Selected work

Representative Papers

LAVA: Logic-Aware Validation and Augmentation Framework for Large-Scale Financial Document Auditing

Aug 17, 2026

This study addresses verification challenges in financial document auditing arising from heterogeneous formats and embedded rules by proposing a logic-aware enhancement framework based on multimodal large language models. The framework employs a four-stage modular pipeline integrating document retrieval, layout-preserving extraction, metadata augmentation, and symbolic verification to enable fine-grained error attribution and end-to-end traceability. Experiments on real-world benchmarks demonstrate that this approach significantly outperforms baselines by effectively mitigating hallucinations and accurately handling edge cases while maintaining efficient token utilization. Consequently, the proposed method satisfies the stringent requirements for high trustworthiness and interpretability in high-frequency auditing scenarios, offering a robust solution for complex financial compliance tasks.

0 citationsRead paper

FRAMES: Guarded and Dual-Objective Skill Evolution for Agents in Policy-Governed Enterprise Workflows

Aug 03, 2026

This work addresses the challenges faced by large language model (LLM) agents in policy-constrained enterprise workflows—such as document auditing—including sparse feedback, performance degradation due to frequent rule updates, and the need to balance accuracy, reasoning cost, and auditability. To tackle these issues, the authors propose FRAMES, a framework that bootstraps deployable skills during cold-start and enables continuous skill evolution through consensus-based skill mutation, Pareto-optimal trade-offs between precision and computational cost, and a degradation-aware validation mechanism within a closed-loop system. FRAMES is the first approach to achieve auditable, efficient, and robust skill iteration in policy-intensive settings, demonstrating state-of-the-art precision–cost trade-offs on both an internal production system and the tau-bench benchmark, significantly outperforming existing baselines.

0 citationsRead paper

Dual Signal Decomposition of Stochastic Time Series

Aug 07, 2025

This work addresses heteroscedastic stochastic time series by proposing a dual-signal decomposition framework that disentangles the original series into three components: mean, dispersion (i.e., time-varying volatility), and stationary white noise. Methodologically, it employs a dual-output neural network—or equivalently, a nonlinear optimization model—to jointly model the dynamics of mean and dispersion. An adaptive regularization weighting mechanism, grounded in statistical process control, is introduced to balance learning objectives. Crucially, first- and second-order temporal derivative regularizers are integrated to enforce signal smoothness, suppress noise, and preserve structural integrity. Two complementary learning paradigms—joint learning and sequential learning—are developed to enable collaborative modeling, cross-effect analysis, and multi-series structural comparison in a 2D signal space. Experiments demonstrate accurate separation of uncorrelated noise, robust handling of abrupt and smooth regime shifts, substantial improvements in predictive accuracy and interpretability for both signals, and strong generalization and extensibility across diverse applications.

0 citationsRead paper

Recent publications

Latest Papers

LAVA: Logic-Aware Validation and Augmentation Framework for Large-Scale Financial Document Auditing

Aug 17, 2026

This study addresses verification challenges in financial document auditing arising from heterogeneous formats and embedded rules by proposing a logic-aware enhancement framework based on multimodal large language models. The framework employs a four-stage modular pipeline integrating document retrieval, layout-preserving extraction, metadata augmentation, and symbolic verification to enable fine-grained error attribution and end-to-end traceability. Experiments on real-world benchmarks demonstrate that this approach significantly outperforms baselines by effectively mitigating hallucinations and accurately handling edge cases while maintaining efficient token utilization. Consequently, the proposed method satisfies the stringent requirements for high trustworthiness and interpretability in high-frequency auditing scenarios, offering a robust solution for complex financial compliance tasks.

0 citationsRead paper

FRAMES: Guarded and Dual-Objective Skill Evolution for Agents in Policy-Governed Enterprise Workflows

Aug 03, 2026

This work addresses the challenges faced by large language model (LLM) agents in policy-constrained enterprise workflows—such as document auditing—including sparse feedback, performance degradation due to frequent rule updates, and the need to balance accuracy, reasoning cost, and auditability. To tackle these issues, the authors propose FRAMES, a framework that bootstraps deployable skills during cold-start and enables continuous skill evolution through consensus-based skill mutation, Pareto-optimal trade-offs between precision and computational cost, and a degradation-aware validation mechanism within a closed-loop system. FRAMES is the first approach to achieve auditable, efficient, and robust skill iteration in policy-intensive settings, demonstrating state-of-the-art precision–cost trade-offs on both an internal production system and the tau-bench benchmark, significantly outperforming existing baselines.

0 citationsRead paper

Dual Signal Decomposition of Stochastic Time Series

Aug 07, 2025

This work addresses heteroscedastic stochastic time series by proposing a dual-signal decomposition framework that disentangles the original series into three components: mean, dispersion (i.e., time-varying volatility), and stationary white noise. Methodologically, it employs a dual-output neural network—or equivalently, a nonlinear optimization model—to jointly model the dynamics of mean and dispersion. An adaptive regularization weighting mechanism, grounded in statistical process control, is introduced to balance learning objectives. Crucially, first- and second-order temporal derivative regularizers are integrated to enforce signal smoothness, suppress noise, and preserve structural integrity. Two complementary learning paradigms—joint learning and sequential learning—are developed to enable collaborative modeling, cross-effect analysis, and multi-series structural comparison in a 2D signal space. Experiments demonstrate accurate separation of uncorrelated noise, robust handling of abrupt and smooth regime shifts, substantial improvements in predictive accuracy and interpretability for both signals, and strong generalization and extensibility across diverse applications.

0 citationsRead paper