Fangzhou Xie

Yuanhe Zhang, Fangzhou Xie, Zhenhong Zhou et al.

Large Language Diffusion Models (LLDMs) exhibit comparable performance to LLMs while offering distinct advantages in inference speed and mathematical reasoning tasks.The precise and rapid generation capabilities of LLDMs amplify concerns of harmful generations, while existing jailbreak methodologies designed for Large Language Models (LLMs) prove limited effectiveness against LLDMs and fail to expose safety vulnerabilities.Successful defense cannot definitively resolve harmful generation concerns, as it remains unclear whether LLDMs possess safety robustness or existing attacks are incompatible with diffusion-based architectures.To address this, we first reveal the vulnerability of LLDMs to jailbreak and demonstrate that attack failure in LLDMs stems from fundamental architectural differences.We present a PArallel Decoding jailbreak (PAD) for diffusion-based language models. PAD introduces Multi-Point Attention Attack, which guides parallel generative processes toward harmful outputs that inspired by affirmative response patterns in LLMs. Experimental evaluations across four LLDMs demonstrate that PAD achieves jailbreak attack success rates by 97%, revealing significant safety vulnerabilities. Furthermore, compared to autoregressive LLMs of the same size, LLDMs increase the harmful generation speed by 2x, significantly highlighting risks of uncontrolled misuse.Through comprehensive analysis, we provide an investigation into LLDM architecture, offering critical insights for the secure deployment of diffusion-based language models.

Fangzhou Xie

In this study, a new R package, \texttt{rethnicity} is provided for predicting ethnicity based on names. The Bidirectional LSTM and Florida Voter Registration were used as the model and training data, respectively. Special care was given for the accuracy of minority groups, by adjusting the imbalance in the dataset. The models were trained and exported to C++ and then integrated with R using Rcpp. Additionally, the availability, accuracy, and performance of the package were compared with other solutions.

Fangzhou Xie

Recent proposal of Wasserstein Index Generation model (WIG) has shown a new direction for automatically generating indices. However, it is challenging in practice to fit large datasets for two reasons. First, the Sinkhorn distance is notoriously expensive to compute and suffers from dimensionality severely. Second, it requires to compute a full $N\times N$ matrix to be fit into memory, where $N$ is the dimension of vocabulary. When the dimensionality is too large, it is even impossible to compute at all. I hereby propose a Lasso-based shrinkage method to reduce dimensionality for the vocabulary as a pre-processing step prior to fitting the WIG model. After we get the word embedding from Word2Vec model, we could cluster these high-dimensional vectors by $k$-means clustering, and pick most frequent tokens within each cluster to form the "base vocabulary". Non-base tokens are then regressed on the vectors of base token to get a transformation weight and we could thus represent the whole vocabulary by only the "base tokens". This variant, called pruned WIG (pWIG), will enable us to shrink vocabulary dimension at will but could still achieve high accuracy. I also provide a \textit{wigpy} module in Python to carry out computation in both flavor. Application to Economic Policy Uncertainty (EPU) index is showcased as comparison with existing methods of generating time-series sentiment indices.

Fangzhou Xie

I propose a novel method, the Wasserstein Index Generation model (WIG), to generate a public sentiment index automatically. To test the model`s effectiveness, an application to generate Economic Policy Uncertainty (EPU) index is showcased.