert generator Used to determine the hidden state after editing Generate multiple expert parameters, which can be represented as a set of candidate experts, such as expert 1, expert 2, ..., expert... Gated networks Used to determine the hidden state after editing Generate the gating probability, which can be expressed as: Then, active experts can be selected from multiple candidate experts based on gating probabilities. Active experts can include expert 1 parameter, expert 2 parameter, expert 3 parameter, and expert... Parameters, etc., among which Active experts are assigned numbers. Subsequently, the outputs of these active experts are fused, which can be done using a weighted summation method, where the weights can be determined by gating probabilities. The fusion results are used to generate candidate responses. .

[0035] In this embodiment, the edited hidden state Used to characterize the semantic state formed by the model under current editing knowledge and multi-hop pressure cases; expert generator Expert parameters used to generate parameters adapted to the current editing task; gating network Gating probabilities used to output the required expert selection or expert fusion. Candidate Response This is used to represent the response result obtained after fusing the outputs of active experts. Through the above structure, the model can dynamically call different experts for different editing knowledge and different stress cases, and control the expert contributions through gating probabilities, thereby improving the rule preservation ability in multi-hop inference.

[0036] In one embodiment, the edited hidden state can be represented as .in, Indicates the first The visual modal information input at any time, Indicates the first The text modal information or status information input at any time. Represents the target information or output constraints corresponding to the current input. Expert generator Can be based on the hidden state after editing Generate candidate expert parameters; Gated network Based on the current gating state and editing status Generate gate probability Gating probability Used to identify the set of active experts from multiple candidate experts. This is used to determine the weight of each active expert's output in the fusion process.

[0037] In an exemplary representation, the output hidden state corresponding to a candidate response can be represented as:

[0038] in, This indicates the hidden output state corresponding to the candidate response. Indicates the hidden state after editing. This represents the set of active experts determined under the current input conditions. Indicates the first The gating probability for each expert This indicates that the gating probabilities or expert contributions of at least some active experts are aggregated. Indicates the first Each expert adjusts the hidden state based on the edited hidden state. As expressed above, the candidate response is not directly generated from the original edited hidden state, but rather obtained by modifying the edited hidden state using the weighted output of the active expert set.

[0039] In one embodiment, the difference term of the gating probability can be expressed as:

[0040] in, This indicates the difference in gating or routing adjustments after editing. Indicates the first The gating probability for each expert This indicates the current gating state or baseline gating probability. This difference term can be used to characterize the degree of deviation of a specific expert relative to the current gating state, thus providing a reference for subsequent updates to the hybrid expert routing policy or expert parameters via GRPO.

[0041] In step S5, each candidate response is evaluated for compliance with the rule set on a rule-by-rule basis to obtain the rule satisfaction result. Specifically, for each candidate response, it can be determined whether the candidate response satisfies the corresponding rule according to each rule in the rule set. Rule-by-rule compliance evaluation can be implemented using a structured judge, or it can be implemented using a rule checking program, template matching program, field parsing program, or model-assisted evaluation program. The evaluation output can include rule identifier, whether it is satisfied, violation location, violation reason, and confidence information. Compared with only judging whether the final answer is correct, rule-by-rule compliance evaluation can locate which specific rule is violated in the candidate response, thus providing fine-grained feedback for subsequent indicator calculation and reward construction.

[0042] In step S6, case-level accuracy, rule-level accuracy, target rule accuracy, and hierarchy preservation score are calculated. For each multi-hop stress case, the case judgment result can be determined based on the rule satisfaction results of a set of candidate responses corresponding to that multi-hop stress case. If the case judgment result indicates that all rules of the corresponding multi-hop stress case are satisfied, the multi-hop stress case can be judged as correct, and the case-level accuracy is determined accordingly. Rule-level accuracy can be determined based on the proportion of satisfied rules among all rules. Target rule accuracy can be determined based on the proportion of satisfied target rules among all target rules. The hierarchy preservation score is obtained by weighted summation of case-level accuracy, rule-level accuracy, and target rule accuracy, and is used to comprehensively characterize the ability of the edited model to maintain edited knowledge and rule constraints in the multi-hop inference chain.

[0043] In one embodiment, a first-round compliance filtering diagnostic can also be introduced. Specifically, for multimodal knowledge editing samples, it can be first determined whether the model satisfies the rule set in the first-round response. If the first-round response no longer satisfies the rule set, the sample may reflect insufficient basic instruction following ability of the model. For samples whose first-round response satisfies the rule set, a first-round compliance subset can be formed, and the hierarchical preservation score can be recalculated on the first-round compliance subset. If the hierarchical preservation score is low on the complete sample, but significantly improved on the first-round compliance subset, it indicates that some errors originate from basic instruction following defects; if there is still a significant decrease on the first-round compliance subset, it indicates that the model has hierarchical degradation in subsequent multi-hop inference processes.

[0044] In step S7, a multi-dimensional reward is constructed, and the hybrid expert routing strategy or expert parameters are updated via GRPO. Specifically, editing performance rewards, expert co-activation rewards, and sequence consistency rewards can be constructed based on rule satisfaction results and hierarchical preservation scores. Editing performance rewards can be used to encourage candidate responses to satisfy the rule set and target rule subset corresponding to the edited knowledge; expert co-activation rewards can be used to regulate the collaborative relationship among active experts, avoiding excessive concentration of expert selection or imbalance of expert contributions; sequence consistency rewards can be used to constrain candidate responses to maintain stable rule compliance results under multi-round, multi-hop, or restatement conditions.

[0045] In one embodiment, the model reasoning reinforcement process can be represented as a state. Next action The strategy optimization process involves constructing the total reward based on the results of rule satisfaction and hierarchical preservation scores. The total reward can be expressed as:

[0046] in, Indicates the state Next action The total reward received at that time This indicates an editing performance bonus. This indicates that the experts jointly activated the reward. Indicates the first The sequence consistency reward corresponding to time step 1. and This indicates the weighting coefficient for the corresponding reward item. Status This can include at least one of the following: current multi-hop stress case, edited hidden state, rule set, and historical response information; action. It may include at least one of the following: expert selection action, route adjustment action, or expert parameter update action.

[0047] Among them, editing performance rewards It can be constructed based on locality, reliability, generalization, and multi-hop portability loss, and is expressed as:

[0048] in, Indicates localized loss. Indicates a loss of reliability. Indicates generalization loss. This indicates multi-hop portability loss. and This represents the weighting coefficient of the corresponding loss term. By adopting a negative loss form, the performance reward can be increased while reducing locality of decay, improving reliability, enhancing generalization ability, and improving multi-hop portability.

[0049] Experts activate rewards It can be used to suppress repeated co-activation among functionally similar experts, as shown below:

[0050] in, This indicates the number of active experts under the current input conditions. and They represent the first One active expert and the first The capacity-normalized weights of active experts, and These represent the amount of hidden state adjustment generated by the corresponding active expert. This represents the similarity between two hidden state adjustments. Because this reward item has a negative sign, the reward value decreases when the outputs of two high-weighted experts are highly similar, thereby incentivizing different active experts to undertake complementary edit-preservation, relationship-preservation, or response-generation tasks.

[0051] In one embodiment, the expert capacity normalization weight can be expressed as:

[0052] in, Indicates the first The capacity-normalized weights of individual experts, Indicates the first A parameter vector or capacity representation vector for each expert. This represents the L2 norm of the parameter vector or capacity representation vector. Through the above normalization method, expert contributions can be smoothly allocated based on the capacity representations of different experts.

[0053] Sequence Consistency Reward It can be constructed based on the similarity of the expert sparsity distribution between the current step and historical steps, and is represented as:

[0054] in, Indicates the first The expert sparsity distribution vector of the step. Indicates the first The expert sparsity distribution vector of the step. Indicates the callback depth or history window depth. Indicates the attenuation factor. This represents the similarity between two expert sparsity distribution vectors. A decay factor is introduced to... This can allow more recent historical steps to have a greater impact on sequence consistency rewards, thereby modulating the historical continuity and variability of expert routes in multi-round, multi-hop, or restatement stress cases.

[0055] Subsequently, based on the total reward corresponding to a set of candidate responses, the relative advantage within the group can be calculated using GRPO, and the hybrid expert routing strategy or expert parameters can be updated according to the relative advantage within the group, so that the model is more inclined to generate candidate responses that satisfy the rule set and maintain editing knowledge in subsequent multi-hop stress cases.

[0056] Figure 4 This is a schematic diagram of the two-stage random SVD sampling and expert verification process provided for this application. Figure 4 As shown, before updating the hybrid expert routing policy or expert parameters via GRPO, the expert parameter matrix can be processed in two stages. The first stage is the random SVD sampling stage, with the expert parameter matrix as the input.

[0057] In one embodiment, the stochastic SVD decomposition of the expert parameter matrix can be expressed as:

[0058] in, Indicates the first The expert parameter matrix to be optimized at any given time or current moment. and Represents the components of the decomposed matrix. Represents a singular value diagonal matrix. Indicates reservation There are singular values. By performing singular value decomposition on the expert parameter matrix, the perturbation of the expert parameters can be transformed from the complete matrix space to the singular value space and the matrix component space, thereby reducing the computational complexity of the candidate expert search process.

[0059] In the early stages ( , Indicates the current iteration number. (Representing the stage switching threshold), normalized noise can be injected into the singular values ​​to obtain the perturbed singular values. The perturbed singular values ​​are... A singular value can be represented as:

[0060] in, Indicates the first The perturbation-induced singular values ​​obtained in the next iteration Indicates the first [number] before the disturbance A singular value, This represents the maximum value among the singular values. Indicates the noise intensity factor. Indicates random noise, and It follows a mean of 0 and a variance parameter of . The Gaussian distribution. By following... Normalizing the noise amplitude can make the perturbation intensity of larger and smaller singular values ​​match their scale, avoiding excessive perturbation of small singular values ​​and thus compromising the stability of expert parameters.

[0061] In the later stages, full-spectrum perturbation can be performed, and candidate expert parameters can be reconstructed based on the perturbed singular values ​​and the decomposed matrix components. Full-spectrum perturbation refers to applying perturbation to all retained singular values ​​as a whole, rather than only to a subset. The reconstructed candidate expert parameters can be used to generate candidate experts and further validated through deterministic rule checks. Thus, two-stage stochastic SVD sampling enables stable local exploration in the early stages and expands the candidate expert search range in the later stages, thereby improving the adaptability of candidate expert parameters to different multi-hop stress cases.

[0062] By performing stochastic SVD decomposition on the expert parameter matrix, the matrix components and singular values ​​are obtained. Normalized noise is then injected into the singular values ​​to obtain the perturbed parameter representation. By injecting normalized noise at the singular value level, different candidate expert directions can be explored with lower parameter perturbation costs, reducing the computational burden caused by directly performing random perturbations on the complete expert parameters.

[0063] The second stage is the reconstruction and verification stage. First, candidate expert parameters are reconstructed based on the decomposed matrix components and the singular values ​​after injecting normalized noise, and candidate experts are generated based on these parameters. Then, deterministic rule checking conditions are constructed based on the rule set, and it is determined whether the responses generated by the candidate experts meet the rule verification conditions. If the rule verification conditions are met, the candidate expert is retained; otherwise, the candidate expert is discarded or sampling continues. Through these two stages, candidate experts that are more conducive to rule preservation and multi-hop inference stability can be selected from the expert parameter search space.

[0064] Figure 5 This is a schematic diagram illustrating the comparison results of multi-hop portability provided in this application. Figure 5 As shown, on two types of MLLMs, BLIP2-OPT and LLaVA-v1.5, the portability of methods such as FT, MEND, SERAC, RECIPE, RLEdit, LEMoE, and LiveEdit in average 2- to 4-hop inference tasks is compared under test conditions including Orig, Input, and Reph. The vertical axis in the figure represents portability; a higher value indicates that the edited knowledge can be preserved and transferred in multi-hop inference chains. The different colored bars in the figure correspond to the portability data under the MoE-GRPO full benchmark, the LEMoE control method, the MoE-GRPO first-round compliance subset, and the MoE-GRPO control setting, respectively. Figure 5It can be seen that different editing methods show significant differences in multi-hop portability, indicating that relying solely on single-hop editing accuracy is insufficient to fully reflect the ability of the edited model to maintain hierarchy under real inference pressure.

[0065] Furthermore, Figure 5 The full baseline and the first-round compliance subset can be used to illustrate the necessity of the first-round compliance filtering diagnostics. Low scores on the full baseline may simultaneously contain both basic instruction following failures and subsequent multi-hop inference degradation errors, while the first-round compliance subset can, to some extent, exclude samples that do not follow the rules in the first round, thus allowing for a more focused observation of hierarchy preservation capabilities within the multi-hop inference chain. Therefore, this application improves the accuracy and interpretability of the model diagnostic results by recalculating the hierarchy preservation score on the first-round compliance subset.

[0066] Figure 6 This is a schematic diagram illustrating the experimental results of the impact of chain-like reasoning on multi-hop accuracy provided in this application. Figure 6 As shown, for models such as Qwen2.5-VL, Qwen2.5-7B-Instruct, and LLaVA-Next-7B, the μ accuracy can be compared under three conditions: baseline results, with CoT, and without CoT. Figure 6 In the 2-hop, 3-hop, and 4-hop scenarios, using CoT showed varying degrees of improvement compared to not using CoT, such as +2.5%, +1.8%, and +3.2% as marked in the figure. These experimental results demonstrate that explicitly unfolding the intermediate inference process helps the model maintain better inference consistency in some multi-hop scenarios. However, CoT itself cannot replace rule extraction, rule-by-rule compliance judgment, and reward feedback mechanisms. Therefore, in this application, CoT can be used as an optional inference enhancement method, combined with the aforementioned multi-hop stress case generation, rule judgment, and policy optimization processes.

[0067] Figure 7 A schematic diagram of the structure of the multimodal knowledge-edited model reasoning ability enhancement system provided in this application. (See diagram for example.) Figure 7 As shown, the system may include a sample acquisition module 11, a rule normalization module 12, a stress case generation module 13, a response generation module 14, a rule-by-rule judgment module 15, an indicator calculation module 16, and a strategy optimization module 17.

[0068] The sample acquisition module 11 is used to acquire multimodal knowledge editing samples. These samples may include edited knowledge, source hints, and textual and visual modal information associated with the edited knowledge. The sample acquisition module 11 can obtain this information from a knowledge base, training dataset, edited sample library, multimodal question-answering dataset, or manually configured profiles, and output the edited knowledge samples to the rule normalization module 12.

[0069] The rule normalization module 12 is used to extract and normalize objective, detectable rules from the edited knowledge, resulting in a rule set. The rule normalization module 12 can identify precise numerical restrictions, fixed output formats, required fields, prohibited fields, and structured serialization constraints in candidate rules, and exclude subjective rules, purely descriptive rules, and rules that cannot be verified through structured judgment. The rule normalization module 12 can then output the rule set to the stress test generation module 13.

[0070] The stress case generation module 13 generates multi-hop stress cases based on preset failure types, source hints, and rule sets, and labels the multi-hop stress cases with a subset of target rules. The stress case generation module 13 can construct multi-hop inference chains based on the rule set and generate multi-hop stress cases through methods such as single-round drift, multi-round drift, inference-to-answer drift, completion deviation, role boundary expansion, partial compliance fallacies, and adversarial hint rewriting. The stress case generation module 13 can output the multi-hop stress cases to the response generation module 14.

[0071] The response generation module 14 is used to input multi-hop stress cases into a knowledge-edited MLLM and generate a set of candidate responses. The response generation module 14 may include a hybrid expert structure and an expert generator. Gating network And an active expert pool. Response generation module 14 can be based on the edited hidden state. Generate multiple candidate expert parameters and base them on gating probabilities. Select active experts, and then fuse their outputs to obtain candidate responses. Candidate responses can be output to the rule-by-rule judgment module 15.

[0072] The rule-by-rule judgment module 15 performs rule-by-rule compliance judgment on each candidate response and the rule set to obtain the rule satisfaction result. The rule-by-rule judgment module 15 can check the fields, format, values, prohibited content, and structured output of each candidate response and determine whether it satisfies each rule in the rule set. The rule-by-rule judgment module 15 can output the rule satisfaction result to the indicator calculation module 16.

[0073] The metric calculation module 16 is used to calculate case-level accuracy, rule-level accuracy, target rule accuracy, and hierarchy preservation score based on rule fulfillment results. The metric calculation module 16 can also retain samples that satisfy the rule set in the first round of responses, forming a first-round compliance subset, and recalculate the hierarchy preservation score on this first-round compliance subset to distinguish between basic instruction following defects and hierarchy degradation in multi-hop inference processes. The metric calculation module 16 can output the hierarchy preservation score and reward signal to the policy optimization module 17.

[0074] The strategy optimization module 17 is used to construct editing performance rewards, expert co-activation rewards, and sequence consistency rewards based on rule satisfaction results and hierarchical preservation scores, and to update the hybrid expert routing strategy or expert parameters via GRPO. The strategy optimization module 17 can feed back the updated routing strategy or expert parameter results to the response generation module 14, enabling the response generation module 14 to select a more suitable combination of active experts for the current editing knowledge and rule preservation task in subsequent multi-hop stress cases.

[0075] Figure 8 A schematic diagram of the structure of the computer device provided in this application. Figure 8 As shown, a computer device may include a processor, internal memory, non-volatile storage media, a system bus, a network interface, a display screen, and input devices. The processor can be connected to the internal memory, non-volatile storage media, network interface, display screen, and input devices via the system bus. The non-volatile storage media can store the operating system and computer programs, and the internal memory provides the environment for the operation of the operating system and computer programs. The network interface can be used to communicate with external terminals, servers, model service platforms, or knowledge base systems. The display screen can be used to display multi-hop stress cases, candidate responses, rule satisfaction results, hierarchy preservation scores, or policy optimization results. The input devices can be used to receive user-configured edit knowledge, rule parameters, model selection information, or test task information.

[0076] When the computer program is executed by the processor, it enables the computer device to implement the method for enhancing the reasoning ability of the multimodal knowledge-edited model provided in this application. For example, the processor can perform the following processing: acquiring multimodal knowledge-edited samples; extracting and standardizing objectively detectable rules from the edited knowledge to obtain a rule set; generating multi-hop stress cases based on preset hierarchical failure types, source hints, and the rule set, and labeling a subset of target rules; inputting the multi-hop stress cases into the knowledge-edited MLLM to generate a set of candidate responses; performing rule-by-rule compliance judgment on each candidate response; calculating case-level accuracy, rule-level accuracy, target rule accuracy, and hierarchical preservation score; constructing a multi-dimensional reward based on the rule satisfaction result and the hierarchical preservation score, and updating the hybrid expert routing strategy or expert parameters through GRPO.

[0077] This application also provides a program product. This program product can be stored in a non-transitory computer-readable storage medium. When the program product is run on a computer device, the computer device can execute the aforementioned method for enhancing model reasoning capabilities after multimodal knowledge editing. The non-transitory computer-readable storage medium may include read-only memory, random access memory, flash memory, disk, optical disk, solid-state drive, or other media capable of storing computer programs. The program product can run locally, or it can be downloaded via a network or deployed on a cloud model service platform.

[0078] Through the above implementation methods, this application can transform edited knowledge in multimodal knowledge editing samples into a set of rules that can be verified one by one, and generate multi-hop stress cases based on hierarchical failure types, enabling the model's rule retention capability under multi-hop, restatement, multi-round, and adversarial stress to be objectively evaluated. Furthermore, this application generates candidate responses through a hybrid expert structure, obtains fine-grained rule satisfaction results through rule-by-rule compliance judgment, and constructs a reward signal based on hierarchical retention scores, updating the hybrid expert routing strategy or expert parameters using GRPO. Therefore, this application can provide a fine-grained, diagnostic, and feedback-based reasoning capability enhancement path for the multimodal knowledge-edited model without relying on full parameter fine-tuning, thereby improving the continuous retention capability of edited knowledge in multi-hop inference chains and the reliability of model deployment.

[0079] The above embodiments are merely illustrative of the technical concept and features of this application, intended to enable those skilled in the art to understand the content of the invention and implement it accordingly, and should not be construed as limiting the scope of protection of the invention. It is obvious to those skilled in the art that the invention is not limited to the details of the above exemplary embodiments, and that the invention can be implemented in other specific forms without departing from the spirit or essential characteristics of the invention. Therefore, the embodiments should be considered exemplary and non-limiting in all respects. The scope of the invention is defined by the appended claims rather than the foregoing description, and thus all changes falling within the meaning and scope of the equivalents of the claims are intended to be included within the invention.

Claims

1. A method for enhancing the reasoning ability of a model after multimodal knowledge editing, characterized in that, include: Obtain multimodal knowledge editing samples, the samples including edited knowledge, source hints, and at least two types of modal information associated with the edited knowledge, the at least two types of modal information including text modal information and visual modal information; Objective and detectable rules are extracted from the edited knowledge and standardized to obtain a rule set; Based on the preset hierarchical failure type, the source prompt, and the rule set, a case-conditionalized multi-hop stress case is generated, and a target rule subset is labeled for the multi-hop stress case; The multi-hop stress case is input into a knowledge-edited multimodal large language model configured with a hybrid expert routing strategy to generate a set of candidate responses; Each candidate response is evaluated for compliance with the rule set on a rule-by-rule basis to obtain the rule satisfaction result. Based on the results obtained from the rules, calculate the case-level accuracy, rule-level accuracy, target rule accuracy, and hierarchy preservation score. Based on the rules, the model constructs editing performance rewards, expert co-activation rewards, and sequence consistency rewards according to the results and hierarchical preservation scores. The model also optimizes and updates the hybrid expert routing strategy or expert parameters through a group relative strategy to enhance the model's multi-hop inference capability.

2. The method for enhancing the reasoning ability of a model after multimodal knowledge editing according to claim 1, characterized in that, The step of extracting and standardizing objectively detectable rules from the edited knowledge to obtain a rule set includes: Extract candidate rules from the knowledge triples corresponding to the edited knowledge; The candidate rules are screened for detectability, and candidate rules including at least one of precise numerical restrictions, fixed output formats, required fields, prohibited fields and structured serialization constraints are retained; Exclude subjective rules, purely descriptive rules, and rules that cannot be verified through structured judgment; The retained candidate rules are standardized in terms of fields, constraint expressions, and target objects to obtain the rule set.

3. The method for enhancing the reasoning ability of a model after multimodal knowledge editing according to claim 1, characterized in that, The preset failure types include at least two of the following: forgetting type, misunderstanding type, and attack type; The forgetting types include at least one of single-round drift, multi-round drift, and reasoning-to-answer drift; The types of misunderstandings include at least one of completion deviation, role boundary expansion, and partial compliance fallacy; The attack types include at least one of explicit overriding, role reassignment, task restructuring, indirect execution, nested injection, reference injection, constraint trade-offs, progressive inducement, and authority overriding.

4. The method for enhancing the reasoning ability of a model after multimodal knowledge editing according to claim 1, characterized in that, The process of generating conditional multi-hop stress cases based on preset failure types, source hints, and rule sets includes: A multi-hop inference chain is determined based on the source hints and the rule set; In the multi-hop reasoning chain, multiple sequentially associated knowledge triples are identified, each knowledge triple including a subject, a relation, and an object; From the plurality of sequentially associated knowledge triples, determine the target knowledge triple that is hit by the fact editing, and replace the object in the target knowledge triple with the edited object; Update the reasoning object of subsequent jumps in the multi-hop reasoning chain based on the edited object; The updated multi-hop inference chain is rewritten with hints according to the preset level of failure type to obtain the case conditional multi-hop stress case.

5. The method for enhancing the reasoning ability of a model after multimodal knowledge editing according to claim 1, characterized in that, The process involves inputting the multi-hop stress case into a knowledge-edited multimodal large language model configured with a hybrid expert routing strategy to generate a set of candidate responses, including: Obtain the edited hidden state of the multimodal large language model after knowledge editing when processing the multi-hop stress case; Multiple expert parameters are generated based on the edited hidden state using an expert generator; Gating probabilities corresponding to the multiple expert parameters are generated through a gating network; Active experts are determined from among multiple experts based on the gating probability; The outputs of the active experts are combined to generate the set of candidate responses.

6. The method for enhancing the reasoning ability of a model after multimodal knowledge editing according to claim 5, characterized in that, Before updating the hybrid expert routing policy or expert parameters through the group-relative strategy, the method further includes: Perform random singular value decomposition on the expert parameter matrix to obtain the matrix components and singular values; Normalized noise is injected into the singular values ​​to obtain perturbed singular values. Candidate expert parameters are reconstructed based on the decomposed matrix components and the perturbed singular values; Candidate experts are generated based on the candidate expert parameters. Deterministic rule checking conditions are constructed according to the rule set. The responses generated by the candidate experts are verified based on the deterministic rule checking conditions, so as to retain candidate experts that meet the rule verification conditions.

7. The method for enhancing the reasoning ability of a model after multimodal knowledge editing according to claim 1, characterized in that, The calculation of case-level accuracy, rule-level accuracy, target rule accuracy, and hierarchy preservation score based on the results of the rule satisfaction includes: For each multi-hop stress case, the case judgment result of the multi-hop stress case is determined according to the rule satisfaction result of a set of candidate responses corresponding to the multi-hop stress case. When the case judgment result indicates that all rules are satisfied, the multi-hop stress case is judged as correct. The case-level accuracy is determined based on the percentage of correct multi-hop stress cases among all multi-hop stress cases. The rule-level accuracy is determined based on the percentage of satisfied rules among all rules. The accuracy of the target rule is determined based on the proportion of the target rule that is satisfied among all target rules. The case-level accuracy, rule-level accuracy, and target rule accuracy are weighted and summed according to preset weights to obtain the hierarchical preservation score; The method further includes: retaining multimodal knowledge editing samples that satisfy the rule set in the first round of response to form a first-round compliance subset, and recalculating the hierarchical preservation score on the first-round compliance subset to distinguish between baseline instruction following defects and hierarchical degradation in multi-hop inference processes.

8. A system for enhancing the reasoning ability of a multimodal knowledge-edited model, characterized in that, include: The sample acquisition module is used to acquire multimodal knowledge editing samples, the samples including editing knowledge, source prompts, and at least two types of modal information associated with the editing knowledge, the at least two types of modal information including text modal information and visual modal information; The rule normalization module is used to extract and normalize objective and detectable rules from the edited knowledge to obtain a rule set. The stress case generation module is used to generate conditional multi-hop stress cases based on preset hierarchical failure types, the source prompts, and the rule set, and to label the multi-hop stress cases with a subset of target rules. The response generation module is used to input the multi-hop stress case into a knowledge-edited multimodal large language model configured with a hybrid expert routing strategy, and generate a set of candidate responses; The rule-by-rule judgment module is used to perform rule-by-rule compliance judgment on each candidate response and the rule set to obtain the rule satisfaction result; The indicator calculation module is used to calculate case-level accuracy, rule-level accuracy, target rule accuracy, and hierarchical preservation score based on the results of the rule satisfaction. The strategy optimization module is used to construct editing performance rewards, expert co-activation rewards, and sequence consistency rewards based on the results and hierarchical retention scores of the rules, and to update the hybrid expert routing strategy or expert parameters through group relative strategy optimization to enhance the multi-hop inference capability of the model.

Images