Method and system for constructing multi-agent and thought chain

Abstract

The application relates to the technical field of computer systems and discloses a multi-agent and thought chain construction method and system, which comprises the following steps: acquiring a forward propagation hidden state feature vector of a first-level agent node, constructing a semantic feature evolution sequence by calculating the spatial cosine distance between adjacent network layers, extracting a second-order difference value of the sequence on a depth axis as a curvature measure index, and identifying a semantic vector offset point caused by logical forced fitting; in response to the offset point triggering cooperative topology reconstruction, constructing a local entropy value sequence according to the joint information entropy of the output probability distribution under historical reasoning steps, and dynamically correcting a reasoning convergence boundary threshold value. The application positions a logical deviation through inherent tensor geometric properties, realizes real-time self-repairing of a reasoning path, solves the problems of logical divergence and illusion in a long-link reasoning process, and enhances the reasoning stability of a multi-agent system.

Description

Technical Field

[0001] This invention belongs to the field of computer system technology, and in particular relates to a method and system for constructing multi-agent and thought chain-based systems. Background Technology

[0002] Currently, in complex reasoning tasks supported by large language models, the mainstream approach consists of thought chain technology based on specific computational models and multi-agent collaborative architecture. This approach uses autoregressive generation of a single model or serial cascading of multiple homogeneous agent nodes to determine the reasoning path. While this approach is effective in handling short-path logic, it exhibits significant fundamental flaws when facing deep decision-making scenarios with long paths and stringent fault tolerance requirements. The logical fragility within the computational model increases significantly with the depth of reasoning.

[0003] Conventional computational models heavily rely on the output data of preceding nodes when generating subsequent logical features, forming a strongly coupled link relationship at the semantic flow level. Due to the lack of an inherent confidence decoupling mechanism, subtle deviations generated in the initial stage of the model can be amplified cascaded along the linear thought chain. Existing communication topologies generally follow static distribution protocols, using the output tensor of upstream nodes as deterministic prior inputs for downstream nodes. When the computational model generates logical illusions with high confidence features, subsequent stages cannot achieve self-correction, causing the entire computing cluster to fall into logical collapse and generating a large amount of ineffective computing power loss. In addition to the limitations at the hardware and protocol levels, such as the aforementioned static topology and semantic coupling, existing technologies have also explored dynamic control algorithms for the thought chain reasoning process, but still have not been able to address the inherent limitations of the computational model. Feature evolution can effectively address the problem of logical stability. For example, Chinese invention patent application CN120106122A discloses a method and system for constructing multi-agent thought chains based on reinforcement learning. It guides each agent to plan tasks through reinforcement learning strategies and attempts to use thought chains for small-step reasoning to enhance global collaboration. However, this type of technical solution essentially treats thought chains as an external behavioral logic of high-level task decomposition. It lacks real-time monitoring of the semantic feature evolution law inside the model feedforward chain and cannot perceive the intrinsic illusion caused by deep parameter fitting deviation. Moreover, because it is highly dependent on the reward feedback of the external reward function, it has obvious defensive lag when dealing with long-link reasoning and it is difficult to achieve in-situ repair of the reasoning path before semantic distortion accumulates to the output layer.

[0004] Therefore, how to utilize the tensor features in the feedforward link of a specific computational model to construct a method that can detect logical collapse and perform dynamic topology reconstruction has become the technical problem to be solved by this invention. Summary of the Invention

[0005] This invention provides a method for constructing a multi-agent and thought chain framework, comprising the following steps: Step S101: During the process of the multi-agent system performing long-link thinking chain reasoning for complex logical tasks, the hidden state feature vectors generated by the first-level agent node during the forward propagation operation and located in the hidden layer of the corresponding neuron in the two adjacent network computing layers are extracted in real time. Step S102: Perform unit norm mapping on the hidden state feature vectors in adjacent network layers, calculate the spatial cosine distance of the hidden state feature vectors after mapping in the hyperspherical projection state, and construct a semantic feature evolution sequence that reflects the evolution law of reasoning logic in the depth dimension of the model based on the mapping relationship between the spatial cosine distance and the network layer index. Step S103: Extract the second-order difference value of the semantic feature evolution sequence on the depth index axis, use the second-order difference value as a curvature measure index to characterize the rate of change of semantic offset trend, and identify the calculation nodes in the semantic feature evolution sequence that meet the condition that the second-order difference value exceeds the preset offset threshold as semantic vector offset points caused by logical forced fitting. Step S104: In response to the identified semantic vector offset point, the multi-agent system is reconstructed from a static serial structure to a dynamic cooperative topology. By extracting the joint information entropy of the historical inference steps in the multi-agent system under the probability distribution of the full vocabulary in the output layer, a local entropy value sequence reflecting the uncertainty of the inference process is constructed. Based on the polarity and magnitude of the differential gradient of the local entropy value sequence, the inference convergence boundary threshold is dynamically corrected so that the internal inference logic state of the multi-agent system is maintained within the pre-set entropy value safety range.

[0006] Preferably, step S102 includes: obtaining the network computation layer depth index corresponding to the first-level agent node in the forward propagation inference task, and performing tensor dimension normalization transformation on the hidden state feature vector; by calculating the cosine value of the angle between the current network layer hidden state feature vector and the previous network layer hidden state feature vector in the tensor space, defining the calculation result as the offset representing the fluctuation of semantic features; arranging the offset according to the temporal relationship of the depth index to generate a semantic feature evolution curve reflecting the convergence characteristics of data flow within the model.

[0007] Preferably, the rule for calculating the second-order difference value in step S103 is as follows: ,in, The curvature feature value at the network layer depth index k. This represents the offset corresponding to the k-th layer in the semantic feature evolution curve; when When the offset exceeds a pre-set threshold, the k-th layer is determined to be a semantic vector offset point.

[0008] Preferably, after step S104, the method further includes: step S401, during the collaborative topology reconstruction process, calculating the spatial cosine distance between the feature vectors output by the heterogeneous bypass nodes that are parallel to the primary agent node; step S402, when a polarity conflict is detected in the inference results output by the heterogeneous bypass nodes, applying a decay weight mask to the conflicting feature dimensions, filtering out the conflicting feature dimensions and retaining the feature dimensions within the consensus interval for feature fusion.

[0009] Preferably, in step S402, a filtering mechanism for dimensional divergence is constructed by using a decay weight mask, which causes the logical deduction state of the multi-agent system to regress to a pre-set semantic baseline value when adversarial data input is detected.

[0010] Preferably, the dynamic correction of the inference convergence boundary threshold includes: correcting the trigger criterion based on the evolution slope of the local entropy value sequence; when the logical divergence of the computation model exceeds the preset divergence limit, correcting the inference convergence boundary threshold upwards to maintain the output stability of the inference main chain of the multi-agent system. Before triggering the cooperative topology reconstruction in step S104, the method also includes executing a verification path for the anti-disturbance of the computation model.

[0011] Preferably, the verification path includes: injecting a pre-defined Gaussian noise tensor into the input side of the primary agent node; calculating the relative entropy divergence between the original output probability distribution and the response output probability distribution after noise injection; quantifying the stability score of the current inference path based on the relative entropy divergence; and performing cooperative topology reconstruction when the stability score is lower than a pre-defined safety limit.

[0012] The preferred rule for calculating relative entropy divergence is as follows: Where S is the relative entropy divergence, P(x) is the original output probability distribution, and Q(x) is the response output probability distribution; the relative entropy divergence characterizes the sensitivity of the computational model to input disturbances.

[0013] Preferably, the collaborative topology reconstruction obtains the evolution trend of the reasoning logic by the polarity and magnitude of the differential gradient of the local entropy sequence, and adjusts the computing power allocation ratio within the multi-agent system according to the evolution trend.

[0014] A system based on multi-agent and thought chain construction includes: The data extraction unit is used to extract, in real time, the hidden state feature vectors generated by the first agent node during the forward propagation operation in the multi-agent system, which are located in two adjacent network computing layers. The sequence processing unit is used to perform unit norm mapping on the hidden state feature vectors in adjacent network layers, calculate the spatial cosine distance of the mapped hidden state feature vectors in the hyperspherical projection state, and construct a semantic feature evolution sequence that reflects the evolution law of reasoning logic in the depth dimension of the model based on the mapping relationship between the spatial cosine distance and the network layer index. The offset determination unit is used to extract the second-order difference value of the semantic feature evolution sequence on the depth index axis, use the second-order difference value as a curvature measure index to characterize the rate of change of semantic offset trend, and identify the calculation nodes in the semantic feature evolution sequence that meet the condition that the second-order difference value exceeds the preset offset threshold as semantic vector offset points. The topology control unit is used to trigger the reconstruction transformation of the multi-agent system from a static serial structure to a dynamic cooperative topology structure in response to the identified semantic vector offset point. By extracting the joint information entropy of the historical inference steps in the multi-agent system under the probability distribution of the full vocabulary in the output layer, it constructs a local entropy value sequence that reflects the uncertainty of the inference process. Based on the polarity and magnitude of the differential gradient of the local entropy value sequence, it dynamically corrects the inference convergence boundary threshold, so that the internal inference logic state of the multi-agent system is maintained within the pre-set entropy value safety range.

[0015] Compared to existing technologies, this invention, based on a multi-agent and thought chain construction method, has the following advantages: 1. In the construction of multi-agent and thought chain, by extracting the hidden state feature vectors of the output of continuous hidden network layers and calculating the spatial cosine distance between adjacent layers, a trajectory smoothness sequence distributed along the network depth axis is constructed, enabling the computational model to detect the evolution law of deep features. This technical path based on second-order difference operation to extract trajectory curvature index can quantify the nonlinear deformation of feature vectors in feedforward flow, thereby identifying the spatial leap caused by forced fitting of causal relationship before the logical fault reaches the output layer. This mechanism avoids the limitation of relying solely on the probability distribution of the output layer for judgment, ensuring that even in the low-entropy illusion state, the system can still trigger correction by monitoring the abnormal deviation of the hidden layer semantic manifold, improving the logical determinism of complex reasoning tasks in the deep spatial evolution process.

[0016] 2. By constructing a two-dimensional decision space using the global spatial dispersion of the joint information entropy and attention weight matrix, the inherent divergence attribute of the task and the reasoning collapse state are essentially decoupled. When the joint information entropy exceeds the convergence boundary threshold and the global spatial dispersion exceeds the preset threshold, asynchronous topology reconstruction is triggered. When the global spatial dispersion is low, it is judged as normal logical divergence. This dual-gating defense mechanism reuses the accompanying information in the computation process and can accurately identify real logical vulnerabilities without introducing an external verification network. It reduces redundant path building and computing power squeeze caused by conventional complex reasoning, and optimizes the scheduling efficiency and computing throughput of multi-agent clusters in long-cycle thinking chain tasks.

[0017] 3. During asynchronous topology reconstruction, the spatial cosine distance between the dimensionality-reduced feature vectors output by heterogeneous bypass nodes is calculated, and a decay mask is applied to conflicting feature dimensions to build a defense barrier against dimensionality-level disagreements. When polarity opposition is detected in the bypass inference conclusions, the feature data that causes opposition is filtered out by masking, and only the feature dimensions within the consensus safety range are retained for dimensionality reduction fusion. This processing method avoids semantic pollution caused by feature neutralization in conventional fusion operations, ensuring that the system can retreat to the basic consensus semantic baseline when facing extreme ambiguity or adversarial context injection, and ensuring the antifragility of the core inference state machine under extreme conditions. Attached Figure Description

[0018] Figure 1 This is a flowchart of the multi-agent thinking chain logic evolution and adaptive reconstruction method of the present invention; Figure 2 This is a diagram of the multi-agent data flow system architecture supported by the present invention for constructing thought chains. Detailed Implementation

[0019] The technical solutions of the embodiments of this application will be clearly described below with reference to the accompanying drawings. Obviously, the described embodiments are only some embodiments of this application, not all embodiments. All other embodiments obtained by those skilled in the art based on the embodiments of this application are within the scope of protection of this application.

[0020] It should be noted that all directional and positional terms used in this invention, such as: up, down, left, right, front, back, vertical, horizontal, inner, outer, top, low, lateral, longitudinal, center, etc., are only used to explain the relative positional relationship and connection between components in a specific state (as shown in the accompanying drawings). They are only for the convenience of describing this invention and do not require that this invention be constructed and operated in a specific orientation. Therefore, they should not be construed as limiting this invention. In addition, the descriptions of "first," "second," etc., in this invention are for descriptive purposes only and should not be construed as indicating or implying their relative importance or implicitly indicating the number of technical features indicated.

[0021] In the description of this invention, unless otherwise explicitly specified and limited, the terms installation, connection, and linking should be interpreted broadly. For example, they can refer to fixed connections, detachable connections, or integral connections; they can refer to mechanical connections; they can refer to direct connections or indirect connections through an intermediate medium; they can refer to the internal connection of two components. For those skilled in the art, the specific meaning of the above terms in this invention can be understood according to the specific circumstances.

[0022] In the description of this specification, references to the terms "an embodiment," "some embodiments," "illustrative embodiments," "examples," "specific examples," or "some examples," etc., indicate that a specific feature, structure, material, or characteristic described in connection with that embodiment or example is included in at least one embodiment or example of the present invention. In this specification, the illustrative expressions of the above terms do not necessarily refer to the same embodiment or example, and the specific features, structures, materials, or characteristics described may be combined in any suitable manner in one or more embodiments or examples.

[0023] A method for constructing a multi-agent and thought chain framework includes the following steps: Step S101: During the long-link thinking chain reasoning process for complex logical tasks in the multi-agent system, the hidden state feature vectors generated by the first-level agent node when performing forward propagation operations are extracted in real time and located in the hidden layer of the corresponding neuron in the two adjacent network computing layers. Step S102: Perform unit norm mapping on the hidden state feature vectors in adjacent network layers, calculate the spatial cosine distance of the hidden state feature vectors after mapping in the hyperspherical projection state, and construct a semantic feature evolution sequence that reflects the evolution law of reasoning logic in the depth dimension of the model based on the mapping relationship between the spatial cosine distance and the network layer index. Step S103: Extract the second-order difference value of the semantic feature evolution sequence on the depth index axis, use the second-order difference value as a curvature measure index to characterize the rate of change of semantic offset trend, and identify the calculation nodes in the semantic feature evolution sequence that meet the condition that the second-order difference value exceeds the preset offset threshold as semantic vector offset points caused by logical forced fitting. Step S104: In response to the identified semantic vector offset point, the multi-agent system is reconstructed from a static serial structure to a dynamic cooperative topology. By extracting the joint information entropy of the historical inference steps in the multi-agent system under the probability distribution of the full vocabulary in the output layer, a local entropy value sequence reflecting the uncertainty of the inference process is constructed. Based on the polarity and magnitude of the differential gradient of the local entropy value sequence, the inference convergence boundary threshold is dynamically corrected so that the internal inference logic state of the multi-agent system is maintained within the pre-set entropy value safety range.

[0024] Preferably, step S102 includes: obtaining the network computation layer depth index corresponding to the first-level agent node in the forward propagation inference task, and performing tensor dimension normalization transformation on the hidden state feature vector; by calculating the cosine value of the angle between the current network layer hidden state feature vector and the previous network layer hidden state feature vector in the tensor space, defining the calculation result as the offset representing the fluctuation of semantic features; arranging the offset according to the temporal relationship of the depth index to generate a semantic feature evolution curve reflecting the convergence characteristics of data flow within the model.

[0025] Preferably, the rule for calculating the second-order difference value in step S103 is as follows: ,in, The curvature feature value at the network layer depth index k. This represents the offset corresponding to the k-th layer in the semantic feature evolution curve; when When the offset exceeds a pre-set threshold, the k-th layer is determined to be a semantic vector offset point.

[0026] Preferably, after step S104, the method further includes: step S401, during the collaborative topology reconstruction process, calculating the spatial cosine distance between the feature vectors output by the heterogeneous bypass nodes that are parallel to the primary agent node; step S402, when a polarity conflict is detected in the inference results output by the heterogeneous bypass nodes, applying a decay weight mask to the conflicting feature dimensions, filtering out the conflicting feature dimensions and retaining the feature dimensions within the consensus interval for feature fusion.

[0027] Preferably, in step S402, a filtering mechanism for dimensional divergence is constructed by using a decay weight mask, which causes the logical deduction state of the multi-agent system to regress to a pre-set semantic baseline value when adversarial data input is detected.

[0028] Preferably, the dynamic correction of the inference convergence boundary threshold includes: correcting the trigger criterion based on the evolution slope of the local entropy value sequence; when the logical divergence of the computation model exceeds the preset divergence limit, correcting the inference convergence boundary threshold upwards to maintain the output stability of the inference main chain of the multi-agent system. Before triggering the cooperative topology reconstruction in step S104, the method also includes executing a verification path for the anti-disturbance of the computation model.

[0029] Preferably, the verification path includes: injecting a pre-defined Gaussian noise tensor into the input side of the primary agent node; calculating the relative entropy divergence between the original output probability distribution and the response output probability distribution after noise injection; quantifying the stability score of the current inference path based on the relative entropy divergence; and performing cooperative topology reconstruction when the stability score is lower than a pre-defined safety limit.

[0030] The preferred rule for calculating relative entropy divergence is as follows: Where S is the relative entropy divergence, P(x) is the original output probability distribution, and Q(x) is the response output probability distribution; the relative entropy divergence characterizes the sensitivity of the computational model to input disturbances.

[0031] Preferably, the collaborative topology reconstruction obtains the evolution trend of the reasoning logic by the polarity and magnitude of the differential gradient of the local entropy sequence, and adjusts the computing power allocation ratio within the multi-agent system according to the evolution trend.

[0032] A system based on multi-agent and thought chain construction includes: The data extraction unit is used to extract, in real time, the hidden state feature vectors generated by the first agent node during the forward propagation operation in the multi-agent system, which are located in two adjacent network computing layers. The sequence processing unit is used to perform unit norm mapping on the hidden state feature vectors in adjacent network layers, calculate the spatial cosine distance of the mapped hidden state feature vectors in the hyperspherical projection state, and construct a semantic feature evolution sequence that reflects the evolution law of reasoning logic in the depth dimension of the model based on the mapping relationship between the spatial cosine distance and the network layer index. The offset determination unit is used to extract the second-order difference value of the semantic feature evolution sequence on the depth index axis, use the second-order difference value as a curvature measure index to characterize the rate of change of semantic offset trend, and identify the calculation nodes in the semantic feature evolution sequence that meet the condition that the second-order difference value exceeds the preset offset threshold as semantic vector offset points. The topology control unit is used to trigger the reconstruction transformation of the multi-agent system from a static serial structure to a dynamic cooperative topology structure in response to the identified semantic vector offset point. By extracting the joint information entropy of the historical inference steps in the multi-agent system under the probability distribution of the full vocabulary in the output layer, it constructs a local entropy value sequence that reflects the uncertainty of the inference process. Based on the polarity and magnitude of the differential gradient of the local entropy value sequence, it dynamically corrects the inference convergence boundary threshold, so that the internal inference logic state of the multi-agent system is maintained within the pre-set entropy value safety range.

[0033] Example 1: In a distributed computing cluster scenario where a logical deduction task based on a multi-agent and thought chain construction method is being executed and is in continuous operation, when a specific computing model is processing long-link thought chain deductions with a depth exceeding a preset threshold, its internal neuron nodes experience spatial jumps when dealing with out-of-bounds knowledge distributions. This logical shift caused by forced parameter fitting manifests as the output layer probability distribution remaining within the convergence interval while the hidden state tensor undergoes orthogonal shift, forming a unidirectional cumulative logical distortion. To avoid the defensive lag caused by relying solely on the output layer to determine the divergent state, the system extends the monitoring dimension to the deep space of the feedforward link of the specific computing model. The system extracts the first-level agent node in real time during the thought chain reasoning process of the multi-agent system before execution. During the propagation operation, hidden state feature vectors are generated and located in the hidden layers of corresponding neurons in two adjacent network computation layers. The system performs unit norm mapping on the hidden state feature vectors in adjacent network layers, calculates the spatial cosine distance of the mapped hidden state feature vectors in the hyperspherical projection state, and constructs a semantic feature evolution sequence that reflects the reasoning logic and the evolution law in the model depth dimension based on the mapping relationship between the spatial cosine distance and the network layer index. The second-order difference value of the semantic feature evolution sequence on the depth index axis is extracted and used as a curvature measure index to characterize the rate of change of semantic offset trend. By extracting the nonlinear curvature distortion of the deep hidden state trajectory, the system captures and locates the early causal relationship fitting points where the joint information entropy of the surface layer has not changed.

[0034] The system follows the tensor geometric property monitoring criterion when determining the logical evolution state, and the rule for calculating the second-order difference value is set as follows: ,in, The curvature feature value at the depth index k of the network layer is a dimensionless pure number, used to quantify the degree of nonlinear deformation in the semantic evolution of this network layer. Let be the offset corresponding to the k-th layer in the semantic feature evolution curve, and let be a dimensionless pure number. It is defined as the cosine of the angle between the hidden state feature vector of the current network layer and the hidden state feature vector of the previous network layer in the tensor space. This is the offset corresponding to the (k+1)th layer. For the offset corresponding to the (k-1)th layer, the system identifies computational nodes in the semantic feature evolution sequence that satisfy the condition that the second-order difference value exceeds the preset offset threshold as semantic vector offset points caused by logical forced fitting. In response to the identified semantic vector offset points, the system triggers the reconstruction transformation of the multi-agent system from a static serial structure to a dynamic cooperative topology. It extracts the joint information entropy of the historical inference steps in the multi-agent system under the probability distribution of the full vocabulary in the output layer, constructs a local entropy value sequence reflecting the uncertainty of the inference process, and dynamically corrects the inference convergence boundary threshold based on the polarity and magnitude of the differential gradient of the local entropy value sequence, so that the internal inference logic state of the multi-agent system is maintained within the preset entropy value safety range. The trajectory curvature extraction mechanism of the deep hidden state space and the dynamic mapping mechanism of the surface joint information entropy boundary interact with each other. The two work together to drive the adaptive tilt of computing resources and block the linear transmission path of non-convergent logic states in the static Markov chain. 1000 sets of offline test samples with standard logic chain annotations are selected to perform forward propagation tests on the first-level agent nodes, and the curvature measurement index generated by each network computation layer is recorded. The numerical distribution is used, and the 95th percentile value corresponding to the correct inference path is written into a hardware register as a baseline threshold. The system processor introduces an exponential moving average filter with a decay factor of 0.8 to smooth the semantic feature evolution sequence within three consecutive inference steps, filtering out tensor numerical disturbances caused by instantaneous voltage fluctuations in the underlying hardware. The preferred working window for the offset threshold is set to 0.08 to 0.12, and the second-order difference value is used. If the inference steps exceed the upper limit of the window for two consecutive steps, a semantic vector offset point is determined.

[0035] Before triggering cooperative topology reconstruction, the system intervenes in the disturbance rejection verification path of the computational model by injecting a pre-defined normalized Gaussian noise tensor into the input side of the primary agent node. It then calculates the relative entropy divergence between the original output probability distribution and the response output probability distribution after the noise injection. The calculation rule for this relative entropy divergence is as follows: Where S is the relative entropy divergence, a dimensionless pure number used to characterize the sensitivity of the computational model to input perturbations; x is the lexical index variable in the full vocabulary space; P(x) is the original output probability distribution before noise injection; and Q(x) is the response output probability distribution after noise injection. The system measures the stability score of the current inference path based on the relative entropy divergence and performs cooperative topology reconstruction when the stability score is lower than a pre-set safety threshold. During cooperative topology reconstruction, the system calculates the spatial cosine distance between the feature vectors output by heterogeneous bypass nodes parallel to the primary agent node. When polarity conflicts exist in the inference results output by the path nodes, a decay weight mask is applied to the conflicting feature dimensions to filter out the conflicting feature dimensions and retain the feature dimensions within the consensus interval for feature fusion. This causes the logical inference state of the multi-agent system to revert to a pre-set semantic baseline value. The system transforms multi-dimensional logical conflicts into low-dimensional feature decay operations, suppressing non-consensus feature data from participating in the main link computation. Heterogeneous bypass nodes perform 4-bit and 8-bit weight quantization processing on the pre-trained weights of the primary agent nodes to generate weights, or change the kernel sampling temperature parameter Temperature between 0.2 and 0.8 in a gradient manner. The degree distribution construction triggers the transformation of the multi-agent system from a static serial structure to a dynamic cooperative topology structure. This is achieved by modifying the routing weight matrix of the multi-agent orchestration layer. The tensor flow direction pointing to a single subsequent node is adjusted to have equal weights pointing to that node and heterogeneous bypass nodes. In the feature fusion stage, the system controller reads the relative entropy divergence S value, calculates the reciprocal 1 / S as the baseline attenuation coefficient, and performs hardware multiplication with conflicting feature dimensions whose spatial cosine distance is greater than 0.4. The output is an attenuation weight mask for the divergence of specific dimensions. The hardware multiplier multiplies the mask with the original tensor data of the conflicting feature dimensions, reducing the influence of that dimension in the feature fusion stage. The activation weight of the section, the multi-agent computing cluster forms a closed loop of information flow based on the quantitative physical parameters provided by the deep curvature feature monitoring and relative entropy divergence detection mechanism in the long-link thinking chain deduction. The semantic offset tensor located in the deep network is identified and stripped in place by the system according to the above threshold comparison logic before it reaches the output layer. The system maintains its adaptability to open divergent contexts. The feature dimensionality reduction and fusion action of the main thinking chain eliminates the opposing conflict dimensions. The inference computing power inside the multi-agent cluster is asynchronously reconstructed and allocated when the semantic collapse critical point is detected. The final output probability distribution of the core inference logic main chain is strictly maintained within the structured boundary.

[0036] Example 2: A verification system was constructed in a distributed heterogeneous computing array deploying a specific computing model cluster with a 96-layer network architecture. This computing array contains 128 computing nodes and deploys an external dynamic cooperative topology routing gateway. To reproduce semantic ambiguity interference in an industrial environment, a Gaussian white noise sequence with a signal-to-noise ratio of 15dB is continuously injected into the underlying vocabulary image space of the specific computing model on the input side to form the basic perturbation condition. The preset offset threshold required to identify semantic vector offset points is determined by a trade-off between the timeliness of logical correction and the computational cost of cluster topology reconstruction. When the preset offset threshold approaches its lower limit, the system... Excessive sensitivity to quantitative deformation triggers excessive topology reconstruction actions. When the threshold approaches the upper limit, the system delays and truncates the nonlinear causal error chain. Based on the threshold optimization function composed of the joint information entropy volatility and the bypass consensus delay time, the preferred working window for the preset offset threshold is in the range of 0.08 to 0.12. Accordingly, the experiment sets 0.05 as the lower limit control group for exceeding the range, 0.10 as the baseline parameter for the experimental group of this invention, and 0.15 as the upper limit control group for exceeding the range. In addition, a basic control group that relies on the joint information entropy of the output layer for monitoring is set, as well as a local control group that has a topology reconstruction module but lacks a deep curvature measurement index.

[0037] After initiating the long-link reasoning task, the basic control group, upon reaching the 42nd reasoning step, experienced a slight fluctuation in the joint information entropy of the output layer from the initial 1.21 to 1.25, without triggering the interception condition. Ultimately, the reasoning accuracy dropped to 41.3%. In contrast, the experimental group, when processing the same reasoning link, extracted the hidden state feature vectors of the primary agent node at layers 39, 40, 41, and 42 in real time and performed unit norm mapping. Measurements showed that the spatial cosine distance from layer 39 to layer 40 was 0.011, from layer 40 to layer 41 was 0.013, and from layer 41 to layer 42 suddenly increased to 0.128. The system, based on the formula... The second-order difference value of the depth index at layer 41 was calculated to reach 0.113. This curvature metric quantifies the semantic shift in the deeper layers of the network when the joint information entropy at the surface layer is 1.25. At this point, the local control group missed the interception window, while the experimental group of this invention, with the offset threshold set to 0.10, identified the second-order difference value of 0.113 as exceeding the threshold boundary of 0.10, determined that the calculation node was a semantic vector offset point, triggered cooperative topology reconstruction, and after injecting Gaussian noise tensors into the input side, the calculated relative entropy divergence increased to 4.72. Since it was below the set safety limit, the system mobilized three heterogeneous bypass nodes to intervene in the inference. The spatial cosine distance between the feature vectors output by the heterogeneous bypass nodes was measured to be in the range of 0.45 to 0.62, which was confirmed as a polarity conflict. The system applied a decay weight mask to the feature dimensions that caused the conflict, filtered out the conflicting feature dimensions, and retained the feature dimensions within the consensus range for feature fusion. The reasoning logic state of the multi-agent system reverted to the semantic baseline value. The final accuracy of the long-link reasoning task of the experimental group of this invention rebounded to 89.7%.

[0038] Experimental data on the parameter gradient of the offset threshold confirmed the objective reality of the aforementioned effective working window. When the offset threshold was set to the lower limit of the out-of-range control group at 0.05, the system triggered 47 topology reconstructions in a single long-link inference, increasing the computational cost by 241.6% and the inference latency to 1452.3ms. When the offset threshold was set to the upper limit of the out-of-range control group at 0.15, the initial semantic offset residuals in the deep network accumulated to the output layer, and the final accuracy dropped to 52.4%. Multiple experimental groups of this invention with offset thresholds set to 0.08, 0.10, and 0.12 respectively maintained an accuracy of over 88.4% with no more than 5 topology reconstructions. The data shows that by extracting the second-order difference value of the hidden state feature vector in the depth dimension to locate the semantic vector offset point, and by combining relative entropy divergence evaluation and applying attenuation weight mask to filter out conflict feature dimensions, the logic output accuracy of the multi-agent system under long-link inference conditions with noise interference is not affected by the fitting deviation of the underlying parameters, and the transmission continuity of structured causal features is maintained.

[0039] Example 3: In a distributed computing cluster that deploys a specific computational model and executes long-link thought chain inference tasks, for the execution steps of injecting a standardized Gaussian noise tensor into the input side of the first-level agent node and applying attenuation weight masks to conflicting feature dimensions, the system enables a physical quantization procedure based on tensor states; the memory caches the hidden state feature vectors of the first-level agent node for the 10 consecutive historical inference steps before the current inference step; the processor reads this batch of feature vectors and extracts the variance numerical sequence on each tensor dimension; a basic Gaussian distribution matrix with a mean of 0 and an initial variance of 1 is constructed in memory; the extracted variance numerical sequence is multiplied with the basic Gaussian distribution matrix as a spatial scaling factor to generate a standardized Gaussian noise tensor that matches the local semantic fluctuation amplitude of the current inference step; the noise intensity is determined by the historical variance data generated by the specific computational model; and the generated standardized Gaussian noise tensor is injected into the input side of the first-level agent node.

[0040] The detection module acquires the original output probability distribution and the response output probability distribution after injecting a normalized Gaussian noise tensor. The system controller calculates the relative entropy divergence between the original output probability distribution and the response output probability distribution according to the aforementioned preset relative entropy divergence formula, and uses the calculated relative entropy divergence value as the stability score of the current inference path. The system sets the safety limit judgment threshold to 0.5. When the stability score is lower than 0.5, the system controller generates an instruction signal to trigger cooperative topology reconstruction. In the reconstruction process, the system calculates the spatial cosine distance between the feature vectors output by the heterogeneous bypass nodes. For conflict feature dimensions with a measured value greater than 0.4, the system reads the value of the relative entropy divergence and calculates its reciprocal as the reference attenuation coefficient. It then sends the coefficient and the corresponding spatial cosine distance value into the hardware register to perform a multiplication operation, and outputs the mask attenuation value for that specific conflict dimension.

[0041] The underlying arithmetic logic unit reads the mask attenuation value and uses a hardware multiplier to multiply this value with the original tensor data of the conflicting feature dimension to obtain the Hadamard product. This linear compression operation reduces the activation weight of the conflicting feature dimension in the feature fusion process. For feature dimensions with a spatial cosine distance less than or equal to 0.4, the system controller assigns a weight mask with a value of 1 to preserve them. The system uses arithmetic mapping logic that multiplies the reciprocal of the relative entropy divergence with the spatial cosine distance to convert the feature conflict state between nodes into tensor-scalar multiplication operations of the arithmetic logic unit, filtering out mutually exclusive feature data between heterogeneous bypass nodes. The computing power allocation within the multi-agent cluster is asynchronously reconfigured based on the above feature fusion results, and the final output probability distribution of the main link is maintained within the numerical boundary calculated based on the historical inference step variance and relative entropy divergence.

[0042] Example 4: When the system faces the objective condition of a newly deployed computing cluster and a lack of historical simulation records for a specific computing model, the system feeds a pre-verified benchmark test sequence into the first-level agent node to initiate the pre-calibration procedure. The first-level agent node performs a single forward propagation operation based on the benchmark test sequence. The detection module extracts the hidden state feature vectors of adjacent network layers and performs unit norm mapping processing. The system calculates the cosine of the angle between the hidden state feature vectors in the tensor space as the offset of the corresponding network layer space, and calculates the absolute value of the second difference of the difference between the offsets of consecutive network layers to obtain the corresponding curvature feature value. Then, it calculates the statistical mean of the curvature feature value of the entire link, and performs a multiplication operation of the statistical mean with the scaling factor 1.2 through a hardware multiplier. The final product value is written into the hardware register as the pre-set offset threshold required for subsequent topology reconstruction judgment.

[0043] The system processor synchronously extracts the variance of each tensor dimension generated during the forward propagation operation of the benchmark test sequence, and writes the variance value sequence into the memory cache to construct the initial historical inference record. When the long-link thinking chain inference task is officially started, the variance value sequence is read as a spatial scaling factor, and multiplied with the basic Gaussian distribution matrix in the system memory with a mean of 0 and an initial variance of 1 to generate a standardized Gaussian noise tensor of the initial state and inject it into the input side of the first-level intelligent agent node. The detection module obtains the original output probability distribution and the response output probability distribution, and sends them to the controller to calculate the relative entropy divergence. The system controller reads the relative entropy divergence and calculates its reciprocal as the benchmark attenuation coefficient, and performs a hardware multiplication operation with the spatial cosine distance between the feature vectors of the heterogeneous bypass nodes to output the mask attenuation value. The multi-agent computing cluster completes the operation flow from the system cold start stage to the dynamic topology reconstruction monitoring state based on the above-mentioned quantization calibration parameters.

[0044] Example 5: When the system faces objective conditions where the task complexity fluctuates drastically and the statically set inference convergence boundary threshold cannot adapt to the changing semantic environment, a dynamic mapping adjustment procedure for the inference convergence boundary threshold is initiated. The detection module collects the joint information entropy of the probability distribution of the full vocabulary of the output layer of the multi-agent system in five consecutive historical inference steps in real time. The acquired continuous joint information entropy is stored in the hardware register in chronological order to construct a local entropy value sequence. The system's arithmetic logic unit reads the local entropy value sequence and calculates the first-order forward difference value of the joint information entropy between adjacent inference steps. The algebraic sign of the first-order forward difference value is extracted as the polarity of the difference gradient, and the absolute value of the first-order forward difference value is extracted as the magnitude of the difference gradient.

[0045] The system's arithmetic logic unit determines the threshold compensation parameters based on the differential gradient polarity and amplitude output. When it detects that the differential gradient polarity is positive and the differential gradient amplitude is greater than the preset baseline fluctuation tolerance of 0.02, the system determines that the logic divergence of the calculation model is in a state of continuous expansion. It reads the differential gradient amplitude and multiplies it by the proportional compensation coefficient of 1.5 calibrated in the read-only memory through a hardware multiplier to obtain the corresponding compensation increment. It adds this increment to the inference convergence boundary threshold of the current cycle and outputs the updated inference convergence boundary threshold. The quantitative adjustment logic based on the derivative characteristics of the local entropy value sequence transforms the logic divergence state into specific numerical compensation instructions, enabling the multi-agent cluster to scale the trigger boundary of feature fusion in real time according to the uncertainty evolution law of the output layer when processing long-link thinking chain inference.

[0046] The embodiments of this application have been described above with reference to the accompanying drawings. Unless otherwise specified, the embodiments and features in the embodiments of this application can be combined with each other. This application is not limited to the specific embodiments described above. The specific embodiments described above are merely illustrative and not restrictive. Those skilled in the art can make many other forms under the guidance of this application without departing from the spirit of this application and the scope of protection of this invention, and all of these forms are within the protection scope of this application.

Claims

1. A method for constructing a multi-agent and thought chain framework, characterized in that, Includes the following steps: Step S101: During the process of the multi-agent system performing long-link thinking chain reasoning for complex logical tasks, the hidden state feature vectors generated by the first-level agent node during the forward propagation operation and located in the hidden layer of the corresponding neuron in the two adjacent network computing layers are extracted in real time. Step S102: Perform unit norm mapping on the hidden state feature vectors in adjacent network layers, calculate the spatial cosine distance of the hidden state feature vectors after mapping in the hyperspherical projection state, and construct a semantic feature evolution sequence that reflects the evolution law of reasoning logic in the depth dimension of the model based on the mapping relationship between the spatial cosine distance and the network layer index. Step S103: Extract the second-order difference value of the semantic feature evolution sequence on the depth index axis, use the second-order difference value as a curvature measure index to characterize the rate of change of semantic offset trend, and identify the calculation nodes in the semantic feature evolution sequence that meet the condition that the second-order difference value exceeds the preset offset threshold as semantic vector offset points caused by logical forced fitting. Step S104: In response to the identified semantic vector offset point, the multi-agent system is reconstructed from a static serial structure to a dynamic cooperative topology. By extracting the joint information entropy of the historical inference steps in the multi-agent system under the probability distribution of the full vocabulary in the output layer, a local entropy value sequence reflecting the uncertainty of the inference process is constructed. Based on the polarity and magnitude of the differential gradient of the local entropy value sequence, the inference convergence boundary threshold is dynamically corrected so that the internal inference logic state of the multi-agent system is maintained within the pre-set entropy value safety range.

2. The method for constructing a multi-agent and thought chain according to claim 1, characterized in that, Step S102 includes: obtaining the network computation layer depth index corresponding to the first-level agent node in the forward propagation inference task, and performing tensor dimension normalization transformation on the hidden state feature vector; by calculating the cosine value of the angle between the current network layer hidden state feature vector and the previous network layer hidden state feature vector in the tensor space, defining the calculation result as the offset representing the fluctuation of semantic features; arranging the offset according to the temporal relationship of the depth index to generate a semantic feature evolution curve reflecting the convergence characteristics of data flow within the model.

3. The method for constructing a multi-agent and thought chain according to claim 2, characterized in that, The rules for calculating the second-order difference value in step S103 are as follows: ,in, The curvature feature value at the network layer depth index k. This represents the offset corresponding to the k-th layer in the semantic feature evolution curve; when When the offset exceeds a pre-set threshold, the k-th layer is determined to be a semantic vector offset point.

4. The method for constructing a multi-agent and thought chain according to claim 1, characterized in that, Step S104 is followed by: Step S401, during the collaborative topology reconstruction process, calculating the spatial cosine distance between the feature vectors output by the heterogeneous bypass nodes that run parallel to the primary agent node; Step S402, when a polarity conflict is detected in the inference results output by the heterogeneous bypass nodes, applying a decay weight mask to the conflicting feature dimensions, filtering out the conflicting feature dimensions and retaining the feature dimensions within the consensus interval for feature fusion.

5. The method for constructing a multi-agent and thought chain according to claim 4, characterized in that, In step S402, a filtering mechanism for dimensional divergence is constructed by using a decay weight mask. When adversarial data input is detected, the logical deduction state of the multi-agent system is regressed to a pre-set semantic baseline value.

6. The method for constructing a multi-agent and thought chain according to claim 1, characterized in that, The dynamic correction of the inference convergence boundary threshold includes: correcting the trigger criterion based on the evolution slope of the local entropy value sequence; when the logical divergence of the computation model exceeds the preset divergence limit, the inference convergence boundary threshold is corrected upward to maintain the output stability of the inference main chain of the multi-agent system. Before triggering the cooperative topology reconstruction in step S104, the verification path for the anti-disturbance of the computation model is also executed.

7. The method for constructing a multi-agent and thought chain according to claim 6, characterized in that, The verification path includes: injecting a pre-defined Gaussian noise tensor into the input side of the primary agent node; calculating the relative entropy divergence between the original output probability distribution and the response output probability distribution after noise injection; quantifying the stability score of the current inference path based on the relative entropy divergence; and performing cooperative topology reconstruction when the stability score is lower than a pre-defined safety limit.

8. The method for constructing a multi-agent and thought chain according to claim 7, characterized in that, The rules for calculating relative entropy divergence are as follows: Where S is the relative entropy divergence, P(x) is the original output probability distribution, and Q(x) is the response output probability distribution; the relative entropy divergence characterizes the sensitivity of the computational model to input disturbances.

9. The method for constructing a multi-agent and thought chain according to claim 1, characterized in that, Collaborative topology reconstruction obtains the evolution trend of inference logic by measuring the polarity and magnitude of the differential gradient of the local entropy sequence, and adjusts the computing power allocation ratio within the multi-agent system according to the evolution trend.

10. A system for constructing multi-agent and thought chain frameworks, used to implement the method for constructing multi-agent and thought chain frameworks as described in claim 1, characterized in that, include: The data extraction unit is used to extract, in real time, the hidden state feature vectors generated by the first agent node during forward propagation operations and located in two adjacent network computation layers during the thought chain reasoning process of a multi-agent system. The sequence processing unit is used to perform unit norm mapping on the hidden state feature vectors in adjacent network layers, calculate the spatial cosine distance of the mapped hidden state feature vectors in the hyperspherical projection state, and construct a semantic feature evolution sequence that reflects the evolution law of reasoning logic in the depth dimension of the model based on the mapping relationship between the spatial cosine distance and the network layer index. The offset determination unit is used to extract the second-order difference value of the semantic feature evolution sequence on the depth index axis, use the second-order difference value as a curvature measure index to characterize the rate of change of semantic offset trend, and identify the calculation nodes in the semantic feature evolution sequence that meet the condition that the second-order difference value exceeds the preset offset threshold as semantic vector offset points. The topology control unit is used to trigger the reconstruction transformation of the multi-agent system from a static serial structure to a dynamic cooperative topology structure in response to the identified semantic vector offset point. By extracting the joint information entropy of the historical inference steps in the multi-agent system under the probability distribution of the full vocabulary in the output layer, it constructs a local entropy value sequence that reflects the uncertainty of the inference process. Based on the polarity and magnitude of the differential gradient of the local entropy value sequence, it dynamically corrects the inference convergence boundary threshold, so that the internal inference logic state of the multi-agent system is maintained within the pre-set entropy value safety range.

Images