Enhanced decision-making method and device based on thinking chain labeling, equipment and medium

Abstract

The invention relates to the technical field of artificial intelligence, can be applied to business scenes of financial science and technology, medical health and the like, and discloses an enhanced decision method, device, equipment and medium based on thinking chain annotation, which comprises the following steps: extracting core information features to generate a structured data set, loading a basic language model and executing supervision fine tuning to generate a fine-tuned model, fusing multi-modal input to generate fusion features, constructing a state observation space to receive the fusion features as input, generating reward signals based on a double reward mechanism and optimizing model parameters to generate an optimized model, deploying a monitoring module to dynamically adjust parameter configuration to generate an adaptive decision model, and outputting a decision response result. According to the method, multi-source data extraction, structured expression, multi-modal fusion, reinforcement learning optimization and dynamic adaptive mechanism fusion are carried out, so that the understanding ability of the model to complex data, reasoning transparency and the adaptive ability of the model to coping with environmental changes are remarkably improved.

Description

Technical Field

[0001] The present invention relates to the field of artificial intelligence technology, and in particular to a method, device, equipment and storage medium for enhancing decision-making based on thought chain annotation. Background Art

[0002] In the fintech sector, data is highly fragmented. Raw data from various financial institutions, trading markets, and data providers exhibit significant differences in structure, format, and naming conventions. This not only complicates data integration and preprocessing but also easily introduces redundancy or omissions during data alignment, thereby weakening the model's ability to fully understand financial information. Furthermore, the highly dynamic nature of financial markets necessitates extremely high data update frequencies. Traditional models often lack rapid adaptation mechanisms and are unable to respond to environmental changes in real time. This leads to delayed decision-making and difficulty in identifying market trends and risk signals in a timely manner, limiting the model's practicality in key scenarios such as investment judgment and risk monitoring.

[0003] In the healthcare sector, patient data, medical images, clinical texts, monitoring signals, and more exhibit complex multimodal heterogeneity. These data originate from diverse devices and systems, with fragmented data formats and inconsistent information representation, creating difficulties for models in cross-modal understanding and comprehensive reasoning. Existing general-purpose models often fail to establish accurate multimodal associations when faced with such highly complex data structures, limiting their adaptability to tasks such as clinical decision support and disease prediction. Furthermore, medical data is characterized by high timeliness and sensitivity, and traditional models lack the ability to dynamically adjust. They are not responsive enough to real-time changes in patient status or rapidly evolving medical scenarios, making it difficult to provide stable and accurate decision recommendations.

[0004] Furthermore, current mainstream reasoning models are highly complex in their internal structure and algorithms, lacking transparency in their reasoning processes, often resulting in "black box" decision-making. This opacity poses a serious challenge in highly regulated sectors such as finance and healthcare, as regulators or professional users often require clear traceability of decision-making basis to ensure the compliance and rationality of the results. Existing models fail to provide clear reasoning paths or intermediate explanations, making them difficult to trust in real-world applications requiring explainability and accountability.

[0005] Furthermore, current general-purpose models often lack the ability to effectively integrate industry-specific knowledge. When applied to complex business environments like finance or healthcare, they are unable to deeply model specialized terminology, domain rules, or contextual logic, demonstrating a significant lack of business generalization capabilities. Faced with novel business scenarios or non-templated problem descriptions, these models struggle to flexibly adapt their response strategies to the context, severely hindering their widespread adoption and implementation in specialized scenarios. Summary of the Invention

[0006] The main purpose of the present invention is to provide an enhanced decision-making method, device, equipment and storage medium based on thought chain annotation, aiming to solve the technical problem that the existing technology cannot simultaneously achieve structured expression, explainable reasoning and dynamic adaptability when facing multi-source heterogeneous data and complex business scenarios, thereby limiting the effective application of the model in high-requirement fields.

[0007] To achieve the above objectives, the present invention provides an enhanced decision-making method based on thought chain annotation, comprising:

[0008] Extract core information features from multi-source heterogeneous data, and annotate the core information features with thought chains to generate a structured data set;

[0009] Select a base language model to load pre-trained weights, and use the structured dataset to perform supervised fine-tuning to generate a fine-tuned model;

[0010] Integrating at least two data types to generate multimodal input data, and performing a multimodal fusion strategy on the multimodal input data to generate fusion features;

[0011] Constructing a simulation environment including a state observation space, wherein the state observation space receives the fused features as state input;

[0012] Designing a dual reward mechanism including a format reward and an accuracy reward, determining a reward signal based on the dual reward mechanism, and iteratively updating the parameters of the fine-tuned model according to the reward signal in the simulation environment using a policy optimization module to generate an optimized model;

[0013] Deploy a real-time data monitoring module to obtain environmental state change data, and dynamically adjust the parameter configuration of the optimized model according to the environmental state change data to generate an adaptive decision model;

[0014] Decision request data is received through the adaptive decision model, and a decision response result is generated.

[0015] Furthermore, to achieve the above-mentioned purpose, the present invention provides an enhanced decision-making device based on thought chain annotation, comprising:

[0016] The thought chain annotation module is used to extract core information features from multi-source heterogeneous data and perform thought chain annotation on the core information features to generate a structured data set;

[0017] A supervised fine-tuning module is used to select a basic language model, load pre-trained weights, and perform supervised fine-tuning using the structured dataset to generate a fine-tuned model;

[0018] A multimodal fusion module, configured to integrate at least two data types to generate multimodal input data, and perform a multimodal fusion strategy on the multimodal input data to generate fusion features;

[0019] A simulation environment construction module, configured to construct a simulation environment including a state observation space, wherein the state observation space receives the fused feature as a state input;

[0020] a policy optimization module, configured to design a dual reward mechanism including a format reward and an accuracy reward, determine a reward signal based on the dual reward mechanism, and iteratively update the parameters of the fine-tuned model according to the reward signal in the simulation environment through the policy optimization module to generate an optimized model;

[0021] A dynamic adaptation module is used to deploy a real-time data monitoring module to obtain environmental state change data, and dynamically adjust the parameter configuration of the optimized model according to the environmental state change data to generate an adaptive decision model;

[0022] The decision generating module is used to receive decision request data through the adaptive decision model and generate a decision response result.

[0023] Furthermore, to achieve the above-mentioned purpose, the present invention also provides a computer device, which includes a memory, a processor, and an enhanced decision-making program based on thought chain annotation stored in the memory and runnable on the processor. When the enhanced decision-making program based on thought chain annotation is executed by the processor, the steps of the enhanced decision-making method based on thought chain annotation as described above are implemented.

[0024] Furthermore, to achieve the above-mentioned purpose, the present invention also provides a computer-readable storage medium, on which an enhanced decision-making program based on thought chain annotation is stored. When the enhanced decision-making program based on thought chain annotation is executed by a processor, the steps of the enhanced decision-making method based on thought chain annotation as described above are implemented.

[0025] Beneficial effects: The present invention relates to the field of artificial intelligence technology and can be applied to business scenarios such as financial technology and medical health. It discloses an enhanced decision-making method, device, equipment and medium based on thought chain annotation, including: extracting core information features from multi-source heterogeneous data and performing thought chain annotation to generate a structured data set, loading the pre-trained weights of the basic language model and using the structured data set to perform supervised fine-tuning to generate a fine-tuned model, integrating at least two data types to generate multimodal input data and performing multimodal fusion to generate fusion features, constructing a state observation space that receives fusion features as state input, designing a dual reward mechanism consisting of format reward and accuracy reward to generate a reward signal, and iteratively updating the parameters of the fine-tuned model in a simulation environment through a strategy optimization module to generate an optimized model, deploying a real-time data monitoring module to obtain environmental state change data and dynamically adjust the parameter configuration of the optimized model to generate an adaptive decision model, and using the adaptive decision model to generate a decision response result. By integrating multi-source data extraction, structured expression, multimodal fusion, reinforcement learning optimization and dynamic adaptive mechanisms, the present invention realizes a complete closed loop of generative models from static supervised training to dynamic real-time decision-making, significantly improving the model's ability to understand complex data, reasoning transparency and adaptability to environmental changes. BRIEF DESCRIPTION OF THE DRAWINGS

[0026] The present invention will be further described below with reference to the accompanying drawings and embodiments, in which:

[0027] Figure 1 Schematic diagram of an application environment of an enhanced decision-making method based on thought chain annotation in one embodiment of the present invention;

[0028] Figure 2 1 is a flow chart of an embodiment of an enhanced decision-making method based on thought chain annotation according to the present invention;

[0029] Figure 3 Schematic diagram of functional modules of a preferred embodiment of an enhanced decision-making device based on thought chain annotation of the present invention;

[0030] Figure 4 A schematic diagram of the structure of a computer device according to an embodiment of the present invention;

[0031] Figure 5 FIG. 2 is another structural diagram of a computer device according to an embodiment of the present invention. DETAILED DESCRIPTION

[0032] It should be understood that the specific embodiments described herein are only used to explain the present invention and are not intended to limit the present invention.

[0033] The enhanced decision-making method based on thought chain annotation provided by the embodiment of the present invention can be applied in the following situations: Figure 1In an application environment, the user end communicates with the server end through a network. The server end can extract the core information features in the multi-source heterogeneous data through the user end and perform thought chain annotation to generate a structured data set, load the pre-trained weights of the basic language model and use the structured data set to perform supervised fine-tuning to generate a fine-tuned model, integrate at least two data types to generate multimodal input data and perform multimodal fusion to generate fusion features, construct a state observation space that receives fusion features as state input, design a dual reward mechanism consisting of format reward and accuracy reward to generate a reward signal, and iteratively update the parameters of the fine-tuned model in a simulation environment through a strategy optimization module to generate an optimized model, deploy a real-time data monitoring module to obtain environmental state change data and dynamically adjust the parameter configuration of the optimized model to generate an adaptive decision model, and use the adaptive decision model to generate a decision response result. The present invention realizes a complete closed loop of the generative model from static supervised training to dynamic real-time decision-making by integrating multi-source data extraction, structured expression, multimodal fusion, reinforcement learning optimization and dynamic adaptive mechanism, which significantly improves the model's ability to understand complex data, reasoning transparency and adaptability to environmental changes. The user end may be, but is not limited to, various personal computers, laptops, smartphones, tablet computers, and portable wearable devices. The server end may be implemented as an independent server or a server cluster consisting of multiple servers. The present invention is described in detail below using specific embodiments.

[0034] See also Figure 2 , Figure 2 This is a flow chart of an embodiment of the enhanced decision-making method based on thought chain annotation provided by the present invention. It should be noted that although a logical order is shown in the flow chart, in some cases, the steps shown or described may be performed in a different order than here.

[0035] like Figure 2 As shown, the enhanced decision-making method based on thought chain annotation proposed in the present invention includes the following steps:

[0036] S10, extracting core information features from multi-source heterogeneous data, and annotating the core information features with thought chains to generate a structured data set;

[0037] In this embodiment, multi-source heterogeneous data refers to a data set that comes from multiple data sources and has inconsistencies in structure, format, semantics or temporal attributes, such as structured databases, unstructured text, image information, audio recordings, form messages or API interaction results. In order to achieve subsequent model capability enhancement, it is necessary to extract key fragments that contribute information to downstream tasks from these different types of data. In actual processing, based on information concentration theory, task relevance measurement or context aggregation rules, atomic fields, semantic fragments or knowledge tags that directly support specific tasks can be identified as core information features.

[0038] The selection of core information features relies not only on static content but also requires evaluation based on their contextual frequency, association strength, and consistency distribution across different data sources. Once extracted, they cannot be used directly for supervised learning; instead, a structured annotation mechanism is required to enhance their interpretability and semantic mapping capabilities. To this end, we introduce thought chain annotation, which explicitly breaks down the original decision chain into four stages: data observation, fact analysis, rule interpretation, and outcome-oriented analysis. Each stage maps a set of decision actions or reasoning paths and establishes a binding relationship with the corresponding core information features.

[0039] This structured expression constructs an explicit chain with transparent reasoning logic that can be used in the subsequent supervised fine-tuning phase. To ensure the consistency and stability of the annotations, the generated annotations can be verified based on logical consistency detection rules, semantic constraint templates, or weakly supervised verification strategies. Chains that are logically incomplete or semantically contradictory are screened out and ultimately integrated into the structured dataset. Structured datasets are organized according to field templates, including but not limited to original input fields, extracted information fields, annotation step sequence fields, and consistency verification status fields, with clear input-output alignment.

[0040] In practical implementation, multiple data access channels can be pre-configured, such as database synchronization interfaces, web crawler modules, log pipelines, and API collection modules, to uniformly import data into a heterogeneous data cache. Subsequently, the data distillation module is scheduled to perform feature screening, including keyword evaluation based on TF-IDF, cross-source information aggregation based on GNNs, representative sample identification based on cluster centers, or extraction of high-weighted fragments through a trained selective attention mechanism. Each algorithm can be configured to adapt to the stability, semantic strength, and computing resource requirements of the data source.

[0041] The extracted core information features are fed into the logic parsing module, where the reasoning process is broken down through template rule matching, inference graph construction, or a command control mechanism generated based on a large language model. The results are organized into a multi-layered sequence, forming a decision chain and mapping it to the corresponding feature nodes. This sequence is then passed through the rule determination module for consistency checks. For example, this can be done by checking for causal closure based on first-order logic, assessing semantic coherence using a language model, or performing small-scale spot checks through manual labeling.

[0042] All validated thought chains are automatically numbered, encapsulated as structured entries, and stored in a structured dataset along with their source fields. The structure definition can use JSON Schema or table field templates, covering elements such as field ID, content representation, annotated chain ID, reasoning segment, and verification status. It also supports versioning of some chains to facilitate subsequent training and updates.

[0043] Example: In the healthcare sector, core diagnosis-related symptoms, indicator trends, and past medication information can be extracted from electronic medical records, imaging reports, and test data. Structured annotation can be achieved by breaking down the reasoning path from observation to analysis to judgment and conclusion. For example, a patient's "recurrent fever with lung shadows" can be mapped to an observation node, "CT scan showing consolidation with bronchial gas" can be mapped to an analysis node, and "Considering bacterial pneumonia" can be formed into an inference node, ultimately leading to a treatment recommendation node.

[0044] In the fintech sector, credit reports, bank records, and conversation logs can be used to extract income changes, repayment behavior, and asset holdings relevant to credit assessments, and annotate them to form a risk assessment chain. For example, mapping "multiple overdue payments in the past three months" to risk factor analysis, combined with "income decline and debt growth" to form a closed-loop reasoning loop, yielding a structured output of "credit score downgrade."

[0045] This embodiment not only improves the structural quality of the input data, but also establishes a clear mapping relationship between information and reasoning paths by constructing a process that extracts core information features from multi-source heterogeneous data and combines it with a thought chain annotation mechanism. This method converts low-quality, unstructured input into training data with semantic chains and logical explanatory power, providing stronger supervision signals for subsequent model training. The addition of logical consistency verification provides a systematic guarantee for the quality of data annotation and significantly reduces the risk of misleading model learning. The structured data set thus formed has a clear decision chain at the semantic level, which significantly improves the model's fitting ability and transparency for reasoning tasks.

[0046] S20, selecting a basic language model to load pre-trained weights, and performing supervised fine-tuning using the structured dataset to generate a fine-tuned model;

[0047] In this embodiment, the basic language model refers to a deep neural network architecture with general natural language understanding and generation capabilities, which is often implemented based on the Transformer structure and pre-trained on a large-scale corpus. The selection of this model requires comprehensive consideration of the model scale, structural depth, language coverage capabilities, and adaptation requirements of downstream tasks. For example, you can choose an open source multilingual model, or you can choose a language model optimized for structured question answering or reasoning tasks. The pre-trained weights represent the set of parameters obtained by training the model on a large-scale corpus dataset, covering word embeddings, encoder layer parameters, attention matrix weights, etc., and have the ability to generalize semantics.

[0048] The process of loading pre-trained weights typically involves initializing the model structure, aligning the model parameter dimensions, and loading the weight values ​​into the matching layers. After loading the pre-trained weights, supervised fine-tuning is required to tailor the model to the semantic structure, logical patterns, and output style of the task at hand. Supervised fine-tuning involves performing supervised learning on training examples with clear input and target output pairs. A loss function is used to define the gap between the model output and the true label, and backpropagation is used to optimize the model parameters.

[0049] The dataset used for supervised fine-tuning is a structured dataset, consisting of input-output pairs extracted from heterogeneous data and annotated with thought chains, with explicit logical relationships and clear objective expressions. During training, the structured dataset is generally divided into a training set and a validation set. The training set is used to learn model parameters, while the validation set is used to evaluate generalization performance. Loss can be calculated using various methods, such as cross-entropy, KL divergence, and ranking loss, depending on the task type and output structure. Model parameter updates are performed using a gradient descent algorithm, often combined with an optimizer such as Adam or LAMB for scheduling and control to accelerate convergence.

[0050] When the model's performance on the validation set reaches a preset threshold, such as accuracy, BLEU score, or inference chain matching rate, supervised fine-tuning is considered complete, resulting in a fine-tuned model. This model has the ability to map structured inputs to target decision outputs while retaining the general language understanding capabilities developed during the pre-training phase.

[0051] In actual deployment, Transformer language models, such as BERT, GPT, and T5, can be used as the base language model. The pre-trained weight loading phase initializes the weights by calling the model framework's built-in parameter loading API. Some parameters can be frozen to control training scale. Each sample in a structured dataset consists of input fields, target output fields, and inference path fields. The format conversion module converts these data into a unified encoding format acceptable to the model, such as a TokenID sequence.

[0052] During training, a supervised learning framework is used. Inputs are encoded and fed into the model, and the outputs undergo linear mapping and softmax processing to produce predictions. A loss function is defined to calculate the deviation between the model output and the structured labels. Backpropagation is used to calculate gradients and update parameters. Dropout and early stopping mechanisms are introduced to prevent overfitting. After training, a validation set is used to evaluate the model's performance on metrics such as structured reasoning and output logical coherence, ensuring that the model not only completes the language task but also maintains the consistency of the original reasoning path.

[0053] During fine-tuning, a multi-task training strategy can be introduced, combining auxiliary objectives (such as structural analysis and text sorting) to optimize the model's representation capabilities. During training, the log output and visualization module can record changes in training loss and model performance trends in real time, facilitating iterative adjustments.

[0054] Example: In the healthcare business field, you can choose a basic language model with pre-trained medical knowledge corpus, such as BioBERT or ClinicalT5. After loading its pre-trained weights, fine-tune it using a structured clinical question-and-answer dataset. For example, map "The patient had a fever and cough three days ago" to a diagnostic recommendation or examination plan, and train the model to deduce reasonable medical treatment logic from the symptom description.

[0055] In the field of financial technology business, a general pre-trained language model such as T5 or GLM can be selected as the basic language model. After loading the weights, it can be fine-tuned in combination with structured credit assessment annotation data. For example, if the input is "There are 3 credit card overdue times in the past six months, and the monthly income has decreased by 20%", and the target output is "It is recommended to conduct a prudent assessment and attach guarantee terms", the model can grasp the financial risk control decision logic by learning the pairing relationship between the input and the reasoning chain.

[0056] This embodiment significantly improves the model's adaptability to the logical structure and semantic preferences of the current task by loading a basic language model with general comprehension capabilities and performing supervised fine-tuning in conjunction with a structured dataset. This process, while retaining the model's original language generalization capabilities, injects information flows and logical chains specific to domain-specific tasks, enhancing the model's reasoning consistency and output standardization. The fine-tuned model can effectively process structured inputs and generate logically coherent and interpretable outputs, reducing training costs while accelerating downstream deployment.

[0057] S30, integrating at least two data types to generate multimodal input data, and performing a multimodal fusion strategy on the multimodal input data to generate fusion features;

[0058] In this embodiment, integrating at least two data types means that when constructing the model input, it is not limited to a single modal information, such as using only text, but other modal information such as images, voice, time series, structured tables, etc. is introduced to make the input have richer semantic expression capabilities and cross-dimensional feature connections. Two or more data types have different sources, different structures, and different encoding forms, so they require independent preprocessing and feature extraction modules. The original data types may include but are not limited to: clinical text and imaging data, financial news and K-line chart data, user interaction logs and product attribute tables, etc.

[0059] Multimodal input data encapsulates the feature representations of different modalities into a unified, integrated input format that can be recognized and processed by subsequent models. This encapsulation typically requires a modality-aware module to extract embedded features from each modality. For example, this involves converting text into semantic vectors, images into visual feature maps, and structured tables into embedding matrices.

[0060] Multimodal fusion strategies organically combine feature representations from multiple modalities. This strategy not only integrates static features but also captures collaborative patterns and dynamic semantic matching relationships between modalities. Common fusion methods include feature-level concatenation, attention weighting, aligned interactions, and collaborative modeling. Dimensionality unification maps the fused multimodal representations to a tensor representation of uniform length and structure to facilitate subsequent model processing, such as feeding into the state observation space or decision module.

[0061] Fusion features are the output of fusion strategies. They possess cross-modal information representation capabilities, enabling them to integrate the semantics of data from multiple sources within the same representation space while maintaining logical connections between them. For example, text descriptions in medical records and imaging patterns in X-rays can be mapped together into a patient state vector, or user financial data and sentiment text can be mapped together into an investment intent vector.

[0062] In practical implementation, a modular design can be used to deploy multiple feature extraction submodules. Text data is fed into a natural language processing module, which performs word segmentation, word vector mapping, and position encoding, outputting semantic feature vectors. Image data is fed into a visual neural network module, such as a CNN or ViT, to extract high-dimensional image feature vectors. Structured table data is fed into an embedding conversion module, such as a row-column embedding or tree structure encoding module, to generate a dense representation.

[0063] Multimodal input data can be constructed by unifying features from different modalities into an input tensor through a feature concatenation module, using modality identifiers to distinguish sources. For example, a learnable modality vector label can be added to each modality to enhance the model's ability to perceive relationships between modalities.

[0064] For multimodal fusion strategies, a multi-head cross-attention mechanism based on the Transformer architecture can be used to enable the model to focus on the interactions between different modalities during the fusion process. Alternatively, a graph neural network can be employed to combine node representations of different modalities into a graph structure, modeling inter-modal dependencies through graph convolution.

[0065] The output fused features must be linearly mapped to a unified dimension. For example, text features can be mapped from 768 dimensions to 512 dimensions, aligned with the 512-dimensional image features, and then concatenated, weighted summed, or gated fused to form a unified dimensional fused feature tensor.

[0066] Example: In the healthcare sector, text descriptions in electronic medical records can be integrated with corresponding medical images. The text is fed into a BERT model to extract language representations, while the image is fed into a ResNet model to extract visual features. A cross-modal attention mechanism is then used to fuse the two into a unified diagnostic vector, which can be used to assist in disease identification or recommend treatment pathways.

[0067] In the fintech sector, users' text answers to risk questionnaires can be input alongside structured historical transaction records. The text portion is encoded using a language model, while the structured data undergoes numerical normalization and embedding. This is then fused into a user risk profile vector using a multimodal alignment strategy, which can then be used to recommend personalized financial products or adjust automated risk control strategies.

[0068] This embodiment achieves a unified representation of cross-modal semantic information by constructing multimodal input data and implementing a multimodal fusion strategy, enhancing the model's adaptability to complex scenarios at the perception, understanding, and decision-making levels. Compared to single-modal input, this approach can capture potential connections and complementary information between modalities, improving the semantic consistency and reasoning accuracy of the model after fusion of heterogeneous data, while preserving the unique strengths of each modality and enhancing the interpretability and task adaptability of the final output.

[0069] S40, constructing a simulation environment including a state observation space, wherein the state observation space receives the fusion feature as a state input;

[0070] In this embodiment, the purpose of building a simulation environment is to conduct policy exploration and model optimization under controllable conditions, providing a continuous feedback loop for model behavior. This simulation environment can be a data environment for offline modeling or a container for policy evolution during virtual execution. It accepts input, outputs behavior, and provides feedback rewards, forming an indispensable foundation for reinforcement learning or decision optimization.

[0071] The state observation space represents an abstract modeling of the external input state within a simulation environment, recording all perceptual elements of the current environmental state. The state observation space is typically designed to contain several slots or vector fields for accommodating input signals, historical context, environmental indicators, and other content. In this step, dedicated input slots are set in the state observation space for fused features, allowing the fused features generated in the previous step to serve as direct input to the current environmental state, thereby driving the dynamic evolution of the environmental state.

[0072] Fusion features as state inputs inject a joint semantic representation containing multimodal information into the state space to characterize the current environmental perception dimension. This injection requires dimensionality standardization and ensures consistency with the state observation space definition format. For example, tensor dimensions and encoding order must be set synchronously during environment initialization. The fused feature input drives state updates within the simulated environment and triggers the reward-feedback mechanism, influencing state transitions during the decision-making process.

[0073] In addition to fusion feature slots, the state observation space can also be configured with slots for other dynamic state indicators, such as historical decision trajectories, time window information, and policy probability distributions, to expand the observation dimension and express behavioral context. The definition of the state observation space should be scalable and modular to adapt to different behavioral optimization tasks.

[0074] In practice, we first define a multidimensional state tensor structure to represent the state observation space. This tensor is divided into multiple named regions, each of which is used to accommodate a specific type of state information. A high-priority slot is reserved for fusion feature injection, defined as a vector structure of standard dimensions, such as a floating-point vector of length 512.

[0075] During the environment construction phase, the dimensions, naming, and formatting of each slot in the state observation space are set through the parameter configuration interface. The state container is then instantiated and initialized to an empty value. After the fused features are generated by the preceding module, they are sent to the simulation environment via the feature routing mechanism. The state update engine then parses them and injects them into the corresponding slots, completing the state population.

[0076] Once the state observation space is initialized and populated, the simulation environment's behavioral rule system or policy module can be activated, beginning the interactive process of state transition and reward feedback. In the reinforcement learning framework, this space is periodically read by the policy network to drive the next behavioral decision. In the decision graph deduction framework, this space serves as an input reference for each round of simulation, participating in path evaluation and reasoning correction.

[0077] To ensure efficient operation, shared memory or tensorflow mechanisms can be used to accelerate read and write operations on state slots. State slots can also be set as traceable variables to record state evolution trajectories and historical behavior links, supporting strategy backtracking analysis.

[0078] Example: In the healthcare sector, by constructing a state observation space that incorporates fused features from patient images and medical records, it is possible to simulate a doctor's clinical judgment process under multi-information conditions. For example, the fused features reflect the degree of abnormality in a patient's X-ray image and the symptom characteristics described in the text. Based on this, the state observation space activates auxiliary diagnosis simulation logic, providing a continuous feedback environment for the model's learning and judgment logic.

[0079] In the fintech sector, by building a simulation environment that incorporates a fusion of user behavior logs and financial indicators, we can simulate the evolution of user risk preferences under the influence of multiple factors. The system records a user's historical trading patterns and integrates them with real-time public opinion characteristics in a state observation space. Based on this, it generates investment portfolio recommendations and uses return fluctuations as feedback signals to optimize the model.

[0080] This embodiment constructs a state observation space that incorporates fused features and establishes a simulation environment within it, providing a unified contextual expression and interactive foundation for model optimization and behavioral learning. In this state space with fused information input, the model not only fully understands the current input information but also perceives changes in the correlations between different modalities, enabling more flexible and adaptive state response strategies, thereby improving the efficiency and strategy quality of downstream model iterative optimization.

[0081] S50, designing a dual reward mechanism including a format reward and an accuracy reward, determining a reward signal based on the dual reward mechanism, and iteratively updating the parameters of the fine-tuned model according to the reward signal in the simulation environment using a policy optimization module to generate an optimized model;

[0082] In this embodiment, the purpose of constructing a dual reward mechanism is to make the model training process subject to the dual constraints of output structure standardization and semantic decision accuracy. Format reward refers to scoring based on the degree of consistency between the model output and the expected output specification in terms of structure, grammar or format. This score can come from a structure validator or rule matching engine to evaluate whether the output conforms to the expression template, data arrangement specifications, paragraph format, code writing specifications, etc. set by the task. Accuracy reward measures the degree of semantic consistency between the model output and the reference answer, true annotation or target behavior, which can be measured by indicators such as accuracy, recall, BLEU score, and F1 value.

[0083] The reward signal is generated by combining the two scoring results described above. This combination can take the form of weighted summation, minimum selection, gated aggregation, and other methods. The final output is a single numerical value that serves as a feedback signal to drive the policy optimization process. The reward signal must be monotonic, differentiable, and decomposable to ensure that it accurately guides model parameter updates during the gradient propagation process.

[0084] The policy optimization module performs parameter updates in a simulated environment. This module consists of a policy evaluator, a gradient estimator, and a parameter updater. The policy evaluator receives input from the current state observation space, as well as model output and reward signals, to evaluate the performance of the current policy. The gradient estimator calculates the model's loss gradient under the current policy and uses it in conjunction with policy update formulas (such as PPO, REINFORCE, and DDPG) to determine the optimization direction. The parameter updater adjusts the parameters of the fine-tuned model based on the gradient information.

[0085] This update process is iterative until the reward signal converges or the preset number of optimization rounds is reached. The final updated model is the optimized model, which has stronger structural standardization and semantic adaptability.

[0086] In practical implementation, format rewards can be automatically verified by setting a regular expression template, outputting a format tree structure, or a structure rule graph. Each round of model output is fed into a structure verifier, which outputs a format matching score as the format reward. Accuracy rewards can be evaluated using a manually annotated reference dataset, calculating the accuracy rate or F1 value based on the degree of match between the predicted result and the reference answer, which serves as the accuracy score.

[0087] When generating reward signals, weight parameters α and β can be set for format reward and accuracy reward respectively. The reward signal R = α × format reward + β × accuracy reward. The specific weight value can be determined through cross-validation or priority can be set according to actual business needs.

[0088] The policy optimization module uses a policy gradient-based reinforcement learning algorithm, such as the Proximal Policy Optimization (PPO) method, to evaluate policy performance in each round and perform a pruned policy gradient optimization process based on the reward signal. In this case, the model acts as a behavior policy network, and its output influences the behavior sampling in the next round.

[0089] During the optimization process, after each round of parameter updates, the model output is re-entered into the simulation environment for the next round of trials, and reward signals are repeatedly collected for feedback learning. This continues until the optimization goal meets a preset condition, such as when the cumulative average reward improvement rate drops below a certain threshold. This results in the final optimized model.

[0090] Example: In healthcare scenarios, model output often needs to conform to standard formats for medical documents or electronic medical records while also providing accurate diagnostic recommendations. Format rewards ensure that output content conforms to the SOAP template (structured like chief complaint, present medical history, physical signs, and diagnostic recommendations), while accuracy rewards ensure that diagnostic recommendations are consistent with the actual condition description, thereby improving the system's practicality and compliance in assisting diagnosis.

[0091] In FinTech scenarios, models may need to automatically generate risk assessment reports or investment recommendations. Format rewards ensure that the output adheres to financial report formatting, paragraphing, and citation standards, while accuracy rewards verify that risk assessments are consistent with historical data and the actual rules system. Through this dual-reward mechanism, optimized models trained on these metrics can generate text outputs that are both compliant and valuable for decision-making.

[0092] This implementation comprehensively improves the model's output quality by incorporating both structural standardization and semantic accuracy into the reward system. Structural rewards enhance output format consistency and system stability, helping the model generate responses that meet business requirements in formal applications. Accuracy rewards ensure the model's ability to effectively express content and correctly execute task objectives. This dual reward mechanism creates orthogonal constraints, improving the model's convergence speed and generalization capabilities in the policy space and preventing overfitting to a single dimension.

[0093] S60, deploying a real-time data monitoring module to obtain environmental state change data, and dynamically adjusting the parameter configuration of the optimized model according to the environmental state change data to generate an adaptive decision model;

[0094] In this embodiment, the real-time data monitoring module is deployed to enable the system to dynamically perceive and rapidly respond to changes in environmental conditions. This module primarily captures external environmental status information relevant to model operation. Environmental conditions can include dynamic factors such as data source update frequency, changes in data distribution, fluctuations in input complexity, system load, and abnormal interactive feedback. This information is compared and analyzed with the current operational status of the model to provide a basis for adjusting model parameter configuration.

[0095] Environmental state change data refers to information items that show differences or trends compared to the previous state. This information must be timely, quantifiable, and mappable to support adaptive model adjustments. To meet monitoring requirements in different scenarios, capturing environmental state change data requires multi-dimensional sensing channels, including but not limited to network status monitoring, user interaction behavior analysis, input sample distribution detection, and system performance statistics.

[0096] Dynamic adjustment of parameter configurations requires responsiveness and stability, typically involving immediate changes to model hyperparameters, control factors, or scheduling rules. Examples include dynamic adjustments to the learning rate, context window length, generation threshold, and policy exploration coefficient. This adjustment process, based on a parameter mapping mechanism and combining the corresponding adjustment policy set with environmental state change data, takes effect in real time during inference.

[0097] An adaptive decision-making model is a model structure that maintains stable and adaptable decision-making performance under varying environmental conditions. It possesses state-awareness and configuration flexibility. This model can automatically adjust its parameters and behavior based on the latest environmental conditions without retraining, thereby continuously generating highly consistent and timely responses.

[0098] In specific applications, the real-time data monitoring module can be integrated with the model call logic through server-side deployment. This module configures a data collection interface to continuously receive external environment status data, such as API access frequency, input data field change rate, and user operation behavior chain changes. The collected data is preprocessed using a caching mechanism and sliding window statistics to identify change trends and mutation characteristics.

[0099] Change pattern analysis uses statistical learning or anomaly detection algorithms to identify whether the current state deviates from the normal operating baseline. Cluster analysis, principal component analysis, or rate of change threshold methods can be used to assess the significance of environmental conditions and generate change pattern analysis results.

[0100] During parameter configuration updates, the system uses rule matching or control model calls to middleware, mapping different types of environmental changes to model parameter configurations based on pre-set mapping relationships. For example, when input data length fluctuates frequently, the context length is automatically scaled; when prediction accuracy shows a downward trend, stronger regularization terms are enabled; and when interactive load increases, the number of generation steps is reduced to improve response efficiency.

[0101] Ultimately, the adaptive decision model reads the current parameter configuration before each round of invocation and dynamically executes the corresponding control strategy during the inference process without restarting the model or reloading weights, thereby maintaining response continuity and result consistency.

[0102] Example: In healthcare scenarios, as patient health data is dynamically updated (such as real-time vital sign monitoring and incoming lab results), the model needs to respond promptly to changes in disease trends. By deploying a real-time data monitoring module, new input dimensions or measurement frequency changes can be detected, and the prediction model's attention mechanism or time window parameters can be automatically adjusted to ensure the continued effectiveness of diagnosis and treatment recommendations.

[0103] In FinTech business scenarios, market fluctuations, policy changes, and shifts in user behavior patterns can all impact model input distribution and forecast targets. By monitoring market data or user behavior logs in real time, the system can adjust the model's risk weight parameters or strategy rules when it detects abnormal growth in input variables or triggers risk factors, generating investment recommendations tailored to the current market state and mitigating losses caused by outdated judgments.

[0104] This embodiment utilizes a real-time data monitoring module to continuously detect subtle changes in the external environment and adaptively adjust parameter configurations in real time before model inference. This enhances the model's ability to cope with dynamic input scenarios and improves the stability and accuracy of output results. In scenarios with frequently changing data environments or multiple concurrent users, this adaptive mechanism can significantly reduce the frequency of manual intervention, enabling the long-term autonomous operation of the decision-making system.

[0105] S70, receiving decision request data through the adaptive decision model and generating a decision response result.

[0106] In this embodiment, after completing parameter configuration adjustments, the adaptive decision model must be able to execute decision reasoning based on external input data and output response results. This process first receives an input data request, typically from an upstream business system or service interface. The request includes a decision task identifier, request context data, parameter constraints, and the desired output structure. This request data must meet structural integrity, field type matching, and semantic consistency requirements. Upon entering the system, the input interface module first performs protocol conversion and data format verification.

[0107] Before data enters the decision model, the input parsing module converts structured or semi-structured request content into a tensor representation that can be processed within the model. If the request contains multiple data formats, such as natural language, numerical metrics, image fragments, or behavioral trajectories, modality discrimination and preprocessing are also required. These data are mapped to their corresponding embedding spaces and then concatenated to generate the input vector.

[0108] The adaptive decision model controls input processing through updated parameter configurations, including adjusting the weight distribution of the attention mechanism, the strategic path for activating feature channels, and controlling the output constraints of the generator. During the model inference phase, forward propagation is performed based on the optimal policy parameters for the current environment state, outputting a decision vector or policy score. This output must be mapped back to the semantic domain or output format of the original task, such as a classification label, numerical score, or structured response sequence.

[0109] The system also needs to complete post-output processing, including confidence screening of output results, response structure encoding, task context binding, response log tagging, and data signing or metadata annotation of results. The completed response data is delivered to the requester through the response interface module, supporting both synchronous and asynchronous interaction modes.

[0110] To ensure consistent and stable output, the system dynamically corrects potential response deviations or abnormal decision paths by combining historical request traces, model state caching, and a policy correction mechanism during response generation. Under multi-task concurrency, the system also utilizes a model context isolation mechanism to decouple parameter states from different requests, ensuring that each decision path is processed independently.

[0111] In addition, to cope with sudden business or temporary high-priority requests, a weighted priority mechanism can be introduced to control the order of request processing, and resources can be dynamically allocated in combination with the task scheduling queue to maintain the response delay within a controllable range and meet business needs with higher real-time requirements.

[0112] In actual deployment, the adaptive decision model can be connected to the business platform as a service interface, receiving external decision request data through HTTP, gRPC, or message queues. After preprocessing at the interface layer, this data enters the input parsing module, which completes field verification, gap filling, and input standardization operations, and constructs the input tensor or embedding vector.

[0113] During the decision-making process, the model reads the parameters configured in the previous step, including the context window size, attention distribution range, and generation control coefficient. These parameters determine how the model processes the input data, thus affecting the final response strategy.

[0114] The model completes reasoning through a forward computation process, generating the optimal decision based on parameters adapted to the current environment. The results are parsed by a post-processing module into a user-readable structure, such as a command sequence, a natural language response, a score list, or a classification label. These are then bound to a request identifier and returned to the caller.

[0115] The system can also integrate a response caching mechanism to accelerate the processing of common or repeated requests, further improving response speed; in high-concurrency scenarios, throughput control can be achieved through response batching or parallel reasoning.

[0116] This embodiment utilizes an adaptive decision model to generate real-time decisions, ensuring that responses closely align with the current environment and significantly improving the decision-making system's ability to perceive and process changes in external input. The system possesses a mechanism for parsing and responding to multiple types of input, enabling it to stably output interpretable and consistent results within a volatile data stream, enhancing the automation of business processing and intelligent decision-making capabilities.

[0117] The present invention relates to the field of artificial intelligence technology and can be applied to business scenarios such as financial technology and medical health. It discloses an enhanced decision-making method, device, equipment and medium based on thought chain annotation, including: extracting core information features from multi-source heterogeneous data and performing thought chain annotation to generate a structured data set, loading the pre-trained weights of the basic language model and using the structured data set to perform supervised fine-tuning to generate a fine-tuned model, integrating at least two data types to generate multimodal input data and performing multimodal fusion to generate fusion features, constructing a state observation space that receives fusion features as state input, designing a dual reward mechanism consisting of format reward and accuracy reward to generate a reward signal, and iteratively updating the parameters of the fine-tuned model in a simulation environment through a strategy optimization module to generate an optimized model, deploying a real-time data monitoring module to obtain environmental state change data and dynamically adjust the parameter configuration of the optimized model to generate an adaptive decision model, and using the adaptive decision model to generate a decision response result. The present invention realizes a complete closed loop of the generative model from static supervised training to dynamic real-time decision-making by integrating multi-source data extraction, structured expression, multimodal fusion, reinforcement learning optimization and dynamic adaptive mechanism, significantly improving the model's ability to understand complex data, reasoning transparency and adaptability to environmental changes.

[0118] In one embodiment, the above step S10 includes:

[0119] S101, collecting raw data from at least two independent heterogeneous data sources;

[0120] S102, performing information condensation processing on the raw data through a data distillation module to extract core information features related to the target task;

[0121] S103, decomposing the decision-making process of the core information feature into an interpretable sequence of steps of data observation, domain analysis, factor evaluation, and suggestion output through a logic parsing module;

[0122] S104, associating the explainable step sequence with the corresponding core information features to form a thought chain annotation;

[0123] S105, verifying the logical coherence of the thought chain annotation through a consistency verification module, and generating a verification result including a pass status or a fail status;

[0124] S106, screening thought chain annotations that meet the logical coherence standard according to the verification result, and generating valid annotation data according to the screened thought chain annotations;

[0125] S107: Integrate the valid annotation data according to a preset data structure to generate a structured data set.

[0126] In this embodiment, it is first necessary to collect raw data from at least two independent data sources with inconsistent data structures or formats. These heterogeneous data sources may include databases, web crawlers, PDF documents, voice recordings, image information or structured interfaces. During the collection process, data type identification and encoding conversion operations are required to ensure the availability and semantic consistency of the data. After the raw data is collected, it is input into the data distillation module, which undertakes the task of information concentration. Its essence is to eliminate noise information that is irrelevant to the target task and extract feature fields or semantic fragments that have a significant effect on decision-making. Specifically, it includes technical processes such as keyword extraction, associated field matching, contextual relationship analysis and topic modeling, and the output is a preliminary focused core information feature set.

[0127] After processing core information features, explicit modeling of the reasoning path is required. This involves explicitly breaking down the reasoning process triggered by these features in actual tasks through a logical parsing module, constructing an interpretable sequence of steps consisting of data observation, domain analysis, factor assessment, and recommendation output. Data observation determines the initial state and boundaries of the input data, while domain analysis incorporates specialized knowledge graphs or rule sets to analyze its contextual attributes. Factor assessment assesses factors influencing decisions and quantifies their importance. Recommendation output forms executable reasoning conclusions based on this analysis.

[0128] The above sequence of steps is bound to the corresponding core information features, forming a one-to-one explicit link, which is the thought chain annotation. Thought chain annotation not only preserves the key features of the input data itself, but also embeds the corresponding logical process and reasoning structure to construct a data-driven reasoning graph. After the annotation is completed, the logical structure of the annotated data must be verified through a consistency verification module to assess the coherence of the internal causal chain and whether there are logical conflicts between the steps. The verification process can be based on symbolic logic parsing, rule engine matching, or counterfactual testing based on fine-tuned models.

[0129] Results from the consistency check are classified as either passed or failed. Only data samples that meet the logical consistency criteria are retained and marked as validly labeled data, while those that do not meet the criteria are discarded or enter the re-labeling process. All validly labeled data that passes the check undergoes field alignment, data format normalization, and semantic tagging standardization according to pre-set structural standards, ultimately consolidating into a data structure set suitable for supervised training, known as a structured dataset.

[0130] This dataset not only contains explicit feature inputs related to the core tasks, but also carries a complete reasoning chain structure, providing a stable and interpretable source of supervision signals for subsequent model training and significantly improving the credibility of the data-driven process.

[0131] This embodiment can effectively improve the semantic expressiveness and reasoning transparency of input data by extracting core information features from heterogeneous data sources and constructing thought chain annotations with causal chains. Further, logically coherent data samples are screened out through a consistency verification mechanism, and standardized and integrated into a structured data set, thereby constructing a data resource with high-quality supervision capabilities. This process not only improves the generalization ability and decision-making credibility of the model in the subsequent training stage, but also effectively avoids reasoning bias caused by incorrect annotations. It is especially suitable for business scenarios with extremely high requirements for decision-making process transparency and data quality.

[0132] In one embodiment, the above step S20 includes:

[0133] S201, selecting a basic language model architecture with natural language understanding capabilities, and loading pre-trained weights into the parameter space of the basic language model;

[0134] S202, dividing the structured data set into a training subset and a validation subset;

[0135] S203, performing supervised fine-tuning training on the basic language model loaded with pre-trained weights based on the training subset, and determining a loss value between the model prediction output and the labeled data;

[0136] S204, updating the parameters of the basic language model through a back propagation module according to the loss value;

[0137] S205, analyzing the model performance indicators of the updated basic language model based on the verification subset;

[0138] S206: When the model performance index reaches a preset index threshold, a fine-tuned model is generated.

[0139] In this example, a basic language model architecture with natural language understanding capabilities is selected as the training starting point. This architecture should be capable of processing long texts, extracting semantic relationships, and performing textual reasoning. Common implementations include language models based on the Transformer architecture, such as BERT, GPT, and T5. The basic language model is selected based on its parameter scale, pre-training corpus sources, context modeling capabilities, and scalability to ensure good adaptability during subsequent supervised fine-tuning.

[0140] Once the model architecture is determined, pre-trained general-purpose weights are loaded into the parameter space of the selected model, enabling the migration of general language knowledge to task-specific adaptability. Loading pre-trained weights preserves the semantic structure of the existing language knowledge by initializing all or part of the model's parameters.

[0141] After loading, the structured dataset needs to be partitioned into training and validation subsets. The partition ratio is determined based on the generalization and stability requirements of the actual task, with a typical ratio of 80% training data and 20% validation data. During the partitioning process, ensure a balanced distribution of sample label categories to improve training effectiveness and evaluation reliability.

[0142] During training, supervised fine-tuning is performed on the base language model loaded with pre-trained weights using the training subset. The core of supervised fine-tuning is to guide the model output using labeled structured data and construct a loss function between the model's predicted output and the labeled data. This loss function measures the difference between the model's output and the true label. Common loss functions such as cross-entropy and mean squared error can be used, combined with regularization factors to mitigate the risk of overfitting.

[0143] Based on the output of the loss function, the backpropagation module is activated to automatically calculate the gradient of each layer's parameters and update the model parameters in the reverse direction. Parameter updates can be performed through multiple rounds of iterations using optimization algorithms such as Adam and SGD to ensure that the model continuously approaches the semantic expression of the annotated data.

[0144] After each round of training or at the end of training, the current model status is evaluated using the validation subset to analyze its performance on unseen samples. Performance metrics may include accuracy, recall, F1 score, or task-specific evaluation metrics such as BLEU and ROUGE to determine whether the current model meets the task requirements.

[0145] Only when the evaluation results meet the pre-set performance threshold standards, such as an accuracy rate of not less than 95% and an F1 value higher than 0.85, will the current model state be confirmed for use in the subsequent stages. At this time, the model is output as a stable and convergent training result, that is, the fine-tuned model.

[0146] This implementation significantly improves the model's ability to perceive task semantics by loading a basic language model with natural language understanding capabilities and performing supervised fine-tuning based on a structured dataset. It also effectively integrates general pre-trained knowledge with target task-specific patterns. By partitioning training and validation subsets, constructing a loss function, and performing an iterative training and evaluation process, the resulting fine-tuned model ensures good generalization, stability, and interpretability, laying a solid foundation for providing accurate and controllable language output in complex reasoning scenarios.

[0147] In one embodiment, the above step S30 includes:

[0148] S301, obtaining original multi-source data including at least text data type and non-text data type;

[0149] S302, processing the text data type by a text feature extraction module to generate a text feature vector;

[0150] S303, processing the non-text data type by a non-text feature extraction module to generate a non-text feature vector;

[0151] S304, integrating the text feature vector and the non-text feature vector to generate multimodal input data;

[0152] S305: Execute a cross-modal fusion strategy on the multimodal input data through a multimodal fusion module, and output the fusion result as a dimensionally aligned fusion feature.

[0153] In this embodiment, the construction of multimodal input relies on the acquisition and processing of at least two different modal types of information from the original data. The data types involved include textual data and non-textual data. Non-textual data can include images, audio, structured tables, time series, etc. The acquisition sources of the original multi-source data should cover different data dimensions and information sources, such as business logs, medical images, financial transaction forms, conversation records, etc.

[0154] For text data types, the text feature extraction module extracts semantic information. This module typically includes processes such as word segmentation, word embedding encoding, and syntactic structure analysis. Based on the Transformer model or LSTM architecture, it can extract contextual dependencies and semantic hierarchies within language sequences. The final output is a fixed-length or dynamic-length text feature vector containing semantic features such as key information entities, contextual relationships, and sentiment.

[0155] Non-text data types are processed by the non-text feature extraction module, employing different modeling approaches based on data modality. For example, convolutional neural networks can be used to extract local and global visual features for image data; numerical embeddings can be generated for tabular data using numerical normalization, missing value imputation, and feature engineering tools; and frequency structure features can be extracted from audio data using mel-spectrograms. During the generation process, non-text feature vectors must maintain semantic integrity and temporal consistency to ensure semantic alignment with text features.

[0156] Integrating these text feature vectors with non-text feature vectors requires constructing a unified input format. Common approaches include vector concatenation, cross-modal attention mechanisms, or common embedding space mapping. The goal of data integration is to construct multimodal input data that can simultaneously express heterogeneous information from multiple sources in a unified dimension.

[0157] The multimodal fusion module receives integrated multimodal input data and implements a cross-modal fusion strategy. This module can employ an early fusion strategy to concatenate input vectors and feed them into a unified model, or a late fusion strategy to process them separately and then perform high-level semantic alignment. More complex fusion strategies include dynamic weighting based on a cross-modal attention mechanism, graph neural networks to process the relationship graphs between different modalities, or the construction of multi-pathway interaction structures to achieve deep information exchange between modalities. The fusion module outputs dimensionally aligned fusion features, ensuring complete, unified, and parseable information representation capabilities for subsequent processing.

[0158] This embodiment extracts structured features from text and non-text data and integrates them into a unified input. The multimodal fusion module can model the semantic interactions between different modalities in a unified vector space. This process significantly improves the model's ability to express heterogeneous information and effectively expands the perceptual range of input information. The resulting fused features possess cross-modal consistency and semantic aggregation capabilities, providing a stable, high-dimensional, and highly semantically relevant foundational representation for subsequent state input modeling and reward feedback calculations.

[0159] In one embodiment, the above step S40 includes:

[0160] S401, defining a state observation space including dynamic state indicator slots;

[0161] S402, configuring a dedicated fusion feature slot in the state observation space;

[0162] S403, loading the fusion feature into the dedicated fusion feature slot;

[0163] S404, initializing other dynamic state indicator slots in the state observation space;

[0164] S405 , constructing a simulation environment based on the state observation space in which the fusion feature is loaded and other dynamic state indicator slots are initialized.

[0165] In this embodiment, the construction of the state observation space aims to provide a high-dimensional, structured information input interface for the decision-making process. First, the structure of the state observation space is defined, which must include several dynamic state indicator slots. Each slot corresponds to a key variable in the system. These variables may come from business rules, environmental feedback, external context, or historical decision-making behavior. The definition process of dynamic state indicator slots generally relies on domain knowledge or historical data statistical features. For example, in medical scenarios, this may include patient risk index, disease stage, intervention frequency, etc.; in the financial field, it may include asset volatility, risk exposure, investment preference labels, etc.

[0166] A dedicated fused feature slot is set in the state observation space to receive the fused features generated in the previous step. This slot must be structured to align with the output dimensions of the fused features and capable of variable length support or dimensionality compression to accommodate the dimensionality differences that may exist among multimodal features. This slot is typically defined as a high-priority input field to ensure that the fused features have a central influence in decision state modeling.

[0167] When loading fused features into fused feature slots, the system performs feature dimension consistency checks and embedding mapping transformations to ensure the data's semantic location within the state space structure is accurate and to preserve intermodal interaction information. On this basis, the system initializes other dynamic state indicator slots in the state observation space. This initialization logic involves setting initial values, calculating sliding averages, and performing normalization or filling in missing values ​​to ensure the availability and stability of the overall state vector.

[0168] Once the fused features are successfully loaded into the corresponding slots and the other indicator slots are initialized, the overall state observation space is constructed. Based on this, the system generates a simulation environment capable of accepting state inputs, providing feedback on reward signals, and supporting policy updates. This simulation environment can be constructed based on a reinforcement learning framework, such as by constructing an interactive state-action feedback loop, or by building a static test environment based on a rule-driven model to evaluate the model's response to input states under controlled conditions.

[0169] This embodiment constructs a state observation space containing slots for dynamic indicators and fused features, enabling the fused features to be embedded in the simulation environment in a structured manner, achieving semantic unification and high-dimensional representation of state inputs. Furthermore, by combining dynamic indicator modeling of the business environment, the realism and adversarial nature of the environment simulation are enhanced, providing a highly abstract state representation foundation for subsequent strategy optimization and model updates, and improving the model's adaptability and responsiveness to complex inputs and uncertain environments.

[0170] In one embodiment, the above step S50 includes:

[0171] S501, establish format scoring criteria for output structure compliance verification;

[0172] S502, establishing an accuracy scoring standard for decision validity verification;

[0173] S503, determining a compliance score of the output action of the fine-tuned model in the simulation environment according to the format scoring standard, and generating a format reward;

[0174] S504, determining the effectiveness score of the decision-making result of the fine-tuned model in the simulation environment according to the accuracy scoring standard, and generating an accuracy reward;

[0175] S505, combining the format reward and the accuracy reward to generate a reward signal;

[0176] S506, receiving state data and a reward signal of a state observation space in the simulation environment through a policy optimization module;

[0177] S507, determining a policy gradient based on the state data and the reward signal;

[0178] S508: Update the parameters of the fine-tuned model according to the policy gradient until a preset iteration termination condition is met, thereby generating an optimized model.

[0179] In this embodiment, the reward mechanism is designed to enhance the model's performance in generating structure and decision-making content, employing a dual strategy of formatting rewards and accuracy rewards. The formatting scoring criteria are used to measure the structural compliance of the model output. The construction of this scoring criteria generally relies on output template rules, format specifications, or syntax tree structures. For example, in financial contract generation, the legal order of clauses and paragraphs must be followed, or in medical report output, standard diagnostic items and data formats must be included. The system can parse the model output's structural tags, punctuation rules, or syntactic hierarchy, calculate its match with the expected template, and generate a formatting reward as a measure of structural compliance.

[0180] Accuracy scoring criteria are designed to assess the degree of agreement between model decisions and real-world or expert results. These criteria are typically based on historically annotated data or known optimal decision samples, and can be evaluated using metrics such as accuracy, F1 score, and match rate. When executing a strategy in a simulated environment, the system compares the model's decision output with the reference answer to determine its effectiveness score, which in turn generates an accuracy reward reflecting the model's degree of achievement of the task objectives.

[0181] Format rewards and accuracy rewards constitute two independent but complementary evaluation dimensions. The system numerically combines these two types of rewards to form the final reward signal. This combination can be weighted summation, normalized fusion, or maximum selection, depending on the specific task's preference for format and content. This reward signal is input into the policy optimization module along with the state data in the state observation space.

[0182] The policy optimization module calculates policy gradients based on state data and reward signals. This process can employ proximal policy optimization (PPO), deep deterministic policy gradients (DDPG), or other policy learning algorithms to ensure that the gradients converge toward a more optimal policy. The policy gradient represents the optimal direction in the model parameter space. This gradient is used to update the parameters of the fine-tuned model, allowing it to evolve towards the goal of satisfying structural specifications while improving output accuracy with each round of interaction. This iterative process continues until the policy optimization module detects that the model performance meets or exceeds a preset termination condition, completing the generation of the optimized model.

[0183] This embodiment strengthens the model's multi-objective learning capabilities in a simulated environment by constructing a dual evaluation system of format rewards and accuracy rewards, providing comprehensive feedback on model output from the two dimensions of structural compliance and content correctness. The introduction of structural rules into the reward mechanism effectively constrains the output form of the model's generated results, while accuracy feedback promotes model updates towards more accurate task completion, thereby achieving a dual improvement in output quality and task effectiveness. Leveraging the iterative update capabilities of the policy optimization module, an optimized model with both structural standardization and content validity is ultimately obtained, enhancing the model's generalization adaptability and reasoning reliability in real-world scenarios.

[0184] In one embodiment, the above step S60 includes:

[0185] S601, configuring a real-time data acquisition interface, and establishing a real-time data monitoring module based on the real-time data acquisition interface;

[0186] S602, continuously capturing environmental status change data through the real-time data monitoring module;

[0187] S603, performing a change pattern analysis operation on the environmental state change data to generate a change pattern analysis result;

[0188] S604, determining the need for adjusting model parameters based on the change pattern analysis result;

[0189] S605, determining a parameter adjustment amount according to the parameter adjustment requirement;

[0190] S606: Update the parameter configuration of the optimized model based on the parameter adjustment amount to generate an adaptive decision model.

[0191] In this embodiment, the deployment of a real-time data monitoring module presupposes the construction of a real-time data acquisition interface for interaction with external systems. This interface must support multiple data access protocols and format conversion capabilities to accommodate the structural differences of heterogeneous data sources in complex business systems, including message queues, log streams, API streams, or sensor sampling streams. Interface configuration typically includes parameters such as data refresh frequency, caching mechanism, and abnormal interruption reconnection mechanism to ensure that environmental state change data can be stably captured even under high-frequency fluctuations.

[0192] The real-time data monitoring module is typically implemented using an event stream processing architecture or streaming computing framework, such as Kafka, Flink, or Spark Streaming, to continuously receive and cache data on environmental state changes. This data includes multi-dimensional indicators such as changes in input data volume, request frequency fluctuations, user behavior pattern changes, and scenario constraint adjustments, reflecting the dynamic characteristics of the current model's operating environment.

[0193] Change pattern analysis models time series trends based on this data. Sliding window analysis, clustering algorithms, or predictive models can be used to extract potential change patterns. The results can be presented as data structures such as behavioral pattern categories, fluctuation level labels, anomaly factors, or trend directions, serving as the basis for subsequent parameter adjustments.

[0194] Determining the need for model parameter adjustments relies on strategies that correlate the change pattern with the current state of the model parameters. For example, in a financial transaction scenario, if a rapid increase in transaction request frequency is detected, the system might determine that the model response threshold needs to be increased or the inference window needs to be compressed to improve real-time performance. This decision is made through mapping rules or an online rule engine, which outputs specific adjustment requirements.

[0195] Calculating parameter adjustments typically involves three key elements: the target, the adjustment magnitude, and the acceptable error range. This can be achieved through incremental update strategies, dynamic optimizer adjustment strategies (such as adaptive learning rate adjustments), or hyperparameter reconfiguration modules. In multi-objective scenarios, coordinated adjustments across multiple parameter dimensions are also necessary to avoid performance regressions or decreased stability.

[0196] After applying the parameter adjustments to the currently optimized model's configuration, the system generates an adaptive decision model. This model retains the original model structure and dynamically updates only key parameters, balancing update efficiency with deployment stability. The adaptive model can instantly respond to changes in the external environment, ensuring stable and dynamically adaptable decision-making results.

[0197] Example: In the fintech sector, to improve the investment advisory system's responsiveness to macroeconomic events and the effectiveness of recommending personalized asset allocation plans for clients, an adaptive decision-making model system based on multimodal fusion and reinforcement learning optimization can be constructed. This system is designed to process heterogeneous financial data from multiple sources, including brokerage trading systems, financial news, historical client operations, and macroeconomic data, and generate stable, real-time, and interpretable asset allocation recommendations.

[0198] First, we collect raw data from multiple heterogeneous financial data sources. This data includes structured historical transaction flows, credit reports, and fund net asset value records, as well as unstructured content such as financial news, analyst reports, and policy announcements. Using the data distillation module, we unify these disparate data, stripping away noise while retaining core information features strongly related to client investment intentions, high-frequency trading signals, or market warning signals.

[0199] To enhance the model's decision-making transparency and compliance, each extracted core information feature undergoes a logical analysis of the decision-making process. For example, identifying a "RRR cut" policy from news reports involves breaking down the decision path into "observing policy signals - analyzing monetary environment changes - evaluating interest-sensitive assets - and providing a recommendation to increase holdings," forming an explanatory chain. Each link is annotated with its corresponding information feature to form a thought chain. A consistency check module then evaluates the integrity of the logical chain, selecting logically closed samples to generate a structured dataset.

[0200] Next, a financial language model based on the Transformer architecture is selected and loaded with its weights, pre-trained on financial corpus. The structured dataset is divided into a training set and a validation set, and the model is fine-tuned through supervision. The system uses the training set to update the model parameters and evaluates its predictive ability on the validation set until convergence to the fine-tuned model.

[0201] To enhance the model's ability to understand complex data sources, we further constructed a multimodal input pathway. The text feature extraction module encodes natural language data such as analyst reports and news into vectors. The non-text feature extraction module processes non-linguistic data such as candlestick chart patterns and user trading behavior vectors. These two types of features are aligned and fused at the vector level to form a unified fusion feature representation.

[0202] Based on this foundation, a simulation environment is constructed to simulate investor behavior and market feedback. Fusion features are used as state inputs, and slots in the state observation space are configured for volatility, capital flows, risk appetite, and other factors, encoding the dynamic state of multiple market factors. Through this simulation environment, the system evaluates the performance of the model's output asset allocation recommendations under different market conditions.

[0203] To further optimize model performance and compliance, a dual reward mechanism was implemented within the simulation environment. This included format rewards, such as whether the model output met regulatory requirements for a structured recommendation template; and accuracy rewards, such as whether the model's recommendations resulted in returns that outperformed the benchmark index. The policy optimization module received input from the state space and dual reward signals, executing policy gradient iterative optimization to continuously adjust model parameters and generate the optimized model.

[0204] Considering the ever-changing financial market environment, the system deploys a real-time data monitoring module connected to the real-time market data interface, news push platform, and user interaction recording system. This module continuously acquires data on environmental changes, such as interest rate fluctuations, index gaps, and sudden increases in trading volume. Using trend recognition algorithms, it analyzes potential fluctuations and determines whether dynamic adjustments to model parameters are necessary.

[0205] Once the change pattern meets the reconfiguration conditions, the system generates parameter adjustments based on the analysis results and injects them into the current optimized model, updating its parameter configuration and generating an adaptive decision-making model that is adaptable to the current market. This model can continue to maintain effective reasoning in the new environment state.

[0206] Ultimately, the investment advisory system receives the client's investment request data through an adaptive decision-making model, including asset type preferences, tolerable risk levels, and current holdings, and generates real-time asset allocation recommendations, such as "it is recommended to convert 25% of positions from money market funds to short-term credit bond ETFs." It can also explain the recommendation generation path to meet compliance disclosure requirements.

[0207] The system supports high-frequency fluctuation perception, multi-dimensional data fusion, regulatory-oriented explainable recommendations and adaptive adjustments in financial technology scenarios, significantly improving the intelligence level of customer service and the strategic robustness of financial product configuration.

[0208] In the medical and health business, in order to achieve differentiated treatment strategy recommendations and dynamic disease management for individual patients, a medical decision-making support system that integrates multimodal health data and has adaptive capabilities can be built to assist doctors in improving the accuracy of disease identification, the targeted nature of intervention recommendations, and the timeliness of plan updates.

[0209] The system first extracts core information features from multi-source heterogeneous data, including electronic medical records (EMRs), medical imaging data, gene sequencing reports, data collected by vital sign monitoring devices, patient follow-up records, and publicly available medical knowledge bases. These data vary significantly in structure, standards, accuracy, and format. Using a data distillation module, the system eliminates redundant and conflicting information, retaining core information features closely related to disease identification, such as the patient's disease progression, symptom presentation, and treatment response.

[0210] To ensure the model's interpretability and meet clinical review requirements, the system transforms these core information features into a structured decision chain. For example, an abnormal signal in a lung CT image can be broken down into: image recognition → density analysis → localization of suspected lesions → matching historical pneumonia cases → generation of anti-infection treatment recommendations. The system extracts these decision paths through a logical parsing module and performs structured chain annotation to ensure that each recommendation can be traced back to the underlying data. Once the annotation is complete, a consistency check module verifies the logical chain's coherence, screens for reliable annotations, and ultimately generates a structured medical dataset.

[0211] Next, the system selects and loads a language model architecture pre-trained on Chinese medical data (such as MedGPT or a medical BERT variant), partitions the structured dataset into training and validation subsets, and performs a supervised fine-tuning process. During the training phase, the model learns the mapping rules for generating treatment recommendations from feature chains. During the validation phase, the recommendations are evaluated for consistency with real-world case plans, determining whether the target accuracy has been achieved and generating a fine-tuned model.

[0212] Taking into account the modal diversity of medical data, the system further integrates at least two data types: text (such as medical records), images (such as MRI scans), and time series (such as electrocardiograms). A language model extracts semantic vectors from text data, a CNN model extracts lesion regional features from image data, and a recurrent neural network encodes time series data. The system combines these feature vectors to form a joint representation of the patient's multidimensional diagnostic and treatment data, achieving cross-modal information integration and improving the model's understanding of complex cases.

[0213] Next, a simulated clinical environment is constructed, containing a state observation space with multiple slots, such as the patient's current pathological status, drug sensitivity indicators, and laboratory indicator trends. The fused features are loaded into the corresponding slots in this space, enabling the system to simulate decisions based on the current medical status. This simulation environment can be used to simulate medical reasoning tasks such as treatment outcome assessment and plan risk prediction.

[0214] A dual reward mechanism is introduced within the simulation environment to guide model optimization. On the one hand, a format reward is defined to evaluate whether the model's recommendations conform to medical documentation standards and treatment recommendation expression standards; on the other hand, an accuracy reward is defined to assess whether the recommendations are consistent with expert annotations and have achieved good results in historical case studies. Through the policy optimization module, the system continuously performs policy gradient updates within the simulation environment based on state input and reward feedback, optimizing model parameters and generating a more effective optimized model.

[0215] Considering that patients' conditions can change at any time, the system deploys a real-time data monitoring module, connecting to the hospital information system (HIS), wearable devices, smart beds, remote follow-up systems, and other channels. This module collects dynamic health status data such as blood pressure, heart rate, respiration, and pain scores in real time, and combines it with external medical warning information to analyze potential deterioration trends.

[0216] The system analyzes the change patterns and determines whether to trigger a model adjustment process. If the original model parameters are no longer appropriate for the current state, such as when acute fluctuations in the patient's condition or an abnormal medication response are detected, the system uses a parameter adjustment mechanism to readjust the model parameters and generate a new model that adapts in real time—an adaptive decision-making model—to maintain sensitivity to changes in the patient's condition.

[0217] Ultimately, doctors can use this adaptive decision model to input a specific patient's diagnosis and treatment background and current data, and the system will return personalized treatment recommendations, such as "recommended adjustment of aspirin dosage to 75mg / day due to observed elevated liver enzymes and prolonged clotting time." This response process also features a clear chain of reasoning, supporting medical documentation and multidisciplinary consultations, enhancing the controllability and credibility of the medical process.

[0218] The system can be widely applied to hospital decision-making support platforms, telemedicine service platforms and chronic disease management systems. While ensuring the scientific nature and transparency of clinical decision-making, it can significantly improve the intelligence level of individualized diagnosis and treatment and the efficiency of medical services.

[0219] This embodiment implements the model's rapid adaptability to external dynamic factors by deploying a real-time data monitoring module to continuously track state changes in the model's operating environment, and combines change pattern analysis with parameter reconfiguration operations. This avoids the problem of delayed response of traditional static models to environmental changes, and improves the model's stability and decision-making robustness in highly volatile or complex business scenarios. In particular, when user behavior switches frequently or input data patterns fluctuate dramatically, the system can dynamically adjust model parameters to generate new adaptive models, effectively enhancing the model's practicality and continuous reasoning capabilities.

[0220] In one embodiment, a method for enhancing decision-making based on thought chain annotation is provided. The method for enhancing decision-making based on thought chain annotation corresponds to the method for enhancing decision-making based on thought chain annotation in the above embodiment. Figure 3 , Figure 3 This is a schematic diagram of the functional modules of a preferred embodiment of the enhanced decision-making device based on thought chain annotation according to the present invention. These modules include a thought chain annotation module 10, a supervised fine-tuning module 20, a multimodal fusion module 30, a simulation environment construction module 40, a strategy optimization module 50, a dynamic adaptation module 60, and a decision generation module 70. Each functional module is described in detail below:

[0221] The thought chain annotation module 10 is used to extract core information features from multi-source heterogeneous data and perform thought chain annotation on the core information features to generate a structured data set;

[0222] A supervised fine-tuning module 20 is configured to select a basic language model, load pre-trained weights, and perform supervised fine-tuning using the structured dataset to generate a fine-tuned model;

[0223] A multimodal fusion module 30 is configured to integrate at least two data types to generate multimodal input data, and to perform a multimodal fusion strategy on the multimodal input data to generate fusion features;

[0224] A simulation environment construction module 40 is used to construct a simulation environment including a state observation space, wherein the state observation space receives the fused feature as a state input;

[0225] a policy optimization module 50 for designing a dual reward mechanism including a format reward and an accuracy reward, determining a reward signal based on the dual reward mechanism, and iteratively updating the parameters of the fine-tuned model according to the reward signal in the simulation environment through the policy optimization module to generate an optimized model;

[0226] A dynamic adaptation module 60 is configured to deploy a real-time data monitoring module to obtain environmental state change data, and dynamically adjust the parameter configuration of the optimized model according to the environmental state change data to generate an adaptive decision model;

[0227] The decision generating module 70 is configured to receive decision request data through the adaptive decision model and generate a decision response result.

[0228] In one embodiment, the thought chain annotation module 10 is specifically used to:

[0229] Collecting raw data from at least two independent heterogeneous data sources;

[0230] The raw data is subjected to information condensation processing by a data distillation module to extract core information features related to the target task;

[0231] Decomposing the decision-making process of the core information features into an interpretable sequence of steps including data observation, domain analysis, factor evaluation and recommendation output through a logic parsing module;

[0232] Associating the explainable step sequence with the corresponding core information features to form a thought chain annotation;

[0233] Verifying the logical coherence of the thought chain annotation through a consistency verification module, and generating a verification result including a pass status or a fail status;

[0234] Screening thought chain annotations that meet the logical coherence standard according to the verification results, and generating valid annotation data based on the screened thought chain annotations;

[0235] The valid annotated data are integrated according to a preset data structure to generate a structured data set.

[0236] In one embodiment, the supervisory fine-tuning module 20 is specifically configured to:

[0237] Select a basic language model architecture with natural language understanding capabilities and load pre-trained weights into the parameter space of the basic language model;

[0238] Dividing the structured dataset into a training subset and a validation subset;

[0239] Performing supervised fine-tuning training on a base language model loaded with pre-trained weights based on the training subset, and determining a loss value between the model's predicted output and the labeled data;

[0240] Updating the parameters of the basic language model through a back-propagation module according to the loss value;

[0241] Analyzing model performance indicators of the updated basic language model based on the validation subset;

[0242] When the model performance index reaches a preset index threshold, a fine-tuned model is generated.

[0243] In one embodiment, the multimodal fusion module 30 is specifically configured to:

[0244] Acquire original multi-source data including at least text data types and non-text data types;

[0245] Processing the text data type through a text feature extraction module to generate a text feature vector;

[0246] Processing the non-text data type by a non-text feature extraction module to generate a non-text feature vector;

[0247] Performing data integration on the text feature vector and the non-text feature vector to generate multimodal input data;

[0248] A cross-modal fusion strategy is performed on the multimodal input data through a multimodal fusion module, and the fusion result is output as a dimensionally aligned fusion feature.

[0249] In one embodiment, the simulation environment construction module 40 is specifically configured to:

[0250] Define a state observation space containing dynamic state indicator slots;

[0251] configuring a dedicated fusion feature slot in the state observation space;

[0252] loading the fusion feature into the dedicated fusion feature slot;

[0253] Initializing other dynamic state indicator slots in the state observation space;

[0254] Based on the state observation space in which the fusion feature is loaded and other dynamic state indicator slots are initialized, a simulation environment is constructed.

[0255] In one embodiment, the policy optimization module 50 is specifically configured to:

[0256] Establishing format scoring criteria for output structure compliance verification;

[0257] Establish accuracy scoring criteria for decision validity verification;

[0258] Determining the compliance score of the output action of the fine-tuned model in the simulation environment according to the format scoring standard, and generating a format reward;

[0259] Determining the effectiveness score of the decision-making results of the fine-tuned model in the simulation environment according to the accuracy scoring criteria, and generating an accuracy reward;

[0260] combining the format reward and the accuracy reward to generate a reward signal;

[0261] receiving, via a policy optimization module, state data and a reward signal in a state-observation space within the simulation environment;

[0262] determining a policy gradient based on the state data and the reward signal;

[0263] The parameters of the fine-tuned model are updated according to the policy gradient until a preset iteration termination condition is met to generate an optimized model.

[0264] In one embodiment, the dynamic adaptation module 60 is specifically configured to:

[0265] Configuring a real-time data acquisition interface, and establishing a real-time data monitoring module based on the real-time data acquisition interface;

[0266] Continuously capture environmental status change data through the real-time data monitoring module;

[0267] Performing a change pattern analysis operation on the environmental state change data to generate a change pattern analysis result;

[0268] Determining model parameter adjustment requirements based on the change pattern analysis results;

[0269] Determining a parameter adjustment amount according to the parameter adjustment requirement;

[0270] The parameter configuration of the optimized model is updated based on the parameter adjustment amount to generate an adaptive decision model.

[0271] In one embodiment, a computer device is provided. The computer device may be a server, and its internal structure diagram may be as follows: Figure 4 As shown. The computer device includes a processor, a memory, a network interface and a database connected via a system bus. The processor of the computer device is used to provide determination and control capabilities. The memory of the computer device includes a non-volatile and / or volatile storage medium and an internal memory. The non-volatile storage medium stores an operating system, a computer program and a database. The internal memory provides an environment for the operation of the operating system and the computer program in the non-volatile storage medium. The network interface of the computer device is used to communicate with an external user terminal via a network connection. When the computer program is executed by the processor, it realizes the functions or steps on the service side of an enhanced decision-making method based on thought chain annotation.

[0272] In one embodiment, a computer device is provided. The computer device may be a user terminal, and its internal structure diagram may be as follows: Figure 5 As shown. The computer device includes a processor, a memory, a network interface, a display screen and an input device connected via a system bus. Among them, the processor of the computer device is used to provide determination and control capabilities. The memory of the computer device includes a non-volatile storage medium and an internal memory. The non-volatile storage medium stores an operating system and a computer program. The internal memory provides an environment for the operation of the operating system and the computer program in the non-volatile storage medium. The network interface of the computer device is used to communicate with an external server via a network connection. When the computer program is executed by the processor, it realizes the functions or steps on the user side of an enhanced decision-making method based on thought chain annotation

[0273] In one embodiment, a computer device is provided, including a memory, a processor, and a computer program stored in the memory and executable on the processor. When the processor executes the computer program, the following steps are performed:

[0274] Extract core information features from multi-source heterogeneous data, and annotate the core information features with thought chains to generate a structured data set;

[0275] Select a base language model to load pre-trained weights, and use the structured dataset to perform supervised fine-tuning to generate a fine-tuned model;

[0276] Integrating at least two data types to generate multimodal input data, and performing a multimodal fusion strategy on the multimodal input data to generate fusion features;

[0277] Constructing a simulation environment including a state observation space, wherein the state observation space receives the fused features as state input;

[0278] Designing a dual reward mechanism including a format reward and an accuracy reward, determining a reward signal based on the dual reward mechanism, and iteratively updating the parameters of the fine-tuned model according to the reward signal in the simulation environment using a policy optimization module to generate an optimized model;

[0279] Deploy a real-time data monitoring module to obtain environmental state change data, and dynamically adjust the parameter configuration of the optimized model according to the environmental state change data to generate an adaptive decision model;

[0280] Decision request data is received through the adaptive decision model, and a decision response result is generated.

[0281] In one embodiment, a computer-readable storage medium is provided, on which a computer program is stored. When the computer program is executed by a processor, the following steps are implemented:

[0282] Extract core information features from multi-source heterogeneous data, and annotate the core information features with thought chains to generate a structured data set;

[0283] Select a base language model to load pre-trained weights, and use the structured dataset to perform supervised fine-tuning to generate a fine-tuned model;

[0284] Integrating at least two data types to generate multimodal input data, and performing a multimodal fusion strategy on the multimodal input data to generate fusion features;

[0285] Constructing a simulation environment including a state observation space, wherein the state observation space receives the fused features as state input;

[0286] Designing a dual reward mechanism including a format reward and an accuracy reward, determining a reward signal based on the dual reward mechanism, and iteratively updating the parameters of the fine-tuned model according to the reward signal in the simulation environment using a policy optimization module to generate an optimized model;

[0287] Deploy a real-time data monitoring module to obtain environmental state change data, and dynamically adjust the parameter configuration of the optimized model according to the environmental state change data to generate an adaptive decision model;

[0288] Decision request data is received through the adaptive decision model, and a decision response result is generated.

[0289] It should be noted that the above functions or steps that can be implemented by the computer-readable storage medium or computer device can be found in the relevant descriptions of the server side and the user side in the aforementioned method embodiment. To avoid repetition, they will not be described one by one here.

[0290] Those skilled in the art will appreciate that all or part of the processes in the above-mentioned embodiments can be implemented by instructing the relevant hardware through a computer program. The computer program can be stored in a non-volatile computer-readable storage medium. When the computer program is executed, it can include the processes of the embodiments of the above-mentioned methods. Among them, any reference to memory, storage, database or other media used in the embodiments provided in this application can include non-volatile and / or volatile memory. Non-volatile memory can include read-only memory (ROM), programmable ROM (PROM), electrically programmable ROM (EPROM), electrically erasable programmable ROM (EEPROM) or flash memory. Volatile memory can include random access memory (RAM) or external cache memory. By way of illustration and not limitation, RAM is available in various forms, such as static RAM (SRAM), dynamic RAM (DRAM), synchronous DRAM (SDRAM), double data rate SDRAM (DDRSDRAM), enhanced SDRAM (ESDRAM), synchronous link (Synchlink) DRAM (SLDRAM), memory bus (Rambus) direct RAM (RDRAM), direct memory bus dynamic RAM (DRDRAM), and memory bus dynamic RAM (RDRAM).

[0291] Those skilled in the art will clearly understand that for the sake of convenience and brevity of description, only the division of the above-mentioned functional units and modules is used as an example. In actual applications, the above-mentioned functions can be distributed and completed by different functional units and modules as needed, that is, the internal structure of the device can be divided into different functional units or modules to complete all or part of the functions described above.

[0292] It should be noted that if any software tools or components other than those of the Company appear in the embodiments of this application, they are merely for illustration and do not represent actual use. The above embodiments are intended only to illustrate the technical solutions of the present invention, not to limit them. Although the present invention has been described in detail with reference to the above embodiments, those skilled in the art should understand that they can still modify the technical solutions described in the above embodiments, or replace some of the technical features therein with equivalents. These modifications or replacements do not deviate the essence of the corresponding technical solutions from the spirit and scope of the technical solutions of the embodiments of the present invention, and should all be included in the scope of protection of the present invention.

Claims

1. An enhanced decision-making method based on thought chain annotation, characterized in that: The following steps are involved: Extract core information features from multi-source heterogeneous data, and annotate the core information features with thought chains to generate a structured data set; Select a base language model to load pre-trained weights, and use the structured dataset to perform supervised fine-tuning to generate a fine-tuned model; Integrating at least two data types to generate multimodal input data, and performing a multimodal fusion strategy on the multimodal input data to generate fusion features; Constructing a simulation environment including a state observation space, wherein the state observation space receives the fused features as state input; Designing a dual reward mechanism including a format reward and an accuracy reward, determining a reward signal based on the dual reward mechanism, and iteratively updating the parameters of the fine-tuned model according to the reward signal in the simulation environment using a policy optimization module to generate an optimized model; Deploy a real-time data monitoring module to obtain environmental state change data, and dynamically adjust the parameter configuration of the optimized model according to the environmental state change data to generate an adaptive decision model; Decision request data is received through the adaptive decision model, and a decision response result is generated.

2. The enhanced decision-making method based on thought chain annotation according to claim 1 is characterized in that: Extract core information features from multi-source heterogeneous data and annotate the core information features with thought chains to generate a structured data set, including: Collecting raw data from at least two independent heterogeneous data sources; The raw data is subjected to information condensation processing by a data distillation module to extract core information features related to the target task; Decomposing the decision-making process of the core information features into an interpretable sequence of steps including data observation, domain analysis, factor evaluation and recommendation output through a logic parsing module; Associating the explainable step sequence with the corresponding core information features to form a thought chain annotation; Verifying the logical coherence of the thought chain annotation through a consistency verification module, and generating a verification result including a pass status or a fail status; Screening thought chain annotations that meet the logical coherence standard according to the verification results, and generating valid annotation data based on the screened thought chain annotations; The valid annotated data are integrated according to a preset data structure to generate a structured data set.

3. The enhanced decision-making method based on thought chain annotation according to claim 1 is characterized in that: Select a base language model to load pre-trained weights, and use the structured dataset to perform supervised fine-tuning to generate a fine-tuned model, including: Select a basic language model architecture with natural language understanding capabilities and load pre-trained weights into the parameter space of the basic language model; Dividing the structured dataset into a training subset and a validation subset; Performing supervised fine-tuning training on a base language model loaded with pre-trained weights based on the training subset, and determining a loss value between the model's predicted output and the labeled data; Updating the parameters of the basic language model through a back-propagation module according to the loss value; Analyzing model performance indicators of the updated basic language model based on the validation subset; When the model performance index reaches a preset index threshold, a fine-tuned model is generated.

4. The enhanced decision-making method based on thought chain annotation according to claim 1 is characterized in that: Integrating at least two data types to generate multimodal input data, and performing a multimodal fusion strategy on the multimodal input data to generate fusion features, including: Acquire original multi-source data including at least text data types and non-text data types; Processing the text data type through a text feature extraction module to generate a text feature vector; Processing the non-text data type by a non-text feature extraction module to generate a non-text feature vector; Performing data integration on the text feature vector and the non-text feature vector to generate multimodal input data; A cross-modal fusion strategy is performed on the multimodal input data through a multimodal fusion module, and the fusion result is output as a dimensionally aligned fusion feature.

5. The enhanced decision-making method based on thought chain annotation according to claim 1 is characterized in that: Constructing a simulation environment including a state observation space, wherein the state observation space receives the fused feature as a state input, including: Define a state observation space containing dynamic state indicator slots; configuring a dedicated fusion feature slot in the state observation space; loading the fusion feature into the dedicated fusion feature slot; Initializing other dynamic state indicator slots in the state observation space; Based on the state observation space in which the fusion feature is loaded and other dynamic state indicator slots are initialized, a simulation environment is constructed.

6. The enhanced decision-making method based on thought chain annotation according to claim 1 is characterized in that: Designing a dual reward mechanism including a format reward and an accuracy reward, determining a reward signal based on the dual reward mechanism, and iteratively updating the parameters of the fine-tuned model according to the reward signal in the simulation environment using a policy optimization module to generate an optimized model, including: Establishing format scoring criteria for output structure compliance verification; Establish accuracy scoring criteria for decision validity verification; Determining the compliance score of the output action of the fine-tuned model in the simulation environment according to the format scoring standard, and generating a format reward; Determining the effectiveness score of the decision-making results of the fine-tuned model in the simulation environment according to the accuracy scoring criteria, and generating an accuracy reward; combining the format reward and the accuracy reward to generate a reward signal; receiving, via a policy optimization module, state data and a reward signal in a state-observation space within the simulation environment; determining a policy gradient based on the state data and the reward signal; The parameters of the fine-tuned model are updated according to the policy gradient until a preset iteration termination condition is met to generate an optimized model.

7. The enhanced decision-making method based on thought chain annotation according to claim 1 is characterized in that: Deploy a real-time data monitoring module to obtain environmental state change data, and dynamically adjust the parameter configuration of the optimized model according to the environmental state change data to generate an adaptive decision model, including: Configuring a real-time data acquisition interface, and establishing a real-time data monitoring module based on the real-time data acquisition interface; Continuously capture environmental status change data through the real-time data monitoring module; Performing a change pattern analysis operation on the environmental state change data to generate a change pattern analysis result; Determining model parameter adjustment requirements based on the change pattern analysis results; Determining a parameter adjustment amount according to the parameter adjustment requirement; The parameter configuration of the optimized model is updated based on the parameter adjustment amount to generate an adaptive decision model.

8. An enhanced decision-making device based on thought chain annotation, characterized in that: The enhanced decision-making device based on thought chain annotation includes: The thought chain annotation module is used to extract core information features from multi-source heterogeneous data and perform thought chain annotation on the core information features to generate a structured data set; A supervised fine-tuning module is used to select a basic language model, load pre-trained weights, and perform supervised fine-tuning using the structured dataset to generate a fine-tuned model; A multimodal fusion module, configured to integrate at least two data types to generate multimodal input data, and perform a multimodal fusion strategy on the multimodal input data to generate fusion features; A simulation environment construction module, configured to construct a simulation environment including a state observation space, wherein the state observation space receives the fused feature as a state input; a policy optimization module, configured to design a dual reward mechanism including a format reward and an accuracy reward, determine a reward signal based on the dual reward mechanism, and iteratively update the parameters of the fine-tuned model according to the reward signal in the simulation environment through the policy optimization module to generate an optimized model; A dynamic adaptation module is used to deploy a real-time data monitoring module to obtain environmental state change data, and dynamically adjust the parameter configuration of the optimized model according to the environmental state change data to generate an adaptive decision model; The decision generating module is used to receive decision request data through the adaptive decision model and generate a decision response result.

9. A computer device, characterized in that: The computer device includes a memory, a processor, and an enhanced decision-making program based on thought chain annotation stored in the memory and capable of running on the processor. When the enhanced decision-making program based on thought chain annotation is executed by the processor, the steps of the enhanced decision-making method based on thought chain annotation as described in any one of claims 1 to 7 are implemented.

10. A computer-readable storage medium, characterized in that The storage medium stores an enhanced decision-making program based on thought chain annotation. When the enhanced decision-making program based on thought chain annotation is executed by the processor, the steps of the enhanced decision-making method based on thought chain annotation as described in any one of claims 1 to 7 are implemented.

Images