Selected work

Representative Papers

Sampling for Region-Aggregated Spatial Scan Statistics

Jul 01, 2026

This study addresses the mismatch between areal-aggregated geographic data and point-based spatial scan statistics, where representing regions by their centroids discards critical spatial information and reduces statistical power. To mitigate this limitation, the authors propose a simple yet scalable preprocessing strategy: uniformly sampling 20–50 points within each region’s geometry and distributing the region’s observed count equally among these points. This approach better preserves the underlying spatial distribution while remaining computationally tractable. Empirical evaluations demonstrate that the method substantially enhances the detection performance of spatial scan statistics on aggregated regional data across diverse scenarios. The authors advocate its adoption as a standard preprocessing step for analyzing areal-aggregated datasets in spatial anomaly detection tasks.

0 citationsRead paper

Mobile Gamer Lifetime Value Prediction via Objective Decomposition and Reconstruction

Oct 09, 2025

Accurate user lifetime value (LTV) prediction in real-time bidding (RTB) remains challenging due to the heavy-tailed distribution of LTV and high sensitivity to outliers, leading to poor predictive stability and accuracy. Method: This paper proposes a target decomposition and reconstruction framework that decouples LTV into two orthogonal subtasks—purchase frequency and average transaction amount—thereby avoiding direct modeling of the skewed LTV distribution. Leveraging purchase behavior patterns specific to mobile gaming users, we design a decomposition-based model, train it on an industrial-scale dataset, and deploy it within TapTap’s RTB system for online A/B testing. Contribution/Results: Experimental results demonstrate that our approach significantly outperforms state-of-the-art baselines—including the Zero-Inflated Log-Normal (ZILN) model—in both prediction accuracy and robustness under real-world advertising conditions. The improved LTV estimation enables more effective dynamic ad inventory allocation and enhances platform-level return on investment (ROI).

0 citationsRead paper

Elucidating the Design Space of Decay in Linear Attention

Sep 05, 2025

This work systematically investigates the design space of decay mechanisms in linear-complexity sequence models, focusing on four key dimensions: parameterization strategies, parameter-sharing schemes, decay granularity (scalar vs. vector), and compatibility with relative positional encodings (e.g., RoPE). Through large-scale empirical evaluation across diverse language modeling tasks using linear attention architectures, we make four novel findings: (1) distinct parameterization strategies exhibit well-defined regimes of effectiveness; (2) parameter sharing frequently induces performance degradation; (3) vector-level decay significantly outperforms scalar decay in long-range modeling—reversing their relative ranking; and (4) RoPE generally fails to improve—and often harms—performance across most linear attention variants. Our study establishes reproducible theoretical insights and practical guidelines for decay mechanism design in efficient sequence modeling.

0 citationsRead paper

Learning Spatial Decay for Vision Transformers

Aug 13, 2025

Vision Transformers (ViTs) lack explicit spatial inductive bias in self-attention, resulting in weak spatial structure modeling. Existing spatial decay methods rely on fixed distance metrics—ignoring image content—and thus suffer from limited generalizability. To address this, we propose a data-dependent spatial decay mechanism and introduce Context-Aware Gating (CAG), the first such gating module for 2D ViTs. CAG jointly encodes spatial priors (Manhattan distance) and learnable content representations to enable dynamic, adaptive inter-patch attention modulation. This establishes a novel paradigm unifying spatial and content-aware dynamic decay. Extensive experiments on ImageNet-1K classification and generative modeling demonstrate significant improvements over strong baselines, validating the method’s effectiveness, robustness, and cross-task generalization capability.

0 citationsRead paper

AlphaGo Moment for Model Architecture Discovery

Jul 23, 2025

AI research is constrained by human cognitive biases that impose prior assumptions on the neural architecture search space. Method: We propose ASIA-Arch, a system enabling a paradigm shift from automated optimization to automated innovation—achieving autonomous hypothesis generation, code implementation, and training-validation闭环 via large-scale experimentation (1,773 trials, 20,000 GPU-hours) and empirical learning. Contribution/Results: For the first time, we empirically establish that scientific breakthroughs in architecture design obey a computational scaling law. Without any human-imposed architectural priors, ASIA-Arch discovers 106 novel linear-attention architectures outperforming state-of-the-art baselines, uncovering emergent design principles beyond human intuition. This constitutes the first demonstration of AI-driven, scientifically grounded, autonomous neural architecture innovation—heralding a self-accelerating era in AI research and development.

0 citationsRead paper

Recent publications

Latest Papers

Sampling for Region-Aggregated Spatial Scan Statistics

Jul 01, 2026

This study addresses the mismatch between areal-aggregated geographic data and point-based spatial scan statistics, where representing regions by their centroids discards critical spatial information and reduces statistical power. To mitigate this limitation, the authors propose a simple yet scalable preprocessing strategy: uniformly sampling 20–50 points within each region’s geometry and distributing the region’s observed count equally among these points. This approach better preserves the underlying spatial distribution while remaining computationally tractable. Empirical evaluations demonstrate that the method substantially enhances the detection performance of spatial scan statistics on aggregated regional data across diverse scenarios. The authors advocate its adoption as a standard preprocessing step for analyzing areal-aggregated datasets in spatial anomaly detection tasks.

0 citationsRead paper

Mobile Gamer Lifetime Value Prediction via Objective Decomposition and Reconstruction

Oct 09, 2025

Accurate user lifetime value (LTV) prediction in real-time bidding (RTB) remains challenging due to the heavy-tailed distribution of LTV and high sensitivity to outliers, leading to poor predictive stability and accuracy. Method: This paper proposes a target decomposition and reconstruction framework that decouples LTV into two orthogonal subtasks—purchase frequency and average transaction amount—thereby avoiding direct modeling of the skewed LTV distribution. Leveraging purchase behavior patterns specific to mobile gaming users, we design a decomposition-based model, train it on an industrial-scale dataset, and deploy it within TapTap’s RTB system for online A/B testing. Contribution/Results: Experimental results demonstrate that our approach significantly outperforms state-of-the-art baselines—including the Zero-Inflated Log-Normal (ZILN) model—in both prediction accuracy and robustness under real-world advertising conditions. The improved LTV estimation enables more effective dynamic ad inventory allocation and enhances platform-level return on investment (ROI).

0 citationsRead paper

Elucidating the Design Space of Decay in Linear Attention

Sep 05, 2025

This work systematically investigates the design space of decay mechanisms in linear-complexity sequence models, focusing on four key dimensions: parameterization strategies, parameter-sharing schemes, decay granularity (scalar vs. vector), and compatibility with relative positional encodings (e.g., RoPE). Through large-scale empirical evaluation across diverse language modeling tasks using linear attention architectures, we make four novel findings: (1) distinct parameterization strategies exhibit well-defined regimes of effectiveness; (2) parameter sharing frequently induces performance degradation; (3) vector-level decay significantly outperforms scalar decay in long-range modeling—reversing their relative ranking; and (4) RoPE generally fails to improve—and often harms—performance across most linear attention variants. Our study establishes reproducible theoretical insights and practical guidelines for decay mechanism design in efficient sequence modeling.

0 citationsRead paper

Learning Spatial Decay for Vision Transformers

Aug 13, 2025

Vision Transformers (ViTs) lack explicit spatial inductive bias in self-attention, resulting in weak spatial structure modeling. Existing spatial decay methods rely on fixed distance metrics—ignoring image content—and thus suffer from limited generalizability. To address this, we propose a data-dependent spatial decay mechanism and introduce Context-Aware Gating (CAG), the first such gating module for 2D ViTs. CAG jointly encodes spatial priors (Manhattan distance) and learnable content representations to enable dynamic, adaptive inter-patch attention modulation. This establishes a novel paradigm unifying spatial and content-aware dynamic decay. Extensive experiments on ImageNet-1K classification and generative modeling demonstrate significant improvements over strong baselines, validating the method’s effectiveness, robustness, and cross-task generalization capability.

0 citationsRead paper

AlphaGo Moment for Model Architecture Discovery

Jul 23, 2025

AI research is constrained by human cognitive biases that impose prior assumptions on the neural architecture search space. Method: We propose ASIA-Arch, a system enabling a paradigm shift from automated optimization to automated innovation—achieving autonomous hypothesis generation, code implementation, and training-validation闭环 via large-scale experimentation (1,773 trials, 20,000 GPU-hours) and empirical learning. Contribution/Results: For the first time, we empirically establish that scientific breakthroughs in architecture design obey a computational scaling law. Without any human-imposed architectural priors, ASIA-Arch discovers 106 novel linear-attention architectures outperforming state-of-the-art baselines, uncovering emergent design principles beyond human intuition. This constitutes the first demonstration of AI-driven, scientifically grounded, autonomous neural architecture innovation—heralding a self-accelerating era in AI research and development.

0 citationsRead paper