Geometry of Divergence: Tracking Hidden-State Trajectories for Adaptive Multi-Turn Reasoning
研究通过跟踪隐藏状态轨迹和分析几何信号来解决多轮对话中语言模型的推理一致性问题，提高任务成功率并降低成本。
0 citationsRead paper
Selected work
Representative Papers
The Little Book of Generative AI Foundations: An Intuitive Mathematical Primer
This work addresses the lack of intuitive and systematic mathematical introductions in generative AI, which hinders researchers’ understanding of the intrinsic connections and derivation logic among diverse models. By constructing a coherent theoretical pathway, it unifies mainstream approaches—including PCA, variational autoencoders, diffusion models, normalizing flows, autoregressive models, GANs, and their variants—within a common probabilistic and optimization framework. The exposition integrates core tools such as variational inference, optimal transport, energy-based models, and diffusion processes. Presented with both rigor and accessibility, this study elucidates the mathematical foundations of generative modeling, filling a critical gap in pedagogical resources for mathematically novice readers and substantially enhancing the accessibility of foundational principles in generative AI.
0 citationsRead paper
Recommendation Algorithms: A Comparative Study in Movie Domain
This study addresses the limited personalization performance in movie recommendation systems by formulating the recommendation task as a regression problem. Leveraging the Netflix dataset, the authors conduct an exploratory analysis that integrates multi-source features, including aggregated statistics, matrix factorization embeddings, and user–item similarity measures. They innovatively combine these heterogeneous feature types into unified regression models and systematically compare the predictive performance of XGBoost, k-nearest neighbors (KNN), and matrix factorization (implemented via the Surprise library). Using root mean squared error (RMSE) as the evaluation metric, experimental results demonstrate that matrix factorization achieves the best performance on this dataset, thereby validating both the effectiveness of the proposed feature fusion strategy and the practical utility of comparative algorithmic evaluation in recommender systems.
0 citationsRead paper
Reproducibility in Event-Log Research: A Parametrised Generator and Benchmark for Event-based Signatures
This study addresses the challenge of evaluating signature-based cybersecurity detection methods due to the scarcity of publicly available, real-world event logs, which are often restricted for privacy and sensitivity reasons. To overcome this limitation, the authors propose a parameterized synthetic event log generation approach that models known attack signatures to produce labeled, configurable log data reflecting realistic scenarios. Complementing this generator, they introduce a benchmarking framework specifically designed for signature-based detection evaluation. This framework establishes, for the first time, a reproducible and comparable testing environment. Experimental results demonstrate that on the generated benchmark datasets, the DBSCAN clustering algorithm achieves an Adjusted Rand Index exceeding 0.95 in most scenarios, thereby validating both the effectiveness and practical utility of the proposed methodology.
0 citationsRead paper
Why Commodity WiFi Sensors Fail at Multi-Person Gait Identification: A Systematic Analysis Using ESP32
This study investigates the fundamental limitations of low-cost commercial Wi-Fi devices—specifically the ESP32 platform—in multi-person gait recognition. Through a systematic evaluation of six signal separation techniques (FastICA, SOBI, PCA, NMF, wavelet transform, and tensor decomposition) across scenarios involving 1 to 10 individuals, and by introducing a novel set of multidimensional diagnostic metrics, the work demonstrates for the first time that the primary bottleneck lies in the insufficient quality of channel state information (CSI) captured by the hardware, rather than in algorithmic shortcomings. Experimental results show that all methods achieve only 45–56% accuracy (σ = 3.74%), with performance degrading significantly as the number of subjects increases, thereby confirming that the ESP32 is ill-suited for reliable multi-person gait recognition.
0 citationsRead paper
Recent publications
Latest Papers
Geometry of Divergence: Tracking Hidden-State Trajectories for Adaptive Multi-Turn Reasoning
研究通过跟踪隐藏状态轨迹和分析几何信号来解决多轮对话中语言模型的推理一致性问题，提高任务成功率并降低成本。
0 citationsRead paper
The Little Book of Generative AI Foundations: An Intuitive Mathematical Primer
This work addresses the lack of intuitive and systematic mathematical introductions in generative AI, which hinders researchers’ understanding of the intrinsic connections and derivation logic among diverse models. By constructing a coherent theoretical pathway, it unifies mainstream approaches—including PCA, variational autoencoders, diffusion models, normalizing flows, autoregressive models, GANs, and their variants—within a common probabilistic and optimization framework. The exposition integrates core tools such as variational inference, optimal transport, energy-based models, and diffusion processes. Presented with both rigor and accessibility, this study elucidates the mathematical foundations of generative modeling, filling a critical gap in pedagogical resources for mathematically novice readers and substantially enhancing the accessibility of foundational principles in generative AI.
0 citationsRead paper
Recommendation Algorithms: A Comparative Study in Movie Domain
This study addresses the limited personalization performance in movie recommendation systems by formulating the recommendation task as a regression problem. Leveraging the Netflix dataset, the authors conduct an exploratory analysis that integrates multi-source features, including aggregated statistics, matrix factorization embeddings, and user–item similarity measures. They innovatively combine these heterogeneous feature types into unified regression models and systematically compare the predictive performance of XGBoost, k-nearest neighbors (KNN), and matrix factorization (implemented via the Surprise library). Using root mean squared error (RMSE) as the evaluation metric, experimental results demonstrate that matrix factorization achieves the best performance on this dataset, thereby validating both the effectiveness of the proposed feature fusion strategy and the practical utility of comparative algorithmic evaluation in recommender systems.
0 citationsRead paper
Reproducibility in Event-Log Research: A Parametrised Generator and Benchmark for Event-based Signatures
This study addresses the challenge of evaluating signature-based cybersecurity detection methods due to the scarcity of publicly available, real-world event logs, which are often restricted for privacy and sensitivity reasons. To overcome this limitation, the authors propose a parameterized synthetic event log generation approach that models known attack signatures to produce labeled, configurable log data reflecting realistic scenarios. Complementing this generator, they introduce a benchmarking framework specifically designed for signature-based detection evaluation. This framework establishes, for the first time, a reproducible and comparable testing environment. Experimental results demonstrate that on the generated benchmark datasets, the DBSCAN clustering algorithm achieves an Adjusted Rand Index exceeding 0.95 in most scenarios, thereby validating both the effectiveness and practical utility of the proposed methodology.
0 citationsRead paper
Why Commodity WiFi Sensors Fail at Multi-Person Gait Identification: A Systematic Analysis Using ESP32
This study investigates the fundamental limitations of low-cost commercial Wi-Fi devices—specifically the ESP32 platform—in multi-person gait recognition. Through a systematic evaluation of six signal separation techniques (FastICA, SOBI, PCA, NMF, wavelet transform, and tensor decomposition) across scenarios involving 1 to 10 individuals, and by introducing a novel set of multidimensional diagnostic metrics, the work demonstrates for the first time that the primary bottleneck lies in the insufficient quality of channel state information (CSI) captured by the hardware, rather than in algorithmic shortcomings. Experimental results show that all methods achieve only 45–56% accuracy (σ = 3.74%), with performance degrading significantly as the number of subjects increases, thereby confirming that the ESP32 is ill-suited for reliable multi-person gait recognition.
0 citationsRead paper