Xiaokai Wei

Danilo Ribeiro, Shen Wang, Xiaofei Ma et al. · amazon-science

Large language models have achieved high performance on various question answering (QA) benchmarks, but the explainability of their output remains elusive. Structured explanations, called entailment trees, were recently suggested as a way to explain and inspect a QA system's answer. In order to better generate such entailment trees, we propose an architecture called Iterative Retrieval-Generation Reasoner (IRGR). Our model is able to explain a given hypothesis by systematically generating a step-by-step explanation from textual premises. The IRGR model iteratively searches for suitable premises, constructing a single entailment step at a time. Contrary to previous approaches, our method combines generation steps and retrieval of premises, allowing the model to leverage intermediate conclusions, and mitigating the input size limit of baseline encoder-decoder models. We conduct experiments using the EntailmentBank dataset, where we outperform existing benchmarks on premise retrieval and entailment tree generation, with around 300% gain in overall correctness.

Xiaokai Wei, Sujan Gonugondla, Wasi Ahmad et al. · amazon-science, ibm-research

ML-powered code generation aims to assist developers to write code in a more productive manner, by intelligently generating code blocks based on natural language prompts. Recently, large pretrained deep learning models have substantially pushed the boundary of code generation and achieved impressive performance. Despite their great power, the huge number of model parameters poses a significant threat to adapting them in a regular software development environment, where a developer might use a standard laptop or mid-size server to develop her code. Such large models incur significant resource usage (in terms of memory, latency, and dollars) as well as carbon footprint. Model compression is a promising approach to address these challenges. Several techniques are proposed to compress large pretrained models typically used for vision or textual data. Out of many available compression techniques, we identified that quantization is mostly applicable for code generation task as it does not require significant retraining cost. As quantization represents model parameters with lower-bit integer (e.g., int8), the model size and runtime latency would both benefit from such int representation. We extensively study the impact of quantized model on code generation tasks across different dimension: (i) resource usage and carbon footprint, (ii) accuracy, and (iii) robustness. To this end, through systematic experiments we find a recipe of quantization technique that could run even a $6$B model in a regular laptop without significant accuracy or robustness degradation. We further found the recipe is readily applicable to code summarization task as well.

Yuantong Li, Xiaokai Wei, Zijian Wang et al. · amazon-science, stanford

People frequently interact with information retrieval (IR) systems, however, IR models exhibit biases and discrimination towards various demographics. The in-processing fair ranking methods provide a trade-offs between accuracy and fairness through adding a fairness-related regularization term in the loss function. However, there haven't been intuitive objective functions that depend on the click probability and user engagement to directly optimize towards this. In this work, we propose the In-Batch Balancing Regularization (IBBR) to mitigate the ranking disparity among subgroups. In particular, we develop a differentiable \textit{normed Pairwise Ranking Fairness} (nPRF) and leverage the T-statistics on top of nPRF over subgroups as a regularization to improve fairness. Empirical results with the BERT-based neural rankers on the MS MARCO Passage Retrieval dataset with the human-annotated non-gendered queries benchmark \citep{rekabsaz2020neural} show that our IBBR method with nPRF achieves significantly less bias with minimal degradation in ranking performance compared with the baseline.

Weizhi Zhang, Xiaokai Wei, Wei-Chieh Huang et al.

Recent advancements in Large Language Models (LLMs) have expanded context windows to million-token scales, yet benchmarks for evaluating memory remain limited to short-session synthetic dialogues. We introduce \textsc{MemoryCD}, the first large-scale, user-centric, cross-domain memory benchmark derived from lifelong real-world behaviors in the Amazon Review dataset. Unlike existing memory datasets that rely on scripted personas to generate synthetic user data, \textsc{MemoryCD} tracks authentic user interactions across years and multiple domains. We construct a multi-faceted long-context memory evaluation pipeline of 14 state-of-the-art LLM base models with 6 memory method baselines on 4 distinct personalization tasks over 12 diverse domains to evaluate an agent's ability to simulate real user behaviors in both single and cross-domain settings. Our analysis reveals that existing memory methods are far from user satisfaction in various domains, offering the first testbed for cross-domain life-long personalization evaluation.

Reza Shirkavand, Xiaokai Wei, Chen Wang et al.

While collaborative filtering delivers predictive accuracy and efficiency, and Large Language Models (LLMs) enable expressive and generalizable reasoning, modern recommendation systems must bring these strengths together. Growing user expectations, such as natural-language queries and transparent explanations, further highlight the need for a unified approach. However, doing so is nontrivial. Collaborative signals are often token-efficient but semantically opaque, while LLMs are semantically rich but struggle to model implicit user preferences when trained only on textual inputs. This paper introduces Item-ID + Oral-language Mixture-of-Experts Language Model (IDIOMoE), which treats item interaction histories as a native dialect within the language space, enabling collaborative signals to be understood in the same way as natural language. By splitting the Feed Forward Network of each block of a pretrained LLM into a separate text expert and an item expert with token-type gating, our method avoids destructive interference between text and catalog modalities. IDIOMoE demonstrates strong recommendation performance across both public and proprietary datasets, while preserving the text understanding of the pretrained model.

Se-eun Yoon, Xiaokai Wei, Yexi Jiang et al.

In this paper, we present a systematic effort to design, evaluate, and implement a realistic conversational recommender system (CRS). The objective of our system is to allow users to input free-form text to request recommendations, and then receive a list of relevant and diverse items. While previous work on synthetic queries augments large language models (LLMs) with 1-3 tools, we argue that a more extensive toolbox is necessary to effectively handle real user requests. As such, we propose a novel approach that equips LLMs with over 10 tools, providing them access to the internal knowledge base and API calls used in production. We evaluate our model on a dataset of real users and show that it generates relevant, novel, and diverse recommendations compared to vanilla LLMs. Furthermore, we conduct ablation studies to demonstrate the effectiveness of using the full range of tools in our toolbox. We share our designs and lessons learned from deploying the system for internal alpha release. Our contribution is the addressing of all four key aspects of a practicable CRS: (1) real user requests, (2) augmenting LLMs with a wide variety of tools, (3) extensive evaluation, and (4) deployment insights.

Chen Wang, Xiaokai Wei, Yexi Jiang et al.

With the vast and dynamic user-generated content on Roblox, creating effective game recommendations requires a deep understanding of game content. Traditional recommendation models struggle with the inconsistent and sparse nature of game text features such as titles and descriptions. Recent advancements in large language models (LLMs) offer opportunities to enhance recommendation systems by analyzing in-game text data. This paper addresses two challenges: generating high-quality, structured text features for games without extensive human annotation, and validating these features to ensure they improve recommendation relevance. We propose an approach that extracts in-game text and uses LLMs to infer attributes such as genre and gameplay objectives from raw player interactions. Additionally, we introduce an LLM-based re-ranking mechanism to assess the effectiveness of the generated text features, enhancing personalization and user satisfaction. Beyond recommendations, our approach supports applications such as user engagement-based integrity detection, already deployed in production. This scalable framework demonstrates the potential of in-game text understanding to improve recommendation quality on Roblox and adapt recommendations to its unique, user-generated ecosystem.

Zheng Hui, Xiaokai Wei, Yexi Jiang et al.

Conversational recommender systems (CRS) have advanced with large language models, showing strong results in domains like movies. These domains typically involve fixed content and passive consumption, where user preferences can be matched by genre or theme. In contrast, games present distinct challenges: fast-evolving catalogs, interaction-driven preferences (e.g., skill level, mechanics, hardware), and increased risk of unsafe responses in open-ended conversation. We propose MATCHA, a multi-agent framework for CRS that assigns specialized agents for intent parsing, tool-augmented retrieval, multi-LLM ranking with reflection, explanation, and risk control which enabling finer personalization, long-tail coverage, and stronger safety. Evaluated on real user request dataset, MATCHA outperforms six baselines across eight metrics, improving Hit@5 by 20%, reducing popularity bias by 24%, and achieving 97.9% adversarial defense. Human and virtual-judge evaluations confirm improved explanation quality and user alignment.

Xisen Jin, Dejiao Zhang, Henghui Zhu et al.

Pretrained language models (PTLMs) are typically learned over a large, static corpus and further fine-tuned for various downstream tasks. However, when deployed in the real world, a PTLM-based model must deal with data distributions that deviate from what the PTLM was initially trained on. In this paper, we study a lifelong language model pretraining challenge where a PTLM is continually updated so as to adapt to emerging data. Over a domain-incremental research paper stream and a chronologically-ordered tweet stream, we incrementally pretrain a PTLM with different continual learning algorithms, and keep track of the downstream task performance (after fine-tuning). We evaluate PTLM's ability to adapt to new corpora while retaining learned knowledge in earlier corpora. Our experiments show distillation-based approaches to be most effective in retaining downstream performance in earlier domains. The algorithms also improve knowledge transfer, allowing models to achieve better downstream performance over the latest data, and improve temporal generalization when distribution gaps exist between training and evaluation because of time. We believe our problem formulation, methods, and analysis will inspire future studies towards continual pretraining of language models.

Xiaokai Wei, Shen Wang, Dejiao Zhang et al.

Pretrained Language Models (PLM) have established a new paradigm through learning informative contextualized representations on large-scale text corpus. This new paradigm has revolutionized the entire field of natural language processing, and set the new state-of-the-art performance for a wide variety of NLP tasks. However, though PLMs could store certain knowledge/facts from training corpus, their knowledge awareness is still far from satisfactory. To address this issue, integrating knowledge into PLMs have recently become a very active research area and a variety of approaches have been developed. In this paper, we provide a comprehensive survey of the literature on this emerging and fast-growing field - Knowledge Enhanced Pretrained Language Models (KE-PLMs). We introduce three taxonomies to categorize existing work. Besides, we also survey the various NLU and NLG applications on which KE-PLM has demonstrated superior performance over vanilla PLMs. Finally, we discuss challenges that face KE-PLMs and also promising directions for future research.

Dejiao Zhang, Feng Nan, Xiaokai Wei et al.

Unsupervised clustering aims at discovering the semantic categories of data according to some distance measured in the representation space. However, different categories often overlap with each other in the representation space at the beginning of the learning process, which poses a significant challenge for distance-based clustering in achieving good separation between different categories. To this end, we propose Supporting Clustering with Contrastive Learning (SCCL) -- a novel framework to leverage contrastive learning to promote better separation. We assess the performance of SCCL on short text clustering and show that SCCL significantly advances the state-of-the-art results on most benchmark datasets with 3%-11% improvement on Accuracy and 4%-15% improvement on Normalized Mutual Information. Furthermore, our quantitative analysis demonstrates the effectiveness of SCCL in leveraging the strengths of both bottom-up instance discrimination and top-down clustering to achieve better intra-cluster and inter-cluster distances when evaluated with the ground truth cluster labels.

Lichao Sun, Lifang He, Zhipeng Huang et al.

Meta-graph is currently the most powerful tool for similarity search on heterogeneous information networks,where a meta-graph is a composition of meta-paths that captures the complex structural information. However, current relevance computing based on meta-graph only considers the complex structural information, but ignores its embedded meta-paths information. To address this problem, we proposeMEta-GrAph-based network embedding models, called MEGA and MEGA++, respectively. The MEGA model uses normalized relevance or similarity measures that are derived from a meta-graph and its embedded meta-paths between nodes simultaneously, and then leverages tensor decomposition method to perform node embedding. The MEGA++ further facilitates the use of coupled tensor-matrix decomposition method to obtain a joint embedding for nodes, which simultaneously considers the hidden relations of all meta information of a meta-graph.Extensive experiments on two real datasets demonstrate thatMEGA and MEGA++ are more effective than state-of-the-art approaches.

Lichao Sun, Xiaokai Wei, Jiawei Zhang et al.

A recent report indicates that there is a new malicious app introduced every 4 seconds. This rapid malware distribution rate causes existing malware detection systems to fall far behind, allowing malicious apps to escape vetting efforts and be distributed by even legitimate app stores. When trusted downloading sites distribute malware, several negative consequences ensue. First, the popularity of these sites would allow such malicious apps to quickly and widely infect devices. Second, analysts and researchers who rely on machine learning based detection techniques may also download these apps and mistakenly label them as benign since they have not been disclosed as malware. These apps are then used as part of their benign dataset during model training and testing. The presence of contaminants in benign dataset can compromise the effectiveness and accuracy of their detection and classification techniques. To address this issue, we introduce PUDROID (Positive and Unlabeled learning-based malware detection for Android) to automatically and effectively remove contaminants from training datasets, allowing machine learning based malware classifiers and detectors to be more effective and accurate. To further improve the performance of such detectors, we apply a feature selection strategy to select pertinent features from a variety of features. We then compare the detection rates and accuracy of detection systems using two datasets; one using PUDROID to remove contaminants and the other without removing contaminants. The results indicate that once we remove contaminants from the datasets, we can significantly improve both malware detection rate and detection accuracy

Xiaokai Wei, Bokai Cao, Philip S. Yu

Multi-view high-dimensional data become increasingly popular in the big data era. Feature selection is a useful technique for alleviating the curse of dimensionality in multi-view learning. In this paper, we study unsupervised feature selection for multi-view data, as class labels are usually expensive to obtain. Traditional feature selection methods are mostly designed for single-view data and cannot fully exploit the rich information from multi-view data. Existing multi-view feature selection methods are usually based on noisy cluster labels which might not preserve sufficient information from multi-view data. To better utilize multi-view information, we propose a method, CDMA-FS, to select features for each view by performing alignment on a cross diffused matrix. We formulate it as a constrained optimization problem and solve it using Quasi-Newton based method. Experiments results on four real-world datasets show that the proposed method is more effective than the state-of-the-art methods in multi-view setting.

Weixiang Shao, Lifang He, Chun-Ta Lu et al.

In the era of big data, it is becoming common to have data with multiple modalities or coming from multiple sources, known as "multi-view data". Multi-view data are usually unlabeled and come from high-dimensional spaces (such as language vocabularies), unsupervised multi-view feature selection is crucial to many applications. However, it is nontrivial due to the following challenges. First, there are too many instances or the feature dimensionality is too large. Thus, the data may not fit in memory. How to select useful features with limited memory space? Second, how to select features from streaming data and handles the concept drift? Third, how to leverage the consistent and complementary information from different views to improve the feature selection in the situation when the data are too big or come in as streams? To the best of our knowledge, none of the previous works can solve all the challenges simultaneously. In this paper, we propose an Online unsupervised Multi-View Feature Selection, OMVFS, which deals with large-scale/streaming multi-view data in an online fashion. OMVFS embeds unsupervised feature selection into a clustering algorithm via NMF with sparse learning. It further incorporates the graph regularization to preserve the local structure information and help select discriminative features. Instead of storing all the historical data, OMVFS processes the multi-view data chunk by chunk and aggregates all the necessary information into several small matrices. By using the buffering technique, the proposed OMVFS can reduce the computational and storage cost while taking advantage of the structure information. Furthermore, OMVFS can capture the concept drifts in the data streams. Extensive experiments on four real-world datasets show the effectiveness and efficiency of the proposed OMVFS method. More importantly, OMVFS is about 100 times faster than the off-line methods.