An anti-illusion large model method and system for special child education rehabilitation

Abstract

This invention relates to the field of special education and rehabilitation technology, and particularly to a large-scale anti-hallucination model method and system for the education and rehabilitation of children with special needs, comprising: S1, fusing input data or instructions with context to construct a global state feature vector; S2, reasoning about and verifying the target action: if valid, calling knowledge retrieval or data analysis tools to obtain professional evidence and observation results and looping; if invalid, activating the automatic regression channel and switching to a preset template as a fallback output for each domain score; S3, accumulating the scores to obtain a total score and determining the developmental level, outputting a diagnostic report containing professional evidence constraints; S4, constructing a capability profile, extracting indicators of weak domains and interference parameters, weighting and quantifying them to obtain an intervention priority score; S5, reasoning progressively along the vertical logical chain according to priority, with downstream nodes embedding the feature constraints of upstream nodes, and outputting a personalized intervention plan.

Description

Technical Field

[0001] This invention relates to the field of special education rehabilitation technology, and in particular to a large-scale anti-hallucination model method and system for the education and rehabilitation of children with special needs. Background Technology

[0002] Special needs education and rehabilitation is a systematic project that highly relies on multidisciplinary expertise in medicine, developmental psychology, and applied behavior analysis. Its entire process encompasses the analysis of multi-source, heterogeneous assessment data on children, developmental level assessment, diagnostic report writing, and the development of individualized intervention plans. With the rapid development of large language models in natural language understanding and intelligent decision support, utilizing large models to assist in the analysis of special education data and the generation of structured reports has gradually become an important technical path to improve the intelligence level of education and rehabilitation. Under this technical architecture, the system typically drives the large model to generate the next target action and external tool invocation instructions through an intelligent agent reasoning mechanism, thereby gradually retrieving different file parsing or knowledge retrieval components to assist in completing complex assessment tasks.

[0003] In existing technologies, large language models are prone to planning-level illusions as the reasoning process deepens under long-term, high-information-density special education assessment tasks. This leads to format distortions, incorrect tool names, or out-of-bounds parameters in the generated tool invocation instructions. Such illegal action instructions, resulting from the loss of control during the execution of large models, directly cause logical collapses in the underlying executors and related functional modules in existing technologies. Consequently, the entire long pipeline assessment task is abnormally interrupted, the system hangs, and it cannot stably output the final scores for each core domain. Summary of the Invention

[0004] To overcome the above deficiencies, this invention provides a method and system for anti-hallucination large modeling for the education and rehabilitation of children with special needs. It aims to improve the problem that the large model generates illegal action instructions due to planning hallucinations, which leads to the collapse of the underlying actuator logic and causes abnormal interruption of the entire pipeline evaluation task and system hang.

[0005] In a first aspect, the present invention provides the following technical solution: a large-scale anti-hallucination model method for the education and rehabilitation of children with special needs, comprising: S1. Receive children's rehabilitation data or user instructions, integrate current input with historical context, and dynamically construct a global state feature vector; S2. Based on the global state feature vector, infer the target action and perform a legality check: If legal, call the matching knowledge retrieval or data analysis tool to obtain professional evidence and observation results, and return to step S1 until the score calculation for each field is completed in a loop; if illegal, activate the automatic degradation control channel and switch to the fallback generation mode of the preset template to output the score for each field. S3. The scores of each area are summed to obtain the overall assessment score, and the development level is determined by mapping and matching. The result is a diagnostic assessment report for children that includes professional evidence. S4. Construct a capability profile based on the child's diagnostic assessment report, extract multidimensional indicators and interference parameters of the weak areas to be intervened, perform weighted quantitative calculation using preset weights, and obtain the intervention priority score for the weak areas. S5. Establish the implementation order based on the intervention priority scores of each weak area, and use this order to carry out progressive large model reasoning along the vertical logical chain. The generation of each downstream node is embedded with the feature constraints generated by the upstream node at the previous level, and finally outputs a personalized education and rehabilitation intervention plan.

[0006] Secondly, the present invention provides the following technical solution: a large-scale anti-hallucination model system for the education and rehabilitation of children with special needs, the system comprising: The feature vector construction module is used to receive children's rehabilitation data or user instructions, integrate the current input with the historical context, and dynamically construct a global state feature vector. The action verification and degradation control module is used to infer the target action based on the global state feature vector and perform legality verification: if legal, it calls the matching knowledge retrieval or data analysis tool to obtain professional evidence and observation results, and returns to the feature vector construction module until the score calculation of each field is completed in a loop; if illegal, it activates the automatic degradation control channel and switches to the fallback generation mode of the preset template to output the score of each field. The assessment report generation module is used to accumulate the scores of each domain output by the action verification and regression control module to obtain the overall assessment score, and determine the development level through mapping and matching, and output a child diagnostic assessment report containing professional evidence constraints or the preset template constraints. The ability profile and priority calculation module is used to construct an ability profile based on the child's diagnostic assessment report, extract multi-dimensional indicators and interference parameters of the weak areas to be intervened, and perform weighted quantitative calculation using preset weights to obtain the intervention priority score for the weak area. The intervention plan cascade generation module is used to establish the implementation order based on the intervention priority scores of each weak area. Following this order, a large model reasoning is performed step by step along the vertical logical chain. The generation of each downstream node embeds the feature constraints generated by the upstream node at the previous level, and finally outputs a personalized education and rehabilitation intervention plan.

[0007] The present invention has the following beneficial effects: 1. In this invention, by introducing a rigid action legality verification algorithm and activating an automatic degradation control channel in the agent reasoning process, a soft landing transformation of the system to a fallback generation mode is achieved when the agent's generated instructions are distorted or the tool call fails. This solves the problem of abnormal interruption and task suspension of the entire system flow due to instruction parsing errors and non-standard formats when the general large model agent faces high-complexity special education assessment data.

[0008] 2. In this invention, by constructing the AgenticRAG anti-hallucination retrieval enhancement link oriented towards professional knowledge constraints, the function of accurately recalling the Top-K professional knowledge evidence set from the preset special education and rehabilitation knowledge base using semantic similarity algorithm and using it as a hard fact constraint when decoding the large language model is solved. This solves the problem that the general large language model frequently fabricates false professional questions and answers and erroneous assessment conclusions due to the lack of factual basis when directly applied to the education and rehabilitation scenarios of special children.

[0009] 3. In this invention, by introducing an intervention priority quantification algorithm based on multidimensional indicators and operations research interference parameters on the basis of the profile construction unit, the invention realizes the multi-channel weighted quantification summation of the actual deficit residual ratio in the core domain of children with the risk score of performance behavior, the importance coefficient of daily life adaptation, and the intensity of family needs, etc. This solves the problem that traditional special education rehabilitation program formulation relies too much on the subjective experience of rehabilitation personnel, resulting in the lack of scientific quantitative basis and insufficient targeting of intervention plans.

[0010] 4. In this invention, by designing a multi-level business logic chain feature cascading constraint mechanism that is vertically progressive and nested, the search space and probability distribution of the large language model when decoding text at downstream nodes are physically limited by the text features generated by upstream nodes. This solves the problem in the special education rehabilitation system that the lack of automatic digital connection between assessment and diagnosis conclusions and subsequent specific training plan formulation leads to the generalization of the generated rehabilitation plan content and a serious disconnect between the assessment report and the actual training task. Attached Figure Description

[0011] Figure 1 The flowchart shows a large-scale anti-hallucination model method for the education and rehabilitation of children with special needs proposed in this invention. Figure 2 This is a general framework diagram of a large-scale anti-hallucination model method for the education and rehabilitation of children with special needs proposed in this invention; Figure 3 This is a flowchart of the AgenticRAG anti-hallucination question-and-answer process for an anti-hallucination large-scale model method for the education and rehabilitation of children with special needs proposed in this invention. Figure 4The flowchart of the ReAct educational rehabilitation assessment agent is a large-scale anti-hallucination model method for the education and rehabilitation of children with special needs proposed in this invention. Figure 5 This is a diagram showing the cascading generation of a child's ability profile and rehabilitation plan for an anti-hallucination large-scale model method for the education and rehabilitation of children with special needs, as proposed in this invention. Figure 6 This invention presents a flowchart for generating a structured diagnostic assessment report for a large-scale anti-hallucination model method for the education and rehabilitation of children with special needs. Figure 7 This is a flowchart illustrating the intervention prioritization and rehabilitation plan generation process of a large-scale anti-hallucination model method for the education and rehabilitation of children with special needs proposed in this invention. Figure 8 This is a system deployment architecture diagram of a large-scale model for the education and rehabilitation of children with special needs, proposed in this invention, which aims to combat hallucinations. Figure 9 This is an architecture diagram of a large-scale anti-hallucination model system for the education and rehabilitation of children with special needs, as proposed in this invention. Detailed Implementation

[0012] The technical solutions in the embodiments of the present invention will be clearly and completely described below with reference to the accompanying drawings. Obviously, the described embodiments are only some embodiments of the present invention, and not all embodiments. Based on the embodiments of the present invention, all other embodiments obtained by those skilled in the art without creative effort are within the scope of protection of the present invention.

[0013] Example 1: In a first embodiment of the present invention, the present invention provides a large-scale anti-hallucination model method for the educational rehabilitation of children with special needs, such as... Figures 1-8 As shown, it includes the following steps: S1. Receive children's rehabilitation data or user instructions, integrate current input with historical context, and dynamically construct a global state feature vector; Furthermore, in S1, constructing the global state feature vector includes: Extract multimodal features from user instructions and children's rehabilitation data, and obtain multi-turn dialogue history context interaction features and agent short-term memory features; Multimodal features, interaction features, and agent short-term memory features are combined and fused to dynamically construct the global state feature vector for the current inference step.

[0014] Specifically, during system operation, the data input layer and data parsing layer maintain a dual-channel input monitoring stream to capture real-time input data for the current inference step. When a diagnostic assessment report generation or knowledge question-and-answer task is triggered, the first monitoring channel receives natural language commands input by the user to express specific rehabilitation questions, assessment analysis requests, or specific intervention fine-tuning actions. The second monitoring channel receives rehabilitation data for children uploaded by the user or retrieved from the system database, including the child's historical medical records, behavioral observation records, tabular data, and score records from standardized assessment scales such as the Autism Developmental Behavior Assessment Scale.

[0015] To ensure a stable and consistent understanding of multi-source heterogeneous data by the subsequent large language model and agent scheduling layer, the system first sends the data to the data parsing layer for preprocessing. The data parsing layer performs multi-format recognition and content extraction on the uploaded files: for unstructured medical records, text feedback, or observation records, the system uses text cleaning and key information extraction algorithms to remove garbled characters and unnecessary noise, extracting standardized text information; for tabular scales or spreadsheet data, the system uses a table-to-text algorithm to align row and column levels, standardizing and flattening the data. This preprocessing ultimately converts the multi-source heterogeneous data into a standardized, directly readable unstructured dataset.

[0016] After the basic data is normalized and parsed, the system sends the unified text dataset and real-time user commands to the feature extraction unit of the model service layer. The feature extraction unit calls the deployed text embedding model to map the received user commands and standardized unstructured dataset into a high-dimensional dense vector space for semantic representation. In this way, the system extracts task intent feature vectors representing the user's core intent and standardized text feature vectors that comprehensively map the child's current baseline abilities and historical diagnostic background at the current inference step. Together, these two constitute the static input feature set for the current inference step.

[0017] To achieve complex contextual understanding in long-cycle evaluation tasks, the agent scheduling layer initiates a feature retrieval mechanism when constructing the current state. On one hand, the system retrieves historical context interaction feature vectors containing multi-turn question-and-answer records from the historical session buffer in real time to pinpoint the global business flow. On the other hand, the agent's internal short-term memory module outputs short-term memory feature vectors accumulated after the previous inference step, target action invocation, and tool execution. These short-term memory feature vectors faithfully record the agent's previous thinking state, trial-and-error records, and intermediate parsing progress in various domains, effectively preventing the agent from getting stuck in a deadlock of repeatedly calling the same tool or parameters in subsequent planning.

[0018] To integrate static, real-time task intent, in-depth multi-source textual data, dynamic historical conversational cues, and the agent's own short-term decision trajectory, the system utilizes a feature concatenation operator to horizontally concatenate the task intent feature vector, normalized text feature vector, historical context interaction feature vector, and agent short-term memory feature vector of the current inference step along the vector dimension. Subsequently, the system performs dimensionality reduction and multi-channel feature alignment on the jointly concatenated features using a layer normalization algorithm and feature weight matrix, dynamically constructing the... The agent's state representation at each step is the global state feature vector. The mathematical calculation and construction logic of this global state feature vector strictly follows the following formula: ; In the above formula, This represents the current inference step obtained through dynamic construction. The global state feature vector of the agent will be input into the agent's inference layer as the core context support for subsequent thought chain thinking and action planning. This represents the task intent feature vector obtained after vectorizing the current real-time user command; This represents the standardized text feature vector obtained after structured cleaning and entity extraction of uploaded multi-source heterogeneous children's rehabilitation data; This represents a historical context interaction feature vector retrieved from the historical session buffer, reflecting the evolution of multiple rounds of dialogue; Indicates the current inference step At that time, the agent's short-term memory feature vector is dynamically derived from the previous decision and tool execution trajectory. This represents a vector concatenation operator in the feature space, used to... as well as These four independent feature vectors are horizontally concatenated and stitched together in terms of dimension. This represents a preset feature fusion weight matrix, used to perform linear transformation, spatial transformation, and dimension alignment on the concatenated high-dimensional joint spliced ​​features; This represents the preset fusion bias vector, which is used to provide translational degrees of freedom for feature space mapping transformation to avoid feature overfitting. The representation layer normalization operator is used to perform hierarchical smoothing and normalization on the global state feature vector after fusion transformation, balancing the numerical scale of different feature channels under long text processing in large models, thereby greatly improving the convergence speed and feature representation stability of agents under long-term complex evaluation and inference.

[0019] Through the complete calculation of the above formula, the global state feature vector S_t dynamically constructed by the system not only fully preserves the child's multi-source baseline ability data, but also seamlessly integrates the user's current real-time control intention with the agent's own historical execution trajectory, eliminating the problem of fragmented evaluation elements in traditional schemes. This global state feature vector is then fed into the next step of agent scheduling and inference decision-making process, providing complete, highly robust, and strongly causal underlying state feature support for subsequent target action generation, rigid legality verification, and anti-hallucination adaptive degradation control.

[0020] S2. Reason the target action based on the global state feature vector and perform legality verification: If legal, call the matching knowledge retrieval or data analysis tool to obtain professional evidence and observation results, and return to step S1 until the score calculation of each field is completed in a loop; if illegal, activate the automatic degradation control channel, switch to the fallback generation mode of the preset template to output the score of each field. Furthermore, in S2, the validity check includes: The global state feature vector is input into the large language model for intent reasoning, generating a target action that includes tool identification and input parameters. Extract the tool identifier and input parameters from the target action, and match and verify them against the preset tool whitelist, preset parameter format, and boundary constraints to determine whether the target action is legal or illegal.

[0021] Furthermore, in S2, obtaining expert evidence and observations includes: If the tool type is knowledge retrieval, the semantic similarity between the semantic vector corresponding to the user command and the knowledge fragment vector in the preset special education rehabilitation knowledge base is calculated, and the knowledge fragments with matching scores are sorted and selected to form a set of professional evidence as professional evidence. If the tool type is data parsing, the corresponding parsing tool will be called to perform format recognition and content parsing on the uploaded multi-source heterogeneous data, and extract standardized text content as the observation result.

[0022] Furthermore, in S2, the output scores for each domain include: When the target action is determined to be illegal, the current tool call is blocked, and a structured report generation template composed of feature slots from various domains is retrieved; The contextual feature data before the current inference step is aligned and concatenated with the corresponding feature slots in the generated template. The large language model is used to complete the text for the incomplete feature slots, and quantitative simulation scores for each domain are output.

[0023] Specifically, after receiving the globally constructed feature vector from step S1 at the agent inference layer or control core, the system uses it as a deep context-driven input. The large language model uses this feature vector as a basis for multi-step intent inference, predicting the most appropriate action to be performed, and generating structured text content as the target action. This target action is typically formatted as a specific agent action flow, which explicitly encapsulates the tool identifier to be invoked and its associated input parameters.

[0024] Because large language models are prone to planning illusions in complex special education assessment tasks, generating uncontrollable non-preset tool names, malformed syntax, or out-of-bounds parameters, the system does not directly allow the generated target action to be passed to the underlying tool execution layer. Instead, it immediately triggers an action validity verification mechanism for rigid interception. The system first uses parsing operators to extract the tool identifier and input parameters from the target action, introduces an action validity verification algorithm, and hard-determines the safety boundary of the target action by calculating the action validity judgment result.

[0025] In this embodiment, variables are defined. For the first Define the legality verification status value of the target action. For large language models in the first The step-by-step reasoning generates a target action containing the planned intent. Simultaneously, symbols are defined. The system pre-stores a set of allowed tools, known as the system tool whitelist. The system uses the following hard rules and logical formulas to perform a validity check on the target action of the current step: ; In the above formula, Indicates the action from the current target. The tool name corresponding to the action separated from the text. This represents the motion parameters that are synchronously extracted from it. This represents a preset parameter validity constraint judgment operator, used to verify whether the action parameters fully meet the preset parameter format and boundary constraints. When the extracted tool name... Exists in the preset toolset In, and motion parameters When the parameter format and boundary constraints are fully consistent with the system's preset parameters, the result of the action legality judgment is... If the value is assigned to the first state value 1, the system determines that the current target action is legal. If any of the above conditions are not met, the action legality judgment result is lost. When assigned the second state value of 0, the system determines that the current target action is illegal.

[0026] When the legality judgment result of the action When the system determines that the target action is valid, it executes the tool call normally and routes the workflow accordingly. Based on the tool type declared in the target action, the system adaptively routes the workflow to the corresponding execution chain. If the corresponding tool type is knowledge retrieval, the intelligent agent's anti-hallucination retrieval enhancement link is triggered. The system first uses a feature transformation unit to calculate the semantic similarity between the semantic vector converted from the user's command and the vectors of various knowledge fragments in the preset special education rehabilitation knowledge base. In this retrieval calculation, the system employs a semantic similarity calculation algorithm, the mathematical formula of which is as follows: ; In the above formula, Indicates user input problem The semantic vectors of the special education rehabilitation knowledge base A knowledge fragment Semantic similarity scores between high-dimensional vectors; Represents the problem vector. This represents a knowledge fragment vector. After traversing and calculating the semantic relevance of all available knowledge fragments in the knowledge base, the system sorts the knowledge fragments in descending order based on their semantic similarity scores and performs threshold filtering, selecting the fragments with the highest scores. A collection of knowledge fragments forms a set of professional evidence addressing the current issue. : ; In the above formula, This represents the complete set of all segmented and vectorized knowledge fragments in the predefined special education and rehabilitation knowledge base. This indicates a descending truncation operator. The final extracted set of professional evidence. It will be immediately incorporated into the agent's context as a hard, professional evidence constraint for subsequent large model-generated answers and to avoid factual illusions.

[0027] If the corresponding tool type is data parsing, the system triggers the underlying file parsing tool to perform multi-format file recognition and content parsing on the multi-source heterogeneous data uploaded by the user, including scales, medical records, or tables. The file parsing tool automatically identifies the file encoding and format, breaking down the barriers between multiple sources and heterogeneous data, stripping away and extracting standardized text content with high information density, and outputting it as the current observation result. After successfully obtaining professional evidence or observation results, the system records the newly generated data feedback as new environmental observations, performs memory updates, and adds it to short-term memory. Subsequently, the system determines whether to continue reasoning. If the loop condition is met, it returns to step S1, driving the agent to dynamically construct the next global state feature vector, until the score calculation for each core domain such as perception, communication, and language is completed through repeated iterations.

[0028] When the legality judgment result of the action When the system determines that the current target action is illegal, it immediately activates the automatic degradation control channel to prevent the large model from causing the underlying tool components to crash due to the generation of incorrect instructions, thus leading to the abnormal interruption of the entire rehabilitation assessment task. The system first issues a rigid interrupt command, forcibly blocking the execution of the current erroneous tool call, and directly switches to degradation generation mode. The system retrieves the internally preset stored structured report generation template, composed of feature slots from various domains, and extracts existing contextual feature data (such as user instructions, parsed data, and historical sessions) successfully accumulated in previous inference steps from historical sessions, parsed data, and previous normal loops, aligning and splicing it with the feature slots. For the remaining blank assessment slots where data could not be retrieved through the external toolchain due to the action being blocked due to illegality, the system uses the filled report template as a rigid decoding boundary and calls the large language model to perform basic text completion and targeted completion generation for the incomplete feature slots. Under the dual constraints of the known context and template slots, the large language model performs basic text completion, forcibly outputting the quantitative simulation score of the incomplete core domain in the blank slots. Through this automatic degradation mechanism, the system provides a highly secure soft landing control when the agent's reasoning goes out of control, ensuring that the evaluation pipeline can still output scores for each domain stably and continuously when errors occur in intermediate steps, thus avoiding program interruption.

[0029] S3. The scores of each area are summed to obtain the overall assessment score, and the development level is determined by mapping and matching. The result is a diagnostic assessment report for children that includes professional evidence. Furthermore, in S3, the overall evaluation score includes: The scores of children with special needs in multiple core developmental areas are linearly summed to obtain the overall assessment score. The overall assessment score is mapped and matched with the preset grading intervals to determine the grading results of the ability development level of children with special needs.

[0030] Furthermore, in S3, the child diagnostic assessment report includes: The system invokes a pre-defined diagnostic report structure framework to generate a multi-dimensional diagnostic assessment text based on the grading results. The user instructions, professional evidence, and anti-hallucination constraint prompts are combined as the decoding context of the large language model, which restricts the large language model from making suggested text inferences based on professional evidence, and outputs a child diagnostic assessment report accordingly.

[0031] Specifically, after the system completes the score calculation for each core area or triggers the automatic degradation control channel to output quantitative simulated scores, the data stream will seamlessly converge and flow to the report generation interface and the corresponding processing unit. In order to quantitatively and scientifically assess the overall developmental status of children with special needs, the system first extracts the corresponding scores of children with special needs in six core areas: sensory development, communication and language development, social interaction development, cognitive development, motor development, and self-care development.

[0032] After obtaining the corresponding scores for each of the aforementioned core developmental areas, the system inputs these scores as the underlying mathematical foundation into the assessment calculation unit. The assessment calculation unit then calls a preset total assessment score calculation algorithm to linearly accumulate the quantitative scores, dynamically constructing a total child development assessment score to represent a macro-based baseline for the child's overall ability development. This total child development assessment score... The mathematical weighted summation formula is as follows: ; In the above formula, This represents the calculated total developmental assessment score for children with special needs. It is the core data element at the top level for subsequent cross-disciplinary grading and multi-dimensional diagnostic text deduction. It indicates that special needs children are in the first The actual capability scores obtained through intelligent agent toolchain analysis or degradation fallback completion within each core development area, among which As domain index variables, each variable uniquely corresponds to one of the six core domains: sensory development, communication and language development, social interaction development, cognitive development, motor development, and self-care development. Through linear summation operations of the operators, the system eliminates the limitations of assessing each ability domain in a separate, isolated manner.

[0033] After dynamically constructing the total score for the child's developmental assessment, the system sends this score to the mapping and matching unit to determine the current developmental and ability level of the child with special needs. The mapping and matching unit retrieves the pre-stored intervals of the developmental delay grading scale within the system, performs interval mapping and rule matching between the child's total developmental assessment score and the boundary values ​​of the grading scale intervals, and determines the current ability development level of the child with special needs, thus realizing the transformation from quantitative values ​​to qualitative professional diagnostic conclusions.

[0034] After obtaining the ability development level assessment results, the system formally triggers the structured diagnostic report generation process. The system first invokes a pre-defined structured diagnostic report framework. This framework strictly regulates the report's length, logic, and section boundaries, requiring it to fully cover four multi-dimensional diagnostic assessment texts: assessment results and diagnostic judgments, analysis of weaknesses in six core areas, personalized education and rehabilitation intervention recommendations, and family support and daily care guidance. The system uses the calculated core area scores, total assessment score, and level assessment results as baseline feature data, automatically filling them into the framework's basic sections such as assessment results and weakness analysis, driving the large language model to generate corresponding standardized diagnostic narrative text.

[0035] To completely eliminate the professional illusion problem generated by the large language model when drafting rehabilitation intervention suggestions and care guidance texts, the system forcibly cuts off the model's free decoding permissions when generating personalized education and rehabilitation intervention suggestion modules, and instead activates an intelligent anti-illusion generation mechanism based on professional knowledge evidence constraints. When generating intervention suggestion text, the system combines and encapsulates the received immediate user instructions, the set of professional knowledge evidence retrieved and persisted using knowledge retrieval tools in the preceding step S2, and the system's built-in anti-illusion constraint prompts, serving as the unique input context for the large language model in the current decoding step. The decoding process for this evidence constraint generation strictly follows the following large language model generation function: ; In the above formula, This represents the final decoded output of the large language model, which includes personalized education and rehabilitation intervention suggestions for the child with special needs, covering dimensions such as institutional training and family cooperation. This represents the large language model generation function; This indicates the user's initial natural language question or intervention analysis instruction, used to anchor the task topic; This refers to the set of professional knowledge evidence retrieved and precisely recalled from the special education and rehabilitation professional knowledge base based on semantic similarity in step S2. This indicates a pre-set, rigid anti-hallucination constraint prompt. Logically, it establishes strict factual boundaries for the attention mechanism of large language models, restricting and mandating that the model must and can only make inferences and continuations of suggested texts based on factual foundations contained in the professional knowledge evidence set. For blind spots that cannot be covered by evidence, it must provide conservative answers or prompt users to supplement further information, thus blocking the proliferation of professional illusions from the bottom up.

[0036] The intervention suggestions generated by the large language model based on formula calculations, along with the simultaneously generated family daily care guidance text, are uniformly assembled into a pre-set report structure. The system ultimately outputs a diagnostic assessment report for children that incorporates professional evidence. This report is not only structurally standardized and data-rich, but each intervention and care suggestion also possesses strong professional knowledge traceability. This provides a scientific, rigorous, and seamless digital starting point for the subsequent cascading generation of personalized rehabilitation plans.

[0037] S4. Construct a capability profile based on the child's diagnostic assessment report, extract multidimensional indicators and interference parameters for the weak areas to be intervened, and perform weighted quantitative calculations using preset weights to obtain the intervention priority score for the weak areas. Furthermore, in S4, the weighted quantization calculation includes: Identify the weak areas requiring intervention from the capability profile, and calculate the defect residual ratio based on the actual score and the highest valid full score in that area; The scores for behavioral risk level, daily life adaptation importance coefficient, and caregiver family demand intensity corresponding to the weak area were extracted as interference parameters. The defect residual ratio and interference parameters are weighted and summed using preset weight parameters to obtain the intervention priority score for the weak area.

[0038] Specifically, after the system outputs a complete diagnostic assessment report for the child, including professional evidence constraints or preset template constraints, at the application output layer, the data flow immediately transfers to the downstream personalized rehabilitation plan cascading generation mechanism. To eliminate the problem of the disconnect between the assessment report and the subsequent actual training plan in traditional programs, the system first activates the profile construction unit. From the generated child diagnostic assessment report, it accurately extracts key feature dimensions such as the child's overall developmental level, scores in the six core areas, clearly identified weak areas, problematic behaviors, strengths and interests, and support conditions from the caregiver's family through text parsing and entity extraction algorithms. These features are then aggregated at the data level to construct a digital profile of the child's abilities that objectively reflects the child's current overall developmental status.

[0039] Building upon the completed digital profiling of children's abilities, to ensure that the subsequently generated individualized rehabilitation plans are not merely generalities but precisely address the most pressing developmental needs of children with special needs, the system employs a priority calculation unit to deeply mine the data within the profiling to determine the order of interventions for each vulnerable area. The system first scans and identifies the vulnerable areas requiring rehabilitation intervention from the constructed profiling. Then, for each vulnerable area to be intervened in, the priority calculation unit simultaneously extracts multidimensional indicators and multiple interference parameters.

[0040] In terms of multidimensional indicator extraction, the system extracts the actual ability score obtained by the weak area in the scale assessment, as well as the highest valid full score corresponding to the weak area within the system. In order to quantify the severity of skill deficiency in the core area, the system calls the preset residual calculation logic to calculate the ratio between the actual ability score and the highest valid full score, and obtains the defect residual ratio of the weak area, which is used to characterize the proportion of the defect space of the weak area under the full score baseline.

[0041] Regarding interference parameter extraction, the system further extracts three key assessment parameters from the child's ability profile and diagnostic assessment report as interference parameters: the first is the behavioral risk level score corresponding to the weak area, which is used to characterize whether the lack of skills in this area will lead to highly dangerous problem behaviors such as self-harm and aggression; the second is the importance coefficient of the weak area to the child's daily life adaptation, which is used to characterize the practical value of the skills in this area for the child's independent life and community integration; and the third is the demand intensity score raised by the caregiver family for the weak area, which is used to characterize the caregiver's urgent desire to intervene in this skill in family care practice.

[0042] To transform the aforementioned multimodal, cross-dimensional indicators and interference parameters into qualitative scores that can be directly ranked, the system introduces an intervention priority quantification algorithm. This algorithm retrieves pre-set weighted parameters composed of multiple different influencing factors and uses these parameters to dynamically weight and quantify the calculated defect residual ratio, extracted behavioral risk level score, daily life adaptation importance coefficient, and caregiver family need intensity score. A final weighted sum is then performed to calculate the intervention priority score for the weak area. The operations research mathematical formula for this intervention priority score is as follows: ; In the above formula, Represents the calculated first... Weak areas The specific intervention priority score directly determines the order of rehabilitation training in the subsequent cascading reasoning plan. This indicates that the special needs child is in a currently weak area. The actual ability score calculated by the internal analysis; This indicates the vulnerable area that the system allows or pre-sets. The highest valid full score. A mathematical term constructed by subtracting 1 from the ratio of the two. This represents the percentage of deficiencies in that weak area. The higher the percentage, the more severe the skill gap in that area for children with special needs.

[0043] At the same time, in the formula This indicates that the images extracted show children with special needs in this vulnerable area. The corresponding risk level score for the performance behavior; This indicates the system's pre-defined weak point. Importance coefficient of children's adaptation to daily life; This indicates that caregiver families are addressing this vulnerable area. The score for the intensity of subjective needs. as well as These correspond to four core weights in the preset weight parameter set, specifically and uniquely corresponding to the domain defect residual ratio weight, special behavior risk weight, daily life adaptation importance weight, and caregiver family demand intensity weight. The values ​​of each weight parameter are configured during system initialization and satisfy the normalization constraint that the sum is 1.

[0044] Through the complete matrix multiplication and addition operations of the above formula, the system completely transforms the originally vague and subjective special education intervention decisions into intervention priority scoring with rigorous mathematical support. Because the scoring algorithm not only meticulously considers the child's actual ability deficiencies but also incorporates real-world production conditions such as special behavioral risks, daily life adaptability, and family care pressures through interference parameters, it achieves a highly scientific extraction of special education assessment conclusions. This intervention priority score is then used as a core input entity label, seamlessly flowing to the lower-level intervention plan cascading generation module. This provides a precise control chain and ranking basis for subsequent cascading prompt word large-scale model reasoning involving multi-level vertical logical chains.

[0045] S5. Establish the implementation order based on the intervention priority scores of each weak area, and use this order to carry out progressive large model reasoning along the vertical logical chain. The generation of each downstream node is embedded with the feature constraints generated by the upstream node at the previous level, and finally outputs a personalized education and rehabilitation intervention plan.

[0046] Furthermore, in S5, the output of personalized educational rehabilitation intervention programs includes: The order of intervention implementation is determined based on the intervention priority score, and the child's diagnostic assessment report is used as the initial feature input into the large language model based on this order. The large language model is driven to derive the next level of stage target text along the vertical multi-level business logic chain, and the text generated at the previous level is cascaded and embedded into the context of the prompt words at the next level as feature constraints, until the final personalized education and rehabilitation intervention plan is output.

[0047] Specifically, after receiving the intervention priority scores for each vulnerable area calculated in step S4, the system uses a dynamic sorting operator to arrange them in descending order of scores, thus establishing the priority order for intervention implementation in each vulnerable area. This sorting mechanism based on operational quantification scores ensures that the most risky vulnerable areas, those most impactful on children's daily life adaptation, and those with the strongest family needs are given priority in the intervention implementation queue.

[0048] After establishing the sequence of interventions, the system formally enters a vertically progressive multi-level business logic chain, driving the large language model to perform anti-generalization, strongly coherent cascading prompt word reasoning. The system inputs the child's diagnostic assessment report generated in step S3 as the initial feature into the large language model. Using this initial feature as the control boundary, the large language model first performs reasoning at the first-level business logic node for the weak areas with the highest intervention priority, generating long-term rehabilitation target text that matches the developmental level of the special child.

[0049] After successfully generating the long-term rehabilitation target text, the core control mechanism of the cascade generation mechanism begins to function. To avoid divergence or generalization illusions in subsequent generation stages, the system embeds the long-term rehabilitation target text generated by the first-level node as a top-level feature constraint into the input prompt context of the second-level business logic node before moving to the next level of business logic node. The system then drives the large language model to perform derivational reasoning in the second-level business logic node, decoding and generating the corresponding short-term target text under the strict constraint of the long-term rehabilitation target text feature. Since the upstream target has been forcibly written into the current attention mechanism decoding context, the vocabulary probability distribution of the large language model when generating the short-term target text is rigidly constrained within the causal framework of the long-term target.

[0050] Following this logic, the system cascades downwards along a multi-level, vertically progressive business logic chain. The system uses the short-term target text generated in the previous step as new feature constraints, incorporating it into the context of the prompt words to progressively generate text containing specific daily or weekly training tasks. Subsequently, the current training task text again becomes the input feature, driving the model to continue progressively generating family generalization strategy suggestion text under its constraints. Finally, at the last level of business logic, based on the aforementioned intervention context, the system drives the model to generate text regarding the dynamic adjustment of the review cycle and review suggestions.

[0051] This multi-level, vertically linked business logic chain mechanism—comprising assessment conclusions, long-term goals, short-term goals, training tasks, family generalization, and reassessment recommendations—ensures that the generation space of each downstream node is physically constrained by the text features generated by its upstream node. After completing the cascading model inference across all levels of the vertical logic chain, the system standardizes and assembles the texts of long-term rehabilitation goals, short-term stage goals, daily or weekly training tasks, family generalization strategies, and dynamic adjustment recommendations for reassessment cycles. The final output is a highly structured, personalized educational rehabilitation intervention plan with strong causal logic. This plan achieves continuous, automatic transformation and perfect digital integration from the initial assessment conclusion to the lowest-level training task.

[0052] Example 2: In a specific scenario involving the assessment of a child with special needs in core areas such as sensory development, communication, and language development, rehabilitation institutions or caregivers need to upload the child's autism developmental behavior assessment scale score, historical medical records, and daily behavioral observation records through a front-end interactive interface. However, due to the complexity of the uploaded assessment data and the presence of a large amount of high-density tables and text data, the large language model, when driving the educational rehabilitation assessment agent to perform long-pipeline analysis and generate tool call instructions, is prone to generating illegal actions such as non-preset tool names, distorted grammar, or out-of-bounds parameters due to illusions at the planning level. Without an action verification mechanism, these illegal instructions will directly cause the logic of the underlying executor and related functional modules to collapse, resulting in abnormal interruption of the entire long-cycle structured report generation task, system hang, and inability to stably output the final assessment conclusion. To solve the above problems, this invention provides an anti-illusion large model system for the educational rehabilitation of children with special needs, the structure of which is as follows: Figure 9 As shown. The specific implementation process of this system is as follows: The feature vector construction module is used to receive children's rehabilitation data or user instructions, integrate the current input with the historical context, and dynamically construct a global state feature vector. The action verification and degradation control module is used to infer the target action based on the global state feature vector and perform legality verification: if legal, it calls the matching knowledge retrieval or data analysis tool to obtain professional evidence and observation results, and returns to the feature vector construction module until the score calculation of each field is completed in a loop; if illegal, it activates the automatic degradation control channel, switches to the fallback generation mode of the preset template, and outputs the score of each field. The assessment report generation module is used to accumulate the scores from various areas output by the action verification and regression control module to obtain the overall assessment score, and to determine the development level through mapping and matching, and output a child diagnostic assessment report containing professional evidence constraints or preset template constraints. The competency profile and priority calculation module is used to construct a competency profile based on the child's diagnostic assessment report, extract multi-dimensional indicators and interference parameters of the weak areas to be intervened, and perform weighted quantitative calculation using preset weights to obtain the intervention priority score for the weak area. The intervention plan cascade generation module is used to establish the implementation order based on the intervention priority scores of each weak area. Following this order, a large model reasoning is carried out step by step along the vertical logical chain. The generation of each downstream node is embedded with the feature constraints generated by the upstream node of the previous level, and finally outputs a personalized education and rehabilitation intervention plan.

[0053] Specifically, the feature vector construction module receives input natural language analysis instructions and uploaded child assessment data. It uses text cleaning and table-to-text algorithms to perform pre-structural processing on multi-source heterogeneous data, and combines multi-round historical conversation information with the agent's short-term memory accumulated from the previous inference step to dynamically construct the global state feature vector of the current inference step in strict accordance with the state representation formula.

[0054] The data stream flows to the action verification and degradation control module. This module uses the global state feature vector as the driving context for intent reasoning of the large language model, generating a target action containing the tool name and action parameters. Before executing the tool call, the action verification and degradation control module immediately triggers a rigid action legality verification mechanism for the target action, extracting the tool name and action parameters, and strictly performing boundary checks using the action legality judgment formula. If the tool name matches the preset tool set and the action parameter format is valid, the system allows and executes the corresponding external tool call normally.

[0055] If, during multi-round loop reasoning, the large language model experiences planning illusions due to interference from complex data, generating illegal action instructions that miss the whitelist or exceed parameter limits, the output of the action legality judgment formula will be a second state value of zero. At this point, the action verification and degradation control module instantly issues a rigid interrupt command, forcibly blocking the execution of the currently erroneous external tool call and preventing the transmission of the corrupted action. Subsequently, the system immediately activates the automatic degradation control channel, directly switching to degradation generation mode. This module automatically retrieves the system's preset structured report generation template, composed of six core domain feature slots, and aligns and splices the existing contextual feature data successfully parsed and accumulated in previous sessions and normal loops with the template slots. For the remaining blank evaluation slots where data could not be retrieved through the external toolchain due to the illegal action being blocked, this module uses the spliced ​​report template as a rigid decoding boundary, utilizing the large language model's basic text completion capabilities for targeted continuation, and forcibly outputs the quantitative simulation score of the incomplete core domain in the blank slots.

[0056] Through the soft landing processing of this degradation control module, the system provides a hard safety guarantee when the agent's reasoning goes out of control. This allows the assessment report generation module to take over the scores of each core area output by the action verification and degradation control modules in normal or degradation modes, successfully calculate the total score of the child's development assessment through the linear accumulation formula, and complete the developmental level determination. This ensures that the entire pipeline can still stably and continuously output child diagnostic assessment reports containing professional evidence constraints or preset template constraints when errors occur in intermediate steps, completely eliminating the problems of abnormal system interruption and suspension.

[0057] Finally, it should be noted that the above description is only a preferred embodiment of the present invention and is not intended to limit the present invention. Although the present invention has been described in detail with reference to the foregoing embodiments, those skilled in the art can still modify the technical solutions described in the foregoing embodiments or make equivalent substitutions for some of the technical features. Any modifications, equivalent substitutions, improvements, etc., made within the spirit and principles of the present invention should be included within the protection scope of the present invention.

Claims

1. A large-scale anti-hallucination model method for the education and rehabilitation of children with special needs, characterized in that, include: S1. Receive children's rehabilitation data or user instructions, integrate current input with historical context, and dynamically construct a global state feature vector; S2. Based on the global state feature vector, infer the target action and perform a legality check: If legal, call the matching knowledge retrieval or data analysis tool to obtain professional evidence and observation results, and return to step S1 until the score calculation for each field is completed in a loop; if illegal, activate the automatic degradation control channel and switch to the fallback generation mode of the preset template to output the score for each field. S3. The scores of each area are summed to obtain the overall assessment score, and the development level is determined by mapping and matching. The result is a diagnostic assessment report for children that includes professional evidence. S4. Construct a capability profile based on the child's diagnostic assessment report, extract multidimensional indicators and interference parameters of the weak areas to be intervened, perform weighted quantitative calculation using preset weights, and obtain the intervention priority score for the weak areas. S5. Establish the implementation order based on the intervention priority scores of each weak area, and use this order to carry out progressive large model reasoning along the vertical logical chain. The generation of each downstream node is embedded with the feature constraints generated by the upstream node at the previous level, and finally outputs a personalized education and rehabilitation intervention plan.

2. The method for an anti-hallucination large-scale model for the education and rehabilitation of children with special needs, as described in claim 1, is characterized in that... In S1, constructing the global state feature vector includes: Extract the multimodal features of the user instructions and the child's rehabilitation data, and obtain the multi-turn dialogue history context interaction features and agent short-term memory features; The multimodal features, interaction features, and agent short-term memory features are combined and fused in a multidimensional manner to dynamically construct the global state feature vector for the current inference step.

3. The anti-hallucination large-scale model method for the education and rehabilitation of children with special needs according to claim 1, characterized in that, In S2, the validity check includes: The global state feature vector is input into a large language model for intent reasoning to generate the target action containing tool identifiers and input parameters; Extract the tool identifier and input parameters from the target action, and match and verify them against a preset tool whitelist, preset parameter format, and boundary constraints to determine whether the target action is legal or illegal.

4. The method for an anti-hallucination large-scale model for the education and rehabilitation of children with special needs, as described in claim 1, is characterized in that... In S2, the acquisition of professional evidence and observation results includes: If the tool type is knowledge retrieval, then the semantic similarity between the semantic vector corresponding to the user instruction and the knowledge fragment vector in the preset special education rehabilitation knowledge base is calculated, and the knowledge fragments with matching scores are sorted and selected to form a professional evidence set as the professional evidence. If the tool type is data parsing, then the corresponding parsing tool is invoked to perform format recognition and content parsing on the uploaded multi-source heterogeneous data, and the standardized text content is extracted as the observation result.

5. The method for an anti-hallucination large-scale model for the education and rehabilitation of children with special needs, as described in claim 1, is characterized in that... In S2, the output scores for each domain include: When the target action is determined to be illegal, the current tool call is blocked, and a structured report generation template composed of feature slots from various domains is retrieved; The contextual feature data before the current inference step is aligned and concatenated with the corresponding feature slots in the generated template. The large language model is used to complete the text for the incomplete feature slots, and the quantitative simulation scores for each domain are output.

6. The method for an anti-hallucination large-scale model for the education and rehabilitation of children with special needs, as described in claim 1, is characterized in that... In S3, the overall evaluation score includes: The overall assessment score is obtained by linearly summing the corresponding scores of children with special needs in multiple core developmental areas. The overall assessment score is mapped and matched with a preset grading interval to determine the grading result of the special needs child's ability development level.

7. The method for an anti-hallucination large-scale model for the education and rehabilitation of children with special needs, as described in claim 1, is characterized in that... In S3, the child diagnostic assessment report includes: The system invokes a pre-defined diagnostic report structure framework to generate a multi-dimensional diagnostic assessment text based on the grading results. The user instructions, professional evidence, and anti-hallucination constraint prompts are combined as the decoding context of the large language model, which restricts the large language model from making suggested text inferences based on professional evidence, and outputs a child diagnostic assessment report accordingly.

8. The method for a large-scale anti-hallucination model for the education and rehabilitation of children with special needs, as described in claim 1, is characterized in that... In S4, the weighted quantization calculation includes: Weak areas to be intervened in are identified from the capability profile, and the defect residual ratio is calculated based on the actual score and the highest valid full score in that area. The performance risk level score, daily life adaptation importance coefficient, and caregiver family demand intensity score corresponding to the weak area are extracted as the interference parameters. The defect residual ratio and the interference parameter are weighted and summed using preset weight parameters to obtain the intervention priority score for the weak area.

9. The anti-hallucination large-scale model method for the education and rehabilitation of children with special needs according to claim 1, characterized in that, In S5, the output personalized education and rehabilitation intervention plan includes: The order of intervention implementation is determined based on the intervention priority score, and the child's diagnostic assessment report is used as the initial feature input into the large language model based on this order. The large language model is driven to derive the next level of stage target text along the vertical multi-level business logic chain, and the text generated at the previous level is cascaded and embedded into the context of prompt words at the next level as feature constraints, until the personalized education and rehabilitation intervention plan is finally output.

10. A large-scale anti-hallucination model system for the education and rehabilitation of children with special needs, characterized in that, The system for an anti-hallucination large-scale model method for the educational rehabilitation of children with special needs, as described in any one of claims 1-9, comprises: The feature vector construction module is used to receive children's rehabilitation data or user instructions, integrate the current input with the historical context, and dynamically construct a global state feature vector. The action verification and degradation control module is used to infer the target action based on the global state feature vector and perform legality verification: if legal, it calls the matching knowledge retrieval or data analysis tool to obtain professional evidence and observation results, and returns to the feature vector construction module until the score calculation of each field is completed in a loop; if illegal, it activates the automatic degradation control channel and switches to the fallback generation mode of the preset template to output the score of each field. The assessment report generation module is used to accumulate the scores of each domain output by the action verification and regression control module to obtain the overall assessment score, and determine the development level through mapping and matching, and output a child diagnostic assessment report containing professional evidence constraints or the preset template constraints. The ability profile and priority calculation module is used to construct an ability profile based on the child's diagnostic assessment report, extract multi-dimensional indicators and interference parameters of the weak areas to be intervened, and perform weighted quantitative calculation using preset weights to obtain the intervention priority score for the weak area. The intervention plan cascade generation module is used to establish the implementation order based on the intervention priority scores of each weak area. Following this order, a large model reasoning is performed step by step along the vertical logical chain. The generation of each downstream node embeds the feature constraints generated by the upstream node at the previous level, and finally outputs a personalized education and rehabilitation intervention plan.

Images