113 results about "Cost matrix" patented technology

The invention discloses a multi-target tracking method based on threshold separation and feature enhancement, and the method achieves the association between a target and a track in a current frame through a data association algorithm based on a target detection result through the motion features and appearance features of the target. The method comprises the following steps: dividing a high-score detection frame and a low-score detection frame according to a result of a detector and a detection confidence score threshold value; extracting motion features and appearance features of the high-score detection frame, and predicting positions of all tracks in the track set in the current frame; the first association occurs between a high-score detection frame and a track set, motion features and appearance features are fused and used, the second association occurs between a low-score detection frame and a track which is not matched, and only the motion features are used for calculating a cost matrix; trajectory management includes initializing new trajectories, deleting expired trajectories, and removing repeated trajectories. According to the method, a feature extraction mode and a feature fusion mode are improved, and the high-order accuracy of multi-target tracking is improved.

The invention discloses a reward model training method, a large language model optimization method and correlation, and the reward model training method comprises the steps: obtaining a preference training sample pair and a to-be-trained reward model, the preference training sample pair comprising a preferred response sample and a non-preferred response sample; calculating an award score difference between the preferred response sample and the non-preferred response sample based on a to-be-trained award model; constructing a cost matrix based on the reward score difference and the semantic association degree between the preferred response sample and the non-preferred response sample; calculating a loss margin based on the cost matrix; and based on the loss margins, carrying out calculation to obtain paired preference loss values of the band margins, and updating parameters of the to-be-trained reward model by taking minimization of the loss values based on the band margins as an optimization target to obtain a trained reward model. The learning ability and overall generalization performance of the model for difficult samples are improved, excessive dependence on simple samples is avoided, and then the generation quality of the large language model in complex tasks is improved.

The invention discloses a method for applying linear programming to CDN (Content Delivery Network) scheduling, which relates to the technical field of content delivery networks and comprises the steps of data preparation, strategy layer version smooth configuration, macroscopic layer and microscopic layer linear solution and online execution. Basic data are collected, cleaned and repaired, and a version change rule is set; the macroscopic layer constructs a linear programming model, and the cross-provincial bearing quota is solved with the aim of minimizing the cross-provincial cost; the micro layer takes the quota as a boundary and generates domain name class-node weight vectors in parallel; and adapting a routing request online through weighted rendezvous hashing and request features. According to the method, a dynamic cost matrix and a weight granularity control technology are integrated, the engineering problem of linear programming is solved, second-level response, approximate global optimal scheduling and accurate execution of floating-point-level weight are realized, memory overhead is reduced, smooth updating of a strategy and system stability are guaranteed, and CDN service quality and operation efficiency are improved.

The invention discloses a new energy power station inspection method based on a multi-modal large model small sample open set, and the method constructs a multivariate text learnable prompt, and guides a detection model to form a finer-grained category decision boundary through the text information related to an aggregation task. Due to the lack of real unknown class samples in the training process, the mining of unknown class pseudo samples is regarded as a bipartite graph matching task for the first time, and the model is optimized to form a compact unknown class decision boundary by adding unknown class virtual nodes, constructing a cost matrix and mining the unknown class pseudo samples. In order to solve the problem that known classes and unknown classes of small samples are prone to confusion, the method proposes unknown class optimization loss based on cost perception, considers the classification and positioning quality of the unknown classes, and improves the open set detection performance of the model. According to the method, transformation from a single-mode vision small model to a multi-mode vision large model is realized, and good generalized small sample open set target detection performance can be obtained only by a small amount of training data.

The invention relates to the technical field of acceleration signal recognition, in particular to an acceleration signal recognition method and device, and the method comprises the following steps: carrying out the first-order numerical differential operation of a three-axis acceleration signal based on a collected original three-axis acceleration signal, and obtaining a movement speed sequence of the signal; according to the method, first-order and second-order numerical differential operation is carried out on an original three-axis acceleration signal, and the speed change and the impact change of the signal are extracted at the same time and combined to form a kinematic feature set, so that the dynamic features of the motion state can be fully expressed in the two dimensions of speed and impact. The Euclidean distance between the to-be-identified signal and the motion template is calculated based on the kinematic feature set to form the basic distance matrix, and the constraint cost matrix is constructed in combination with the speed sequence and the impact signal, so that the constraint on the kinematic consistency in the signal comparison process is enhanced, and the matching path is more reasonable.

The invention relates to a multi-sensor fusion updating method and device, equipment and a storage medium. According to the main technical scheme, the method comprises the following steps: acquiring single-source track data, wherein the single-source track data comprises multiple types of single-source track information; sorting the multi-class single-source track information according to timestamps, sequentially converting the single-source track information of the previous frame in two adjacent frames into global track information, and performing fusion calculation on the global track information and the unit track information of the next frame to obtain a multi-class cost matrix; screening and matching the various cost matrixes according to a preset matching model to obtain an optimal matching result of the various cost matrixes; obtaining category attributes of global track information in the optimal matching result of each type of cost matrix, and determining a prediction motion model according to the category attributes; and according to the prediction motion model, performing prediction updating on global track information in the optimal matching result of each type of cost matrix to obtain a corresponding optimal prediction updating result. According to the invention, the effect of improving the perception capability of the system to the surrounding environment can be achieved.

The invention belongs to the technical field of artificial intelligence and information retrieval, and particularly relates to a group-level core set selection method for a large model recommendation system. The method comprises the following steps: defining a core set selection task, and taking test loss after minimizing subset fine tuning as an optimization target; the method comprises the following steps: constructing an agent optimization target: based on an optimal transmission theory and gradient norm analysis, considering the distribution difference between a candidate subset and a verification set, and constructing an upper bound of agent test loss; initialization-re-refining calculation: by introducing a special cost matrix, converting an original problem into a p-median problem, and optimizing by using a greedy and sample exchange strategy; and label enhancement: according to a joint distribution category decomposition principle, further improving the category coverage capability and the performance stability of the core set in a label recommendation scene. According to the method, the data scale and computing resource consumption required by fine tuning can be remarkably compressed while the recommendation performance is kept, and key support is provided for efficient deployment of a large language model in a recommendation system.

The invention provides a fine-grained three-dimensional model classification method and system based on dynamic prototype learning, and belongs to the technical field of three-dimensional geometric analysis. Comprising the steps of performing multi-view orthogonal projection rendering on a three-dimensional model, performing feature extraction and projection operation on an image sequence of each view after projection rendering based on an encoder, and fusing multi-view features; a shared prototype pool is constructed, a prototype cost matrix is generated, a structured measurement space is constructed by calculating the cosine similarity of the multi-view features and the prototype cost matrix, and dynamic soft allocation is executed; sequentially executing a dynamic updating strategy and joint loss optimization on the shared prototype pool; and measuring the distance between the test sample features and the prototypes in the shared prototype pool to realize the probabilistic decision-making of the fine-grained category. Therefore, a dynamic prototype learning framework with geometric perception capability is constructed, and the precision bottleneck of a fine-grained classification task can be broken through by analyzing interpretable features and prototype mapping relations.

Embodiments of the present disclosure relate to a network of federated computing nodes in which a cost matrix is distributed. Each node has stored thereon distribution vectors, and create an optimal transportation transfer function using the Sinkhorn-Knopp theorem. In other words, the nodes are each configured to generate updated distribution vectors using the cost matrix shared amongst the nodes and their own distribution vectors. The updated distribution vectors of every node is thereafter used to create the optimal transportation transfer function by concatenating the updated distribution vectors received from each node. This approach can be centralized in a star-network architecture, or in a peer-to-peer-network architecture. Furthermore, the approach may be synchronous or asynchronous depending on the implementation.

The invention discloses an underwater robot formation control method based on a virtual structure and a cost equalization method, and the method comprises the following steps: designing a polymorphic formation library, and importing the polymorphic formation library into a main AUV; establishing a random coordinate system by taking the master AUV as a center, generating a three-dimensional virtual matrix, and introducing a virtual AUV between the slave AUV and the corresponding virtual mass point; selecting a corresponding formation from the polymorphic formation library of the main AUV according to task requirements; calculating the cost of each member to the target point through a cost model according to the state of the current formation member based on the selected formation, and obtaining a cost matrix; task allocation is carried out on the target points through a cost equalization method; and the slave AUV realizes formation reconstruction by tracking the motion trail of the virtual AUV. According to the method, the flexibility and adaptability of the formation are enhanced, the safety and reliability in the task execution process are effectively improved, and a brand new solution is provided for application of the AUV in complex operation.

The invention provides an electrical equipment operation state real-time statistical analysis method and system, and belongs to the field of electrical equipment operation state monitoring, and the method comprises the steps: obtaining multi-dimensional time sequence sensor data during the operation of electrical equipment, selecting a proper wavelet basis function based on Shannon information entropy to carry out wavelet packet decomposition, and extracting time-frequency domain features; constructing a dynamic relation graph which takes a sensor as a node and takes a Granger causal relationship as a weight, and learning node space features by using a graph convolutional network in combination with time-frequency features; the node features are input into a gating circulation unit, and electrical equipment state sequence codes fused with space-time dependence are generated; and carrying out Viterbi decoding by using the state transition cost matrix, and reasoning an optimal electrical equipment operation state path. According to the method, the uncertainty of a prediction result can be quantified, clear confidence evaluation is provided for a final analysis conclusion, and the reliability of decision making is improved.

The invention relates to a deep learning task resource allocation method and device, equipment and a medium. The method comprises the following steps: respectively analyzing a computational graph structure and a historical resource monitoring log corresponding to a deep learning task, and generating a tensor dependency graph and a resource use time sequence matrix; performing time-varying demand prediction on the basis of the matrix, generating a time-phased resource constraint table, performing memory allocation processing on the basis of a tensor dependency graph, and generating a tensor memory partitioning scheme and an inter-partition communication cost matrix; and performing static resource pre-allocation based on the time-phased resource constraint table and the tensor memory partitioning scheme, generating pre-allocated resource configuration, and performing resource scheduling and outputting real-time resource configuration through a deep reinforcement learning model according to the time-phased resource constraint table, the inter-partition communication cost matrix and the pre-allocated resource configuration. According to the method, by means of dynamic resource allocation, cross-partition communication cost optimization, reinforcement learning optimization and the like, the resource utilization rate and task execution efficiency of a deep learning task are remarkably improved.

The invention discloses a multi-language large model dialogue optimization method and system fused with a knowledge graph, relates to the technical field of natural language processing and artificial intelligence dialogues, and is used for solving the problems of named entity recognition, time-varying attribute processing and dynamic updating of the knowledge graph in cross-language dialogues. And performing word-by-word scanning on each round of dialogue text through a multi-language naming mention extractor to form a traceable mention index. And performing cross-language retrieval and candidate entity positioning in the knowledge graph through a multi-language pre-training semantic model and an alias matching path. And for time-varying attributes and title changes, priority graph adjustment is carried out based on time slice windows and tense evidences, and smooth transition between new titles and old titles is guaranteed. And the accuracy and the stability of the final dialogue response are ensured by constructing a sparse cost matrix and an uncertainty re-discrimination process. According to the method, the problem of title switching in a dialogue system under multiple contexts is effectively solved, and the application effect of the knowledge graph in dialogues is improved.

The invention discloses a PC component BIM twinborn collaborative exchange method based on deep learning, and aims to solve the problems that unstructured multi-modal data of a site or a factory is difficult to automatically identify the identity, the state and the quality of a PC component and accurately map the unstructured multi-modal data with BIM twinborn objects one by one, and increment updating capable of realizing multi-party collaborative exchange is difficult to form. According to the method, time synchronization and coordinate calibration preprocessing is carried out on image data, video data and point cloud data, component instance recognition and feature extraction are carried out by using a deep learning model, and component state information and component quality information are generated; analyzing a BIM twinborn model to obtain priori features such as a unique identifier of a component, a component type, a size parameter, a spatial position and component geometry, performing cross-modal feature fusion under priori guidance, generating a candidate matching relationship, and constructing an optimal transmission cost matrix containing feature difference and constraint penalty to obtain a soft matching matrix; one-to-one mapping is obtained by adopting matching confidence rejection and assignment solution, an identification result and an evidence data index are written into a twinborn object, component-level difference is carried out on the twinborn object and a previous version to generate evidence increment exchange data, and the technical effects of high-reliability automatic updating and traceable cooperative exchange of the component-level twinborn model are achieved.

A method for calculating an optimal route includes selecting first and second optimization targets; distributing optimal route calculation of the first and second optimization targets to first and second thread modules, respectively; calculating a cost matrix between nodes or receiving the cost matrix from a database when calculating an optimal route according to the LNS by applying the first and second optimization targets; randomly arranging nodes of a starting point, an arrival point and stopovers, and executing the optimal route calculation according to the LNS based on the cost matrix between the arranged nodes; and selecting an optimal route that satisfies a predetermined condition according to the first and second optimization targets calculated in the optimal route calculation. The optimal route calculation is executed by updating the cost matrix between two nodes connected by destruction and repair of the LNS into a cost matrix corresponding to the optimization target.

The invention discloses a cold and hot data identification and evaluation method based on cost-sensitive learning, and the method is characterized in that the method comprises the steps: constructing a power grid data feature sample set; based on the power grid data feature sample set, a cost sensitive learning algorithm is adopted to construct a classification model, and the cost sensitive learning algorithm applies differential penalty to cold and hot data misclassification by introducing an asymmetric cost matrix; collecting power grid operation data in real time, and generating a to-be-identified sample feature vector; inputting the to-be-identified sample feature vector into the classification model, and outputting a cold and hot data classification result and classification confidence; and dynamically adjusting a data storage strategy and a computing resource allocation scheme according to the output classification result and the classification confidence thereof. According to the method, the asymmetric cost matrix is introduced, the probability that hot data is misjudged as cold data is effectively reduced, the accuracy of cold and hot data classification is improved, and therefore a data storage strategy and computing resource allocation are more accurately optimized.

The invention relates to the technical field of computers. The target interruption trajectory association method based on the fusion strategy comprises the following steps: carrying out average distance calculation on a predicted position sequence and a position sequence of a new trajectory set to obtain an average distance value, and constructing a distance loss matrix based on the average distance value; performing feature parameter extraction processing on the historical trajectory and the new trajectory in the coarse correlation trajectory pair set to obtain a speed feature, an acceleration feature and an angular velocity feature; performing characteristic parameter loss calculation on the speed characteristic, the acceleration characteristic and the angular speed characteristic to generate a characteristic parameter loss matrix; performing weighted fusion processing on the distance loss matrix and the characteristic parameter loss matrix to generate an overall association cost matrix; and carrying out Hungary algorithm optimization processing on the overall association cost matrix, and outputting an association trajectory pair so as to achieve the technical effects of improving association robustness in a long-interruption scene, reducing the probability of misassociation of a high-maneuvering target and enhancing scene generalization ability.

The invention discloses a recommendation model training method and device based on article fairness. The recommendation model training method comprises the steps that ID embedded representations of a user and an article and multi-modal content embedded representations of the article are constructed; a cost matrix is constructed based on the multi-modal semantic similarity between the articles, source distribution and target distribution in optimal transmission are redefined by using a weighting mechanism and a fairness transmission mechanism based on article exposure, and optimal transmission loss is constructed based on the cost matrix and a transmission matrix; and constructing ID representation loss based on the ID embedded representation of the user and the article and the interaction label, constructing multi-modal content representation loss based on the ID embedded representation of the user and the multi-modal content embedded representation of the article and the interaction label, and constructing the three types of loss into joint loss to train an article recommendation model. The method can effectively improve feature representation learning of unpopular articles, realizes collaborative improvement of fairness and performance of a recommendation system, and is especially suitable for cold start and long-tail article recommendation scenes.

This invention discloses a multilingual large-scale model dialogue optimization method and system integrating knowledge graphs, belonging to the fields of natural language processing and artificial intelligence dialogue technology. It addresses the problems of named entity recognition, time-varying attribute processing, and dynamic knowledge graph updates in cross-language dialogues. A multilingual named mention extractor scans each round of dialogue text word-by-word, forming a traceable mention index. Cross-language retrieval and candidate entity localization in the knowledge graph are performed using a multilingual pre-trained semantic model and alias matching paths. For time-varying attributes and title changes, priority graph adjustment is performed based on time slice windows and temporal evidence to ensure a smooth transition between old and new titles. The accuracy and stability of the final dialogue response are ensured by constructing a sparse cost matrix and an uncertainty re-discrimination process. This invention effectively solves the title switching problem in multi-context dialogue systems and improves the application effect of knowledge graphs in dialogue.

In an embodiment, a method for training a detection transducer (DETR) includes initializing, with at least one processor, a parameter of the DETR; propagating, with the at least one processor, an input image through the DETR; determining, with the at least one processor, an error value by comparing an output of the DETR with a known expected output; and iteratively updating, with the at least one processor, a parameter in the DETR based on the error value by minimizing a global loss, where minimizing the global loss includes: generating an allocation cost matrix; solving an allocation optimization problem based on the allocation cost matrix to obtain an optimal allocation cost; a global loss function is then minimized, which is represented as a sum of the optimal allocation cost and the allocation-independent loss.

A method for ascertaining an assignment rule in order to combine test results of different tests of the same semiconductor element. The method includes the following steps: adapting a model, for example, a linear regression model, use of the model for predicting the test data, calculating a cost matrix on the basis of the predictions, applying a greedy implementation to the cost matrix in order to obtain a new assignment rule and repeating these steps multiple times.

The invention provides a self-adaptive distribution system and method for GPU server load prediction, and the method comprises the steps: obtaining a monitoring data stream of a GPU server in real time, and constructing a monitoring sequence matrix according to the monitoring data stream; performing load mutation extraction on each sensitive mutation index in the monitoring sequence matrix to obtain a load pulse factor of each sensitive mutation index, constructing a multi-scale feature set, and generating a context representation vector sequence according to the multi-scale feature set; carrying out load prediction by using the context representation vector sequence to obtain a GPU load prediction sequence, and mapping the GPU load prediction sequence into a resource demand profile; and generating a resource migration cost matrix according to the current resource state information and the resource demand profile, and generating a resource allocation optimization strategy based on the resource migration cost matrix. According to the technical scheme provided by the invention, the abrupt change mode of the load can be accurately captured and represented, and the allocation decision is optimized by coupling the resource migration cost, so that the overall stability of self-adaptive resource allocation is realized.

The invention discloses an intelligent work order scheduling method and system based on multi-factor cost prediction, and belongs to the technical field of work order management. According to the method, a multi-factor dynamic cost model comprehensively considering in-transit time, skill matching degree, work order priority and service time limit SLA risk is constructed by acquiring work order and personnel states in real time, and a cost index is calculated for each potential scheduling scheme. According to the system, a variable neighborhood search VNS optimization algorithm is adopted, and global solution is carried out on the basis of a cost matrix so as to find an optimal task allocation scheme with the lowest total cost. According to the method, the weight can be dynamically adjusted and optimized according to operation states such as real-time traffic and personnel load, self-adaptive intelligent decision making is realized, and finally, an optimization scheme is automatically distributed to a personnel terminal, so that the work order scheduling efficiency, the resource utilization rate and the SLA fulfillment rate are comprehensively improved.

The application is specifically an AI computing power efficiency increasing method and system based on heterogeneous resource pooling and dynamic scheduling, relates to the technical field of distributed AI computing power scheduling, and comprises the following steps: collecting bandwidth, time delay and packet loss rate data of a wide area network link between each regional node in real time, calculating a transmission cost value between any two nodes, and generating a cost matrix.In the application, a standard virtual computing power unit index is introduced, floating point operation capability, video memory bandwidth and operator compatibility coefficient are uniformly included in calculation, different architecture heterogeneous chips obtain unified dimension logical resource representation, thereby providing a quantitative basis for cross-brand task allocation, and the problem that the existing scheme cannot perform equivalent comparison on heterogeneous chips is solved.

The invention relates to the technical field of atlas modeling, in particular to an industrial chain atlas modeling method and system based on deep learning. Comprising the steps of performing time sequence decomposition on a downstream enterprise sales volume and a price sequence, extracting a demand fluctuation component and calculating a demand influence factor to obtain upstream supply-demand relationship strength; determining a supply and demand intensity index of the upstream enterprise; judging whether the connection weight exceeds a threshold value or not, if so, extracting a cost structure parameter and generating a dynamic adjustment factor, updating the simulation cost matrix, and obtaining a classified cost parameter set through a clustering algorithm; iteratively optimizing the set, outputting upstream cost structure details when a deviation from downstream market performance is less than a threshold value, and calculating a bargaining ability score based on a conduction effect path; and generating an adjustment vector, updating the edge weight of the industrial chain atlas, and outputting resource configuration scheme parameters when the trend is stable. According to the invention, intelligence and precision of industrial chain supply-demand relationship modeling, cost dynamic optimization and resource configuration are realized.

A method for determining a pulse sequence for a quantum processor (2) is specified for solving a Quadratic Unconstrained Binary Optimization, QUBO, problem, comprising: - providing an initial coupling matrix characteristic of a coupling of at least some qubits of the quantum processor (2) and an initial cost matrix characteristic of the QUBO problem, - determining a rearranged cost matrix by rearranging at least some elements of the initial cost matrix dependent on a distance to the initial coupling matrix, - determining an adjusted coupling matrix by adjusting at least some elements of the initial coupling matrix dependent on a further distance of the initial coupling matrix to the rearranged cost matrix, - determining several sub-coupling matrices dependent on the adjusted coupling matrix, and - determining the pulse sequence dependent on the sub-coupling matrices. Further, a device (6), a computer program and a computer-readable storage medium are specified.

A non-linear transformation invariant cluster target matching method based on local topological features comprises the steps that multiple groups of source tetrads and target tetrads are screened from a seeker view point set according to the nearest neighbor principle, after effective description of point targets is constructed through a topological structure, the point targets are matched according to non-linear transformation among views, and the target targets are matched with the source tetrads and the target tetrads. A target point set is mapped to a linear space by using a nonlinear compensation factor, and accurate matching of point targets among different views is realized through matching point pair search and cost matrix optimization. According to the method, the non-linearly distorted point set is compensated, and the base vector linear representation is combined to ensure that the point target description has invariance for the non-linear transformation, so that the high-robustness matching effect is realized, and the method can be widely applied to unmanned aerial vehicle cluster matching, air defense missile multi-target identification and multi-view point set matching tasks in complex scenes.

The invention discloses a multi-target tracking method and system based on dynamic weight and multilevel feature fusion, and the method comprises the steps: carrying out the target detection of a video frame, and obtaining a detection box, and the confidence and appearance features of the detection box; predicting the current position based on the existing trajectory, and obtaining a multi-level appearance feature library which is stored in a hierarchical manner according to confidence; calculating position cost and appearance cost between the detection frame and the trajectory, dynamically fusing weights of the detection frame and the trajectory according to a continuous tracking state of the trajectory, and introducing a penalty term based on confidence of the detection frame to generate a comprehensive association cost matrix; matching is carried out, and the state of the successfully matched track and the feature library are updated; and managing the life cycle of the unmatched track and the detection frame. The method adapts to different tracking scenes through dynamic weight adjustment, the appearance information of each confidence detection frame is fully utilized through the multi-level feature library, and the high and low confidence detection frames are associated at the same time through the global optimization strategy, so that the tracking accuracy, robustness and continuity in a complex scene are remarkably improved.

This invention relates to a self-supervised 3D particle tracking and velocity measurement method, comprising the following steps: S1, acquiring a continuous temporal sequence of 3D source particle sets and 3D target particle sets, and performing feature encoding to obtain the feature matrix of the corresponding particle sets; S2, inputting the feature matrices of the source particle sets and target particle sets into a DFCT, and outputting the aligned cross-frame features; S3, constructing a transmission cost matrix between particles, establishing a dense soft correspondence between particles in two frames, and outputting the initial flow field estimation result and matching confidence; S4, constructing a composite self-supervised loss function, and iteratively optimizing the model parameters by minimizing the composite self-supervised loss function; S5, refining the initial flow field estimation result, and outputting the final 3D fluid velocity field. The beneficial effects of this invention are: solving the problems of algorithm dependence on large-scale high-quality labeled data, low efficiency in extracting semantic features from complex flow field point clouds, and matching ambiguity in high-displacement, high-density scenarios.

This invention discloses a training method for a reward model, an optimization method for a large language model, and related methods. The training method for the reward model includes: obtaining preference training sample pairs and a reward model to be trained, wherein the preference training sample pairs include preferred response samples and non-preferred response samples; calculating the reward score difference between the preferred response samples and non-preferred response samples based on the reward model to be trained; constructing a cost matrix based on the reward score difference and the semantic correlation between the preferred response samples and non-preferred response samples; calculating the loss margin based on the cost matrix; calculating the pairwise preference loss value with the margin based on the loss margin, and updating the parameters of the reward model to be trained with the goal of minimizing the loss value with the margin, thereby obtaining a trained reward model. This improves the model's learning ability on difficult samples and its overall generalization performance, avoids over-reliance on simple samples, and thus improves the generation quality of large language models in complex tasks.