Selected work

Representative Papers

Modelling the term-structure of default risk under IFRS 9 within a multistate regression framework

Feb 20, 2025

This paper addresses the core challenge of inaccurate modeling of the term structure of default risk across the loan lifecycle, which undermines the timeliness and accuracy of expected credit loss (ECL) provisioning under IFRS 9. To this end, we propose a multi-state regression modeling paradigm. Methodologically, we integrate semi-Markov processes, macro-micro jointly driven transition probability modeling, Beta regression, and multinomial logistic regression to construct a three-tiered comparative framework with progressively increasing complexity. We further introduce a novel lightweight model diagnostic toolkit to systematically assess sample representativeness and multi-state model performance. Empirical validation on residential mortgage data demonstrates that increased model complexity significantly enhances predictive accuracy. Moreover, the diagnostic toolkit exhibits strong cross-scenario generalizability, thereby improving both the standardization and practical applicability of multi-state credit risk modeling in banking institutions.

1 citationsRead paper

Joint Beamforming Design and Port Selection in Fluid Antenna-Assisted Multi-Cell Networks: A Personalized Federated Learning Approach

Aug 26, 2026

本文通过一种基于位置感知双分支深度神经网络的联邦表示学习方法，解决了多小区流天线辅助网络中的联合波束成形设计和端口选择问题，以最大化加权和速率。

0 citationsRead paper

Blast Radius

Aug 07, 2026

This work addresses the high computational cost and inefficient token consumption in agent programming caused by poor context management. The authors propose a predictive memory management layer that models contexts in a Polish context space, coupling contextual and code channels to assess prompt reachability and incorporating a reversible context eviction mechanism. They introduce an innovative archiving strategy termed NECROPHORESIS, enabling byte-level reversible storage of dead contexts, and integrate a Recurring Dead Matter (RDM) algorithm to identify and bury redundant, useless records. Experimental results across seven OpenAI models demonstrate a 17–26% reduction in token consumption—the lowest overflow rate among compared methods—and reveal that 378 of 450 buried records were duplicate dead content, none of which required recall, thereby validating both the efficacy and reversibility of the proposed approach.

0 citationsRead paper

Harnessing the Potential of Optimizing Data Mixtures via Bayesian Domain Reweighting

Jul 30, 2026

This work addresses the challenge of optimizing domain mixture ratios in large language model pretraining, where existing methods often rely on strong structural assumptions—such as rank preservation or scaling laws—leading to unstable optimization or high computational costs. To overcome these limitations, the authors propose a Bayesian inference–based domain reweighting approach that learns a Gamma prior from observed data and infers optimal domain weights under a Dirichlet distribution, without requiring such restrictive assumptions. This method substantially reduces both data and computational requirements, enabling efficient identification of high-performing mixture strategies with significantly less training data. Empirical results demonstrate its superiority over current search-based function-fitting approaches in both efficiency and performance.

0 citationsRead paper

From Spatial Semantics to Temporal Context: Leveraging Gaze Trajectory for Weakly Supervised Medical Image Segmentation

Jul 29, 2026

This work addresses the high cost of pixel-level annotation in medical image segmentation by proposing TrailNet, a novel model that leverages clinicians’ eye-tracking trajectories as the sole weak supervision signal. TrailNet jointly models the temporal dynamics of gaze patterns and spatial image semantics through a trajectory-guided spatiotemporal encoder that integrates fixation points with their contextual trajectories. To enhance robustness and suppress gaze-related noise, the method incorporates a multi-scale uncertainty-aware decoder, class-exclusive constraints, and a cyclic knowledge distillation mechanism. Notably, the trained model operates without requiring eye-tracking hardware during inference. Evaluated on two public datasets, TrailNet achieves Dice scores of 81.25% and 81.85%, respectively, significantly outperforming current state-of-the-art approaches.

0 citationsRead paper

Recent publications

Latest Papers

Joint Beamforming Design and Port Selection in Fluid Antenna-Assisted Multi-Cell Networks: A Personalized Federated Learning Approach

Aug 26, 2026

本文通过一种基于位置感知双分支深度神经网络的联邦表示学习方法，解决了多小区流天线辅助网络中的联合波束成形设计和端口选择问题，以最大化加权和速率。

0 citationsRead paper

Blast Radius

Aug 07, 2026

This work addresses the high computational cost and inefficient token consumption in agent programming caused by poor context management. The authors propose a predictive memory management layer that models contexts in a Polish context space, coupling contextual and code channels to assess prompt reachability and incorporating a reversible context eviction mechanism. They introduce an innovative archiving strategy termed NECROPHORESIS, enabling byte-level reversible storage of dead contexts, and integrate a Recurring Dead Matter (RDM) algorithm to identify and bury redundant, useless records. Experimental results across seven OpenAI models demonstrate a 17–26% reduction in token consumption—the lowest overflow rate among compared methods—and reveal that 378 of 450 buried records were duplicate dead content, none of which required recall, thereby validating both the efficacy and reversibility of the proposed approach.

0 citationsRead paper

Harnessing the Potential of Optimizing Data Mixtures via Bayesian Domain Reweighting

Jul 30, 2026

This work addresses the challenge of optimizing domain mixture ratios in large language model pretraining, where existing methods often rely on strong structural assumptions—such as rank preservation or scaling laws—leading to unstable optimization or high computational costs. To overcome these limitations, the authors propose a Bayesian inference–based domain reweighting approach that learns a Gamma prior from observed data and infers optimal domain weights under a Dirichlet distribution, without requiring such restrictive assumptions. This method substantially reduces both data and computational requirements, enabling efficient identification of high-performing mixture strategies with significantly less training data. Empirical results demonstrate its superiority over current search-based function-fitting approaches in both efficiency and performance.

0 citationsRead paper

From Spatial Semantics to Temporal Context: Leveraging Gaze Trajectory for Weakly Supervised Medical Image Segmentation

Jul 29, 2026

This work addresses the high cost of pixel-level annotation in medical image segmentation by proposing TrailNet, a novel model that leverages clinicians’ eye-tracking trajectories as the sole weak supervision signal. TrailNet jointly models the temporal dynamics of gaze patterns and spatial image semantics through a trajectory-guided spatiotemporal encoder that integrates fixation points with their contextual trajectories. To enhance robustness and suppress gaze-related noise, the method incorporates a multi-scale uncertainty-aware decoder, class-exclusive constraints, and a cyclic knowledge distillation mechanism. Notably, the trained model operates without requiring eye-tracking hardware during inference. Evaluated on two public datasets, TrailNet achieves Dice scores of 81.25% and 81.85%, respectively, significantly outperforming current state-of-the-art approaches.

0 citationsRead paper

SafeStats: Efficient 2PC Protocols for Data Statistic-Related Functions

Jul 28, 2026

This work addresses the challenge of efficiently performing privacy-preserving statistical analysis on sensitive data in domains such as healthcare and finance, where existing two-party computation (2PC) protocols suffer from poor performance for statistical functions. The authors propose SafeStats, an efficient 2PC-oriented toolkit for secure statistical analysis, which introduces three key innovations: an equality-test-free shifted frequency counting technique, a piecewise indicator-based counting sort, and a range-reduction protocol integrating binary search. These designs collectively reduce both computational and communication overhead. Experimental evaluation demonstrates that SafeStats achieves superior performance across 14 common statistical tasks; for instance, it accelerates chi-squared tests by 1.5× and reduces communication volume by 4.2× compared to generic 2PC libraries.

0 citationsRead paper