A data-parameter coupled orthogonal continual learning method

Abstract

The application discloses a data-parameter coupling orthogonal continual learning method, relates to the technical field of large language models, and comprises the following steps: step one, formally defining a continuous learning scene and establishing a unified symbolic system; step two, using PSGM to perform soft adjustment on a training signal according to perplexity at a sample level, and reducing the negative influence of high-risk samples on training stability; and step three, using TV-SPT to dynamically extract a historical sensitive direction from a weighted feature at a parameter level, and performing orthogonal truncation on a current gradient according to the historical sensitive direction. The application adopts the data-parameter coupling orthogonal continual learning method, designs a perplexity-aware soft gradient adjustment strategy PSGM, adaptively evaluates and processes data, introduces a time-varying latent concept subspace tracking TV-SPT mechanism, explicitly injects a soft modulation weight of a data end into a feature matrix, and realizes seamless linkage between data features and a parameter space.

Description

Technical Field

[0001] This invention relates to the field of large language model technology, and in particular to an orthogonal continuous learning method that couples data and parameters. Background Technology

[0002] With the rapid development of deep learning technology, large language models based on the Transformer architecture have become the core driving force for next-generation artificial intelligence applications. These models, pre-trained on massive and diverse datasets, possess extremely rich knowledge, logical reasoning rules, and language generation models. However, how to utilize and train these models to adapt to the dynamic needs of the real world remains a key challenge. While fine-tuning can improve the performance of models on specific domain tasks, it also brings a problem—catastrophic forgetting (CF). Catastrophic forgetting refers to the phenomenon where a model loses or overwrites knowledge learned during pre-training during fine-tuning. This phenomenon severely limits the generalization ability and knowledge accumulation of LLMs. Models cannot integrate new and old knowledge like humans, making it difficult to build a continuously growing knowledge system, and also affecting the application of models in vertical fields.

[0003] To address this issue, researchers have conducted extensive research over the past two years. Some studies have attempted to circumvent forgetting using Parameter-Efficient Fine-Tuning (PEFT) techniques (such as LoRA). Intuitively, PEFT updates only a small portion of the parameters and freezes the backbone network, which should protect the model's previously learned knowledge. However, recent evaluations have invariably pointed out that when faced with multi-stage continuous fine-tuning, the PEFT method alone cannot solve the problem of knowledge degradation, leading to catastrophic forgetting similar to that of full-parameter fine-tuning. On the other hand, traditional regularization strategies or memory pools based on activation replay encounter insurmountable computational bottlenecks when migrating to LLMs with hundreds of billions of parameters. Summary of the Invention

[0004] The purpose of this invention is to provide an orthogonal continuous learning method with data-parameter coupling. A perplexity-aware soft gradient adjustment PSGM strategy is designed to adaptively evaluate and process data. A time-varying latent concept subspace tracking (TV-SPT) mechanism is introduced to explicitly inject soft modulation weights from the data side into the feature matrix. The two are deeply coupled, achieving seamless linkage between data features and parameter space.

[0005] To achieve the above objectives, this invention provides a data-parameter coupled orthogonal continuous learning method, comprising the following steps: Step 1: Pre-construction unified initialization, formally defining the continuous learning scenario and establishing a unified symbol system; Step 2: PSGM sample-level perplexity adaptive soft weighting. At the sample level, PSGM is used to softly adjust the training signal based on perplexity to reduce the negative impact of high-risk samples on training stability. Step 3: TV-SPT time-varying latent concept subspace tracking and gradient orthogonal truncation. At the parameter level, TV-SPT is used to dynamically extract historical sensitive directions from weighted features, and orthogonal truncation is performed on the current gradient accordingly.

[0006] Preferably, in step one, a continuous learning sequence consisting of T tasks is defined: in, Indicates the first t A dataset for each task. and These are the input and the target output, respectively. The number of samples for this task, in the th... t During each training phase, the model can only access the current task. Data, historical task data Inaccessible.

[0007] Preferably, in step one, the frozen backbone large language model is initialized. Open the low-rank adapter as an updatable parameter; initialize the TV-SPT dynamic feature dictionary, online exponentiation iteration parameters, EMA momentum coefficient, similarity threshold, and PSGM hyperparameter.

[0008] Preferably, in step two, input the current batch of samples and perform the following steps: S21. For each sample within the batch ( ) Calculate the sample perplexity: in, The length of the target sequence. surface The model is shown for the first j The conditional probability of each target word; S22. Calculate the confusion distribution of this batch and the batch mean. Standard deviation and adaptive threshold , in, This is the adjustment coefficient; S23. Substitute the Sigmoid soft adjustment function to solve for the weight of each sample. , ; in, For the Sigmoid function, Temperature coefficient; S24. Output the batch sample weight set, which is used simultaneously for loss weighting calculation and TV-SPT weighted feature covariance construction.

[0009] Preferably, step three includes the following steps: S31. Construct the weighted feature covariance matrix; S32. Online power method for extracting the dominant feature direction of the current batch; S33. Update the dynamic feature dictionary and store historical sensitive atoms; S34. Gradient orthogonal truncation removes update components that disrupt historical knowledge.

[0010] Preferably, the weighted covariance matrix in S31 is: in, This indicates that the Lth layer of the model is in t Activation matrices over each training step , b This represents the number of samples in the current training batch. d For the hidden layer dimension, Construct a diagonal weight matrix for the weights. .

[0011] Preferably, a dynamic feature dictionary is maintained in S33. ,in Indicates the size of the dictionary, each A unit feature vector, called an atom, is extracted for the principal direction in the current training step. First, calculate its similarity to each atom in the dictionary; Maximum similarity below a preset threshold The current direction is not yet fully represented by the existing dictionary, and will Added as a new atom to the dictionary; maximum similarity not less than If the current direction is similar to existing functional atoms, then the exponential moving average (EMA) method is used to smoothly update the most similar atoms: in, m The momentum coefficient is represented by Norm(), which indicates the normalization operation.

[0012] Preferably, S34 includes the following steps: S341. Calculate the original gradient of the current training step. ; S342, the original gradient The gradient after orthogonal truncation, projected onto the orthogonal complement of the history-sensitive subspace, is: in, I It is the identity matrix; It is a dictionary matrix; S343, Use Complete the parameter update operation.

[0013] Therefore, this invention employs a data-parameter coupled orthogonal continuous learning method, designing a perplexity-aware soft gradient adjustment mechanism (PSGM). This breaks away from the traditional hard mask truncation logic, transforming sample perplexity into adaptive continuous gradient weights. This strategy suppresses high-noise signals while preserving the weak gradient flow of complex long-tail knowledge without loss. More importantly, it explicitly injects perplexity weights into downstream subspace updates, establishing a closed loop of data-parameter linkage. A time-varying latent concept subspace tracking mechanism (TV-SPT) is introduced: abandoning computationally expensive static SVD, a core feature dictionary is dynamically constructed in the weighted feature space purified by PSGM using online power-law iteration and exponential moving average (EMA). This mechanism enables the orthogonal projection matrix to accurately track and dynamically evolve with fine-tuning tasks, fundamentally overcoming the functional subspace drift problem in continuous learning.

[0014] The technical solution of the present invention will be further described in detail below with reference to the accompanying drawings and embodiments. Attached Figure Description

[0015] Figure 1 This is a graph showing the evaluation results of the Qwen2.5 algorithm on the MMLU dataset before training with different parameter values. Figure 2 This is a graph showing the change in MMLU accuracy of the model of the present invention before and after fine-tuning by various methods; Figure 3 This is a graph showing the performance of the model after fine-tuning in this invention on the SuperNI dataset. Detailed Implementation

[0016] The technical solution of the present invention will be further described below with reference to the accompanying drawings and embodiments.

[0017] Unless otherwise defined, the technical or scientific terms used in this invention shall have the ordinary meaning understood by one of ordinary skill in the art to which this invention pertains. The terms "first," "second," and similar terms used in this invention do not indicate any order, quantity, or importance, but are merely used to distinguish different components. Terms such as "comprising" or "including" mean that the element or object preceding the word encompasses the elements or objects listed following the word and their equivalents, without excluding other elements or objects. Terms such as "connected" or "linked" are not limited to physical or mechanical connections, but can include electrical connections, whether direct or indirect. Terms such as "upper," "lower," "left," and "right" are used only to indicate relative positional relationships; when the absolute position of the described object changes, the relative positional relationship may also change accordingly.

[0018] Example Please see Figures 1-3 This invention provides a data-parameter coupled orthogonal continuous learning method, comprising the following steps: Step 1: Preliminary unified initialization. First, the continuous learning scenario is formally defined, and a unified symbol system is established.

[0019] Consider a continuous learning sequence consisting of T tasks: in Let represent the dataset for the t-th task. and These are the input and the target output, respectively. This represents the number of samples for this task. In the... t During each training phase, the model can only access the current task. Data, historical task data Inaccessible.

[0020] Let the large language model be ,in This represents all trainable parameters of the model. For the parameter dimension, an efficient fine-tuning method is adopted, updating only the low-rank adapter parameters while keeping the trunk parameters frozen.

[0021] Step 2: Perplexity-aware soft gradient adjustment mechanism (PSGM). At the sample level, PSGM is used to softly adjust the training signal based on perplexity to reduce the negative impact of high-risk samples on training stability.

[0022] The mathematical definition of perplexity.

[0023] In continuous fine-tuning scenarios, different samples have varying impacts on model parameter updates. Some samples may exhibit high perplexity due to high noise, significant distribution shifts, or the current model not yet forming stable predictions. If these samples are trained on the same scale as ordinary samples, they are prone to generating overly strong updates in local stages, thereby exacerbating the perturbation of historical knowledge. Therefore, this paper first introduces sample-level perplexity to measure the model's prediction uncertainty for the current sample.

[0024] For sample ( Its perplexity is defined as follows: in, The length of the target sequence. The model represents the first j The conditional probability of each target word. Higher perplexity indicates greater uncertainty in the model's prediction of that sample. However, high-perplexity samples are not necessarily completely ineffective; they may contain complex knowledge valuable for new tasks. Therefore, this paper does not employ a direct "hard truncation" approach but instead designs a continuously differentiable soft adjustment function.

[0025] Adaptive soft adjustment function.

[0026] For a size of b The training batches, with sample perplexities of respectively This paper maps it to sample weights.

[0027] in, For the Sigmoid function, For temperature coefficient, This is the adaptive threshold for the current training phase. and Let these represent the mean and standard deviation of the confusion level for the current batch, respectively. This is the adjustment coefficient.

[0028] The above design has two implications. First, weights. The threshold is continuously adjusted according to the perplexity. High-perplexity samples are not directly discarded, but only suppressed in terms of update intensity, thus avoiding information loss caused by hard filtering. Secondly, the threshold... The statistical distribution of the current batch is dynamically determined, enabling the model to adaptively distinguish sample types based on the training state, rather than relying on fixed empirical thresholds.

[0029] Step 3: Time-varying latent concept subspace tracking (TV-SPT). At the parameter level, TV-SPT is used to dynamically extract historical sensitive directions from weighted features, and orthogonal truncation is performed on the current gradient accordingly.

[0030] S31, Feature Extraction and Weighting.

[0031] set up This indicates that the Lth layer of the model is in t The activation matrix over each training step, where b This indicates the number of samples in the current training batch. d For the hidden layer dimension, each row of the matrix corresponds to the feature representation of a sample.

[0032] To achieve collaborative modeling between the data and parameters, this paper explicitly injects the sample weights output by PSGM into the feature statistics process to construct a weighted covariance matrix: This design means that the weights assigned to samples at the data end not only affect the loss calculation, but also directly affect the subspace estimation results at the parameter end.

[0033] S32, Main direction extraction and time-varying tracking.

[0034] Obtaining the weighted covariance matrix Subsequently, this paper employs an online exponential method to iteratively extract the dominant feature orientation. The reason for using an online approach, rather than performing full feature decomposition at each training stage, is primarily due to considerations of computational efficiency and feasibility in continuous learning scenarios. The online exponential method can stably extract the dominant orientation in the current feature space with lower overhead, making it more suitable for integration with continuous fine-tuning. In this process, the goal of TV-SPT is not to obtain a fixed historical guarded subspace all at once, but rather to continuously update the dominant orientation representation based on the current weighted features at each training stage. This is because as the task progresses, the internal feature structure of the model will drift, and correspondingly, the functional orientations related to historical knowledge will also change. Through this dynamic tracking mechanism, TV-SPT enables the constructed guarded subspace to adapt to the representation evolution during continuous learning, thereby effectively overcoming the problem of static subspace methods gradually becoming ineffective during long-term training.

[0035] S33. Construction and updating of dynamic feature dictionary.

[0036] To preserve this principal orientation information that evolves with training, this paper maintains a dynamic feature dictionary. ,in Indicates the size of the dictionary, each A unit eigenvector is called an atom. The role of the atom is to characterize the historical functional directions gradually accumulated during continuous learning.

[0037] For the main direction extracted in the current training step First, calculate its similarity to each atom in the dictionary. If the maximum similarity is lower than a preset threshold... If the current direction is not fully represented by the existing dictionary, then... Add it as a new atom to the dictionary; if the maximum similarity is not less than If the current direction is close to the existing functional atoms, then the exponential moving average (EMA) method is used to smoothly update the most similar atoms.

[0038] in, m The momentum coefficient is represented by Norm(), which indicates the normalization operation.

[0039] This update strategy balances stability and adaptability. On the one hand, historical atoms do not change drastically due to random fluctuations in individual batches, thus maintaining a stable memory of the subspace's historical structure. On the other hand, as new task information is continuously injected, the dictionary can gradually absorb new directional information to adapt to the dynamic changes in the model's internal representation. To facilitate subsequent projection calculations, this paper represents the normalized dictionary as a matrix. .

[0040] S43, Orthogonal truncation in coupled space.

[0041] After completing the latent concept subspace tracking and dynamic feature dictionary update, this paper further performs orthogonal truncation on the current gradient based on this subspace. Orthogonal truncation is not an additional module independent of TV-SPT, but rather an execution step based on the subspace tracked by TV-SPT.

[0042] Let the original gradient of the current training step be... dictionary matrix Zhang Cheng's subspace can be approximated as the space of historical functional directions that need to be protected in the current stage. Then, the gradient after orthogonal truncation can be expressed as: This formula shows that the current gradient is projected into the orthogonal complement of the history-sensitive subspace. After this processing, components in the gradient that strongly coincide with the direction of historical knowledge are weakened, while update components that are more beneficial to the current task learning and less likely to disrupt existing capabilities are retained. Finally, using... Complete the parameter update operation.

[0043] Experimental verification.

[0044] Dataset.

[0045] We conducted continuous learning experiments on the SuperNI dataset proposed by Wang et al. SuperNI is a large-scale, comprehensive instruction fine-tuning benchmark dataset containing 1616 diverse NLP tasks, covering 76 task types and 55 languages. To simulate typical continuous learning scenarios, we selected five representative task subsets covering question answering (QA), question generation (QG), sentiment analysis (SA), text summarization (Sum.), and machine translation (Trans.). For each task category, we randomly selected 3000 data points for fine-tuning and 600 data points for evaluation. During the experiments, we adopted the default learning order of QA→QG→Sum.→SA→Trans.

[0046] We also chose the MMLU (Massive Multitask Language Understanding) dataset to evaluate the degree of forgetting in large models before and after fine-tuning. The MMLU dataset covers 57 disciplines, spanning multiple fields such as science, engineering, and technology. Therefore, the MMLU dataset is a good tool for evaluating the performance of models before and after fine-tuning and judging whether they have forgotten the extensive knowledge acquired during the pre-training phase.

[0047] Experimental Model: To systematically analyze the impact of model size on the continuous fine-tuning forgetting phenomenon, this study selected the Qwen2.5Base series as the basic model. The model information is shown in Table 4.1: Table 4.1 Overview of the Qwen2.5 series models

[0048] To simultaneously evaluate the model's resistance to forgetting and its ability to learn new tasks, this paper divides the evaluation metrics into two categories: one category is used to characterize the degree of forgetting of the model during continuous fine-tuning, and the other category is used to measure the model's learning performance on the SuperNI task.

[0049] In terms of forgetting assessment, this paper uses Accuracy (Acc.) to evaluate the changes in the MMLU dataset before and after fine-tuning. Let the model's MMLU accuracy before continuous fine-tuning be _____. ,Finish t The accuracy rate after training for each task is Then the general knowledge forgetting rate is defined as: The larger the value, the more severe the model's forgetting; conversely, the smaller the value, the stronger the model's ability to retain existing knowledge. The evaluation results of Qwen2.5 on the MMLU dataset before training with different parameter sets are shown below. Figure 1 As shown.

[0050] For evaluating new task learning, this paper uses ROUGE-L as the evaluation metric for the SuperNI dataset. Since all the fine-tuning datasets selected in this paper can be uniformly represented in the form of "instruction-text output", using ROUGE-L can measure the learning performance of the model under a unified standard.

[0051] To verify the effectiveness of the proposed CDOCL method, the experimental section will compare it with the following baselines: (1) Full-parameter fine-tuning: In this setting, the model updates all parameters at each task stage without introducing any additional parameter constraints, replay mechanisms or structural designs.

[0052] (2) LoRA is one of the most commonly used efficient parameter fine-tuning methods in current large language models. Its basic idea is to freeze the parameters of the pre-trained model and introduce trainable low-rank increment matrices only in some linear transformation layers to achieve task adaptation by adding a small number of parameters.

[0053] (3) PiSSA. Unlike LoRA, PiSSA uses the principal singular components in the pre-trained weight matrix to initialize the low-rank adapter, so that the adapter can better align with the more important structural information in the original model at the beginning of training.

[0054] (4) O-LoRA: The core idea of ​​this method is to restrict the low-rank updates of different tasks to mutually orthogonal subspaces, thereby reducing the conflict between the update direction of new tasks and the update direction of historical tasks.

[0055] (5) RandSel, a typical baseline replay method. Its basic idea is to remind the model to retain relevant knowledge of the old task by replaying a small number of historical samples while learning a new task, thereby reducing knowledge forgetting. This paper sets it to a fixed replay ratio of 5% to achieve a relative balance between training cost and retention of old knowledge.

[0056] Experimental Results and Analysis: To comprehensively evaluate the performance of each method in continuous fine-tuning scenarios, this paper analyzes the results from two dimensions: general knowledge retention ability and new task learning ability. The model's forgetting ability is measured by comparing the change in accuracy on the MMLU dataset before and after continuous fine-tuning; a smaller decrease in accuracy indicates a stronger ability to retain existing general knowledge during continuous learning and better anti-forgetting performance. The model's learning performance on new tasks is evaluated using the ROUGE-L index on the SuperNI dataset to characterize its fitting ability to the current task; a higher index indicates a stronger adaptability to new tasks. Joint analysis using these two indices comprehensively reflects the trade-off between "stability" and "plasticity" of each method, examining both the model's ability to effectively retain existing knowledge and its ability to fully absorb new task information. Overall results show that the proposed CDOCL exhibits the best knowledge retention ability under different parameter scales, while also maintaining a high level in new task learning, demonstrating better overall performance and potential for continuous optimization. Specific experimental results can be found in [link to experimental results]. Figure 2 As shown in Figure 3 and Table 4.2.

[0057] Table 4.2 shows the decrease in accuracy on the MMLU dataset after model fine-tuning.

[0058] The MMLU results show that the proposed CDOCL achieves the smallest forgetting amount across all five parameter models (0.5B, 1.5B, 3B, 7B, and 14B), significantly outperforming all baseline methods overall. This demonstrates that CDOCL can more effectively retain general knowledge and historical capabilities acquired during pre-training while continuously absorbing new task knowledge, validating the effectiveness of the "data-parameter coupling orthogonal protection" mechanism in suppressing catastrophic forgetting.

[0059] Looking at the specific performance of each baseline, full-parameter fine-tuning consistently resulted in the most severe forgetting. This indicates that during continuous fine-tuning, if all parameters are updated without constraints, the model is more likely to overwrite existing representations to adapt to new tasks, thus disrupting the original knowledge structure. In contrast, although LoRA and PiSSA reduce the parameter update range through low-rank adaptation, their forgetting levels remain high, indicating that "freezing the backbone + local updates" is insufficient to fundamentally solve the knowledge degradation problem. PiSSA, in particular, is even worse than LoRA for most parameter scales, suggesting that initializing only pre-trained weights using the main singular direction cannot guarantee stability in a continuous learning environment; instead, it may exacerbate interference with the historical functional subspace due to over-reliance on existing main directions. In contrast, RandSel and O-LoRA significantly alleviated forgetting. RandSel, by replaying a small number of historical samples, can continuously input old knowledge into the model while learning new tasks, thus outperforming LoRA and PiSSA overall. O-LoRA further reduces gradient conflicts between new and old tasks by imposing orthogonal constraints on the subspace update direction. This demonstrates that, without relying on old data, orthogonal constraints based on geometric space can indeed effectively protect historical knowledge.

[0060] The trends in parameter size do not show a monotonically decreasing effect of catastrophic forgetting. Instead, forgetting is generally more severe in models ranging from 1.5B to 7B, peaking around 7B, and then declining again at 14B. This phenomenon suggests that while medium-sized models possess strong task adaptability, their parameter redundancy and functional isolation capabilities are insufficient to effectively accommodate both new and old knowledge, making them more prone to representational conflicts and functional overriding. Larger-scale 14B models, with their more ample representation space, can mitigate these conflicts to some extent.

[0061] After fine-tuning the large model, a different trend can be observed in the ROUGE-L metric on the SuperNI dataset compared to MMLU: full parameter fine-tuning achieves the highest scores across all parameter scales, reaching 48.4, 55.8, 60.3, 64.7, and 71.6 respectively. This indicates that when focusing only on the current task's fit, full parameter updates offer the greatest optimization freedom, thus adapting more fully to new task objectives. However, the advantage of full parameter fine-tuning on new tasks comes at the cost of preserving historical knowledge. Considering the aforementioned MMLU results, while full parameter fine-tuning performs best on SuperNI, it is also the method most prone to forgetting. In contrast, CDOCL achieves the second-highest ROUGE-L scores across all parameter scales, at 45.7, 52.3, 56.2, 60.2, and 65.8, second only to full parameter fine-tuning, but significantly outperforms the latter on MMLU. This indicates that while CDOCL retains the necessary constraints in fitting new tasks, the cost is not high, but it gains a significantly stronger ability to preserve general knowledge, thus making it more advantageous in the overall sense of continuous learning.

[0062] Combining the results from MMLU and SuperNI, it can be seen that different methods essentially reflect different choices in the trade-off between "stability" and "plasticity." Full parameter fine-tuning maximizes the model's ability to fit new tasks, thus achieving the highest score on SuperNI, but also causes the strongest perturbation to the original parameter space, resulting in the most severe forgetting in MMLU. While LoRA and PiSSA methods narrow the update range, they lack protection for historical functional directions, thus only mitigating forgetting to a certain extent. RandSel enhances the memory of historical knowledge through data replay, while O-LoRA reduces the conflict between new and old tasks through orthogonal constraints, demonstrating stronger resistance to forgetting.

[0063] In contrast, CDOCL employs coordinated control from both the data and parameter sides: on one hand, PSGM utilizes perplexity information to soft-modulate high-noise, high-interference samples, avoiding information loss caused by hard truncation; on the other hand, TV-SPT tracks the time-varying latent concept subspace online, ensuring that orthogonal constraints are no longer statically based on historical directions but are continuously updated as training progresses. It is this closed-loop mechanism that enables CDOCL to achieve the lowest forgetting rate across all parameter scales and maintain near-full-parameter fine-tuning performance on new tasks, thus achieving superior overall performance compared to all baselines.

[0064] To further verify the independent contributions of the two core modules, perplexity-aware soft gradient adjustment mechanism (PSGM) and time-varying latent concept subspace tracking (TV-SPT), ablation experiments were conducted on the Qwen2.5-7B model. The 7B-scale model was chosen as a representative because the forgetting phenomenon is more pronounced at this scale, better reflecting the impact of different modules on continuous learning performance.

[0065] This paper presents two single-module ablation variants: PSGM and TV-SPT. PSGM removes TV-SPT and the orthogonal cutoff, retaining only PSGM; TV-SPT removes the perplexity-based soft gradient adjustment mechanism, retaining only TV-SPT and the orthogonal cutoff. The complete method is denoted as CDOCL. Table 4.3 shows the experimental results for the different variants. Acc. represents the model's accuracy on the MMLU dataset, and ROUGE-L represents the model's evaluation results on the SuperNI dataset.

[0066] Table 4.3 Ablation test results of CDOCL on Qwen2.5-7B

[0067] As shown in Table 4.3, removing any module reduces the model's general capabilities and also impacts its performance on new tasks. This indicates that both PSGM and TV-SPT play a crucial role in balancing the model's stability and plasticity. Experimental data shows that PSGM and TV-SPT, compared to the complete CDOCL, decreased the accuracy on the MMLU dataset by 4.4 and 6.5 percentage points, respectively. This suggests that without a perplexity-based soft gradient adjustment mechanism, the model's ability to retain historical knowledge and its fitting performance on new tasks are more easily weakened during training.

[0068] In summary, PSGM primarily mitigates interference from high-risk updates on the data side through perplexity-based soft gradient adjustment, while TV-SPT protects historical knowledge-related directions on the parameter side through time-varying latent concept subspace tracking and orthogonal truncation. Together, they support the anti-forgetting ability of CDOCL. Experimental results show that removing PSGM results in more significant performance degradation on MMLU and ROUGE-L, indicating that under the current experimental settings, PSGM plays a more prominent role in maintaining the model's generality and adaptability to new tasks. Meanwhile, TV-SPT effectively alleviates knowledge conflicts during continuous learning through parameter space constraints.

[0069] This invention introduces a perplexity-aware soft gradient adjustment mechanism (PSGM) at the data level and a time-varying latent concept subspace tracking mechanism (TV-SPT) at the parameter level, combined with orthogonal truncation to dynamically protect the relevant directions of historical knowledge, thus constructing a continuous learning framework that balances stability and plasticity. Experimental results on the SuperNI and MMLU datasets based on the Qwen2.5 series models show that CDOCL can maintain good learning performance for new tasks while controlling general knowledge forgetting, and its overall performance is better than many baseline methods. Ablation experiments further verify the effectiveness of the two core modules, PSGM and TV-SPT. The performance degradation of the model is more obvious after removing PSGM, indicating that PSGM makes a more significant contribution to the overall performance of the model.

[0070] Therefore, this invention employs the aforementioned data-parameter coupled orthogonal continuous learning method, designing a perplexity-aware soft gradient adjustment mechanism (PSGM), breaking away from the traditional hard mask truncation logic and transforming sample perplexity into adaptive continuous gradient weights. This strategy suppresses high-noise signals while preserving the weak gradient flow of long-tailed complex knowledge without loss; more importantly, it explicitly injects perplexity weights into downstream subspace updates, establishing a closed loop of "data-parameter" linkage. A time-varying latent concept subspace tracking mechanism (TV-SPT) is introduced, abandoning the computationally expensive static SVD. In the weighted feature space purified by PSGM, a core feature dictionary is dynamically constructed using online power-law iteration and exponential moving average (EMA). This mechanism enables the orthogonal projection matrix to accurately track and dynamically evolve with the fine-tuning task, fundamentally overcoming the functional subspace drift problem in continuous learning.

[0071] Finally, it should be noted that the above embodiments are only used to illustrate the technical solutions of the present invention and not to limit them. Although the present invention has been described in detail with reference to preferred embodiments, those skilled in the art should understand that modifications or equivalent substitutions can still be made to the technical solutions of the present invention, and these modifications or equivalent substitutions cannot cause the modified technical solutions to deviate from the spirit and scope of the technical solutions of the present invention.

Claims

1. A data-parameter coupled orthogonal continuous learning method, characterized in that, Includes the following steps: Step 1: Pre-construction unified initialization, formally defining the continuous learning scenario and establishing a unified symbol system; Step 2: PSGM sample-level perplexity adaptive soft weighting. At the sample level, PSGM is used to softly adjust the training signal based on perplexity to reduce the negative impact of high-risk samples on training stability. Step 3: TV-SPT time-varying latent concept subspace tracking and gradient orthogonal truncation. At the parameter level, TV-SPT is used to dynamically extract historical sensitive directions from weighted features, and orthogonal truncation is performed on the current gradient accordingly.

2. The orthogonal continuous learning method with data-parameter coupling according to claim 1, characterized in that: Step one defines a continuous learning sequence consisting of T tasks: in, Indicates the first t A dataset for each task. and These are the input and the target output, respectively. The number of samples for this task, in the th... t During each training phase, the model can only access the current task. Data, historical task data Inaccessible.

3. The orthogonal continuous learning method with data-parameter coupling according to claim 2, characterized in that: In step one, initialize the frozen backbone large language model. Open the low-rank adapter as an updatable parameter; initialize the TV-SPT dynamic feature dictionary, online exponentiation iteration parameters, EMA momentum coefficient, similarity threshold, and PSGM hyperparameter.

4. The data-parameter coupling orthogonal continuous learning method according to claim 3, characterized in that, In step two, input the current batch of samples and perform the following steps: S21. For each sample within the batch ( ), calculate sample perplexity: in, The length of the target sequence. The model represents the first j The conditional probability of each target word; S22. Calculate the confusion distribution of this batch and the batch mean. Standard deviation and adaptive threshold , in, This is the adjustment coefficient; S23. Substitute the Sigmoid soft adjustment function to solve for the weight of each sample. , ； in, For the Sigmoid function, Temperature coefficient; S24. Output the batch sample weight set, which is used simultaneously for loss weighting calculation and TV-SPT weighted feature covariance construction.

5. The orthogonal continuous learning method with data-parameter coupling according to claim 4, characterized in that: Step three includes the following steps: S31. Construct the weighted feature covariance matrix; S32. Online power method for extracting the dominant feature direction of the current batch; S33. Update the dynamic feature dictionary and store historical sensitive atoms; S34. Gradient orthogonal truncation removes update components that disrupt historical knowledge.

6. The data-parameter coupling orthogonal continuous learning method according to claim 5, characterized in that, The weighted covariance matrix in S31 is: in, This indicates that the Lth layer of the model is in t Activation matrices over each training step , b This represents the number of samples in the current training batch. d For the hidden layer dimension, Construct a diagonal weight matrix for the weights. .

7. The orthogonal continuous learning method with data-parameter coupling according to claim 6, characterized in that: S33 maintains a dynamic feature dictionary ,in Indicates the size of the dictionary, each A unit feature vector, called an atom, is extracted for the principal direction in the current training step. First, calculate its similarity to each atom in the dictionary; Maximum similarity below a preset threshold The current direction is not yet fully represented by the existing dictionary, and will Added as a new atom to the dictionary; maximum similarity not less than If the current direction is similar to existing functional atoms, then the exponential moving average (EMA) method is used to smoothly update the most similar atoms: in, m The momentum coefficient is represented by Norm(), which indicates the normalization operation.

8. The data-parameter coupling orthogonal continuous learning method according to claim 7, characterized in that, S34 includes the following steps: S341. Calculate the original gradient of the current training step. ; S342, the original gradient The gradient after orthogonal truncation, projected onto the orthogonal complement of the history-sensitive subspace, is: in, I It is the identity matrix; It is a dictionary matrix; S343, Use Complete the parameter update operation.

Images