research

2026

1. Preprint Liquid error: internal

   LMSM: LLM Security Framework Inspired by Linux Security Modules

   XiuYu Zhang , Bonan Ruan, Junfeng Fang, An Zhang, Tat-Seng Chua, and 1 more author

   arXiv preprint, 2026

   Abs DOI arXiv PDF Code

   Summary: LMSM brings a modular security architecture to language-model serving. It separates model-internal evidence, policy decisions, and permission to release generated output, allowing different monitoring methods to share the same enforcement infrastructure. Experiments with Transformers and vLLM show that this separation supports multiple policies and concurrent requests. On Qwen3-4B, the prototype substantially reduces successful harmful prompts while retaining most of the throughput of an unmonitored serving configuration.

2. EMNLP Findings Liquid error: internal

   Self-Evaluation Is Already There: Eliciting Latent Judge Calibration in Base LLMs with Minimal Data

   XiuYu Zhang , Yi Shan, Junfeng Fang, and Zhenkai Liang

   In Findings of the Association for Computational Linguistics: EMNLP 2026, 2026

   Abs DOI arXiv PDF Code

   Summary: This work studies whether language models can anticipate the quality scores assigned to their own answers by external model judges. Self-Evaluation Elicitation alternates reinforcement learning with masked distillation to improve this calibration using 160 examples. Experiments across three benchmarks find that base models already possess useful self-evaluation ability, and that the proposed training can expose it while maintaining response quality and transferring to judges outside the training setup.

3. Preprint Liquid error: internal

   Cosine Misleads: Auxiliary Losses Reshape Vision Language Models, Not Their Latents

   XiuYu Zhang , Junfeng Fang, and Zhenkai Liang

   arXiv preprint, 2026

   Abs DOI arXiv PDF Code

   Summary: This study tests whether aligning supervised latent tokens with visual targets improves reasoning in vision-language models. Across five variants, stronger cosine alignment instead accompanies lower answer accuracy. PRISM combines linear probing with latent corruption to examine where answer information appears and whether generation depends on the latent tokens. The results suggest that auxiliary supervision primarily changes shared language-model parameters, while the supervised latents themselves contribute relatively little to the final answer.

4. ICLR Liquid error: internal

   AlphaAlign: Incentivizing Safety Alignment with Extremely Simplified Reinforcement Learning

   Yi Zhang, An Zhang,  XiuYu Zhang , Leheng Sheng, Yuxin Chen, and 2 more authors

   In The Fourteenth International Conference on Learning Representations, 2026

   Abs DOI arXiv HTML PDF Code

   Large language models (LLMs), despite possessing latent safety understanding from their vast pretraining data, remain vulnerable to generating harmful content and exhibit issues such as over-refusal and utility degradation after safety alignment. Current safety alignment methods often result in superficial refusal shortcuts or rely on intensive supervision for reasoning-based approaches, failing to fully leverage the model’s intrinsic safety self-awareness. We propose AlphaAlign, a simple yet effective pure reinforcement learning (RL) framework with verifiable safety reward designed to incentivize this latent safety awareness through proactive safety reasoning. AlphaAlign employs a dual-reward system: a verifiable safety reward encourages correctly formatted and explicitly justified refusals for harmful queries while penalizing over-refusals, and a normalized helpfulness reward guides high-quality responses to benign inputs. This allows the model to develop proactive safety reasoning capabilities without depending on supervised safety-specific reasoning data. AlphaAlign demonstrates three key advantages: (1) Simplicity and efficiency, requiring only binary prompt safety labels and minimal RL steps for substantial improvements. (2) Breaking the safety-utility trade-off, by enhancing refusal of harmful content and reducing over-refusals, while simultaneously maintaining or even improving general task performance and robustness to unseen jailbreaks. (3) Deep alignment, fostering proactive safety reasoning that generates explicit safety rationales rather than relying on shallow refusal patterns.

5. Cell Press Blue Liquid error: internal

   An integrated single-cell and spatial proteotranscriptomics atlas of fibroblast-driven immunoregulation within the human adult oral cavity

   Bruno F. Matuck, Khoa L. A. Huynh, Diana Pereira, Quinn T. Easter,  XiuYu Zhang , and 26 more authors

   Cell Press Blue, 2026

   Abs DOI HTML

   Summary: This study combines single-cell RNA sequencing with spatial protein and RNA measurements to map immune organization across adult oral tissues. The atlas identifies fibroblast populations and neighborhoods associated with immune signaling, using AstroSuite to connect cell identities, locations, and predicted interactions. Comparing healthy tissue with chronic periodontitis reveals changes in fibroblast states and their surrounding immune environments. The resulting resource provides a framework for investigating stromal contributions to oral inflammation and identifying candidate therapeutic pathways.

2025

1. Preprint

   STARComm Scalably Detects Emergent Modules of Spatial Cell-Cell Communication in Inflammation and Cancer

   Khoa L. A. Huynh, Bruno F. Matuck, Deiziane Souza,  XiuYu Zhang , Giancarlo Fatobene, and 5 more authors

   bioRxiv preprint, 2025

   Abs DOI HTML

   Summary: STARComm identifies spatial patterns of intercellular signaling from two- and three-dimensional transcriptomic data by grouping co-located ligand-receptor activity into multicellular communication modules. Analysis of more than 14 million cells across eight cancers uncovered shared and tumor-specific signaling programs. In salivary gland samples from chronic graft-versus-host disease, communication modules were associated with survival and highlighted candidate therapeutic pathways, illustrating the potential of spatial communication analysis for biomarker discovery and disease profiling.

2024

1. SFLLM@NeurIPS 2024 Liquid error: internal

   Advancing Conversational Psychotherapy: Integrating Privacy, Dual-Memory, and Domain Expertise with Large Language Models

   XiuYu Zhang , and Zening Luo

   In NeurIPS Workshop on Statistical Frontiers in LLMs and Foundation Models, 2024

   Abs DOI arXiv HTML

   Mental health has increasingly become a global issue that reveals the limitations of traditional conversational psychotherapy, constrained by location, time, expense, and privacy concerns. In response to these challenges, we introduce SoulSpeak, a Large Language Model (LLM)-enabled chatbot designed to democratize access to psychotherapy. SoulSpeak improves upon the capabilities of standard LLM-enabled chatbots by incorporating a novel dual-memory component that combines short-term and long-term context via Retrieval Augmented Generation (RAG) to offer personalized responses while ensuring the preservation of user privacy and intimacy through a dedicated privacy module. In addition, it leverages a counseling chat dataset of therapist-client interactions and various prompting techniques to align the generated responses with psychotherapeutic methods. We introduce two fine-tuned BERT models to evaluate the system against existing LLMs and human therapists: the Conversational Psychotherapy Preference Model (CPPM) to simulate human preference among responses and another to assess response relevance to user input. CPPM is useful for training and evaluating psychotherapy-focused language models independent from SoulSpeak, helping with the constrained resources available for psychotherapy. Furthermore, the effectiveness of the dual-memory component and the robustness of the privacy module are also examined. Our findings highlight the potential and challenge of enhancing mental health care by offering an alternative that combines the expertise of traditional therapy with the advantages of LLMs, providing a promising way to address the accessibility and personalization gap in current mental health services.

2. IEEE TMM Liquid error: internal

   Learning-Based Auction for Matching Demand and Supply of Holographic Digital Twin Over Immersive Communications

   XiuYu Zhang , Minrui Xu, Rui Tan, and Dusit Niyato

   IEEE Transactions on Multimedia, 2024

   Abs DOI HTML

   Digital Twin (DT) technologies create digital models of physical entities frequently in multimedia forms, which are crucial for concurrent simulation and analysis of real-world systems. In displaying DTs, Holographic-Type Communication (HTC) provides immersive multimedia access for users to interact with Holographic DTs (HDTs) by transmitting holographic data such as Light Field (LF) and other multisensory information. HDT has applications in remote education, work, and social interactions. However, the effective matching of demand and supply between HDT users and providers remains a challenge. To address this issue, we propose a hierarchical architecture that integrates the DT and HTC paradigms. This architecture incorporates a marketplace for HDT services, leveraging a formulated Double Dutch Auction (DDA) mechanism to optimize matching and pricing based on user and provider valuation. Furthermore, We employ an actor-critic-based Deep Reinforcement Learning (DRL) algorithm to train a DDA auctioneer that dynamically adjusts auction clocks during the auction process. As an alternative to the Multi-layer Perceptron (MLP), we experiment with a Deep Simplistic Variational Quantum Circuit (DSVQC) to reduce the number of parameters and enhance performance stability. Our simulations reveal that the proposed learning-based auctioneer achieves 92% optimal social welfare at a 37% auction information exchange cost for an MLP-based actor and 99% optimal social welfare at a 77% auction information exchange cost for a DSVQC-based actor.

3. Preprint Liquid error: internal

   Partially Conditioned Patch Parallelism for Accelerated Diffusion Model Inference

   XiuYu Zhang , Zening Luo, and Michelle E. Lu

   Preprint, 2024

   Abs DOI arXiv

   Diffusion models have exhibited exciting capabilities in generating images and are also very promising for video creation. However, the inference speed of diffusion models is limited by the slow sampling process, restricting its use cases. The sequential denoising steps required for generating a single sample could take tens or hundreds of iterations and thus have become a significant bottleneck. This limitation is more salient for applications that are interactive in nature or require small latency. To address this challenge, we propose Partially Conditioned Patch Parallelism (PCPP) to accelerate the inference of high-resolution diffusion models. Using the fact that the difference between the images in adjacent diffusion steps is nearly zero, Patch Parallelism (PP) leverages multiple GPUs communicating asynchronously to compute patches of an image in multiple computing devices based on the entire image (all patches) in the previous diffusion step. PCPP develops PP to reduce computation in inference by conditioning only on parts of the neighboring patches in each diffusion step, which also decreases communication among computing devices. As a result, PCPP decreases the communication cost by around 70% compared to DistriFusion (the state of the art implementation of PP) and achieves 2.36-8.02 X inference speed-up using 4-8 GPUs compared to 2.32-6.72 X achieved by DistriFusion depending on the computing device configuration and resolution of generation at the cost of a possible decrease in image quality. PCPP demonstrates the potential to strike a favorable trade-off, enabling high-quality image generation with substantially reduced latency.

4. Report Liquid error: internal

   CLAS : A Machine Learning Enhanced Framework for Exploring Large 3D Design Datasets

   XiuYu Zhang , Xiaolei Ye, Jui–Che Chang, and Yue Fang

   Try the demo by clicking at the Website bottom. Please note that the first prompt entered may not return any results. Please wait for a few minutes and try again to let the model boot up , 2024

   Abs DOI arXiv Website

   Three-dimensional (3D) objects have wide applications. Despite the growing interest in 3D modeling in academia and industries, designing and/or creating 3D objects from scratch remains time-consuming and challenging. With the development of generative artificial intelligence (AI), designers discover a new way to create images for ideation. However, generative AIs are less useful in creating 3D objects with satisfying qualities. To allow 3D designers to access a wide range of 3D objects for creative activities based on their specific demands, we propose a machine learning (ML) enhanced framework CLAS - named after the four-step of capture, label, associate, and search - to enable fully automatic retrieval of 3D objects based on user specifications leveraging the existing datasets of 3D objects. CLAS provides an effective and efficient method for any person or organization to benefit from their existing but not utilized 3D datasets. In addition, CLAS may also be used to produce high-quality 3D object synthesis datasets for training and evaluating 3D generative models. As a proof of concept, we created and showcased a search system with a web user interface (UI) for retrieving 6,778 3D objects of chairs in the ShapeNet dataset powered by CLAS. In a close-set retrieval setting, our retrieval method achieves a mean reciprocal rank (MRR) of 0.58, top 1 accuracy of 42.27%, and top 10 accuracy of 89.64%.