Ziwei Zhu

Xiangjue Dong, Ziwei Zhu, Zhuoer Wang et al. · amazon-science, microsoft-research

Pre-trained Language Models are widely used in many important real-world applications. However, recent studies show that these models can encode social biases from large pre-training corpora and even amplify biases in downstream applications. To address this challenge, we propose Co$^2$PT, an efficient and effective debias-while-prompt tuning method for mitigating biases via counterfactual contrastive prompt tuning on downstream tasks. Our experiments conducted on three extrinsic bias benchmarks demonstrate the effectiveness of Co$^2$PT on bias mitigation during the prompt tuning process and its adaptability to existing upstream debiased language models. These findings indicate the strength of Co$^2$PT and provide promising avenues for further enhancement in bias mitigation on downstream tasks.

Bowen Wei, Nan Wang, Yuqing Zhou et al.

Self-evolving large language models (LLMs) learn by generating their own training tasks and solutions, reducing reliance on human-curated supervision. However, in many reasoning domains, the model must also validate generated tasks and judge generated answers to obtain training signals. This creates a training-signal challenge: erroneous self-judgments become erroneous gradient updates. Existing approaches either rely on external verifiers, which limits generality, or treat noisy self-generated feedback as supervision. We propose COSE (Confidence-Orchestrated Self-Evolution), which uses the LLM's intrinsic confidence as a lightweight uncertainty signal to modulate learning. COSE introduces confidence-weighted PPO updates and confidence-prioritized replay. Across 19 held-out benchmarks and four Qwen/Llama backbones (0.6B--4B), COSE consistently improves over base models and achieves the best average performance in general reasoning and mathematics, while remaining competitive on code. Code and data are available at https://anonymous.4open.science/r/COSE_-B5C2.

Zheyuan Liu, Guangyao Dou, Yijun Tian et al.

Machine Unlearning (MU) algorithms have become increasingly critical due to the imperative adherence to data privacy regulations. The primary objective of MU is to erase the influence of specific data samples on a given model without the need to retrain it from scratch. Accordingly, existing methods focus on maximizing user privacy protection. However, there are different degrees of privacy regulations for each real-world web-based application. Exploring the full spectrum of trade-offs between privacy, model utility, and runtime efficiency is critical for practical unlearning scenarios. Furthermore, designing the MU algorithm with simple control of the aforementioned trade-off is desirable but challenging due to the inherent complex interaction. To address the challenges, we present Controllable Machine Unlearning (ConMU), a novel framework designed to facilitate the calibration of MU. The ConMU framework contains three integral modules: an important data selection module that reconciles the runtime efficiency and model generalization, a progressive Gaussian mechanism module that balances privacy and model generalization, and an unlearning proxy that controls the trade-offs between privacy and runtime efficiency. Comprehensive experiments on various benchmark datasets have demonstrated the robust adaptability of our control mechanism and its superiority over established unlearning methods. ConMU explores the full spectrum of the Privacy-Utility-Efficiency trade-off and allows practitioners to account for different real-world regulations. Source code available at: https://github.com/guangyaodou/ConMU.

Zhuoer Wang, Yicheng Wang, Ziwei Zhu et al. · amazon-science, microsoft-research

Question generation is a widely used data augmentation approach with extensive applications, and extracting qualified candidate answers from context passages is a critical step for most question generation systems. However, existing methods for candidate answer extraction are reliant on linguistic rules or annotated data that face the partial annotation issue and challenges in generalization. To overcome these limitations, we propose a novel unsupervised candidate answer extraction approach that leverages the inherent structure of context passages through a Differentiable Masker-Reconstructor (DMR) Model with the enforcement of self-consistency for picking up salient information tokens. We curated two datasets with exhaustively-annotated answers and benchmark a comprehensive set of supervised and unsupervised candidate answer extraction methods. We demonstrate the effectiveness of the DMR model by showing its performance is superior among unsupervised methods and comparable to supervised methods.

Chahat Raj, Anjishnu Mukherjee, Aylin Caliskan et al.

Large Language Models (LLMs) perpetuate social biases, reflecting prejudices in their training data and reinforcing societal stereotypes and inequalities. Our work explores the potential of the Contact Hypothesis, a concept from social psychology for debiasing LLMs. We simulate various forms of social contact through LLM prompting to measure their influence on the model's biases, mirroring how intergroup interactions can reduce prejudices in social contexts. We create a dataset of 108,000 prompts following a principled approach replicating social contact to measure biases in three LLMs (LLaMA 2, Tulu, and NousHermes) across 13 social bias dimensions. We propose a unique debiasing technique, Social Contact Debiasing (SCD), that instruction-tunes these models with unbiased responses to prompts. Our research demonstrates that LLM responses exhibit social biases when subject to contact probing, but more importantly, these biases can be significantly reduced by up to 40% in 1 epoch of instruction tuning LLaMA 2 following our SCD strategy. Our code and data are available at https://github.com/chahatraj/breakingbias.

Anjishnu Mukherjee, Chahat Raj, Ziwei Zhu et al.

Human biases are ubiquitous but not uniform: disparities exist across linguistic, cultural, and societal borders. As large amounts of recent literature suggest, language models (LMs) trained on human data can reflect and often amplify the effects of these social biases. However, the vast majority of existing studies on bias are heavily skewed towards Western and European languages. In this work, we scale the Word Embedding Association Test (WEAT) to 24 languages, enabling broader studies and yielding interesting findings about LM bias. We additionally enhance this data with culturally relevant information for each language, capturing local contexts on a global scale. Further, to encompass more widely prevalent societal biases, we examine new bias dimensions across toxicity, ableism, and more. Moreover, we delve deeper into the Indian linguistic landscape, conducting a comprehensive regional bias analysis across six prevalent Indian languages. Finally, we highlight the significance of these social biases and the new dimensions through an extensive comparison of embedding methods, reinforcing the need to address them in pursuit of more equitable language models. All code, data and results are available here: https://github.com/iamshnoo/weathub.

Chahat Raj, Anjishnu Mukherjee, Aylin Caliskan et al.

Existing works examining Vision-Language Models (VLMs) for social biases predominantly focus on a limited set of documented bias associations, such as gender:profession or race:crime. This narrow scope often overlooks a vast range of unexamined implicit associations, restricting the identification and, hence, mitigation of such biases. We address this gap by probing VLMs to (1) uncover hidden, implicit associations across 9 bias dimensions. We systematically explore diverse input and output modalities and (2) demonstrate how biased associations vary in their negativity, toxicity, and extremity. Our work (3) identifies subtle and extreme biases that are typically not recognized by existing methodologies. We make the Dataset of retrieved associations, (Dora), publicly available here https://github.com/chahatraj/BiasDora.

Sina Mansouri, Mohit Marvania, Vibhavari Ashok Shihorkar et al.

Medical large language models (LLMs) achieve impressive performance on standardized benchmarks, yet these evaluations fail to capture the complexity of real clinical encounters where patients exhibit memory gaps, limited health literacy, anxiety, and other communication barriers. We introduce VeriSim, a truth-preserving patient simulation framework that injects controllable, clinically evidence-grounded noise into patient responses while maintaining strict adherence to medical ground truth through a hybrid UMLS-LLM verification mechanism. Our framework operationalizes six noise dimensions derived from peer-reviewed medical communication literature, capturing authentic clinical phenomena such as patient recall limitations, health literacy barriers, and stigma-driven non-disclosure. Experiments across seven open-weight LLMs reveal that all models degrade significantly under realistic patient noise, with diagnostic accuracy dropping 15-25% and conversation length increasing 34-55%. Notably, smaller models (7B) show 40% greater degradation than larger models (70B+), while medical fine-tuning on standard corpora provides limited robustness benefits against patient communication noise. Evaluation by board-certified clinicians demonstrates high-quality simulation with strong inter-annotator agreement (kappa > 0.80), while LLM-as-a-Judge serves as a validated auxiliary evaluator achieving comparable reliability for scalable assessment. Our results highlight a critical Sim-to-Real gap in current medical AI. We release VeriSim as an open-source noise-injection framework, establishing a rigorous testbed for evaluating clinical robustness.

Xiangjue Dong, Yun He, Ziwei Zhu et al.

A key component of modern conversational systems is the Dialogue State Tracker (or DST), which models a user's goals and needs. Toward building more robust and reliable DSTs, we introduce a prompt-based learning approach to automatically generate effective adversarial examples to probe DST models. Two key characteristics of this approach are: (i) it only needs the output of the DST with no need for model parameters, and (ii) it can learn to generate natural language utterances that can target any DST. Through experiments over state-of-the-art DSTs, the proposed framework leads to the greatest reduction in accuracy and the best attack success rate while maintaining good fluency and a low perturbation ratio. We also show how much the generated adversarial examples can bolster a DST through adversarial training. These results indicate the strength of prompt-based attacks on DSTs and leave open avenues for continued refinement.

Anjishnu Mukherjee, Ziwei Zhu, Antonios Anastasopoulos

We present a comprehensive three-phase study to examine (1) the cultural understanding of Large Multimodal Models (LMMs) by introducing DalleStreet, a large-scale dataset generated by DALL-E 3 and validated by humans, containing 9,935 images of 67 countries and 10 concept classes; (2) the underlying implicit and potentially stereotypical cultural associations with a cultural artifact extraction task; and (3) an approach to adapt cultural representation in an image based on extracted associations using a modular pipeline, CultureAdapt. We find disparities in cultural understanding at geographic sub-region levels with both open-source (LLaVA) and closed-source (GPT-4V) models on DalleStreet and other existing benchmarks, which we try to understand using over 18,000 artifacts that we identify in association to different countries. Our findings reveal a nuanced picture of the cultural competence of LMMs, highlighting the need to develop culture-aware systems. Dataset and code are available at https://github.com/iamshnoo/crossroads

Yuqing Zhou, Ruixiang Tang, Ziyu Yao et al.

Language models (LMs), despite their advances, often depend on spurious correlations, undermining their accuracy and generalizability. This study addresses the overlooked impact of subtler, more complex shortcuts that compromise model reliability beyond oversimplified shortcuts. We introduce a comprehensive benchmark that categorizes shortcuts into occurrence, style, and concept, aiming to explore the nuanced ways in which these shortcuts influence the performance of LMs. Through extensive experiments across traditional LMs, large language models, and state-of-the-art robust models, our research systematically investigates models' resilience and susceptibilities to sophisticated shortcuts. Our benchmark and code can be found at: https://github.com/yuqing-zhou/shortcut-learning-in-text-classification.

Jinhao Pan, Chahat Raj, Anjishnu Mukherjee et al.

Large language models (LLMs) exhibit social biases that reinforce harmful stereotypes, limiting their safe deployment. Most existing debiasing methods adopt a suppressive paradigm by modifying parameters, prompts, or neurons associated with biased behavior; however, such approaches are often brittle, weakly generalizable, data-inefficient, and prone to degrading general capability. We propose \textbf{KnowBias}, a lightweight and conceptually distinct framework that mitigates bias by strengthening, rather than suppressing, neurons encoding bias-knowledge. KnowBias identifies neurons encoding bias knowledge using a small set of bias-knowledge questions via attribution-based analysis, and selectively enhances them at inference time. This design enables strong debiasing while preserving general capabilities, generalizes across bias types and demographics, and is highly data efficient, requiring only a handful of simple yes/no questions and no retraining. Experiments across multiple benchmarks and LLMs demonstrate consistent state-of-the-art debiasing performance with minimal utility degradation. Data and code are available at https://github.com/JP-25/KnowBias.

Bowen Wei, Yunbei Zhang, Jinhao Pan et al.

Personal AI agents like OpenClaw run with elevated privileges on users' local machines, where a single successful prompt injection can leak credentials, redirect financial transactions, or destroy files. This threat goes well beyond conventional text-level jailbreaks, yet existing safety evaluations fall short: most test models in isolated chat settings, rely on synthetic environments, and do not account for how the agent framework itself shapes safety outcomes. We introduce CLAWSAFETY, a benchmark of 120 adversarial test scenarios organized along three dimensions (harm domain, attack vector, and harmful action type) and grounded in realistic, high-privilege professional workspaces spanning software engineering, finance, healthcare, law, and DevOps. Each test case embeds adversarial content in one of three channels the agent encounters during normal work: workspace skill files, emails from trusted senders, and web pages. We evaluate five frontier LLMs as agent backbones, running 2,520 sandboxed trials across all configurations. Attack success rates (ASR) range from 40\% to 75\% across models and vary sharply by injection vector, with skill instructions (highest trust) consistently more dangerous than email or web content. Action-trace analysis reveals that the strongest model maintains hard boundaries against credential forwarding and destructive actions, while weaker models permit both. Cross-scaffold experiments on three agent frameworks further demonstrate that safety is not determined by the backbone model alone but depends on the full deployment stack, calling for safety evaluation that treats model and framework as joint variables.

Yuqing Zhou, Zhuoer Wang, Jie Yuan et al.

Large language model (LLM)-based agents are widely deployed in user-facing services but remain error-prone in new tasks, tend to repeat the same failure patterns, and show substantial run-to-run variability. Fixing failures via environment-specific training or manual patching is costly and hard to scale. To enable self-evolving agents in user-facing service environments, we propose WISE-Flow, a workflow-centric framework that converts historical service interactions into reusable procedural experience by inducing workflows with prerequisite-augmented action blocks. At deployment, WISE-Flow aligns the agent's execution trajectory to retrieved workflows and performs prerequisite-aware feasibility reasoning to achieve state-grounded next actions. Experiments on ToolSandbox and $τ^2$-bench show consistent improvement across base models.

Anjishnu Mukherjee, Ziwei Zhu, Antonios Anastasopoulos

Large language models are typically trained by treating text as a single global distribution, often resulting in geographically homogenized behavior. We study metadata conditioning as a lightweight approach for localization, pre-training 31 models (at 0.5B and 1B parameter scales) from scratch on large-scale English news data annotated with verified URLs, country tags, and continent tags, covering 4 continents and 17 countries. Across four controlled experiments, we show that metadata conditioning consistently improves in-region performance without sacrificing cross-region generalization, enables global models to recover localization comparable to region-specific models, and improves learning efficiency. Our ablation studies demonstrate that URL-level metadata alone captures much of the geographic signal, while balanced regional data coverage remains essential, as metadata cannot fully compensate for missing regions. Finally, we introduce a downstream benchmark of 800 localized news MCQs and show that after instruction tuning, metadata conditioned global models achieve accuracy comparable to LLaMA-3.2-1B-Instruct, despite being trained on substantially less data. Together, these results establish metadata conditioning as a practical and compute-efficient approach for localization of language models.

Zhuohan Wang, Ziwei Zhu, Ziniu Li et al.

Formulating optimization problems for industrial applications demands significant manual effort and domain expertise. While Large Language Models (LLMs) show promise in automating this process, evaluating their performance remains difficult due to the absence of robust metrics. Existing solver-based approaches often face inconsistency, infeasibility issues, and high computational costs. To address these issues, we propose ORGEval, a graph-theoretic evaluation framework for assessing LLMs' capabilities in formulating linear and mixed-integer linear programs. ORGEval represents optimization models as graphs, reducing equivalence detection to graph isomorphism testing. We identify and prove a sufficient condition, when the tested graphs are symmetric decomposable (SD), under which the Weisfeiler-Lehman (WL) test is guaranteed to correctly detect isomorphism. Building on this, ORGEval integrates a tailored variant of the WL-test with an SD detection algorithm to evaluate model equivalence. By focusing on structural equivalence rather than instance-level configurations, ORGEval is robust to numerical variations. Experimental results show that our method can successfully detect model equivalence and produce 100\% consistent results across random parameter configurations, while significantly outperforming solver-based methods in runtime, especially on difficult problems. Leveraging ORGEval, we construct the Bench4Opt dataset and benchmark state-of-the-art LLMs on optimization modeling. Our results reveal that although optimization modeling remains challenging for all LLMs, DeepSeek-V3 and Claude-Opus-4 achieve the highest accuracies under direct prompting, outperforming even leading reasoning models.

Jinhao Pan, Ziwei Zhu, Jianling Wang et al.

Collaborative filtering (CF) based recommendations suffer from mainstream bias -- where mainstream users are favored over niche users, leading to poor recommendation quality for many long-tail users. In this paper, we identify two root causes of this mainstream bias: (i) discrepancy modeling, whereby CF algorithms focus on modeling mainstream users while neglecting niche users with unique preferences; and (ii) unsynchronized learning, where niche users require more training epochs than mainstream users to reach peak performance. Targeting these causes, we propose a novel end-To-end Adaptive Local Learning (TALL) framework to provide high-quality recommendations to both mainstream and niche users. TALL uses a loss-driven Mixture-of-Experts module to adaptively ensemble experts to provide customized local models for different users. Further, it contains an adaptive weight module to synchronize the learning paces of different users by dynamically adjusting weights in the loss. Extensive experiments demonstrate the state-of-the-art performance of the proposed model. Code and data are provided at \url{https://github.com/JP-25/end-To-end-Adaptive-Local-Leanring-TALL-}

Beidi Dong, Jin R. Lee, Ziwei Zhu et al.

The United States has experienced a significant increase in violent extremism, prompting the need for automated tools to detect and limit the spread of extremist ideology online. This study evaluates the performance of Bidirectional Encoder Representations from Transformers (BERT) and Generative Pre-Trained Transformers (GPT) in detecting and classifying online domestic extremist posts. We collected social media posts containing "far-right" and "far-left" ideological keywords and manually labeled them as extremist or non-extremist. Extremist posts were further classified into one or more of five contributing elements of extremism based on a working definitional framework. The BERT model's performance was evaluated based on training data size and knowledge transfer between categories. We also compared the performance of GPT 3.5 and GPT 4 models using different prompts: naïve, layperson-definition, role-playing, and professional-definition. Results showed that the best performing GPT models outperformed the best performing BERT models, with more detailed prompts generally yielding better results. However, overly complex prompts may impair performance. Different versions of GPT have unique sensitives to what they consider extremist. GPT 3.5 performed better at classifying far-left extremist posts, while GPT 4 performed better at classifying far-right extremist posts. Large language models, represented by GPT models, hold significant potential for online extremism classification tasks, surpassing traditional BERT models in a zero-shot setting. Future research should explore human-computer interactions in optimizing GPT models for extremist detection and classification tasks to develop more efficient (e.g., quicker, less effort) and effective (e.g., fewer errors or mistakes) methods for identifying extremist content.

Jinhao Pan, Bowen Wei, Ziwei Zhu

Most deep learning recommendation models operate as black boxes, relying on latent representations that obscure their decision process. This lack of intrinsic interpretability raises concerns in applications that require transparency and accountability. In this work, we propose a Logical-rule Interpretable Autoencoder (LIA) for collaborative filtering that is interpretable by design. LIA introduces a learnable logical rule layer in which each rule neuron is equipped with a gate parameter that automatically selects between AND and OR operators during training, enabling the model to discover diverse logical patterns directly from data. To support functional completeness without doubling the input dimensionality, LIA encodes negation through the sign of connection weights, providing a parameter-efficient mechanism for expressing both positive and negated item conditions within each rule. By learning explicit, human-readable reconstruction rules, LIA allows users to directly trace the decision process behind each recommendation. Extensive experiments show that our method achieves improved recommendation performance over traditional baselines while remaining fully interpretable. Code and data are available at https://github.com/weibowen555/LIA.

Jinhao Pan, Chahat Raj, Ziyu Yao et al.

Large Language Models (LLMs) often exhibit social biases inherited from their training data. While existing benchmarks evaluate bias by term-based mode through direct term associations between demographic terms and bias terms, LLMs have become increasingly adept at avoiding biased responses, leading to seemingly low levels of bias. However, biases persist in subtler, contextually hidden forms that traditional benchmarks fail to capture. We introduce the Description-based Bias Benchmark (DBB), a novel dataset designed to assess bias at the semantic level that bias concepts are hidden within naturalistic, subtly framed contexts in real-world scenarios rather than superficial terms. We analyze six state-of-the-art LLMs, revealing that while models reduce bias in response at the term level, they continue to reinforce biases in nuanced settings. Data, code, and results are available at https://github.com/JP-25/Description-based-Bias-Benchmark.

Ziwei Zhu, Changhao Shan, Bihong Zhang et al.

Spoken languages show significant variation across mandarin and accent. Despite the high performance of mandarin automatic speech recognition (ASR), accent ASR is still a challenge task. In this paper, we introduce meta-learning techniques for fast accent domain expansion in mandarin speech recognition, which expands the field of accents without deteriorating the performance of mandarin ASR. Meta-learning or learn-to-learn can learn general relation in multi domains not only for over-fitting a specific domain. So we select meta-learning in the domain expansion task. This more essential learning will cause improved performance on accent domain extension tasks. We combine the methods of meta learning and freeze of model parameters, which makes the recognition performance more stable in different cases and the training faster about 20%. Our approach significantly outperforms other methods about 3% relatively in the accent domain expansion task. Compared to the baseline model, it improves relatively 37% under the condition that the mandarin test set remains unchanged. In addition, it also proved this method to be effective on a large amount of data with a relative performance improvement of 4% on the accent test set.

Liyang He, Yuren Zhang, Ziwei Zhu et al.

Retrieval Augmented Generation (RAG) systems are increasingly vital in dynamic domains like online gaming, yet the lack of a dedicated benchmark has impeded standardized evaluation in this area. The core difficulty lies in Dual Dynamics: the constant interplay between game content updates and the shifting focus of the player community. Furthermore, the necessity of automating such a benchmark introduces a critical requirement for player-centric authenticity to ensure generated questions are realistic. To address this integrated challenge, we introduce ChronoPlay, a novel framework for the automated and continuous generation of game RAG benchmarks. ChronoPlay utilizes a dual-dynamic update mechanism to track both forms of change, and a dual-source synthesis engine that draws from official sources and player community to ensure both factual correctness and authentic query patterns. We instantiate our framework on three distinct games to create the first dynamic RAG benchmark for the gaming domain, offering new insights into model performance under these complex and realistic conditions. Code is avaliable at: https://github.com/hly1998/ChronoPlay.

Jinhao Pan, Chahat Raj, Ziwei Zhu

Social biases embedded in Large Language Models (LLMs) raise critical concerns, resulting in representational harms -- unfair or distorted portrayals of demographic groups -- that may be expressed in subtle ways through generated language. Existing evaluation methods often depend on predefined identity-concept associations, limiting their ability to surface new or unexpected forms of bias. In this work, we present the Bias Association Discovery Framework (BADF), a systematic approach for extracting both known and previously unrecognized associations between demographic identities and descriptive concepts from open-ended LLM outputs. Through comprehensive experiments spanning multiple models and diverse real-world contexts, BADF enables robust mapping and analysis of the varied concepts that characterize demographic identities. Our findings advance the understanding of biases in open-ended generation and provide a scalable tool for identifying and analyzing bias associations in LLMs. Data, code, and results are available at https://github.com/JP-25/Discover-Open-Ended-Generation

Yun He, Haochen Chen, Ziwei Zhu et al.

We propose PsiRec, a novel user preference propagation recommender that incorporates pseudo-implicit feedback for enriching the original sparse implicit feedback dataset. Three of the unique characteristics of PsiRec are: (i) it views user-item interactions as a bipartite graph and models pseudo-implicit feedback from this perspective; (ii) its random walks-based approach extracts graph structure information from this bipartite graph, toward estimating pseudo-implicit feedback; and (iii) it adopts a Skip-gram inspired measure of confidence in pseudo-implicit feedback that captures the pointwise mutual information between users and items. This pseudo-implicit feedback is ultimately incorporated into a new latent factor model to estimate user preference in cases of extreme sparsity. PsiRec results in improvements of 21.5% and 22.7% in terms of Precision@10 and Recall@10 over state-of-the-art Collaborative Denoising Auto-Encoders. Our implementation is available at https://github.com/heyunh2015/PsiRecICDM2018.

Mehrdad Fazli, Bowen Wei, Ziwei Zhu

Hallucinations, generating responses inconsistent with the visual input, remain a critical limitation of large vision-language models (LVLMs), especially in open-ended tasks such as image captioning and visual reasoning. In this work, we probe the layer-wise generation dynamics that drive hallucinations and propose a training-free mitigation strategy. Employing the Logit Lens, we examine how LVLMs construct next-token distributions across decoder layers, uncovering a pronounced commitment-depth gap: truthful tokens accumulate probability mass on their final candidates earlier than hallucinatory ones. Drawing on this discovery, we introduce Context Embedding Injection (CEI), a lightweight method that harnesses the hidden state of the last input token-the context embedding-as a grounding signal to maintain visual fidelity throughout decoding and curb hallucinations. Evaluated on the CHAIR, AMBER, and MMHal-Bench benchmarks (with a maximum token length of 512), CEI outperforms state-of-the-art baselines across three LVLMs, with its dynamic variant yielding the lowest overall hallucination rates. By integrating novel mechanistic insights with a scalable intervention, this work advances the mitigation of hallucinations in LVLMs.

Bowen Wei, Ziwei Zhu

Rule-based neural networks stand out for enabling interpretable classification by learning logical rules for both prediction and interpretation. However, existing models often lack flexibility due to the fixed model structure. Addressing this, we introduce the Normal Form Rule Learner (NFRL) algorithm, leveraging a selective discrete neural network, that treat weight parameters as hard selectors, to learn rules in both Conjunctive Normal Form (CNF) and Disjunctive Normal Form (DNF) for enhanced accuracy and interpretability. Instead of adopting a deep, complex structure, the NFRL incorporates two specialized Normal Form Layers (NFLs) with adaptable AND/OR neurons, a Negation Layer for input negations, and a Normal Form Constraint (NFC) to streamline neuron connections. We also show the novel network architecture can be optimized using adaptive gradient update together with Straight-Through Estimator to overcome the gradient vanishing challenge. Through extensive experiments on 11 datasets, NFRL demonstrates superior classification performance, quality of learned rules, efficiency and interpretability compared to 12 state-of-the-art alternatives. Code and data are available at \url{https://anonymous.4open.science/r/NFRL-27B4/}.

Jinke Ren, Yaping Sun, Hongyang Du et al.

This paper delves into the applications of generative artificial intelligence (GAI) in semantic communication (SemCom) and presents a thorough study. Three popular SemCom systems enabled by classical GAI models are first introduced, including variational autoencoders, generative adversarial networks, and diffusion models. For each system, the fundamental concept of the GAI model, the corresponding SemCom architecture, and the associated literature review of recent efforts are elucidated. Then, a novel generative SemCom system is proposed by incorporating the cutting-edge GAI technology-large language models (LLMs). This system features two LLM-based AI agents at both the transmitter and receiver, serving as "brains" to enable powerful information understanding and content regeneration capabilities, respectively. This innovative design allows the receiver to directly generate the desired content, instead of recovering the bit stream, based on the coded semantic information conveyed by the transmitter. Therefore, it shifts the communication mindset from "information recovery" to "information regeneration" and thus ushers in a new era of generative SemCom. A case study on point-to-point video retrieval is presented to demonstrate the superiority of the proposed generative SemCom system, showcasing a 99.98% reduction in communication overhead and a 53% improvement in retrieval accuracy compared to the traditional communication system. Furthermore, four typical application scenarios for generative SemCom are delineated, followed by a discussion of three open issues warranting future investigation. In a nutshell, this paper provides a holistic set of guidelines for applying GAI in SemCom, paving the way for the efficient implementation of generative SemCom in future wireless networks.

Bowen Wei, Ziwei Zhu

Deep neural networks have achieved remarkable performance in various text-based tasks but often lack interpretability, making them less suitable for applications where transparency is critical. To address this, we propose ProtoLens, a novel prototype-based model that provides fine-grained, sub-sentence level interpretability for text classification. ProtoLens uses a Prototype-aware Span Extraction module to identify relevant text spans associated with learned prototypes and a Prototype Alignment mechanism to ensure prototypes are semantically meaningful throughout training. By aligning the prototype embeddings with human-understandable examples, ProtoLens provides interpretable predictions while maintaining competitive accuracy. Extensive experiments demonstrate that ProtoLens outperforms both prototype-based and non-interpretable baselines on multiple text classification benchmarks. Code and data are available at \url{https://anonymous.4open.science/r/ProtoLens-CE0B/}.

Mamnuya Rinki, Chahat Raj, Anjishnu Mukherjee et al.

Evaluations of Large Language Models (LLMs) often overlook intersectional and culturally specific biases, particularly in underrepresented multilingual regions like South Asia. This work addresses these gaps by conducting a multilingual and intersectional analysis of LLM outputs across 10 Indo-Aryan and Dravidian languages, identifying how cultural stigmas influenced by purdah and patriarchy are reinforced in generative tasks. We construct a culturally grounded bias lexicon capturing previously unexplored intersectional dimensions including gender, religion, marital status, and number of children. We use our lexicon to quantify intersectional bias and the effectiveness of self-debiasing in open-ended generations (e.g., storytelling, hobbies, and to-do lists), where bias manifests subtly and remains largely unexamined in multilingual contexts. Finally, we evaluate two self-debiasing strategies (simple and complex prompts) to measure their effectiveness in reducing culturally specific bias in Indo-Aryan and Dravidian languages. Our approach offers a nuanced lens into cultural bias by introducing a novel bias lexicon and evaluation framework that extends beyond Eurocentric or small-scale multilingual settings.

Bowen Wei, Yuan Shen Tay, Howard Liu et al.

Security Operations Centers (SOCs) are overwhelmed by tens of thousands of daily alerts, with only a small fraction corresponding to genuine attacks. This overload creates alert fatigue, leading to overlooked threats and analyst burnout. Classical detection pipelines are brittle and context-poor, while recent LLM-based approaches typically rely on a single model to interpret logs, retrieve context, and adjudicate alerts end-to-end -- an approach that struggles with noisy enterprise data and offers limited transparency. We propose CORTEX, a multi-agent LLM architecture for high-stakes alert triage in which specialized agents collaborate over real evidence: a behavior-analysis agent inspects activity sequences, evidence-gathering agents query external systems, and a reasoning agent synthesizes findings into an auditable decision. To support training and evaluation, we release a dataset of fine-grained SOC investigations from production environments, capturing step-by-step analyst actions and linked tool outputs. Across diverse enterprise scenarios, CORTEX substantially reduces false positives and improves investigation quality over state-of-the-art single-agent LLMs.

Chahat Raj, Bowen Wei, Aylin Caliskan et al.

While bias in large language models (LLMs) is well-studied, similar concerns in vision-language models (VLMs) have received comparatively less attention. Existing VLM bias studies often focus on portrait-style images and gender-occupation associations, overlooking broader and more complex social stereotypes and their implied harm. This work introduces VIGNETTE, a large-scale VQA benchmark with 30M+ images for evaluating bias in VLMs through a question-answering framework spanning four directions: factuality, perception, stereotyping, and decision making. Beyond narrowly-centered studies, we assess how VLMs interpret identities in contextualized settings, revealing how models make trait and capability assumptions and exhibit patterns of discrimination. Drawing from social psychology, we examine how VLMs connect visual identity cues to trait and role-based inferences, encoding social hierarchies, through biased selections. Our findings uncover subtle, multifaceted, and surprising stereotypical patterns, offering insights into how VLMs construct social meaning from inputs.

Mehrdad Fazli, Bowen Wei, Ahmet Sari et al.

Large vision-language models (LVLMs) achieve impressive performance on multimodal tasks but often suffer from hallucination, and confidently describe objects or attributes not present in the image. Current training-free interventions struggle to maintain accuracy in open-ended and long-form generation scenarios. We introduce the Confidence-Aware Attention Calibration (CAAC) framework to address this challenge by targeting two key biases: spatial perception bias, which distributes attention disproportionately across image tokens, and modality bias, which shifts focus from visual to textual inputs over time. CAAC employs a two-step approach: Visual-Token Calibration (VTC) to balance attention across visual tokens, and Adaptive Attention Re-Scaling (AAR) to reinforce visual grounding guided by the model's confidence. This confidence-driven adjustment ensures consistent visual alignment during generation. Experiments on CHAIR, AMBER, and POPE benchmarks demonstrate that CAAC outperforms baselines, particularly in long-form generations, effectively reducing hallucination.

Bowen Wei, Mehrdad Fazli, Ziwei Zhu

Large language models (LLMs) have demonstrated impressive performance on natural language tasks, but their decision-making processes remain largely opaque. Existing explanation methods either suffer from limited faithfulness to the model's reasoning or produce explanations that humans find difficult to understand. To address these challenges, we propose \textbf{ProtoSurE}, a novel prototype-based surrogate framework that provides faithful and human-understandable explanations for LLMs. ProtoSurE trains an interpretable-by-design surrogate model that aligns with the target LLM while utilizing sentence-level prototypes as human-understandable concepts. Extensive experiments show that ProtoSurE consistently outperforms SOTA explanation methods across diverse LLMs and datasets. Importantly, ProtoSurE demonstrates strong data efficiency, requiring relatively few training examples to achieve good performance, making it practical for real-world applications.

Fardin Ahsan Sakib, Ziwei Zhu, Karen Trister Grace et al.

Social determinants of health (SDOH) extraction from clinical text is critical for downstream healthcare analytics. Although large language models (LLMs) have shown promise, they may rely on superficial cues leading to spurious predictions. Using the MIMIC portion of the SHAC (Social History Annotation Corpus) dataset and focusing on drug status extraction as a case study, we demonstrate that mentions of alcohol or smoking can falsely induce models to predict current/past drug use where none is present, while also uncovering concerning gender disparities in model performance. We further evaluate mitigation strategies - such as prompt engineering and chain-of-thought reasoning - to reduce these false positives, providing insights into enhancing LLM reliability in health domains.

Chahat Raj, Mahika Banerjee, Aylin Caliskan et al.

When a student fails an exam, do we tend to blame their effort or the test's difficulty? Attribution, defined as how reasons are assigned to event outcomes, shapes perceptions, reinforces stereotypes, and influences decisions. Attribution Theory in social psychology explains how humans assign responsibility for events using implicit cognition, attributing causes to internal (e.g., effort, ability) or external (e.g., task difficulty, luck) factors. LLMs' attribution of event outcomes based on demographics carries important fairness implications. Most works exploring social biases in LLMs focus on surface-level associations or isolated stereotypes. This work proposes a cognitively grounded bias evaluation framework to identify how models' reasoning disparities channelize biases toward demographic groups.

Rahul Pandey, Ziwei Zhu, Hemant Purohit

Effective labeled data collection plays a critical role in developing and fine-tuning robust streaming analytics systems. However, continuously labeling documents to filter relevant information poses significant challenges like limited labeling budget or lack of high-quality labels. There is a need for efficient human-in-the-loop machine learning (HITL-ML) design to improve streaming analytics systems. One particular HITL- ML approach is online active learning, which involves iteratively selecting a small set of the most informative documents for labeling to enhance the ML model performance. The performance of such algorithms can get affected due to human errors in labeling. To address these challenges, we propose ORIS, a method to perform Online active learning using Reinforcement learning-based Inclusive Sampling of documents for labeling. ORIS aims to create a novel Deep Q-Network-based strategy to sample incoming documents that minimize human errors in labeling and enhance the ML model performance. We evaluate the ORIS method on emotion recognition tasks, and it outperforms traditional baselines in terms of both human labeling performance and the ML model performance.

Yuqing Zhou, Ziwei Zhu

In text classification tasks, models often rely on spurious correlations for predictions, incorrectly associating irrelevant features with the target labels. This issue limits the robustness and generalization of models, especially when faced with out-of-distribution data where such spurious correlations no longer hold. To address this challenge, we propose the Causally Calibrated Robust Classifier (CCR), which aims to reduce models' reliance on spurious correlations and improve model robustness. Our approach integrates a causal feature selection method based on counterfactual reasoning, along with an unbiased inverse propensity weighting (IPW) loss function. By focusing on selecting causal features, we ensure that the model relies less on spurious features during prediction. We theoretically justify our approach and empirically show that CCR achieves state-of-the-art performance among methods without group labels, and in some cases, it can compete with the models that utilize group labels.

Wen Wang, Shihao Wu, Ziwei Zhu et al.

Fusing regression coefficients into homogenous groups can unveil those coefficients that share a common value within each group. Such groupwise homogeneity reduces the intrinsic dimension of the parameter space and unleashes sharper statistical accuracy. We propose and investigate a new combinatorial grouping approach called $L_0$-Fusion that is amenable to mixed integer optimization (MIO). On the statistical aspect, we identify a fundamental quantity called grouping sensitivity that underpins the difficulty of recovering the true groups. We show that $L_0$-Fusion achieves grouping consistency under the weakest possible requirement of the grouping sensitivity: if this requirement is violated, then the minimax risk of group misspecification will fail to converge to zero. Moreover, we show that in the high-dimensional regime, one can apply $L_0$-Fusion coupled with a sure screening set of features without any essential loss of statistical efficiency, while reducing the computational cost substantially. On the algorithmic aspect, we provide a MIO formulation for $L_0$-Fusion along with a warm start strategy. Simulation and real data analysis demonstrate that $L_0$-Fusion exhibits superiority over its competitors in terms of grouping accuracy.

Jianling Wang, Kaize Ding, Ziwei Zhu et al.

Session-based recommender systems aim to improve recommendations in short-term sessions that can be found across many platforms. A critical challenge is to accurately model user intent with only limited evidence in these short sessions. For example, is a flower bouquet being viewed meant as part of a wedding purchase or for home decoration? Such different perspectives greatly impact what should be recommended next. Hence, this paper proposes a novel session-based recommendation system empowered by hypergraph attention networks. Three unique properties of the proposed approach are: (i) it constructs a hypergraph for each session to model the item correlations defined by various contextual windows in the session simultaneously, to uncover item meanings; (ii) it is equipped with hypergraph attention layers to generate item embeddings by flexibly aggregating the contextual information from correlated items in the session; and (iii) it aggregates the dynamic item representations for each session to infer the general purpose and current need, which is decoded to infer the next interesting item in the session. Through experiments on three benchmark datasets, we find the proposed model is effective in generating informative dynamic item embeddings and providing more accurate recommendations compared to the state-of-the-art.

James Kotary, Ferdinando Fioretto, Pascal Van Hentenryck et al.

The learning-to-rank problem aims at ranking items to maximize exposure of those most relevant to a user query. A desirable property of such ranking systems is to guarantee some notion of fairness among specified item groups. While fairness has recently been considered in the context of learning-to-rank systems, current methods cannot provide guarantees on the fairness of the proposed ranking policies. This paper addresses this gap and introduces Smart Predict and Optimize for Fair Ranking (SPOFR), an integrated optimization and learning framework for fairness-constrained learning to rank. The end-to-end SPOFR framework includes a constrained optimization sub-model and produces ranking policies that are guaranteed to satisfy fairness constraints while allowing for fine control of the fairness-utility tradeoff. SPOFR is shown to significantly improve current state-of-the-art fair learning-to-rank systems with respect to established performance metrics.

Ziwei Zhu, Jianling Wang, James Caverlee

Recommendation algorithms typically build models based on historical user-item interactions (e.g., clicks, likes, or ratings) to provide a personalized ranked list of items. These interactions are often distributed unevenly over different groups of items due to varying user preferences. However, we show that recommendation algorithms can inherit or even amplify this imbalanced distribution, leading to unfair recommendations to item groups. Concretely, we formalize the concepts of ranking-based statistical parity and equal opportunity as two measures of fairness in personalized ranking recommendation for item groups. Then, we empirically show that one of the most widely adopted algorithms -- Bayesian Personalized Ranking -- produces unfair recommendations, which motivates our effort to propose the novel fairness-aware personalized ranking model. The debiased model is able to improve the two proposed fairness metrics while preserving recommendation performance. Experiments on three public datasets show strong fairness improvement of the proposed model versus state-of-the-art alternatives. This is paper is an extended and reorganized version of our SIGIR 2020~\cite{zhu2020measuring} paper. In this paper, we re-frame the studied problem as `item recommendation fairness' in personalized ranking recommendation systems, and provide more details about the training process of the proposed model and details of experiment setup.

Yun He, Ziwei Zhu, Yin Zhang et al.

Knowledge of a disease includes information of various aspects of the disease, such as signs and symptoms, diagnosis and treatment. This disease knowledge is critical for many health-related and biomedical tasks, including consumer health question answering, medical language inference and disease name recognition. While pre-trained language models like BERT have shown success in capturing syntactic, semantic, and world knowledge from text, we find they can be further complemented by specific information like knowledge of symptoms, diagnoses, treatments, and other disease aspects. Hence, we integrate BERT with disease knowledge for improving these important tasks. Specifically, we propose a new disease knowledge infusion training procedure and evaluate it on a suite of BERT models including BERT, BioBERT, SciBERT, ClinicalBERT, BlueBERT, and ALBERT. Experiments over the three tasks show that these models can be enhanced in nearly all cases, demonstrating the viability of disease knowledge infusion. For example, accuracy of BioBERT on consumer health question answering is improved from 68.29% to 72.09%, while new SOTA results are observed in two datasets. We make our data and code freely available.

Ziwei Zhu, Xudong Li, Mengdi Wang et al.

Modeling unknown systems from data is a precursor of system optimization and sequential decision making. In this paper, we focus on learning a Markov model from a single trajectory of states. Suppose that the transition model has a small rank despite of having a large state space, meaning that the system admits a low-dimensional latent structure. We show that one can estimate the full transition model accurately using a trajectory of length that is proportional to the total number of states. We propose two maximum likelihood estimation methods: a convex approach with nuclear-norm regularization and a nonconvex approach with rank constraint. We explicitly derive the statistical rates of both estimators in terms of the Kullback-Leiber divergence and the $\ell_2$ error and also establish a minimax lower bound to assess the tightness of these rates. For computing the nonconvex estimator, we develop a novel DC (difference of convex function) programming algorithm that starts with the convex M-estimator and then successively refines the solution till convergence. Empirical experiments demonstrate consistent superiority of the nonconvex estimator over the convex one.

Ziwei Zhu, Jianling Wang, Yin Zhang et al.

This paper highlights our ongoing efforts to create effective information curator recommendation models that can be personalized for individual users, while maintaining important fairness properties. Concretely, we introduce the problem of information curator recommendation, provide a high-level overview of a fairness-aware recommender, and introduce some preliminary experimental evidence over a real-world Twitter dataset. We conclude with some thoughts on future directions.

Jianqing Fan, Kaizheng Wang, Yiqiao Zhong et al.

Factor models are a class of powerful statistical models that have been widely used to deal with dependent measurements that arise frequently from various applications from genomics and neuroscience to economics and finance. As data are collected at an ever-growing scale, statistical machine learning faces some new challenges: high dimensionality, strong dependence among observed variables, heavy-tailed variables and heterogeneity. High-dimensional robust factor analysis serves as a powerful toolkit to conquer these challenges. This paper gives a selective overview on recent advance on high-dimensional factor models and their applications to statistics including Factor-Adjusted Robust Model selection (FarmSelect) and Factor-Adjusted Robust Multiple testing (FarmTest). We show that classical methods, especially principal component analysis (PCA), can be tailored to many new problems and provide powerful tools for statistical estimation and inference. We highlight PCA and its connections to matrix perturbation theory, robust statistics, random projection, false discovery rate, etc., and illustrate through several applications how insights from these fields yield solutions to modern challenges. We also present far-reaching connections between factor models and popular statistical learning problems, including network analysis and low-rank matrix recovery.