From Black Box to Actionable Insights: An Adaptive Explainable AI Framework for Proactive Tax Risk Mitigation in Small and Medium Enterprises

1. Introduction

The persistent widening of the tax gap poses a severe challenge to national finances. Concurrently, tax regulatory bodies like the Internal Revenue Service (IRS) have intensified audits of businesses, particularly small and medium-sized enterprises (SMEs). Against this backdrop, SMEs face significant pain points during tax filing due to their limited tax expertise and insufficient understanding of complex tax regulations. Existing tax risk detection technologies, largely based on traditional statistical models or machine learning “black box” models, exhibit clear limitations: These models often rely on historical data and fixed rules, failing to respond in real-time to dynamic changes in tax policies (e.g., annual adjustments to R&D expense deduction rates by the IRS) or rapid fluctuations in business operations (e.g., temporary workforce expansion, seasonal income variations). This results in delayed decision-making, high misclassification rates, and a lack of clear interpretability, making it difficult for SMEs to understand and trust the warning results generated by these models. Addressing these issues, this study aims to answer a core research question: How can an explainable AI system be built for SMEs that can identify high-risk tax filing items in real time and provide concrete guidance for correction? To this end, this paper proposes an adaptive explainable artificial intelligence framework for proactive tax risk prevention and control in SMEs. Key contributions include: 1. A model integrating policy awareness, few-shot incremental learning, and dynamic feature iteration to adapt to evolving tax policies and operational changes, fundamentally resolving model lag. 2. Deep integration of explainability techniques like SHAP, transforming black-box predictions into actionable insights that not only flag risks but explain their origins and provide corrective guidance. An interactive interface integrating risk prediction, visual explanations (e.g., radar charts, waterfall charts), and corrective recommendations was developed to lower the adoption barrier for SMEs. Through experiments simulating dynamic tax filing scenarios and detailed case studies, the framework demonstrated outstanding performance in accuracy, robustness, and practicality.

2. Related Work

2.1. Tax Risk Detection Model

Traditional statistical methods were the early mainstream techniques for tax risk detection, with commonly used approaches including linear/logistic regression, decision trees, and clustering algorithms. However, these models rely on fixed rules or historical data, making them unable to respond in real time to changes in tax policies (e.g., the 2023 IRS adjustment to R&D expense deduction rates for SMEs) or dynamic business conditions (e.g., temporary hiring surges, revenue fluctuations). This leads to “rule obsolescence”—for instance, a model still using 2021 deduction thresholds to screen 2024 tax filings would see a significant increase in misclassification rates. Tax risks for SMEs often stem from multi-feature interactions (e.g., “business duration < 3 years + employees < 5 + deduction ratio > 40%”). Traditional statistical methods struggle to capture such nonlinear relationships, performing only one-dimensional detection and resulting in high false-negative rates. This deficiency is directly linked to the absence of feature localization technology: [1] Jin et al. (2022) in “Review of Methods Applying on Facial Alignment” noted that the core value of facial alignment technology lies in achieving multi-feature collaborative matching through precise localization of key features (e.g., facial landmarks). Traditional tax statistical models lack precisely this “feature localization mechanism,” - interaction analysis" mechanism, failing to recognize risk patterns involving multiple feature couplings (e.g., focusing solely on the proportion of deductions while ignoring the synergistic impact of business tenure and employee count); Existing traditional models are predominantly designed for large enterprise data (e.g., considering multinational revenues and complex tax types), failing to optimize for the characteristics of SMEs—which feature “simple tax types and concentrated risk points (e.g., underreported income, non-compliant deductions)”—resulting in redundant detection dimensions and incomplete coverage of critical risk factors.

In recent years, machine learning has gained traction in tax risk detection due to its superior complex data processing capabilities. Mainstream models include ensemble learning and deep learning approaches. Some studies employ neural networks (e.g., LSTM) to process time-series tax filing data (e.g., monthly declarations over the past three years) to capture long-term risk trends. The feasibility of this approach has been validated in other time-series anomaly detection scenarios: [2] Huang and Qiu (2025) in “LSTM-Based Time Series Detection of Abnormal Electricity Usage in Smart Meters” employed LSTM models to capture temporal fluctuation patterns in smart meter electricity consumption (e.g., seasonal usage peaks, daily consumption stability), enabling real-time identification of abnormal usage behavior with an accuracy improvement of 18% over traditional statistical methods. However, the application of LSTM in tax risk detection remains significantly underdeveloped. Monthly tax filing data from small and medium-sized enterprises similarly exhibit temporal patterns such as “seasonal revenue fluctuations” and “quarterly deduction patterns” (e.g., retail enterprises typically report higher fourth-quarter revenues than other quarters). Yet existing models fail to adopt Huang et al.'s temporal quarterly deduction patterns“ (e.g., retail enterprises typically reporting higher fourth-quarter revenues than other quarters). Existing models, however, have not adopted Huang et al.'s time-series analysis approach, failing to uncover sequential correlation risks such as ”three consecutive months of sequential revenue decline with unchanged deductions." Consequently, numerous risk indicators within the time-series dimension remain largely overlooked.

2.2. Applications of Explainable AI (XAI) in Finance

SHAP (SHapley Additive exPlanations) leverages Shapley values from game theory to calculate the “marginal contribution” of each feature to the decision outcome, enabling dual explanations at both global (all samples) and local (individual samples) levels. Its application in finance has expanded from “risk attribution” to “optimization guidance.” In credit risk assessment, banks use SHAP to explain “why a company's loan application was rejected,” such as outputting “the company's debt-to-asset ratio (contribution value -0.8) and overdue payment frequency in the past three years (contribution value -0.5) are the primary negative factors,” helping businesses identify areas for improvement. [3] Qi (2025) in “Enterprise Financial Distress Prediction Based on Machine Learning and SHAP Interpretability Analysis” built a financial distress prediction model using XGBoost. By quantifying the contribution of key features like “debt-to-asset ratio” and " Operating Cash Flow Ratio.“ This approach not only achieved an 89% prediction accuracy but also provided concrete optimization recommendations such as ”reducing the debt-to-asset ratio below 50% can significantly alleviate distress,“ forming a closed-loop of ”prediction-explanation-guidance."

2.3. Research on Continuous Learning in Dynamic Environments

Cross-domain continuous learning research has established a mature “dynamic feature adaptation” approach, offering valuable insights for tax scenarios. [4] Yi et al. (2021) in “DDR-Net: Learning multi-stage multi-view stereo with dynamic depth range” proposed a multi-stage dynamic depth range learning mechanism — — where the model dynamically adjusts the “precision level” of feature extraction (e.g., emphasizing fine-grained features in close-up scenes and global features in distant scenes) based on depth variations across different scenarios. This enables real-time adaptation to dynamic visual environments, preventing performance degradation due to scene changes. This “on-demand feature precision adjustment” approach aligns closely with the need for dynamic adaptation in tax policy. Forgetting is mitigated through regularization (e.g., EWC, Elastic Weight Consolidation) and replay mechanisms (e.g., storing historical key samples). For instance, in credit card fraud detection, the model undergoes incremental updates after processing every 1,000 new transaction records. Concurrently, EWC safeguards the weights of “traditional fraud features (e.g., overseas transactions)” to prevent the learning of new data from invalidating established patterns.

3. Methodology

3.1. Adaptive Learning Model

To more clearly illustrate the actual interaction process between adaptive learning mechanisms and policy changes/feature updates, this paper uses the IRS's 2024 adjustment to the “Additional Deduction for Digital Investments” policy as an example to explain the model's real-time response workflow: Upon the IRS's release of the “Additional 10% Tax Deduction for Small Business Digital Investments in 2024” policy, the system retrieves the policy text in real-time via API. Error! Reference source not found. The NLP module parses key elements: “policy impact dimension” (e.g., “Deduction Category - Digital Investment”), “adjustment value” (10%), and “effective date” (January 1, 2024). The system automatically updates the “Policy-Feature Mapping Table,” associating the “Digital Investment Deduction Rate” with features like “Deduction-to-Revenue Ratio” and “Digital Investment Amount” within the model. For example, after the policy takes effect, the system automatically increases the feature weight of “Digital Investment Amount” from 0.2 to 0.4 while simultaneously reducing the weight of “Traditional Equipment Investment Deduction” (e.g., from 0.3 to 0.1). Incremental training is initiated every 30-50 new SME tax filing records collected. If an enterprise's “deduction ratio exceeds the industry average by 20%,” that sample is flagged as high-risk and prioritized for model updates. Concurrently, the system applies weight protection to key features (e.g., “industry code match”) based on industry benchmarks Error! Reference source not found. (e.g., Industry 7's average risk rate of 6.8%) to prevent model forgetting of common industry risk patterns. Furthermore, the system evaluates feature-risk label correlations every three months using a sliding time window. If the “temporary worker expenditure ratio” maintains a correlation above 0.25 across two consecutive windows, it is automatically added to the core feature set. Conversely, if the “fixed asset depreciation period” falls below a 0.1 correlation in two windows, it is removed.

3.2. Risk Prediction Model

The dynamic feature input layer incorporates fine-grained tax filing features optimized by the adaptive module (e.g., “employee wage deductions,” “R&D investment deductions,” “digital investment deductions”—derived from the original “total deductions”), alongside industry-related features (e.g., “deviation of deductions from industry average,” “alignment of revenue volatility with industry trends”). Feature dynamic weights leverage those generated by the adaptive module based on policy updates (e.g., IRS deduction adjustments) and risk relevance assessments (e.g., the weight for “digital investment deductions” increased from 0.2 to 0.4 after policy implementation), ensuring the model prioritizes high-value risk features. It incorporates compliance benchmarks stored in the adaptive module as core references for risk assessment, preventing one-size-fits-all predictions detached from industry characteristics.

To address the small sample size and real-time requirements of SMEs, a lightweight architecture combining “Enhanced LightGBM + Feature Interaction Boosting” is adopted. LightGBM serves as the core predictor, assigning higher sample weights (1.2–1.5 times) to “high-risk-correlated samples” (e.g., enterprises with historical deduction violations) selected by the adaptive module, thereby boosting the model's learning priority for high-risk patterns. L1 regularization suppresses redundant features (e.g., “fixed asset depreciation period” with low SME risk relevance) to mitigate overfitting in small samples. Experiments confirm this optimization reduces overfitting rates by 15%-20% in SME datasets below 50 samples.

For SME “multi-feature coupled risks” Error! Reference source not found. (e.g., “business duration < 3 years + deductions ratio > 40% + zero-declaration frequency > 2 times”), generate interaction features through a combination of manual design and automated mining. Based on core SME risk points, design interaction features such as “deduction ratio × industry deviation coefficient” and “zero-declaration frequency × number of employees.” Utilize LightGBM's feature importance feedback to dynamically retain second-order interactions among the “Top 10 most important single features,” such as “revenue declaration stability × business frequency × number of zero-declarations × number of employees.” Number of Zero-Declaration Periods × Number of Employees.“ Using LightGBM's feature importance feedback, we dynamically retained second-order interactions among the ”Top 10 Single Feature Importances“ (e.g., ”Income Declaration Stability × Business Duration") to ensure the model captures non-linear risk correlations.

3.3. Explainability Module

Figure 1 displays the “Tax Filing Risk Feature Importance (SHAP Value Distribution)” for Model Version v2. The horizontal axis represents SHAP values (a feature's contribution to risk prediction, where higher values indicate stronger impact), while the vertical axis shows core tax filing features. Industry Code (SHAP value 0.4091) and Tax Deductions (SHAP value 0.3988) are the most critical risk drivers (both with SHAP values near 0.4), indicating that “industry-business alignment” and “deduction rationality” are primary sources of tax risk. Next in influence are Total Income (0.2249) and Total Expenses (0.2171), exerting a secondary impact on risk; Business Age (Years) (0.1020) and Tax Credits (0.0659) exert relatively weaker influence, while Number of Employees (-0.0308) even exhibits a negative contribution (i.e., higher values of this feature correlate with lower risk). The sign of SHAP values indicates the direction of the feature-risk relationship. Error! Reference source not found. A high positive SHAP value for Industry Code implies that “the lower the alignment between the declared industry and actual business operations, the higher the risk” (e.g., a restaurant business declared as retail would significantly elevate risk). The negative SHAP value for Number of Employees indicates that “higher employee counts may imply lower risk” (aligning with the business intuition that “larger enterprises tend to have more standardized compliance management”).

Figure 1. Importance Distribution of Tax Filing Risk Features Based on SHAP Values (Model Version v2).

Figure 1. Importance Distribution of Tax Filing Risk Features Based on SHAP Values (Model Version v2).

4. Experiments and Results

4.1. Model Performance Evaluation

To validate the effectiveness of the proposed “Adaptive-Explainable-Real-Time” tax risk detection framework, experiments employed a dataset simulating dynamic tax filing scenarios for small and medium-sized enterprises (SMEs). Error! Reference source not found. This dataset comprised five incremental batches of tax filing data, each incorporating dynamic factors such as tax policy updates and business fluctuations. The performance of model version v2 in the “tax filing error detection” task was evaluated using core metrics including Accuracy, Precision, and Recall, with results presented in Figure 2 . AUC, Precision, and Recall as core metrics to evaluate the performance of model version v2 in the “tax filing error detection” task. The results are shown in Figure 2 .

The model consistently maintained an overall accuracy above 0.8. Even when processing the fifth batch of data (featuring more complex scenarios and more pronounced shifts in data distribution), accuracy remained stable. This demonstrates the model's robust capability to continuously deliver reliable “risk/compliance” classification results within a dynamic tax environment. The red star markings (Model Updated) in the figure indicate the timing of “adaptive learning model update triggers.” Post-update accuracy showed no significant fluctuation, proving that the adaptive learning mechanism effectively preserves historical compliance knowledge and classification capabilities when updating model parameters. This avoids catastrophic forgetting, delivering a stable risk detection experience for SMEs.

Figure 2. Performance Evaluation Trends of the Adaptive Tax Risk Detection Model on Dynamic Data Streams.

Figure 2. Performance Evaluation Trends of the Adaptive Tax Risk Detection Model on Dynamic Data Streams.

Table 1. the stronger the discrimination capability). Results show the model consistently maintains an AUC above 0.8, peaking near 0.9. This indicates the model effectively distinguishes “high-risk tax filings” from “compliant filings,” accurately capturing core features of high-risk samples even when confronted with dynamically changing tax data. Following model updates, the AUC did not decline but instead showed a slight increase. This validates that the synergistic mechanism between adaptive learning and risk prediction models can continuously optimize feature weights, enhance the ability to delineate the “risk/compliance” boundary, and ensure efficient identification of high-risk filings during policy updates and business fluctuations. The lower-left subplot shows the Precision Trend, where precision measures “the proportion of truly high-risk cases among those predicted as high-risk by the model” (lower values indicate more “false positives”). The trend indicates that during the initial phase (Batch 1 data), precision approached 0.8, demonstrating the model's high “initial identification accuracy” for high-risk cases; As processing batches increased (exposing the model to more complex scenarios and noise), precision declined but remained above 0.4. Following Model Update, precision showed a significant rebound, rising from approximately 0.45 to above 0.5. This change validates that the adaptive learning module can correct the model's tendency to misjudge “high-risk features” through “smallsample incremental training + policy awareness.” This reduces “false risk alerts” for SMEs, allowing businesses to focus on genuinely problematic declaration items and lower compliance costs.

The lower-right subplot shows the Recall Trend. Recall measures “the proportion of actual high-risk samples correctly identified by the model” (lower values indicate more “missed detections”). The results reveal a critical trend: initially, recall was low (close to 0.1), indicating insufficient coverage of “potential high-risk samples” and numerous “missed detections.” As processing batches increased, especially after Model Updates, recall rapidly rose from 0.1 to over 0.6. This significant improvement directly demonstrates that the adaptive learning mechanism effectively captures emerging risk patterns driven by “tax policy updates and business fluctuations” Error! Reference source not found. (e.g., new “digital investment deduction violations” or " Industry Code Mismatch“). This shifts the model from ”passively identifying known risks“ to ”actively uncovering unknown risks,“ substantially reducing ”missed detections" of high-risk filings and providing more comprehensive risk coverage for tax oversight.

4.2. Case Study

Sample 0 is an SME in Industry 7 (e.g., retail). After submitting monthly tax data, the system automatically triggers the "Risk Prediction - Explainable - Corrective Recommendations" process. The core visual outputs are shown in Figure 3 (Industry Benchmark Comparison Radar Chart) and Figure 4 (SHAP Risk Explanation Waterfall Chart).

Figure 3 : The blue polygon represents the industry benchmark feature distribution for Industry 7 (the compliance reference range for each tax reporting dimension), while the red polygon represents Sample 0's actual declared data. The comparison reveals key deviations: Sample 0's total income is significantly below the industry benchmark (the red vertex is positioned far inward on the “Total Income” dimension); Sample 0's total expenses are significantly above the industry benchmark (the red vertex is positioned far outward on the “Total Expenses” dimension); and Sample 0's deductions also exceed the industry benchmark.

Figure 3. Radar Chart Comparing Tax Filing Data of Sample Company (Sample 0) with Industry Benchmark (Industry 7).

Figure 3. Radar Chart Comparing Tax Filing Data of Sample Company (Sample 0) with Industry Benchmark (Industry 7).

Figure 4 presents the waterfall chart for Sample 0 Tax Risk Explanation, illustrating the direction and strength of each tax filing characteristic's contribution to the “Predicted Risk Value (0.348)” (red indicates “pushing risk upward,” blue indicates “pulling risk downward”). The “Total Income” feature value is 13172.968 with a contribution value of +0.6, making it the primary risk driver—corresponding to the radar chart's “Total Income Far Below Industry Benchmark” finding. Error! Reference source not found. This confirms “Significantly Underreported Income” as the primary cause of elevated risk for Sample 0. Total Expenses has a feature value of 1984.908 and a contribution value of +0.03, exerting a slight upward pressure on risk—corresponding to the radar chart's “Total Expenses above industry benchmark,” further amplifying risk; Characteristics such as Number of Employees (contribution -0.23) and Business Age (Years) (contribution -0.09) act as risk-reducing factors (blue bars). However, due to the stronger driving effects of “Total Revenue” and “Total Expenditure,” the overall risk remains above the industry average.

Figure 4. SHAP Explanation Waterfall Plot for Tax Risk Prediction of Sample Enterprise (Sample 0).

Figure 4. SHAP Explanation Waterfall Plot for Tax Risk Prediction of Sample Enterprise (Sample 0).

4.3. Cross-Industry Case Studies

To demonstrate the framework's cross-industry applicability, this paper expands upon the original retail sector (Industry 7) case study by adding examples from manufacturing (Industry 3) and food service (Industry 9). Manufacturing case: Abnormal R&D expense deduction ratio and mismatch between fixed asset depreciation and revenue. SHAP explanations indicate “excessively high R&D expense ratio” as the primary risk driver, recommending enterprises provide R&D project filing documentation to mitigate risk. Food Service Industry Case: Abnormal proportion of ingredient costs and high frequency of zero tax filings. The radar chart reveals “total expenditures” significantly exceeding industry benchmarks. The SHAP waterfall plot further pinpoints “ingredient cost ratio” as the primary risk factor. The system recommends adjusting cost structures or providing supplementary explanations for seasonal procurement patterns.

5. Discussion

5.1. Interpretation of Results

Tax policies impose “customized constraints” on different industries, which is the fundamental reason why “Industry Code” becomes a core risk characteristic. For example, in the case study of Industry 7 in Section 4.2 of this paper, the benchmark risk rate is 6.8%, corresponding to compliance standards such as “upper limit of deductions as a percentage of revenue” and " reasonable total expenditure range“ differ significantly from those in other sectors like catering or manufacturing (e.g., retail typically has lower average deduction ratios than catering). If a company's declared ”Industry Code" misaligns with its actual operations (e.g., a catering firm incorrectly classified as retail), this directly causes structural deviations between its tax filings and the benchmark for its actual industry — — As illustrated by the radar chart in Section 4.2 , the deviation in Sample 0's “Total Revenue - Total Expenses” feature from the Industry 7 benchmark fundamentally reflects risk mapping derived from “Industry Code alignment.” This logical chain—“industry mismatch → data deviation → elevated risk”—establishes “Industry Code” as a foundational risk prediction feature. Its SHAP value (0.4091), ranked first in the feature importance map of Section 3.3 , confirms this attribute's fundamental role in driving risk.

The widespread challenges among SMEs—such as limited tax knowledge and unclear understanding of industry classifications—further highlight the risk implications of “Industry Code.” Unlike large enterprises with specialized tax teams ensuring precise industry code matching, SMEs often file incorrectly due to “unfamiliarity with industry classification standards” or “misclassifying sub-industries under broader categories” (e.g., categorizing “community convenience stores” as “large supermarkets”). The adaptive learning model in Section 3.1 employs a "policy - Feature Mapping“ mechanism, has designated ”Industry Code“ as a ”policy-sensitive feature.“ Its weight dynamically increases with updates to tax policies. When an enterprise misaligns its industry code, the model prioritizes capturing this ”high-weight feature deviation.“ Section 3.2 's risk prediction model amplifies its impact on risk levels, ultimately manifesting as a high risk contribution from ”Industry Code."

The “dynamic feature iteration” and “industry benchmark constraint” designs in this model further amplify the risk contribution of “Industry Code.” The adaptive learning model in Section 3.1 continuously updates compliance benchmarks for each industry (e.g., deductions and revenue thresholds for Industry 7), using “Industry Code” as the “index feature for benchmark retrieval.” Incorrect industry codes cause misaligned benchmark data retrieval, amplifying deviations in subsequent “feature-benchmark comparison” risk assessments. Simultaneously, Section 3.2 's risk prediction model incorporates a “Feature Interaction Enhancement Module” that generates interaction features (e.g., “Industry Code × Deduction Deviation Rate”) by combining “Industry Code” with features like “Deduction Proportion” and “Total Revenue Volatility.” Error! Reference source not found. This amplifies the risk impact of industry codes through multi-feature coupling, ultimately creating a chain reaction effect: “single-feature mismatch → multidimensional risk linkage.”

5.2. Limitations

Although this framework demonstrates strong performance in experiments and case studies, two core limitations persist due to tax scenario characteristics and technical constraints: tax data contains SME commercial secrets (e.g., total revenue, expenditure details) and sensitive information (e.g., corporate legal entity details, employee compensation), while the framework relies on “multi-enterprise data sharing and iteration”— — Section 3.1 's adaptive learning model requires multi-enterprise data updates to refine industry benchmarks (e.g., compliance scope for Industry 7), while Section 3.2 's risk prediction model necessitates multi-enterprise labeled data for parameter optimization. This inherent conflict exists between the “data sharing requirement” and the “privacy protection requirement”: adopting “centralized data storage” risks massive tax data leaks if the system is compromised; opting for “localized data processing” prevents aggregating industry data to update benchmarks, thereby weakening the model's ability to identify common industry risks (e.g., when Industry 7 adjusts its overall deduction standards, localized models cannot promptly access the new benchmarks). The current framework has yet to incorporate privacy-preserving technologies like federated learning, making data privacy a critical barrier to practical implementation.

5.3. Future Work

The current framework focuses on small and medium-sized enterprises. Future work may extend it to individual income tax filing scenarios. The core characteristics of personal tax filing differ fundamentally from those of corporate filing, necessitating the addition of features such as “special additional deductions (e.g., children's education, housing loans)”, " labor remuneration/royalty income,“ and ”personal investment income.“ The adaptive learning model described in Section 3.1 should establish mappings between ”individual characteristics“ and ”policy rules" (e.g., tax rate variations for different income types, deduction limits for specific categories). High-risk points in individual tax filing (e.g., “duplicate reporting of special deductions,” “unreported labor remuneration”) differ from corporate scenarios. Error! Reference source not found. By annotating individual non-compliance data, the risk prediction model in Section 3.2 can be trained to identify “individual-specific risks” (e.g., “duplicate reporting of the same housing loan deduction by both spouses”). Individual users possess weaker tax knowledge, necessitating further simplification of outputs from the explainable module in Section 3.3 (e.g., replacing technical terms with “Your housing loan deduction has been claimed by your spouse; duplicate filing will trigger high risk”) to enhance user comprehension.

6. Conclusion

The proposed integrated tax risk detection framework—“Adaptive Learning - Risk Prediction - Explainability”—effectively achieves “proactive prevention” of SME tax filing errors through the synergistic effects of dynamic adaptation, real-time prediction, and precise explanation. In terms of dynamic adaptability, the adaptive learning model responds in real time to tax policy updates (e.g., adjustments to deduction rules) and business fluctuations (e.g., seasonal income variations). experiments demonstrate that the model completes parameter updates within one hour of policy changes, ensuring risk assessment criteria remain synchronized with the dynamic environment. Regarding real-time predictive capability, the risk prediction model achieves single-sample inference time <50ms. Businesses receive instant risk level and type assessments after submitting tax data. Section 4.1 experiments demonstrate the model maintains stable accuracy >0.8 and AUC >0.8, enabling precise differentiation between compliant and high-risk filings—laying the foundation for “correction upon filing.” In terms of error prevention effectiveness, the “risk-correction” closed-loop integrated with explainable modules (as demonstrated in Section 4.2 where Sample 0's risk significantly decreased after suggested adjustments) enables SMEs to directly rectify high-risk items during filing. This fundamentally reduces the chain of “filing errors → IRS audit → penalty losses,” achieving “preemptive prevention” of tax errors.

Additionally, tax regulators (e.g., IRS) can collaborate with tech companies to embed an “adaptive-explainable AI” framework into official filing platforms, providing SMEs with free “real-time risk detection-explanation-correction” tools. API interfaces connect to policy release platforms to ensure real-time policy awareness for models. An industry benchmark database regularly updates compliance feature ranges (e.g., deductions and revenue benchmarks for Industry 7), providing foundational support for model adaptive learning. At the data level, “federated learning + differential privacy” technology builds an industry data collaboration network where **“data stays within enterprises while models are jointly trained”**. SMEs participate in joint optimization of industry benchmarks and model parameters while retaining local tax data privacy. This approach addresses the model's “small-sample generalization deficiency” while mitigating privacy risks associated with centralized data storage, achieving a balance between “data privacy protection” and “industry model optimization.” At the enterprise level, addressing SMEs' characteristics of “limited tax knowledge and insufficient digital tool proficiency,” tax authorities can collaborate with industry associations to offer free training on “adaptive AI tax filing tools.” Through case-based instruction (e.g., the rectification process in Sample 0 from Section 4.2 ), the tools' "risk identification - explanation - correction“ logic through case studies (e.g., the rectification process in Sample 0, Section 4.2 ). This lowers the usage barrier for SMEs, drives the tool's transition from ”technical feasibility“ to ”actual usage rate," and ultimately elevates tax compliance standards across the entire industry.Tax regulatory agencies (such as the IRS) can collaborate with technology companies to embed this framework into official tax filing platforms, providing small and medium-sized enterprises with free, real-time, and explainable risk detection and remediation tools. Simultaneously, by leveraging federated learning and differential privacy technologies, a secure and compliant industry data collaboration network can be established. This enables continuous model optimization and shared governance of industry risks while safeguarding corporate privacy.