Selected work

Representative Papers

Class-Aware Reinforcement Learning for Counterfactual Explanation Generation

Jul 30, 2026

This work proposes a class-aware reinforcement learning framework to enhance the efficiency and quality of counterfactual explanation generation. By incorporating the model’s predicted class information into the state representation—a novel design in this domain—the proposed approach guides the policy to more effectively explore counterfactual instances that satisfy validity, sparsity, and proximity constraints. The integration of class awareness significantly accelerates policy convergence and improves reward optimization. Empirical evaluation across seven diverse datasets demonstrates that the method generates a greater number of high-quality explanations with fewer training episodes compared to existing approaches. Furthermore, feature importance analyses using SHAP and LIME confirm that class information plays a critical role in guiding action selection during the generation process.

0 citationsRead paper

Memory Efficient Tabular Foundation Models

Jul 29, 2026

This study addresses the high memory footprint of tabular foundation models during deployment by systematically investigating and validating the effectiveness of model compression techniques for this emerging class of models. Integrating multiple compression strategies, the authors optimize representative architectures such as TabPFN, achieving up to a 7.6× (approximately 87%) reduction in memory usage while preserving near-original in-context learning performance. This work substantially enhances the practical deployability of tabular foundation models and offers a viable solution for their application in resource-constrained environments.

0 citationsRead paper

Beyond Generalist LLMs: Specialist Agentic Systems for Structured Code Workflow Execution

Jul 15, 2026

This work addresses the limitations of general-purpose large language models in business process automation, where inconsistent functionality, frequent tool-calling errors, and unstable code quality hinder industrial-grade reliability and maintainability. Focusing on the task of translating BPMN diagrams into executable agent workflows, we propose the first specialized agent system designed for structured code generation. By integrating BPMN control-flow semantics, a deterministic path execution mechanism, and a lightweight code generation strategy, our approach achieves high-precision, low-latency, and zero-repair automation. Experimental results demonstrate that, compared to general-purpose models, our method improves tool-calling accuracy by 9–20 percentage points, reduces latency by 2–4×, decreases calling errors by a factor of three, lowers token-generation costs by over 95%, and entirely eliminates the need for repair iterations.

0 citationsRead paper

Format-Controlled Multi-Scale JPEG Compression Response Analysis for Image-Level Forgery Screening

Jul 07, 2026

This study addresses the high computational cost of existing image forgery detection methods and their susceptibility to format-induced confounding bias, which often leads models to rely on file format artifacts rather than genuine manipulation traces. To overcome these limitations, the authors propose a lightweight, interpretable feature engineering pipeline that enables efficient image-level forgery screening using only CPU resources and gradient-boosted trees. The approach constructs a 405-dimensional feature vector by integrating multi-scale Error Level Analysis (ELA) across seven JPEG quality levels, cross-quality ELA ratios, and additional handcrafted features, effectively capturing inconsistencies in compression history while mitigating format confounding. Evaluated on the CASIA v2.0 JPEG subset, the model achieves an AUC of 0.990 (95% CI: 0.988–0.991) and an F1 score of 0.905; under source-aware partitioning, it maintains strong performance with an AUC of 0.976, and processes each image in under one second.

0 citationsRead paper

Meta-Benchmarks for Financial-Services LLM Evaluation

Jul 02, 2026

Current leaderboards for large language models rely on global average performance, which inadequately captures their specific capabilities in financial services. This work proposes a meta-benchmarking framework that maps 452 publicly available benchmarks to 41 O*NET Generalized Work Activities and aggregates them into 38 BIAN banking business domains. By integrating discriminative power, coverage, and recency through a dynamic multiplicative weighting mechanism, the framework automatically downweights obsolete benchmarks. Coupled with an Elo pairwise competition algorithm, it yields normalized-free, cross-benchmark comparable scores. Validation on a June 2026 snapshot encompassing 288 models from 25 institutions demonstrates that the framework effectively supports model selection and governance decisions within financial institutions.

0 citationsRead paper

Recent publications

Latest Papers

Class-Aware Reinforcement Learning for Counterfactual Explanation Generation

Jul 30, 2026

This work proposes a class-aware reinforcement learning framework to enhance the efficiency and quality of counterfactual explanation generation. By incorporating the model’s predicted class information into the state representation—a novel design in this domain—the proposed approach guides the policy to more effectively explore counterfactual instances that satisfy validity, sparsity, and proximity constraints. The integration of class awareness significantly accelerates policy convergence and improves reward optimization. Empirical evaluation across seven diverse datasets demonstrates that the method generates a greater number of high-quality explanations with fewer training episodes compared to existing approaches. Furthermore, feature importance analyses using SHAP and LIME confirm that class information plays a critical role in guiding action selection during the generation process.

0 citationsRead paper

Memory Efficient Tabular Foundation Models

Jul 29, 2026

This study addresses the high memory footprint of tabular foundation models during deployment by systematically investigating and validating the effectiveness of model compression techniques for this emerging class of models. Integrating multiple compression strategies, the authors optimize representative architectures such as TabPFN, achieving up to a 7.6× (approximately 87%) reduction in memory usage while preserving near-original in-context learning performance. This work substantially enhances the practical deployability of tabular foundation models and offers a viable solution for their application in resource-constrained environments.

0 citationsRead paper

Beyond Generalist LLMs: Specialist Agentic Systems for Structured Code Workflow Execution

Jul 15, 2026

This work addresses the limitations of general-purpose large language models in business process automation, where inconsistent functionality, frequent tool-calling errors, and unstable code quality hinder industrial-grade reliability and maintainability. Focusing on the task of translating BPMN diagrams into executable agent workflows, we propose the first specialized agent system designed for structured code generation. By integrating BPMN control-flow semantics, a deterministic path execution mechanism, and a lightweight code generation strategy, our approach achieves high-precision, low-latency, and zero-repair automation. Experimental results demonstrate that, compared to general-purpose models, our method improves tool-calling accuracy by 9–20 percentage points, reduces latency by 2–4×, decreases calling errors by a factor of three, lowers token-generation costs by over 95%, and entirely eliminates the need for repair iterations.

0 citationsRead paper

Format-Controlled Multi-Scale JPEG Compression Response Analysis for Image-Level Forgery Screening

Jul 07, 2026

This study addresses the high computational cost of existing image forgery detection methods and their susceptibility to format-induced confounding bias, which often leads models to rely on file format artifacts rather than genuine manipulation traces. To overcome these limitations, the authors propose a lightweight, interpretable feature engineering pipeline that enables efficient image-level forgery screening using only CPU resources and gradient-boosted trees. The approach constructs a 405-dimensional feature vector by integrating multi-scale Error Level Analysis (ELA) across seven JPEG quality levels, cross-quality ELA ratios, and additional handcrafted features, effectively capturing inconsistencies in compression history while mitigating format confounding. Evaluated on the CASIA v2.0 JPEG subset, the model achieves an AUC of 0.990 (95% CI: 0.988–0.991) and an F1 score of 0.905; under source-aware partitioning, it maintains strong performance with an AUC of 0.976, and processes each image in under one second.

0 citationsRead paper

Meta-Benchmarks for Financial-Services LLM Evaluation

Jul 02, 2026

Current leaderboards for large language models rely on global average performance, which inadequately captures their specific capabilities in financial services. This work proposes a meta-benchmarking framework that maps 452 publicly available benchmarks to 41 O*NET Generalized Work Activities and aggregates them into 38 BIAN banking business domains. By integrating discriminative power, coverage, and recency through a dynamic multiplicative weighting mechanism, the framework automatically downweights obsolete benchmarks. Coupled with an Elo pairwise competition algorithm, it yields normalized-free, cross-benchmark comparable scores. Validation on a June 2026 snapshot encompassing 288 models from 25 institutions demonstrates that the framework effectively supports model selection and governance decisions within financial institutions.

0 citationsRead paper