617 results about "Inference" patented technology

The invention discloses a multi-modal knowledge graph rule reasoning method and device based on a large model, and the method comprises the steps: carrying out the feature extraction and cross-modal alignment of input text data and image data, and generating a multi-modal feature vector of a unified semantic space; performing knowledge graph storage on the emotion entities and the relationships by adopting an attribute graph model to complete construction of an emotion knowledge graph; generating an interpretable inference rule from the emotion knowledge graph by using a large language model, and eliminating a conflict rule in combination with logic verification; calculating the confidence coefficient of a reasoning path based on an attention mechanism, and carrying out quantitative evaluation on a rule reasoning result; the knowledge graph and the rule base are updated online according to user feedback, and the real-time performance and accuracy of the inference system are optimized through weight adjustment and a forgetting mechanism. According to the method, through innovative technologies such as multi-modal data integration, dynamic knowledge evolution and interpretability reasoning, the limitation of a traditional sentiment analysis method in the aspects of evidence dimension, adaptive capacity, interpretability and the like is broken through.

The invention belongs to the field of artificial intelligence and natural language processing, and provides a method and a system for determining a dynamic retrieval decision scheme based on Monte Carlo tree search in order to solve the problems of redundant retrieval and incapability of multi-step reasoning and dynamic planning of a traditional retrieval enhancement generation method. The dynamic retrieval decision scheme determination method based on Monte Carlo tree search comprises the following steps: receiving an initial query problem and taking the initial query problem as a starting point of a reasoning process to trigger and call a large language model operation; calling a large language model to execute sub-query generation and action selection operation, searching, iteratively and exploring inference actions of sub-query problems based on a Monte Carlo tree, and dynamically screening inference paths by considering self-confident degree scores and honesty degree scores of the large language model to obtain a complete inference track set; and on the basis of the complete reasoning track set, a preference pair is constructed for each sub-query problem, and a reasoning retrieval strategy is optimized, so that the optimal retrieval opportunity of the large language model is autonomously judged. According to the method, invalid retrieval and multi-step reasoning dynamic planning can be reduced.

The invention provides a training data synthesis method and device based on error extrapolation and inference chain analysis, a medium and a program product. The method comprises the steps of obtaining an initial sample set; performing multiple sampling reasoning on the problem of each task sample by using a small language model to generate a plurality of reasoning chains; calculating an overall error score of each reasoning chain based on a preset error evaluation rule, and determining a to-be-corrected reasoning chain; the inference chain to be corrected and the corresponding question are input into the large language model together, and a corrected answer is generated; forming a new task sample by the question and the corrected answer, and finely adjusting partial parameters of the small language model; repeatedly executing the process until the performance index change rate of the model on the task evaluation set is lower than a preset threshold value, and outputting a final task sample; and forming a training sample set by a plurality of final task samples, and performing all-parameter fine tuning on the small language model. According to the method, the training data self-optimization path is constructed by taking the model error as guidance, so that the semantic consistency and the data validity are improved.

The invention relates to the technical field of computers, in particular to a reasoning acceleration method and device, electronic equipment and a storage medium, and is used for improving the reasoning speed when a model executes a task. The method comprises the steps that when a target model is adopted to execute a to-be-reasoned task, a matched reasoning mode is activated according to hardware state information of computing resources occupied by the target model; when the reasoning mode is an instruction acceleration mode, at least one round of reasoning operation is executed, and each round of reasoning operation comprises the steps that storage continuity check is conducted in computing resources for reasoning dependency data related to the input sequence of the round; the reasoning dependency data is generated and stored through the reasoning operation of the previous round; when the storage position of the reasoning dependency data is in a non-continuous state in the computing resources, the storage position is adjusted to be in a continuous state by rearranging the reasoning dependency data; and calling a specified hardware acceleration instruction to load the adjusted reasoning dependency data, and determining the reasoning result of the round based on the loading result.

The invention relates to the technical field of artificial intelligence, can be applied to the fields of financial science and technology and medical health, and discloses a reasoning optimization method, device, equipment and medium for a multi-modal large language model.The method comprises the steps that an input long context sequence is obtained, and key value projection is conducted on the long context sequence to generate an initial key value cache; for each attention layer of the multi-modal large language model, calculating an attention matrix of the attention layer according to the vector dimension of the long context sequence and the initial key value cache; calculating a cross-modal attention entropy according to the attention matrix, and determining a cache size of an attention layer according to the cross-modal attention entropy; optimizing the initial key value cache based on a cumulative attention scoring mechanism and a window strategy to obtain a target key value cache; and reasoning the long context sequence according to the cache size and the target key value cache to generate a long context reasoning result. And the reasoning efficiency and the reasoning accuracy are improved.

The invention relates to the technical field of artificial intelligence, and provides an intelligent dialogue method, system and device based on big language model illusion alleviation and a medium, and the method comprises the steps: outputting multiple rounds of heuristic answer results, stimulation thinking results and self-reflection results based on an input question through a big language model; generating a reasoning chain based on the heuristic answer result, the stimulation thinking result and the self-reflection result of the same round; performing logic semantic relationship detection on each reasoning chain through a small language model to obtain a target reasoning chain with a correct relationship; and extracting and outputting a target answer of the input question from the target reasoning chain. According to the intelligent dialogue method based on big language model illusion relief, the big language model is used for forming an inference chain, the small language model is used for analyzing the output of the big language model, inaccurate information can be accurately recognized and filtered out, staged cooperation between the two models is achieved, the limitation that the big language model conducts illusion detection in a self-reflection mode is overcome, and the intelligent dialogue efficiency is improved. And a result can be accurately output.

The invention provides a deep exploration AI reasoning method and system, and the method comprises the steps: generating a dynamic calculation graph based on a to-be-reasoned task, carrying out the risk assessment of the dynamic calculation graph, generating a risk prediction result, and carrying out the risk prediction of the to-be-reasoned task. The risk assessment at least comprises complexity assessment, memory occupancy assessment and data transmission quantity assessment; based on a risk prediction result, performing optimization processing on the dynamic calculation graph to generate an optimized calculation graph, the optimized calculation graph being used for constraining an inference calculation depth and an inference resource allocation range, the method converting natural language input into a semantic weighted dynamic calculation graph, and identifying a resource risk through node-level complexity assessment; performing self-adaptive pruning in combination with the permission level and reinforcement learning to generate an optimal calculation graph; elastic resource scheduling is implemented based on a multi-target model and real-time monitoring, so that the problem of low collaboration of reasoning depth and resource efficiency is solved.

The invention relates to the technical field of artificial intelligence, in particular to a scene application operation maintenance platform based on an artificial intelligence model, which comprises a reasoning state sensing module, a load balancing scheduling module, an intelligent resource allocation module, a deployment configuration adjustment module and a scene feedback integration module. According to the method, by sensing and reasoning process resource consumption and response time migration trend, abnormal tasks caused by resource fluctuation can be accurately recognized, the state monitoring sensitivity is improved, task allocation records are analyzed to construct conflict structures, the resource conflict recognition and visualization capability is enhanced, the model operation level and the GPU vacancy rate are combined, a task migration path is optimized, and the task migration efficiency is improved. Node resource balance is realized, storage frequency, memory margin and network delay are fused during deployment, migration stability is guaranteed, drift tasks are identified through response logs and call records, an operation and maintenance list is generated, operation and maintenance accuracy and multi-scene matching capability are improved, and reasoning task stability and resource scheduling intelligent level are enhanced.

The invention discloses a big language model-oriented inference operation acceleration method for processing an extremely long text, which belongs to the technical field of large-scale pre-training language model inference optimization, and specifically comprises the following steps of: selecting a big language model, starting a sparse attention mode, inputting different texts, recording execution time of each attention head and an attention mode type in different layers, and calculating the inference operation speed of the large language model; establishing a statistical database; solving a multi-head attention load balanced distribution strategy of the large language model under the actual text; establishing a weight index table by splitting the weight matrix; retrieving a weight sub-matrix corresponding to each attention head, and loading the weight sub-matrix to a corresponding GPU device; by asynchronously preloading the MHA calculation weight and the MLP calculation weight of the adjacent layers and combining KV cache management, load balancing in the reasoning process is realized. According to the method, dynamic segmentation, recombination and scheduling of the attention head are carried out in the reasoning stage, so that the problems of multi-GPU load imbalance and resource idling are effectively avoided, and the system throughput of long sequence processing is remarkably improved.

The invention provides a reasoning acceleration optimization method and system applied to an intelligent dialogue large model, and belongs to the technical field of large models.Firstly, a to-be-reasoned dialogue sequence and reasoning environment configuration information are obtained, and the to-be-reasoned dialogue sequence comprises a user real-time input text and a historical interaction statement chain; the reasoning environment configuration information covers an operation node load state and cache resource occupation information, then joint process deconstruction processing is carried out on the operation node load state and the cache resource occupation information to obtain a reasoning node dependence graph and a resource elasticity demand list, reasoning link optimization processing is carried out based on the result, and a reasoning acceleration execution scheme is generated; the method comprises the steps of reasoning a node parallel scheduling rule and a resource pre-allocation strategy, regulating and controlling a reasoning operation process according to a reasoning acceleration execution scheme, generating a dialogue response sequence after acceleration processing, and finally pushing the dialogue response sequence after acceleration processing to a user interaction terminal to complete intelligent dialogue output. Therefore, the reasoning speed of the intelligent dialogue large model is effectively improved, and the dialogue interaction experience is optimized.

The invention discloses a memory and iterative optimization fused big language model answer reasoning method and system. Firstly, feedback information in a natural language form is generated for answers of sub-questions through an answer verification module, and the situation that following reasoning steps all deviate from a correct reasoning path after a certain step of the model has an error is avoided. Secondly, through a sub-problem verification module, verification pruning is performed on sub-problems generated in thinking chain reasoning, and interference of information of irrelevant sub-problems on the model is avoided. And finally, by utilizing feedback information and iteratively optimizing the output of the model, the self-reflection and correction process of human is simulated, and the situation that the large language model generates the same error in the same reasoning step is avoided. According to the method, the large language model is not finely adjusted, so that resources required for finely adjusting the large language model are saved, and the performance of the large language model in the aspect of reasoning is improved.

The invention provides a large model intelligent reasoning method combining reinforcement learning and retrieval enhancement generation, and belongs to the technical field of artificial intelligence. Comprising the following steps: data input: data preprocessing; constructing a reinforcement learning environment; an RAG mechanism is integrated; training the model; and evaluating and iteratively optimizing. A reinforcement learning framework based on rules is introduced to guide a model to develop advanced reasoning skills such as reflection, verification and summarization. And in combination with a retrieval enhancement generation mechanism, the model can access and utilize a wide background knowledge base before answering questions. The synergistic effect between information retrieval and text generation is optimized. According to the method, the deficiency of knowledge of the model can be made up by introducing the external knowledge base, and the synergistic effect between information retrieval and text generation can be optimized in the reinforcement learning process, so that the capability of the model for processing complex reasoning tasks is remarkably improved, and formation of a more generalization reasoning strategy is promoted.

The embodiment of the invention provides an intelligent agent training method, device and equipment and a computer readable storage medium. The agent is associated with a large language model, the method comprising: obtaining expert trajectory data about an inference chain, the inference chain comprising a plurality of inference steps; generating agent trajectory data about the inference chain by the agent based on the expert trajectory data; the expert trajectory data and the agent trajectory data are progressively compared in an inference step order to generate an inference step-by-inference step reward, and the agent is trained based on the inference step-by-inference step reward. According to the method for training the intelligent agent, the reinforcement learning process of the intelligent agent can be optimized by rewarding one by one reasoning steps. Specifically, the method for training the agent according to the present disclosure can compare expert actions in an expert trajectory with agent actions in an agent trajectory, thereby automatically generating an intermediate reward for fine-grained optimization.

The invention discloses a legal text information extraction and enhancement method and system based on artificial intelligence. The method comprises the steps of data acquisition and processing, semantic segmentation and reconstruction, text logic perception, dynamic knowledge fusion and information extraction and enhancement. The invention relates to the technical field of text information extraction, in particular to a legal text information extraction enhancement method and system based on artificial intelligence, which adopts a joint frame of semantic segmentation reconstruction, text logic perception and dynamic knowledge fusion to realize more accurate recognition of the structural level of a legal text. Understanding of logic association of legal clauses is enhanced, and meanwhile, knowledge among different legal documents is dynamically fused, so that cross reference and intelligent application among different laws and regulations are facilitated; semantic segmentation reconstruction is carried out by adopting a conditional random field bidirectional long-short-term memory model improved by combining hierarchical label reasoning; and carrying out text logic perception by adopting a causal graph convolutional reasoning neural network enhanced by combining a dual-channel logic unit.

The invention provides a question and answer model training method and device, electronic equipment, a storage medium and a program product. The method comprises the steps that multiple candidate reasoning paths are determined based on a first question sample carrying an answer label, and the candidate reasoning paths are used for indicating reasoning steps needing to be executed for solving a question corresponding to the first question sample; classifying the plurality of candidate reasoning paths to obtain a first reasoning path belonging to a first category and a second reasoning path belonging to a second category, the reasoning accuracy of the first reasoning path being greater than the reasoning accuracy of the second reasoning path; determining a first loss value of the question and answer model based on the first reasoning path and the answer label, and determining a second loss value of the question and answer model based on the first reasoning path and the second reasoning path; and training the question and answer model based on the first loss value and the second loss value to obtain a target question and answer model. Through the method, the question solving performance of the question and answer model can be effectively improved.

The invention provides a training method and device of an inference model and a processing method and device of an inference problem, and relates to the field of artificial intelligence, in particular to the technical field of natural language processing, large models and deep learning. Obtaining a training sample set, wherein training samples in the training sample set at least comprise sample reasoning questions, reference reasoning chains corresponding to the sample reasoning questions and reference answers; on the basis of a first training sample in the training sample set, performing supervised fine tuning SFT on a pre-training model to obtain a first inference model after fine tuning; and performing reinforcement learning RL training on the first reasoning model based on the second training sample in the training sample set to obtain a second reasoning model, thereby improving the accuracy and stability of the second reasoning model in the reasoning process.

The invention discloses a knowledge base AI original text accurate reading device containing a DLF dynamic layout file. A data acquisition offline module sets an acquisition website and offline webpages according to the hierarchical relationship of the webpages; the DLF file generation module is used for converting each offline webpage file into an OFD file and assembling a plurality of webpages into a dynamic layout DLF file according to a hierarchical relationship, the content extraction knowledge base storage module is used for extracting content information in each individual webpage and generating a file queue according to the extracted content information, the generated file queue is pushed to the inference model, and the dynamic layout DLF file is obtained. Forming a knowledge base containing the DLF file; the AI reasoning module performs logical reasoning on the input problem and returns a result; the traceability interpretation module is matched with an associated DLF file according to the obtained return data; and the combination comparison module compares and checks multiple page areas of different DLF files at the same time, so that AI original text tracing of various types of files is realized.

Embodiments of the invention provide a resource pre-allocation method and apparatus for a computing device cluster, and an electronic device. The method comprises the steps of obtaining task information of a reasoning task expected to be submitted to a reasoning model for reasoning; dividing the reasoning tasks into a plurality of types according to the input token number and the output token number of the reasoning tasks, and determining an expected concurrency number of each type of reasoning tasks; for each type of reasoning task, determining a first target model in the equipment models of the computing equipment included in the computing equipment cluster according to the number of input tokens and the number of output tokens of the type of reasoning task; and for each type of reasoning task, according to the expected concurrence number, the input token number and the output token number of the type of reasoning task, reserving a first reasoning instance in a computing device of a first target model determined for the type of reasoning task. By applying the embodiment of the invention, the overall reasoning efficiency of the reasoning model can be improved.

The invention provides an interpretable AI inference method and system for self-supervised neural symbol fusion, and the method comprises the steps: S1, a self-supervised learning module: extracting feature information of unlabeled data, and obtaining a feature vector representing symbol knowledge; s2, a neural symbol fusion module: constructing a deep learning framework by using an AI intelligent learning strategy, and training feature vectors in combination with a reward function to obtain structured knowledge; s3, an explainable reasoning module: constructing an explainable prediction model based on neural symbol fusion, and obtaining a solved reasoning result; in the process of explaining the reasoning result, the complex reasoning process is simplified and displayed. Therefore, not only are the advantages of the neural network in processing complex modes and data association kept, but also the interpretability and logical reasoning ability of symbol reasoning are introduced, so that the performance and interpretability of the whole system are improved. Therefore, the model can optimize feature representation and logical reasoning at the same time in the training process.

The invention provides a model training and information replying method and device, a storage medium and a program product, and relates to the technical field of computers. The method comprises the steps of performing continuous pre-training on a base model based on a first training sample to obtain a basic model; performing cold start supervision fine tuning training on the basic model based on the second training sample to obtain a first supervision fine tuning model; performing multiple reasoning based on the third training sample, the target information and the to-be-trained model to obtain a reasoning result; in the Mth reasoning process, the target information comprises information obtained after a target tool determined by previous M-1 reasoning is called; optimizing the to-be-trained model based on the reasoning result to obtain a first reinforcement learning model; and based on the general recognition data and reasoning data output by the first reinforcement learning model, carrying out general recognition alignment training to obtain a target model. According to the method, the target tool can be called to obtain the required target information, so that the information output by the large model is more comprehensive.

The invention relates to the technical field of artificial intelligence, and particularly provides a project file risk analysis method and system, and the method comprises the steps: obtaining a project file, and recognizing a semantic vector of the project file through a large language model; screening out a difference semantic vector inconsistent with the standard semantic vector; performing logical reasoning on the semantic vector of the project file based on the user problem by utilizing a thinking chain reasoning model to obtain an abnormal semantic vector; sequentially inputting the difference semantic vector and the abnormal semantic vector into a dialogue pre-training model to obtain a corresponding risk analysis text and a modification suggestion text; and converting the difference semantic vector and the abnormal semantic vector into statements, and writing the statements and the corresponding risk analysis text and modification suggestion text into a preset report template to obtain a risk analysis report. According to the method, accurate and accurate identification of project file semantics, comprehensive analysis of risk points and accurate generation and supplement of modification suggestions are realized.

The invention discloses a method and device for collaborative reasoning of multiple heterogeneous models, and relates to the technical field of artificial intelligence reasoning prediction.The method comprises the steps that a reasoning request and a cue word submitted by a client side are received; according to the context information of the reasoning request and a preset condition, selecting a plurality of matched heterogeneous models and performing Token prediction at the same time; and selecting an optimal prediction result from the prediction results of all the heterogeneous models, splicing the optimal prediction result and the cue word, if reasoning of the splicing result is not completed, submitting the splicing result as a new cue word to a plurality of matched heterogeneous models, and continuing iterative reasoning until reasoning is completed. According to the method, the overall reasoning performance and efficiency are improved, and the task execution capability of the model in a specific field is enhanced.

The invention relates to the technical field of artificial intelligence, and discloses a large language model training method, a logic problem processing method and computer equipment, and the method comprises the steps: obtaining a first logic problem set, and outputting multiple groups of hyper-parameters of a formatted first large language model; any first logic problem in the first logic problem set is input into a first large language model corresponding to any group of hyper-parameters, thinking chain data comprising a plurality of reasoning steps are obtained, and the large language model generates a current reasoning step based on the first logic problem, the previous reasoning steps and evaluation information of the previous reasoning steps; adjusting model parameters of the first large language model based on the thinking chain data and the evaluation information to obtain a second large language model and an inference step scoring model; and based on the second logic problem set and the reasoning step scoring model, adjusting model parameters of the second large language model to obtain a target large language model with generalization ability. The logical reasoning problem processing effect is improved.

Methods and systems for managing evolving artificial intelligence (AI) models are disclosed. The evolving AI models may be used to generate inferences that may be provided to downstream consumers during a provisioning process. The downstream consumers may rely on the accuracy and consistency of the inferences provided during the provisioning process to provide desired computer-implemented services. The AI models may be updated (e.g., with new training data) automatically and / or frequently over time in order to increase the accuracy of inferences provided by the AI model. However, inferences provided by a newly updated instance of an AI model may be inconsistent with inferences provided by prior instances of the AI model (e.g., due to AI model poisoning). Therefore, to increase the likelihood of providing accurate and consistent (e.g., unpoisoned) inferences to the downstream consumers, an appropriate instance of the AI model may be identified for use in the provisioning process.

The embodiment of the invention relates to the technical field of computers, and discloses a corpus quality evaluation system. The system comprises a feature extraction module used for determining an explicit rule feature and an implicit depth feature of a neural symbol according to an input corpus; the knowledge graph construction module is used for constructing a knowledge graph of the corpus according to the explicit rule features and the implicit depth features; the multi-head self-attention fusion module is used for determining fusion features according to the knowledge graph; the logical reasoning module is used for performing rule reasoning according to the fusion features to obtain a reasoning result; and the multi-task quality scoring module is used for generating a multi-dimensional quality score according to the reasoning result. The technical problem that a traditional method is difficult to meet the current efficiency requirement can be solved at least.

The invention provides a reward model training method and device and an answer evaluation method and device, and relates to the technical field of artificial intelligence. The reward model training method comprises the following steps: inputting an original training sample and a prompt text into an initial reward model to obtain a first reasoning principle and a first score output by the initial reward model; the prompt text comprises a reasoning principle generation requirement text, the reasoning principle generation requirement text comprises reasoning principle generation requirement information, and the reasoning principle is used for indicating reasoning logic which the initial reward model needs to follow during scoring; based on the original training sample, the first reasoning principle and the first score, finely adjusting the initial reward model to obtain a reward model; wherein the initial reward model is a pre-trained large language model. The performance of the reward model can be improved, and finally the performance of the answer generation model enhanced based on the reward model is improved, namely, the generation accuracy of the answer generation model is improved.

The invention relates to the technical field of artificial intelligence, in particular to a task execution method and device based on prompt information, equipment and a medium, and aims to improve the flexibility of a task model prompt project and the accuracy of task processing. The method comprises the following steps: extracting each piece of key information related to a task target from task description information of a task to be reasoned; determining a task scale of the to-be-reasoned task according to the task description information, and selecting a topology type matched with the task scale; based on each piece of key information and the topology type, constructing a reasoning topology structure; reasoning nodes and edges in the topological structure, wherein the nodes and the edges respectively represent the reasoning steps of the task to be reasoned and the dependency relationship of the reasoning steps; generating prompt information of each node based on the reasoning topological structure, wherein the prompt information comprises an execution mode of a reasoning step represented by the corresponding node; and based on each piece of prompt information, guiding the task model to gradually execute each reasoning step according to each node and the dependency relationship thereof to obtain a task execution result.

The invention provides a hierarchical exploration thinking-driven large-model complex graph question and answer processing method and device, relates to the technical field of artificial intelligence, and aims to solve the technical problem that an existing complex graph question and answer method is insufficient in the aspects of reasoning depth, path search strategies and cross-layer information fusion ability. The method comprises the steps of obtaining a to-be-processed complex graph question and answer task; inputting the complex graph question and answer task into a pre-constructed large model, adding an induction mark for a node set by using a first-layer structure, and generating an induction graph; extracting a core entity feature and a trunk logic structure feature in the problem description by using a second-layer structure, and generating a context prompt; according to the induction graph and context prompts, reasoning and exploring are conducted through the third-layer structure, a preliminary question and answer result is generated, and the semantic relation and the logic relation of the preliminary question and answer result are verified; and in response to verification failure, iteratively updating the induction graph and regenerating a question and answer result until verification is passed, and outputting a target question and answer result.

The invention relates to a model training method and device, computer equipment, a computer readable storage medium and a computer program product. The method comprises the following steps: reasoning a sample problem through a reasoning model to obtain a reasoning result; determining a corresponding reasoning length control hyper-parameter according to the difficulty level category of the sample problem; constructing a reasoning length reward function according to the reasoning length control hyper-parameter and the reasoning length of the reasoning result, and constructing a reasoning accuracy reward function according to the reasoning result; and performing model training based on reinforcement learning on the reasoning model according to the reasoning length reward function and the reasoning accuracy reward function. The reasoning model trained by the method can give consideration to reasoning efficiency and accuracy, and a more efficient and accurate reasoning process can be realized.

The invention provides a large model inference chain compression method and system based on confidence guidance, and belongs to the technical field of large models. The method comprises the following steps: constructing a confidence phrase pool for improving the internal confidence coefficient of large model reasoning; obtaining a standard problem set, and generating a corresponding complete reasoning chain for each problem in the standard problem set by using the large model so as to construct an efficient reasoning data set; wherein in the inference chain generation process, confidence injection is carried out on the reflection step in the inference chain by utilizing a confidence phrase pool, and the confidence degree of the current inference chain updated each time is calculated so as to optimize the inference process; and then training the large model by using the efficient reasoning data set, wherein the trained large model is used for automatically generating a reasoning chain which is subjected to effective compression and aims at the input problem. According to the method, redundant steps can be compressed to the greatest extent on the premise of ensuring reasoning integrity and accuracy, and efficient and stable compression of the reasoning chain output by the large model is effectively realized.