Selected work

Representative Papers

Attribution-Based Neuron Utility for Plasticity Restoration in Deep Networks

May 07, 2026

Deep neural networks often suffer from loss of plasticity in continual learning due to neuron saturation and unbounded growth of parameter norms, hindering effective acquisition of new tasks. This work proposes the Gradient versus Reference State Discrepancy (GXD) method, which formulates adaptive resetting as an intervention cost estimation problem for the first time. By leveraging reference-based gradient attribution and a first-order Taylor expansion, GXD precisely quantifies the functional cost of resetting individual neurons, enabling identification of inefficient units and guiding adaptive reinitialization. Experimental results demonstrate that GXD significantly outperforms existing activation- or gradient-based proxy methods across diverse continual learning scenarios, effectively restoring model plasticity and learning capacity.

0 citationsRead paper

De Jure: Iterative LLM Self-Refinement for Structured Extraction of Regulatory Rules

Apr 02, 2026

This study addresses the challenge of structuring legal texts, which traditionally relies heavily on manual annotation, by proposing the first fully automated and domain-agnostic regulatory rule extraction pipeline. The approach comprises four stages: document standardization, semantic decomposition, multidimensional evaluation guided by 19 interpretable criteria, and upstream-prioritized iterative refinement under constrained computational budgets—enabling high-quality rule extraction without any labeled data. Innovatively integrating an LLM-as-a-judge mechanism with an auditable self-iterative optimization strategy, the method demonstrates significant performance gains across financial regulation, healthcare, and AI governance domains. Compliance-oriented question answering based on the extracted rules achieves accuracies of 73.8% in single-rule settings and 84.0% in broad-domain retrieval scenarios.

0 citationsRead paper

Agent WARPP: Workflow Adherence via Runtime Parallel Personalization

Jul 23, 2025

In task-oriented dialogue, large language models (LLMs) frequently deviate from prescribed workflows when handling long-horizon, conditional tasks that depend on external tools and user-specific context. This paper proposes WARPP, a training-free modular framework leveraging multi-agent collaboration and runtime personalization to enhance task adherence. Its core innovation is a parallel Personalizer agent that dynamically prunes conditional branches and customizes execution paths in real time, substantially reducing inference overhead and improving tool-call accuracy. WARPP integrates conditional dependency analysis, synthetic data simulation, and LLM-driven automated evaluation. Experiments across five user intents in banking, flight booking, and healthcare domains demonstrate that as task complexity increases, parameter fidelity and tool accuracy improve significantly, while average token consumption decreases.

0 citationsRead paper

Exploring the In-Context Learning Capabilities of LLMs for Money Laundering Detection in Financial Graphs

Jul 19, 2025

This paper addresses the challenges of detecting complex money laundering patterns and ensuring interpretability in financial graphs. Methodologically, it proposes a language-driven few-shot anti-money laundering (AML) analysis framework: suspicious entities’ k-hop subgraphs are retrieved to capture local topological structures, serialized into structured textual prompts, and fed into large language models (LLMs) for in-context reasoning—yielding both suspiciousness scores and natural-language explanations. Its key contributions include a lightweight graph–language co-processing pipeline enabling LLMs to emulate domain-expert reasoning without fine-tuning, supporting red-flag identification and traceable inference. Experiments on synthetic financial knowledge graphs demonstrate accurate detection of canonical laundering patterns—including layering transfers and shell-company nesting—and generation of coherent, credible, attribution-aware explanations. The approach significantly enhances the interpretability and operational utility of AML systems.

0 citationsRead paper

Recent publications

Latest Papers

Attribution-Based Neuron Utility for Plasticity Restoration in Deep Networks

May 07, 2026

Deep neural networks often suffer from loss of plasticity in continual learning due to neuron saturation and unbounded growth of parameter norms, hindering effective acquisition of new tasks. This work proposes the Gradient versus Reference State Discrepancy (GXD) method, which formulates adaptive resetting as an intervention cost estimation problem for the first time. By leveraging reference-based gradient attribution and a first-order Taylor expansion, GXD precisely quantifies the functional cost of resetting individual neurons, enabling identification of inefficient units and guiding adaptive reinitialization. Experimental results demonstrate that GXD significantly outperforms existing activation- or gradient-based proxy methods across diverse continual learning scenarios, effectively restoring model plasticity and learning capacity.

0 citationsRead paper

De Jure: Iterative LLM Self-Refinement for Structured Extraction of Regulatory Rules

Apr 02, 2026

This study addresses the challenge of structuring legal texts, which traditionally relies heavily on manual annotation, by proposing the first fully automated and domain-agnostic regulatory rule extraction pipeline. The approach comprises four stages: document standardization, semantic decomposition, multidimensional evaluation guided by 19 interpretable criteria, and upstream-prioritized iterative refinement under constrained computational budgets—enabling high-quality rule extraction without any labeled data. Innovatively integrating an LLM-as-a-judge mechanism with an auditable self-iterative optimization strategy, the method demonstrates significant performance gains across financial regulation, healthcare, and AI governance domains. Compliance-oriented question answering based on the extracted rules achieves accuracies of 73.8% in single-rule settings and 84.0% in broad-domain retrieval scenarios.

0 citationsRead paper

Agent WARPP: Workflow Adherence via Runtime Parallel Personalization

Jul 23, 2025

In task-oriented dialogue, large language models (LLMs) frequently deviate from prescribed workflows when handling long-horizon, conditional tasks that depend on external tools and user-specific context. This paper proposes WARPP, a training-free modular framework leveraging multi-agent collaboration and runtime personalization to enhance task adherence. Its core innovation is a parallel Personalizer agent that dynamically prunes conditional branches and customizes execution paths in real time, substantially reducing inference overhead and improving tool-call accuracy. WARPP integrates conditional dependency analysis, synthetic data simulation, and LLM-driven automated evaluation. Experiments across five user intents in banking, flight booking, and healthcare domains demonstrate that as task complexity increases, parameter fidelity and tool accuracy improve significantly, while average token consumption decreases.

0 citationsRead paper

Exploring the In-Context Learning Capabilities of LLMs for Money Laundering Detection in Financial Graphs

Jul 19, 2025

This paper addresses the challenges of detecting complex money laundering patterns and ensuring interpretability in financial graphs. Methodologically, it proposes a language-driven few-shot anti-money laundering (AML) analysis framework: suspicious entities’ k-hop subgraphs are retrieved to capture local topological structures, serialized into structured textual prompts, and fed into large language models (LLMs) for in-context reasoning—yielding both suspiciousness scores and natural-language explanations. Its key contributions include a lightweight graph–language co-processing pipeline enabling LLMs to emulate domain-expert reasoning without fine-tuning, supporting red-flag identification and traceable inference. Experiments on synthetic financial knowledge graphs demonstrate accurate detection of canonical laundering patterns—including layering transfers and shell-company nesting—and generation of coherent, credible, attribution-aware explanations. The approach significantly enhances the interpretability and operational utility of AML systems.

0 citationsRead paper