Selected work

Representative Papers

Kairos: Numerically Robust News Recommendation under Item Cold-Start via Cholesky-based LinUCB

Jul 29, 2026

This work addresses the item cold-start problem in regional news recommendation, which arises from short news lifecycles and shallow article pools. To tackle this challenge, the authors propose an enhanced LinUCB online contextual recommendation method that replaces the Sherman–Morrison formula with a rank-1 update mechanism based on Cholesky decomposition, thereby improving the numerical stability and robustness of covariance matrix estimation. Additionally, Matryoshka representation learning is integrated to reduce inference latency. Experimental results using the Tagesschau API demonstrate that the proposed approach achieves a favorable trade-off between efficiency and accuracy under data-scarce conditions, delivering a 4.85× speedup in inference with negligible degradation in ranking performance—making it well-suited for environments constrained by limited computational resources and sparse data.

0 citationsRead paper

Variational Quantum Conditional Boltzmann Machines for Time-Series Forecasting: Architectures, Symmetric Hyperparameter Evaluation, and a Nonlinear Benchmark

Jul 27, 2026

This study investigates whether quantum computing can surpass classical conditional restricted Boltzmann machines (CRBMs) in time series forecasting and examines the existence of a quantum advantage under fair hyperparameter configurations. To this end, the authors construct and uniformly evaluate four conditional energy-based models: the classical CRBM, a hybrid quantum-classical QCRBM, a fully quantum QQRBM, and a QFeatureQRBM incorporating lagged features. They derive, for the first time, the conditional distributions and training gradients for these models and employ a symmetric hyperparameter grid search to ensure equitable comparison. Experiments on Gaussian process and NARMA-10 benchmarks reveal no significant performance difference between the hybrid QCRBM and the classical CRBM, while fully quantum models perform worse. No systematic quantum advantage is observed, though a marginal benefit cannot be ruled out due to limited sample sizes. This work establishes a comparable framework and theoretical foundation for applying quantum energy models to time series prediction.

0 citationsRead paper

TINY_SCHILLER: A Drop-In German Drama Corpus for Small Language Models

Jul 22, 2026

This work addresses the current lack of ready-to-use German literary datasets suitable for small language models, as existing corpora typically require extensive preprocessing. To bridge this gap, the authors introduce GerDraCor, a compact (2.07 MB), single-file corpus comprising 11 public-domain Schiller dramas, derived from DraCor under a CC0 license and processed through a deterministic pipeline. GerDraCor provides the first plug-and-play German literary dataset analogous to tiny_shakespeare, supporting character-level, GPT-2 BPE, and cl100k_base tokenization schemes. It further enables persona-based segmentation by dramatic character and is fully compatible with Hugging Face’s datasets library, allowing one-line loading. This resource substantially lowers the barrier to incorporating canonical German texts into training workflows for resource-constrained models.

0 citationsRead paper

!Imperio, smolVLA: The Implications of Data Poisoning on Open Source Robotics

Jul 05, 2026

This study addresses a critical vulnerability in open-source robotics communities: the lack of effective validation for contributed data, which renders them susceptible to trigger-based data poisoning attacks. The work presents the first empirical demonstration that injecting only three poisoned samples (0.94% of the training data) into the smolVLA model is sufficient to implant a stealthy, generalizable, and highly destructive backdoor, reducing real-world grasping task success rates to 0%. Even a single poisoned sample can drastically drop performance to 6.7%, while the model maintains approximately 50% success under normal prompts. Conducted on the LeRobot platform, experiments confirm the attack’s robustness across varying trigger positions, highlighting the severe reliance of open-source vision-language-action models on data provenance trustworthiness and exposing their substantial security risks.

0 citationsRead paper

Learning Temporal Patterns in Financial Time Series: A Comparative Study of Quantum LSTM and Quantum Reservoir Computing

May 04, 2026

This study investigates how quantum computing can enhance the accuracy and robustness of financial time series forecasting, particularly in multivariate settings with strong interdependencies. The authors propose a hybrid quantum-classical architecture based on amplitude encoding and systematically compare two quantum models—Quantum Long Short-Term Memory (QLSTM) and Quantum Reservoir Computing (QRC)—augmented with lagged embedding and variational parameter optimization. Empirical evaluation on real-world financial data demonstrates, for the first time, that with carefully designed lag structures and amplitude encoding, both quantum models match the performance of classical LSTM in univariate tasks and achieve modest yet consistent improvements in multivariate scenarios. These findings underscore the critical role of data encoding strategies and dynamic model structure in effective quantum temporal modeling.

0 citationsRead paper

Recent publications

Latest Papers

Kairos: Numerically Robust News Recommendation under Item Cold-Start via Cholesky-based LinUCB

Jul 29, 2026

This work addresses the item cold-start problem in regional news recommendation, which arises from short news lifecycles and shallow article pools. To tackle this challenge, the authors propose an enhanced LinUCB online contextual recommendation method that replaces the Sherman–Morrison formula with a rank-1 update mechanism based on Cholesky decomposition, thereby improving the numerical stability and robustness of covariance matrix estimation. Additionally, Matryoshka representation learning is integrated to reduce inference latency. Experimental results using the Tagesschau API demonstrate that the proposed approach achieves a favorable trade-off between efficiency and accuracy under data-scarce conditions, delivering a 4.85× speedup in inference with negligible degradation in ranking performance—making it well-suited for environments constrained by limited computational resources and sparse data.

0 citationsRead paper

Variational Quantum Conditional Boltzmann Machines for Time-Series Forecasting: Architectures, Symmetric Hyperparameter Evaluation, and a Nonlinear Benchmark

Jul 27, 2026

This study investigates whether quantum computing can surpass classical conditional restricted Boltzmann machines (CRBMs) in time series forecasting and examines the existence of a quantum advantage under fair hyperparameter configurations. To this end, the authors construct and uniformly evaluate four conditional energy-based models: the classical CRBM, a hybrid quantum-classical QCRBM, a fully quantum QQRBM, and a QFeatureQRBM incorporating lagged features. They derive, for the first time, the conditional distributions and training gradients for these models and employ a symmetric hyperparameter grid search to ensure equitable comparison. Experiments on Gaussian process and NARMA-10 benchmarks reveal no significant performance difference between the hybrid QCRBM and the classical CRBM, while fully quantum models perform worse. No systematic quantum advantage is observed, though a marginal benefit cannot be ruled out due to limited sample sizes. This work establishes a comparable framework and theoretical foundation for applying quantum energy models to time series prediction.

0 citationsRead paper

TINY_SCHILLER: A Drop-In German Drama Corpus for Small Language Models

Jul 22, 2026

This work addresses the current lack of ready-to-use German literary datasets suitable for small language models, as existing corpora typically require extensive preprocessing. To bridge this gap, the authors introduce GerDraCor, a compact (2.07 MB), single-file corpus comprising 11 public-domain Schiller dramas, derived from DraCor under a CC0 license and processed through a deterministic pipeline. GerDraCor provides the first plug-and-play German literary dataset analogous to tiny_shakespeare, supporting character-level, GPT-2 BPE, and cl100k_base tokenization schemes. It further enables persona-based segmentation by dramatic character and is fully compatible with Hugging Face’s datasets library, allowing one-line loading. This resource substantially lowers the barrier to incorporating canonical German texts into training workflows for resource-constrained models.

0 citationsRead paper

!Imperio, smolVLA: The Implications of Data Poisoning on Open Source Robotics

Jul 05, 2026

This study addresses a critical vulnerability in open-source robotics communities: the lack of effective validation for contributed data, which renders them susceptible to trigger-based data poisoning attacks. The work presents the first empirical demonstration that injecting only three poisoned samples (0.94% of the training data) into the smolVLA model is sufficient to implant a stealthy, generalizable, and highly destructive backdoor, reducing real-world grasping task success rates to 0%. Even a single poisoned sample can drastically drop performance to 6.7%, while the model maintains approximately 50% success under normal prompts. Conducted on the LeRobot platform, experiments confirm the attack’s robustness across varying trigger positions, highlighting the severe reliance of open-source vision-language-action models on data provenance trustworthiness and exposing their substantial security risks.

0 citationsRead paper

Learning Temporal Patterns in Financial Time Series: A Comparative Study of Quantum LSTM and Quantum Reservoir Computing

May 04, 2026

This study investigates how quantum computing can enhance the accuracy and robustness of financial time series forecasting, particularly in multivariate settings with strong interdependencies. The authors propose a hybrid quantum-classical architecture based on amplitude encoding and systematically compare two quantum models—Quantum Long Short-Term Memory (QLSTM) and Quantum Reservoir Computing (QRC)—augmented with lagged embedding and variational parameter optimization. Empirical evaluation on real-world financial data demonstrates, for the first time, that with carefully designed lag structures and amplitude encoding, both quantum models match the performance of classical LSTM in univariate tasks and achieve modest yet consistent improvements in multivariate scenarios. These findings underscore the critical role of data encoding strategies and dynamic model structure in effective quantum temporal modeling.

0 citationsRead paper