Selected work

Representative Papers

Length Generalization Bounds for Transformers

Feb 13, 2026

This study investigates the length generalization capability of Transformers—specifically, their ability to generalize to arbitrarily long inputs when trained on sequences of bounded length—and the computability of associated generalization bounds. Leveraging formal language theory and computational complexity analysis, and employing C-RASP, a mathematically rigorous abstraction of the Transformer architecture, the work establishes for the first time that C-RASP models with two or more layers admit no computable length generalization bound. In contrast, both the positive fragment of C-RASP and fixed-precision Transformers possess tight, optimal, and computable exponential generalization bounds. These results uncover fundamental limitations inherent in deep Transformer models regarding length generalization while providing an exact characterization of the conditions under which such bounds remain computable.

1 citationsRead paper

Superimposed Transmission for Cooperative Cellular and Cell-Free Massive MIMO Systems

Jul 14, 2026

This work addresses the challenge of balancing spectral efficiency and fairness for cell-edge users in cooperative cellular and cell-free massive MIMO systems. To this end, the authors propose a superposition transmission scheme based on user classification: after distinguishing near-cell and far-cell users, the base station superimposes additional data symbols intended for near-cell users, while distributed access points jointly decode the signals using successive interference cancellation (SIC). This approach represents the first integration of superposition coding into cooperative massive MIMO architectures, achieving substantial gains in peak spectral efficiency without compromising the performance of cell-edge users. Experimental results demonstrate that the proposed scheme outperforms existing network configurations in terms of system capacity, peak spectral efficiency, and fairness at the cell edge.

0 citationsRead paper

Towards Energy Impact on AI-Powered 6G IoT Networks: Centralized vs. Decentralized

Apr 21, 2026

This study addresses the high energy consumption associated with model training and data transmission in AI-driven 6G Internet of Things (IoT) systems by constructing a testbed on real-world German railway infrastructure. It presents the first quantitative comparison between centralized and distributed learning paradigms, explicitly evaluating the trade-off between energy efficiency and predictive performance. Integrating IoT sensing, empirical energy modeling, and machine learning for predictive maintenance, the research demonstrates that distributed architectures can reduce total power consumption by up to 70% while maintaining approximately 90% of the prediction accuracy achieved by centralized approaches. These findings provide compelling empirical evidence and a practical pathway toward energy-efficient, sustainable 6G-IoT system design.

0 citationsRead paper

Conditional Attribution for Root Cause Analysis in Time-Series Anomaly Detection

Apr 19, 2026

Existing methods for root cause analysis in time series anomaly detection often rely on unrealistic feature perturbations and neglect temporal dynamics and cross-variable dependencies, leading to unreliable attributions. This work proposes a conditional attribution framework that identifies root causes by retrieving normal-state contexts similar to the anomalous context while preserving dependency structures, thereby establishing a more faithful baseline for attribution. The approach innovatively combines the latent space of a variational autoencoder with UMAP manifold embeddings to enable efficient, high-fidelity low-dimensional context retrieval, effectively avoiding out-of-distribution artifacts. Furthermore, it incorporates confidence-aware mechanisms and temporal evaluation metrics to enhance attribution reliability. Evaluated on the SWaT and MSDS benchmarks, the method substantially outperforms existing approaches, achieving significant improvements in root cause identification accuracy, temporal localization precision, and cross-model robustness.

0 citationsRead paper

Granularity-Aware Transfer for Tree Instance Segmentation in Synthetic and Real Forests

Apr 15, 2026

This work addresses the domain shift challenge in forest perception arising from the disparity between synthetic data with fine-grained annotations (e.g., trunk and crown) and real-world data limited to coarse-grained tree-level labels. To this end, the authors introduce the Mixed-Granularity Tree Dataset (MGTD) and a four-stage evaluation protocol. The core contribution is a granularity-aware knowledge distillation mechanism that effectively transfers structural priors from a synthetic-domain teacher model to a real-domain student model trained solely on coarse labels, via logit-space fusion and mask unification strategies. By integrating knowledge distillation, instance segmentation, and multi-granularity alignment, the proposed method significantly improves mask AP in tree instance segmentation—particularly for small and distant trees—and establishes a new benchmark for Sim2Real transfer under label granularity constraints.

0 citationsRead paper

Recent publications

Latest Papers

Superimposed Transmission for Cooperative Cellular and Cell-Free Massive MIMO Systems

Jul 14, 2026

This work addresses the challenge of balancing spectral efficiency and fairness for cell-edge users in cooperative cellular and cell-free massive MIMO systems. To this end, the authors propose a superposition transmission scheme based on user classification: after distinguishing near-cell and far-cell users, the base station superimposes additional data symbols intended for near-cell users, while distributed access points jointly decode the signals using successive interference cancellation (SIC). This approach represents the first integration of superposition coding into cooperative massive MIMO architectures, achieving substantial gains in peak spectral efficiency without compromising the performance of cell-edge users. Experimental results demonstrate that the proposed scheme outperforms existing network configurations in terms of system capacity, peak spectral efficiency, and fairness at the cell edge.

0 citationsRead paper

Towards Energy Impact on AI-Powered 6G IoT Networks: Centralized vs. Decentralized

Apr 21, 2026

This study addresses the high energy consumption associated with model training and data transmission in AI-driven 6G Internet of Things (IoT) systems by constructing a testbed on real-world German railway infrastructure. It presents the first quantitative comparison between centralized and distributed learning paradigms, explicitly evaluating the trade-off between energy efficiency and predictive performance. Integrating IoT sensing, empirical energy modeling, and machine learning for predictive maintenance, the research demonstrates that distributed architectures can reduce total power consumption by up to 70% while maintaining approximately 90% of the prediction accuracy achieved by centralized approaches. These findings provide compelling empirical evidence and a practical pathway toward energy-efficient, sustainable 6G-IoT system design.

0 citationsRead paper

Conditional Attribution for Root Cause Analysis in Time-Series Anomaly Detection

Apr 19, 2026

Existing methods for root cause analysis in time series anomaly detection often rely on unrealistic feature perturbations and neglect temporal dynamics and cross-variable dependencies, leading to unreliable attributions. This work proposes a conditional attribution framework that identifies root causes by retrieving normal-state contexts similar to the anomalous context while preserving dependency structures, thereby establishing a more faithful baseline for attribution. The approach innovatively combines the latent space of a variational autoencoder with UMAP manifold embeddings to enable efficient, high-fidelity low-dimensional context retrieval, effectively avoiding out-of-distribution artifacts. Furthermore, it incorporates confidence-aware mechanisms and temporal evaluation metrics to enhance attribution reliability. Evaluated on the SWaT and MSDS benchmarks, the method substantially outperforms existing approaches, achieving significant improvements in root cause identification accuracy, temporal localization precision, and cross-model robustness.

0 citationsRead paper

Granularity-Aware Transfer for Tree Instance Segmentation in Synthetic and Real Forests

Apr 15, 2026

This work addresses the domain shift challenge in forest perception arising from the disparity between synthetic data with fine-grained annotations (e.g., trunk and crown) and real-world data limited to coarse-grained tree-level labels. To this end, the authors introduce the Mixed-Granularity Tree Dataset (MGTD) and a four-stage evaluation protocol. The core contribution is a granularity-aware knowledge distillation mechanism that effectively transfers structural priors from a synthetic-domain teacher model to a real-domain student model trained solely on coarse labels, via logit-space fusion and mask unification strategies. By integrating knowledge distillation, instance segmentation, and multi-granularity alignment, the proposed method significantly improves mask AP in tree instance segmentation—particularly for small and distant trees—and establishes a new benchmark for Sim2Real transfer under label granularity constraints.

0 citationsRead paper

Preliminary analysis of RGB-NIR Image Registration techniques for off-road forestry environments

Mar 12, 2026

This study addresses the challenge of registering RGB and near-infrared (NIR) images in unstructured forest environments by presenting the first systematic evaluation of multi-scale registration methods. It focuses on the GAN-based NeMAR framework—examined under six training configurations—and the MURF feature alignment model, benchmarked against conventional approaches. The findings reveal that NeMAR struggles to maintain geometric consistency due to instability inherent in GAN training, while MURF achieves robust alignment of large-scale structures but fails to preserve fine-grained details. This work elucidates the fundamental trade-off between geometric consistency and detail retention in multimodal image registration within forested scenes, offering empirical insights and clear directions for developing more robust multi-scale registration strategies.

0 citationsRead paper