[







{
  "title": "About Me",
  "url": "/",
  "content": " Hanzhi Zhang is currently a PhD Candidate in Computer Science, directed by Dr. Yunhe Feng in the Department of Computer Science and Engineering (CSE) at the University of North Texas (UNT), Denton, TX, and co-advised by Dr. Kewei Sha in the Department of Data Science, with expected graduation in May 2027. She is a member of the Responsible AI Lab and the LLaVi Lab at UNT.  Her research focuses on Efficient and Sustainable AI Systems, particularly data–model–hardware co-design and hardware-aware, energy-efficient LLM inference. She also explores Reliable and Responsible AI, including multilingual fact verification, hallucination detection and mitigation, trustworthy AI benchmarking, and AI for scientific discovery.  Selected publications                                Dec 2026        B. Fan*, M. Li*, H. Zhang*, S. Dong, N. P. Mareedu, X. Liu, W. Shi, Y. Feng, Y. Huang, and H. Fan.     Towards Visual Query Segmentation in the Wild.     Advances in Neural Information Processing Systems (NeurIPS), 2026.               arXiv   Code  Dataset      BIB     Abstract                     Copy       @misc{fan2026visualquerysegmentationwild,       title={Towards Visual Query Segmentation in the Wild},       author={Bing Fan and Minghao Li and Hanzhi Zhang and Shaohua Dong and Naga Prudhvi Mareedu and Xiaoqiong Liu and Weishi Shi and Yunhe Feng and Yan Huang and Heng Fan},       year={2026},       eprint={2603.08898},       archivePrefix={arXiv},       primaryClass={cs.CV},       url={https://arxiv.org/abs/2603.08898}, }                           In this paper, we introduce visual query segmentation (VQS), a new paradigm of visual query localization (VQL) that aims to segment all pixel-level occurrences of an object of interest in an untrimmed video, given a visual query. Compared with existing VQL locating only the last appearance of a target using bounding boxes, VQS enables more comprehensive (i.e., all object occurrences) and precise (i.e., pixel-level masks) localization, making it more practical for real-world scenarios. To foster research on this task, we introduce VQS-4K, a large-scale benchmark dedicated to VQS. Specifically, VQS-4K consists of 4,111 videos with more than 1.3 million frames, and covers a diverse set of 222 object categories. Each video is paired with a visual query defined by a frame outside search video and its target mask, and annotated with spatial-temporal masklets corresponding to the queried target. To ensure high quality, all videos in VQS-4K are manually labeled with meticulous inspection and iterative refinement. To our best knowledge, VQS-4K is the first benchmark specifically designed for VQS. Furthermore, to stimulate future research, we present a simple yet effective method, named VQ-SAM, which extends SAM 2 by leveraging target-specific and background distractor cues from the video to progressively evolve the memory through a novel multi-stage framework with an adaptive memory generation (AMG) module for VQS, significantly improving the performance. In our experiments on VQS-4K, VQ-SAM achieves promising results and surpasses existing methods, evidencing its effectiveness. With VQS-4K and VQ-SAM, we expect to inspire more future research and applications on VQS. Our benchmark and code are released at https://github.com/VAIL-Fan/VQS.     Source · CC0                                                    Oct 2026        H. Zhang, S. Anjum, H. Fan, W. Zheng, Y. Huang, and Y. Feng.     Poly-FEVER: A Multilingual Hallucination Detection Benchmark.     Findings of the Association for Computational Linguistics (EMNLP Findings), 2026.               Paper    Dataset      BIB     Abstract                     Copy       @inproceedings{zhang-etal-2026-poly,     title = &quot;Poly-{FEVER}: A Multilingual Hallucination Detection Benchmark&quot;,     author = &quot;Zhang, Hanzhi  and       Anjum, Sumera  and       Fan, Heng  and       Zheng, Weijian  and       Huang, Yan  and       Feng, Yunhe&quot;,     editor = &quot;Lee, Hung-Yi  and       Dev, Sunipa  and       Li, Jessy  and       Elliott, Desmond&quot;,     booktitle = &quot;Findings of the {A}ssociation for {C}omputational {L}inguistics: {EMNLP} 2026&quot;,     month = oct,     year = &quot;2026&quot;,     address = &quot;Budapest, Hungary&quot;,     publisher = &quot;Association for Computational Linguistics&quot;,     url = &quot;https://aclanthology.org/2026.findings-emnlp.540/&quot;,     doi = &quot;10.18653/2026.findings-emnlp.540&quot;,     pages = &quot;11353--11369&quot;,     ISBN = &quot;979-8-89176-462-0&quot;,     abstract = &quot;We present Poly-FEVER, an open-source large-scale multilingual benchmark for hallucination detection through controlled factuality judgment in large language models. Poly-FEVER extends FEVER, Climate-FEVER, and SciFact to 77,973 Supported/Refuted claims aligned across 11 languages. We augment each claim with topic metadata from a 22-topic LDA model, enabling topic-aware analysis of multilingual factuality behavior. Translations are produced with Google Cloud Translation and validated by a targeted human audit, GEMBA, and COMETKiwi. We benchmark ChatGPT-3.5, LLaMA-2/3, Phi-3.5, and Mistral across three prompt families and evaluate robustness to model-generated rephrasing. We further analyze resource imbalance using web presence as a proxy for online claim availability, and evaluate retrieval-augmented factuality judgment with DPR over Wikipedia. Results show pronounced cross-lingual disparities under aligned claims, with stronger accuracy in high-resource languages such as English and Chinese and weaker accuracy in lower-resource languages such as Amharic and Tamil. Topic-aware prompting and retrieval provide selective gains, especially in Arabic and Amharic, while their effects vary across languages and can conflict with model priors. Poly-FEVER establishes a rigorous resource for responsible, language-adaptive hallucination detection and language-sensitive factuality evaluation. The proposed Poly-FEVER is available at: https://huggingface.co/datasets/HanzhiZhang/Poly-FEVER.&quot; }                           We present Poly-FEVER, an open-source large-scale multilingual benchmark for hallucination detection through controlled factuality judgment in large language models. Poly-FEVER extends FEVER, Climate-FEVER, and SciFact to 77,973 Supported/Refuted claims aligned across 11 languages. We augment each claim with topic metadata from a 22-topic LDA model, enabling topic-aware analysis of multilingual factuality behavior. Translations are produced with Google Cloud Translation and validated by a targeted human audit, GEMBA, and COMETKiwi. We benchmark ChatGPT-3.5, LLaMA-2/3, Phi-3.5, and Mistral across three prompt families and evaluate robustness to model-generated rephrasing. We further analyze resource imbalance using web presence as a proxy for online claim availability, and evaluate retrieval-augmented factuality judgment with DPR over Wikipedia. Results show pronounced cross-lingual disparities under aligned claims, with stronger accuracy in high-resource languages such as English and Chinese and weaker accuracy in lower-resource languages such as Amharic and Tamil. Topic-aware prompting and retrieval provide selective gains, especially in Arabic and Amharic, while their effects vary across languages and can conflict with model priors. Poly-FEVER establishes a rigorous resource for responsible, language-adaptive hallucination detection and language-sensitive factuality evaluation. The proposed Poly-FEVER is available at: https://huggingface.co/datasets/HanzhiZhang/Poly-FEVER.     Source · CC BY 4.0                                                                 Jul 2025        H. Zhang, H. Fan, K. Sha, Y. Huang, and Y. Feng.     DAM: Dynamic Attention Mask for Long-Context Large Language Model Inference Acceleration.     Findings of the Annual Meeting of the Association for Computational Linguistics (ACL Findings), 2025.               Paper   Code       BIB     Abstract                     Copy       @inproceedings{zhang-etal-2025-dam,     title = &quot;{DAM}: Dynamic Attention Mask for Long-Context Large Language Model Inference Acceleration&quot;,     author = &quot;Zhang, Hanzhi  and       Fan, Heng  and       Sha, Kewei  and       Huang, Yan  and       Feng, Yunhe&quot;,     editor = &quot;Che, Wanxiang  and       Nabende, Joyce  and       Shutova, Ekaterina  and       Pilehvar, Mohammad Taher&quot;,     booktitle = &quot;Findings of the {A}ssociation for {C}omputational {L}inguistics: {ACL} 2025&quot;,     month = jul,     year = &quot;2025&quot;,     address = &quot;Vienna, Austria&quot;,     publisher = &quot;Association for Computational Linguistics&quot;,     url = &quot;https://aclanthology.org/2025.findings-acl.242/&quot;,     doi = &quot;10.18653/v1/2025.findings-acl.242&quot;,     pages = &quot;4663--4676&quot;,     ISBN = &quot;979-8-89176-256-5&quot;,     abstract = &quot;Long-context understanding is crucial for many NLP applications, yet transformers struggle with efficiency due to the quadratic complexity of self-attention. Sparse attention methods alleviate this cost but often impose static, predefined masks, failing to capture heterogeneous attention patterns. This results in suboptimal token interactions, limiting adaptability and retrieval accuracy in long-sequence tasks. This work introduces a dynamic sparse attention mechanism that assigns adaptive masks at the attention-map level, preserving heterogeneous patterns across layers and heads. Unlike existing approaches, our method eliminates the need for fine-tuning and predefined mask structures while maintaining computational efficiency. By learning context-aware attention structures, it achieves high alignment with full-attention models, ensuring minimal performance degradation while reducing memory and compute overhead. This approach provides a scalable alternative to full attention, enabling the practical deployment of large-scale Large Language Models (LLMs) without sacrificing retrieval performance. DAM is available at: \\url{https://github.com/HanzhiZhang-Ulrica/DAM}.&quot; }                           Long-context understanding is crucial for many NLP applications, yet transformers struggle with efficiency due to the quadratic complexity of self-attention. Sparse attention methods alleviate this cost but often impose static, predefined masks, failing to capture heterogeneous attention patterns. This results in suboptimal token interactions, limiting adaptability and retrieval accuracy in long-sequence tasks. This work introduces a dynamic sparse attention mechanism that assigns adaptive masks at the attention-map level, preserving heterogeneous patterns across layers and heads. Unlike existing approaches, our method eliminates the need for fine-tuning and predefined mask structures while maintaining computational efficiency. By learning context-aware attention structures, it achieves high alignment with full-attention models, ensuring minimal performance degradation while reducing memory and compute overhead. This approach provides a scalable alternative to full attention, enabling the practical deployment of large-scale Large Language Models (LLMs) without sacrificing retrieval performance. DAM is available at: https://github.com/HanzhiZhang-Ulrica/DAM.     Source · CC BY 4.0                         All publications &rarr;   Human Elements  Hanzhi Zhang’s personality, as an ENTP-A, is defined by her insatiable curiosity and robust intellectual flexibility. She revels in intellectual challenges and is the quintessential devil’s advocate, skillfully deconstructing arguments and occasionally challenging her own beliefs by exploring alternative viewpoints.  Hanzhi pursues an intriguing blend of interests. She delves into the mystical realms of tarot and astrology, using them as unique tools to influence decisions in programming and research. She also enjoys unraveling the latest gossip and embarks on quests to locate lost items.  Her cultural appetites are satiated by frequent visits to musicals, dramas, museums, galleries, and exhibitions, which enrich her understanding of diverse perspectives.  Photography is another passion of hers, which she humorously describes as a sport. This hobby sees her traversing mountains, laden with heavy gear, tirelessly capturing moments through day and night—an endurance exercise that combines her love for nature and artistry.  "
}





,{
  "title": "Archive Layout with Content",
  "url": "/archive-layout-with-content/",
  "content": " A variety of common markup showing how the theme styles them.  Header one  Header two  Header three  Header four  Header five  Header six  Blockquotes  Single line blockquote:     Quotes are cool.   Tables                 Entry       Item                               John Doe       2016       Description of the item in the list                 Jane Doe       2019       Description of the item in the list                 Doe Doe       2022       Description of the item in the list                          Header1       Header2       Header3                       cell1       cell2       cell3                 cell4       cell5       cell6                       cell1       cell2       cell3                 cell4       cell5       cell6                       Foot1       Foot2       Foot3           Definition Lists     Definition List Title   Definition list division.   Startup   A startup company or startup is a company or temporary organization designed to search for a repeatable and scalable business model.   #dowork   Coined by Rob Dyrdek and his personal body guard Christopher “Big Black” Boykins, “Do Work” works as a self motivator, to motivating your friends.   Do It Live   I’ll let Bill O’Reilly explain this one.   Unordered Lists (Nested)     List item one            List item one                    List item one           List item two           List item three           List item four                       List item two       List item three       List item four           List item two   List item three   List item four   Ordered List (Nested)     List item one            List item one                    List item one           List item two           List item three           List item four                       List item two       List item three       List item four           List item two   List item three   List item four   Buttons  Make any link standout more when applying the .btn class.  Notices  Watch out! You can also add notices by appending {: .notice} to a paragraph.  HTML Tags  Address Tag     1 Infinite Loop Cupertino, CA 95014 United States   Anchor Tag (aka. Link)  This is an example of a link.  Abbreviation Tag  The abbreviation CSS stands for “Cascading Style Sheets”.  Cite Tag  “Code is poetry.” —Automattic  Code Tag  You will learn later on in these tests that word-wrap: break-word; will be your best friend.  Strike Tag  This tag will let you strikeout text.  Emphasize Tag  The emphasize tag should italicize text.  Insert Tag  This tag should denote inserted text.  Keyboard Tag  This scarcely known tag emulates keyboard text, which is usually styled like the &lt;code&gt; tag.  Preformatted Tag  This tag styles large blocks of code.   .post-title {   margin: 0 0 5px;   font-weight: bold;   font-size: 38px;   line-height: 1.2;   and here's a line of some really, really, really, really long text, just to see how the PRE tag handles it and to find out how it overflows; }   Quote Tag  Developers, developers, developers… –Steve Ballmer  Strong Tag  This tag shows bold text.  Subscript Tag  Getting our science styling on with H2O, which should push the “2” down.  Superscript Tag  Still sticking with science and Isaac Newton’s E = MC2, which should lift the 2 up.  Variable Tag  This allows you to denote variables.                                Page Not Found                                                                                     About Me                                           About Me                                                     Archive Layout with Content                                                                                     Posts by Category                                                                                     Posts by Collection                                                                                     CV                                                                                                                                                                                                                                                               Markdown                                                                                     News                                                                                     Page not in menu                                                                                     Page Archive                                                                                     Portfolio                                                                                     Publications                                           Publications and preprints by Hanzhi Zhang on efficient AI systems, language models, and reliable and responsible AI.                                                                                                                                          Professional Services                                           Conference and journal reviewing, external reviewing, and conference and workshop volunteering.                                                     Sitemap                                                                                     Posts by Tags                                                                                     Talk map                                                                                     Talks &amp; Presentations                                           Invited talks, conference and poster presentations, and research seminars.                                                     Teaching &amp; Mentoring                                           Undergraduate and graduate teaching, NSF REU coordination, and student research mentoring at UNT.                                                     Terms and Privacy Policy                                                                                     Blog posts                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                 Editorial academic template assets                                                                                     Jupyter notebook markdown generator                                                                                                                                                                                                                                                                                                                      "
}





,{
  "title": "Posts by Category",
  "url": "/categories/",
  "content": "             "
}





,{
  "title": "Posts by Collection",
  "url": "/collection-archive/",
  "content": "   &lt;h2 id=\"portfolio\" class=\"archive__subtitle\"&gt;portfolio&lt;/h2&gt;                                 IEEEE MOST 2024                                            The 2nd IEEE International Conference on Mobility: Operations, Services, and Technologies                                                     Generative AI Prompt A Thon                                            Microsoft to Co-Host UTA’s Generative AI Prompt-A-Thon                         &lt;h2 id=\"publications\" class=\"archive__subtitle\"&gt;publications&lt;/h2&gt;                                 ByteScience: Bridging Unstructured Scientific Literature and Structured Data with Auto Fine-Tuned Large Language Model in Token Granularity                                            Published in IEEE International Conference on Data Mining Workshops (IEEE ICDMW) , 2024.                    An innovative framework that bridges unstructured scientific literature and structured data through auto fine-tuned Large Language Models operating at token granularity.                       Download Paper                                          D2SO: Detecting Distant and Small Objects for Vision-Based Vehicle Autonomous Systems                                            Published in IEEE International Conference on Mobility: Operations, Services, and Technologies (IEEE MOST) , 2025.                    A specialized framework for detecting distant and small objects in vision-based autonomous vehicle systems, enhancing safety and situational awareness.                       Download Paper                                          HALO: Hallucination Analysis and Learning Optimization to Empower LLMs with Retrieval-Augmented Context for Guided Clinical Decision Making                                            Published in IEEE/ACM Conference on Connected Health: Applications, Systems and Engineering Technologies (IEEE/ACM CHASE) , 2025.                    A comprehensive framework for hallucination analysis and learning optimization that empowers Large Language Models with retrieval-augmented context for guided clinical decision making.                       Download Paper                                          DAM: Dynamic Attention Mask for Long-Context Large Language Model Inference Acceleration                                            Published in Findings of the Annual Meeting of the Association for Computational Linguistics (ACL Findings) , 2025.                    A breakthrough approach to long-context inference in large language models by dynamically learning adaptive attention masks at the granularity of individual attention maps.                       Download Paper                                          Harmful Factuality: LLMs Correcting What They Shouldn’t                                            Published in Findings of the Conference of the European Chapter of the Association for Computational Linguistics (EACL Findings) , 2026.                    A study of harmful factuality in large language models, examining how factual corrections can produce outputs that are true but unfaithful to the source.                       Download Paper                                          TileMix: Tile-Centric Mixed-Precision Attention for LLM Inference Acceleration                                            Published in arXiv preprint , 2026.                                           Download Paper                                          Poly-FEVER: A Multilingual Hallucination Detection Benchmark                                            Published in Findings of the Association for Computational Linguistics (EMNLP Findings) , 2026.                    A comprehensive multilingual benchmark with 77,973 labeled factual claims spanning 11 languages for fact verification and hallucination detection in Large Language Models.                       Download Paper                                          AlignQuant: Tile-Aligned Mixed-Precision Quantization for Efficient LLM Generation                                            Published in arXiv preprint , 2026.                    Tile-aligned mixed-precision post-training quantization for efficient LLM generation.                       Download Paper                                          TileQuant: Tile as the Precision-Allocation Unit for LLM Mixed-Precision Post-Training Quantization                                            Published in arXiv preprint , 2026.                                           Download Paper                                          TriAttention: Unified Sparse Mixed-Precision Attention for Long-Context LLM Inference                                            Published in arXiv preprint , 2026.                    Unified sparse mixed-precision attention for long-context LLM inference.                       Download Paper                                          Towards Visual Query Segmentation in the Wild                                            Published in Advances in Neural Information Processing Systems (NeurIPS) , 2026.                    A benchmark for visual query segmentation in real-world video. Hanzhi co-developed VQS-4K, containing 4,111 videos and more than 1.3 million frames.                       Download Paper              &lt;h2 id=\"talks\" class=\"archive__subtitle\"&gt;talks&lt;/h2&gt;                                 UNT College of Engineering Research Showcase                                            Published: March 27, 2024                    This poster has been selected to receive an outstanding poster award on the 1st place! As a winner of the Engineering Showcase Poster Presentation, Hanzhi Zhang have been selected to represent College of Engineering to present at the University Research Day annual poster session.                          &lt;h2 id=\"teaching\" class=\"archive__subtitle\"&gt;teaching&lt;/h2&gt;                                 CSCE 5300: Introduction to Big Data and Data Science                                            Teaching Assistant · University of North Texas  · Fall 2024 &amp; Spring 2025                    Graduate course, 79 students in Fall 2024 and 64 students in Spring 2025, University of North Texas.                                                     Long-Context Evaluation of Large Language Models                                            Guest Lecturer · University of North Texas  · Summer 2026                    Delivered a guest lecture on Long-Context Evaluation of Large Language Models in UNT’s NSF Research Experiences for Undergraduates (REU) program, alongside her work as Academic Coordinator and Research Mentor.                                                     CSCE 4201: Introduction to Artificial Intelligence                                            Teaching Assistant · University of North Texas  · Fall 2026                    Undergraduate course, 74 students, University of North Texas.                       "
}





,{
  "title": "CV",
  "url": "/#cv-download",
  "content": " "
}









,{
  "title": "Markdown",
  "url": "/markdown/",
  "content": " Locations of key files/directories     Basic config options: _config.yml   Top navigation bar config: _data/navigation.yml   Single pages: _pages/   Collections of pages are .md or .html files in:            _publications/       _portfolio/       _posts/       _teaching/       _talks/           Footer: _includes/footer.html   Static files (like PDFs): /files/   Profile image (can set in _config.yml): images/profile.png   Tips and hints     Name a file “.md” to have it render in markdown, name it “.html” to render in HTML.   Go to the commit list (on your repo) to find the last version Github built with Jekyll.            Green check: successful build       Orange circle: building       Red X: error       No icon: not built           Resources    Liquid syntax guide   MathJax Documentation   MathJax  Support for MathJax Version 3.0 is included in the template:  [\\displaylines{ \\nabla \\cdot E= \\frac{\\rho}{\\epsilon_0} \\ \\nabla \\cdot B=0 \\ \\nabla \\times E= -\\partial_tB \\ \\nabla \\times B  = \\mu_0 \\left(J + \\varepsilon_0 \\partial_t E \\right) }]  The default delimiters of $$...$$ and \\\\[...\\\\] are supported for displayed mathematics, while \\\\(...\\\\) should be used for in-line mathematics (ex., \\(a^2 + b^2 = c^2\\))  Note that since Academic Pages uses Markdown which cases some interference with MathJax and LaTeX for escaping characters and new lines, although some workarounds exist.  Markdown guide  Academic Pages uses kramdown for Markdown rendering, which has some differences from other Markdown implementations such as GitHub’s. In addition to this guide, please see the kramdown Syntax page for full documentation.  Header three  Header four  Header five  Header six  Blockquotes  Single line blockquote:     Quotes are cool.   Tables  Table 1                 Entry       Item                               John Doe       2016       Description of the item in the list                 Jane Doe       2019       Description of the item in the list                 Doe Doe       2022       Description of the item in the list           Table 2                 Header1       Header2       Header3                       cell1       cell2       cell3                 cell4       ce                         ll5       cell6                               cell1       cell2       cell3                 cell4       cell5       cell6                       Foot1       Foot2       Foot3           Definition Lists     Definition List Title   Definition list division.   Startup   A startup company or startup is a company or temporary organization designed to search for a repeatable and scalable business model.   #dowork   Coined by Rob Dyrdek and his personal body guard Christopher “Big Black” Boykins, “Do Work” works as a self motivator, to motivating your friends.   Do It Live   I’ll let Bill O’Reilly explain this one.   Unordered Lists (Nested)     List item one            List item one                    List item one           List item two           List item three           List item four                       List item two       List item three       List item four           List item two   List item three   List item four   Ordered List (Nested)     List item one            List item one                    List item one           List item two           List item three           List item four                       List item two       List item three       List item four           List item two   List item three   List item four   Buttons  Make any link standout more when applying the .btn class.  Notices  Basic notices or call-outs are supported using the following syntax:  **Watch out!** You can also add notices by appending `{: .notice}` to the line following paragraph. {: .notice}   which wil render as:  Watch out! You can also add notices by appending {: .notice} to the line following paragraph.  Footnotes  Footnotes can be useful for clarifying points in the text, or citing information.1 Markdown support numeric footnotes, as well as text as long as the values are unique.2  This is the regular text.[^1] This is more regular text.[^note]  [^1]: This is the footnote itself. [^note]: This is another footnote.   HTML Tags  Address Tag     1 Infinite Loop Cupertino, CA 95014 United States   Anchor Tag (aka. Link)  This is an example of a link.  Abbreviation Tag  The abbreviation CSS stands for “Cascading Style Sheets”.  Cite Tag  “Code is poetry.” —Automattic  Code Tag  You will learn later on in these tests that word-wrap: break-word; will be your best friend.  You can also write larger blocks of code with syntax highlighting supported for some languages, such as Python:  print('Hello World!')   or R:  print(\"Hello World!\", quote = FALSE)   Strike Tag  This tag will let you strikeout text.  Emphasize Tag  The emphasize tag should italicize text.  Insert Tag  This tag should denote inserted text.  Keyboard Tag  This scarcely known tag emulates keyboard text, which is usually styled like the &lt;code&gt; tag.  Preformatted Tag  This tag styles large blocks of code.   .post-title {   margin: 0 0 5px;   font-weight: bold;   font-size: 38px;   line-height: 1.2;   and here's a line of some really, really, really, really long text, just to see how the PRE tag handles it and to find out how it overflows; }   Quote Tag  Developers, developers, developers… –Steve Ballmer  Strong Tag  This tag shows bold text.  Subscript Tag  Getting our science styling on with H2O, which should push the “2” down.  Superscript Tag  Still sticking with science and Isaac Newton’s E = MC2, which should lift the 2 up.  Variable Tag  This allows you to denote variables.    Footnotes  The footnotes in the page will be returned following this line, return to the section on Markdown Footnotes.                 Such as this footnote. &#8617;                 When using text for footnotes markers, no spaces are permitted in the name. &#8617;           "
}





,{
  "title": "News",
  "url": "/news/",
  "content": "    News                2026                        📄Nov 2026: Our research paper, “TriAttention: Unified Sparse Mixed-Precision Attention for Long-Context LLM Inference”, is now available as a preprint on arXiv.                   📄Nov 2026: Our research paper, “TileQuant: Tile as the Precision-Allocation Unit for LLM Mixed-Precision Post-Training Quantization”, is now available as a preprint on arXiv.                   🏆Oct 2026: Selected as a “Top Reviewer” for NeurIPS 2026                   🔍Oct 2026: Served as a reviewer for ICLR 2027                   🔍Oct 2026: Served as a reviewer for ACL Rolling Review 2026 (October cycle)                   📚Oct 2026: Gave lectures on CNNs, NLP, and Transformers for CSCE 4201                   🤝Oct 2026: Started research collaboration with Amazon scientist Qian Qu                   📄Oct 2026: Our research paper, “AlignQuant: Tile-Aligned Mixed-Precision Quantization for Efficient LLM Generation”, is now available as a preprint on arXiv.                   🎤Sep 2026: Gave a presentation on “MixLLM: LLM Quantization with Global Mixed-precision between Output-features and Highly-efficient System Design” at UNT&#39;s LLaVi Lab AI Seminar                   📄Sep 2026: Our research paper, “Towards Visual Query Segmentation in the Wild”, has been accepted for publication in Advances in Neural Information Processing Systems (NeurIPS 2026). I am a co-first author of this work.                   🧪Sep 2026: I contributed to the Amazon Research Awards (ARA) project “TileRoute-Trainium: Tile-Level Mixed-Precision Kernels for Long-Context LLM Inference” as lead student researcher and proposal co-writer.                   🔍Aug 2026: Served as a reviewer for ACL Rolling Review 2026 (August cycle)                   📄Aug 2026: Our research paper, “Poly-FEVER: A Multilingual Hallucination Detection Benchmark”, has been accepted for publication in Findings of the Association for Computational Linguistics (EMNLP 2026).                   📄Aug 2026: Our research paper, “TileMix: Tile-Centric Mixed-Precision Attention for LLM Inference Acceleration”, is now available as a preprint on arXiv.                   🌱Summer 2026: Served as Academic Coordinator and Research Mentor for nine undergraduates in UNT&#39;s NSF REU program                   📚Summer 2026: Gave a guest lecture on “Long-Context Evaluation of Large Language Models” at UNT&#39;s NSF REU program                   🤝Jun 2026: Started research collaboration with ORNL&#39;s AAIMS Group                   🔍May 2026: Served as an external reviewer for ACM SEC 2026                   🔍May 2026: Served as a reviewer for NeurIPS 2026                   🎤May 2026: Gave an invited talk on “Responsible AI” at the Presidential AI Challenge                   🔍May 2026: Served as a reviewer for ACL Rolling Review 2026 (May cycle)                   🎤Apr 2026: Gave a presentation on “Gated Attention for Large Language Models: Non-linearity, Sparsity, and Attention-Sink-Free” at UNT&#39;s LLaVi Lab AI Seminar                   📄Mar 2026: Our research paper, “Towards Visual Query Segmentation in the Wild”, is now available as a preprint on arXiv.                   📄Mar 2026: Our research paper, “Harmful Factuality: LLMs Correcting What They Shouldn’t”, has been published in Findings of the Association for Computational Linguistics (EACL 2026).                   🔍Feb 2026: Served as a reviewer for ICML 2026                        2025                        🎤Nov 2025: Gave a presentation on “SageAttention: Accurate 8-Bit Attention for Plug-and-play Inference Acceleration” at UNT&#39;s LLaVi Lab AI Seminar                   🎤Oct 2025: Gave an invited talk on “Responsible AI: The Hidden Curriculum” at AIxHEART                   🖼️Oct 2025: Presented a poster on “Efficient Long-Context Inference through Dynamic Sparse Attention” at UNT University Research Day                   🎤Jul 2025: Gave a presentation on “DAM: Dynamic Attention Mask for Long-Context Large Language Model Inference Acceleration” at ACL 2025 (Findings)                   🔍Jun 2025: Served as an external reviewer for ICDM 2025                   🔍May 2025: Served as a reviewer for CIKM 2025                   📄May 2025: Our research paper, “DAM: Dynamic Attention Mask for Long-Context Large Language Model Inference Acceleration”, has been accepted for publication in Findings of the Association for Computational Linguistics (ACL 2025).                   🎤May 2025: Gave a presentation on “D2SO: Detecting Distant and Small Objects for Vision-Based Vehicle Autonomous Systems” at IEEE MOST 2025                   🏆Apr 2025: Received a Student Travel Grant Award from IEEE MOST 2025                   🌐Apr 2025: Participated in the Women-in-Computing session at IEEE MOST 2025                   📄Mar 2025: Our research paper, “Poly-FEVER: A Multilingual Fact Verification Benchmark for Hallucination Detection in Large Language Models”, is now available as a preprint on arXiv.                   📄Mar 2025: Our research paper, “D2SO: Detecting Distant and Small Objects for Vision-Based Vehicle Autonomous Systems”, has been accepted for publication in the Proceedings of the IEEE International Conference on Mobility, Operations, Services and Technologies (IEEE MOST 2025).                   🖼️Mar 2025: Presented a poster on “Adaptive Attention Masks for Large Language Model Acceleration” at the North Texas Symposium on Generative AI at UT Arlington                   🎤Feb 2025: Gave a presentation on “Large Language Diffusion Models” at UNT&#39;s LLaVi Lab AI Seminar                   🎓Feb 2025: Passed PhD Qualifying Exam                   📄Feb 2025: Our research paper, “HALO: Hallucination Analysis and Learning Optimization to Empower LLMs with Retrieval-Augmented Context for Guided Clinical Decision Making”, has been accepted for publication in the Proceedings of the IEEE/ACM Conference on Connected Health: Applications, Systems and Engineering Technologies (CHASE 2025).                   📚Spring 2025: Served as a Teaching Assistant for CSCE 5300: Introduction to Big Data and Data Science (64 graduate students)                        2024                        🙋‍♀️Dec 2024: Served as a student volunteer at the NSF CISE MSI Aspiring PI Workshop at UNT                   📄Dec 2024: Our research paper, “ByteScience: Bridging Unstructured Scientific Literature and Structured Data with Auto Fine-Tuned Large Language Model in Token Granularity”, has been published in the Proceedings of the IEEE International Conference on Data Mining Workshops (ICDMW 2024).                   🎤Nov 2024: Presented a proposal on “Efficient Small Language Models for Edge Devices via Task-Aware Sparse Attention” at the NSF IUCRC eCAT meeting                   🎤Oct 2024: Gave a presentation on “MoA: Mixture of Sparse Attention for Automatic Large Language Model Compression” at UNT&#39;s LLaVi Lab AI Seminar                   🖼️Oct 2024: Presented a poster on “Poly-FEVER: A Multilingual Hallucination Detection Benchmark” at UNT University Research Day                   📚Fall 2024: Served as a Teaching Assistant for CSCE 5300: Introduction to Big Data and Data Science (79 graduate students)                   🛠️Jul 2024: Participated in the Argonne Training Program on Extreme-Scale Computing (ATPESC) 2024                   🧪Jul 2024: Started serving as PI for ALCF Polaris computing allocations for “Large Language Models’ Hallucination Detection and Mitigation Ability across Multi Languages” (5,000 node-hours per allocation)                   🔍Jun 2024: Served as a reviewer for ICDM 2024                   🙋‍♀️May 2024: Served as a student volunteer at IEEE MOST 2024                   🌐May 2024: Participated in the AI Prompt-a-Thon at UT Arlington with Microsoft                   🎤Apr 2024: Gave a presentation on “Small Language Model” at UNT&#39;s LLaVi Lab AI Seminar                   🏆Mar 2024: I won First Place for the poster “Investigating Biased Cross-Lingual Hallucination Detection Capabilities of LLMs: POLY-FEVER Benchmark &amp; Mitigating Approaches” at UNT Engineering Showcase.                        2023                        🔍Oct 2023: Served as an external reviewer for WWW 2024                   🔍Oct 2023: Served as an external reviewer for ICNC 2024                   🔍Sep 2023: Served as a reviewer for IEEE ICPADS 2023                   🔍Aug 2023: Served as an external reviewer for AAAI 2024                   🎓Aug 2023: Started PhD in Computer Science at UNT and joined Responsible AI Lab                   🔍Jul 2023: Served as an external reviewer for EISA 2023                   🔍Jul 2023: Served as an external reviewer for Inscrypt 2023                   🔍Jul 2023: Served as an external reviewer for WACV 2024                     "
}





,{
  "title": "Page not in menu",
  "url": "/non-menu-page/",
  "content": " This is a page not in the menu. You can use markdown in this page.  Heading 1  Heading 2  "
}





,{
  "title": "Page Archive",
  "url": "/page-archive/",
  "content": "                               Page Not Found                                                                                     About Me                                           About Me                                                     Archive Layout with Content                                                                                     Posts by Category                                                                                     Posts by Collection                                                                                     CV                                                                                                                                                                                                                                                               Markdown                                                                                     News                                                                                     Page not in menu                                                                                     Page Archive                                                                                     Portfolio                                                                                     Publications                                           Publications and preprints by Hanzhi Zhang on efficient AI systems, language models, and reliable and responsible AI.                                                                                                                                          Professional Services                                           Conference and journal reviewing, external reviewing, and conference and workshop volunteering.                                                     Sitemap                                                                                     Posts by Tags                                                                                     Talk map                                                                                     Talks &amp; Presentations                                           Invited talks, conference and poster presentations, and research seminars.                                                     Teaching &amp; Mentoring                                           Undergraduate and graduate teaching, NSF REU coordination, and student research mentoring at UNT.                                                     Terms and Privacy Policy                                                                                     Blog posts                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                 Editorial academic template assets                                                                                     Jupyter notebook markdown generator                                                                                                                                                                                                                                                                                                                      "
}





,{
  "title": "Portfolio",
  "url": "/portfolio/",
  "content": "                               IEEEE MOST 2024                                            The 2nd IEEE International Conference on Mobility: Operations, Services, and Technologies                                                     Generative AI Prompt A Thon                                            Microsoft to Co-Host UTA’s Generative AI Prompt-A-Thon                       "
}





,{
  "title": "Publications",
  "url": "/publications/",
  "content": "    Publications   For full publication list, please go to my Google Scholar profile.   Filter publications by title, author, or year                                      Preprints                     Nov 2026        H. Zhang, Q. Zhang, Q. Cao, H. Fan, Y. Huang, K. Sha, and Y. Feng.     TriAttention: Unified Sparse Mixed-Precision Attention for Long-Context LLM Inference.     arXiv preprint, 2026.               arXiv   Code       BIB     Abstract                     Copy       @misc{zhang2026triattentionunifiedsparsemixedprecision,       title={TriAttention: Unified Sparse Mixed-Precision Attention for Long-Context LLM Inference},       author={Hanzhi Zhang and Qiao Zhang and Qinglei Cao and Heng Fan and Yan Huang and Kewei Sha and Yunhe Feng},       year={2026},       eprint={2611.07256},       archivePrefix={arXiv},       primaryClass={cs.LG},       url={https://arxiv.org/abs/2611.07256}, }                           Sparsity and quantization reduce the cost of attention by skipping interactions and lowering the precision of those retained. However, useful interactions differ in their precision requirements: some tolerate low-precision scores while remaining sensitive to removal. We therefore introduce TriAttention, which assigns query-key (QK) tiles to skipping, 4-bit, or 8-bit computation, using 4-bit to retain quantization-tolerant interactions at lower cost. To represent compactly, we encode each head&#39;s computation regions with two diagonal-distance boundaries, keeping policy storage independent of context length. Offline calibration then evaluates candidate boundary pairs using normalized projected-output errors and complete-tile execution costs. Using these measurements, a model-wide allocator selects each head&#39;s boundaries to minimize calibration error under a shared computation budget. During prefill, the selected boundaries generate precision-grouped traversals for a fused kernel that skips omitted tiles and combines contributions from both QK precisions into one normalized attention output. The quantized keys and values prepared during prefill are then reused in a compressed key-value (KV) cache for decoding, reducing repeated reads of prompt data. Experiments show that TriAttention achieves end-to-end generation speedups of up to 7.88x over the Torch attention and 2.65x over the BF16 baseline. Meanwhile, TriAttention maintains quality comparable to the original models on general-purpose and long-context tasks. These results show how a compact retention-and-precision policy translates local sensitivities into attention computation and repeated KV-cache access. The implementation is https://github.com/HanzhiZhang-Ulrica/TriAttention.     Source · CC0 1.0 (arXiv metadata)                             Nov 2026        H. Zhang, Q. Zhang, Q. Cao, H. Fan, Y. Huang, K. Sha, and Y. Feng.     TileQuant: Tile as the Precision-Allocation Unit for LLM Mixed-Precision Post-Training Quantization.     arXiv preprint, 2026.               arXiv   Code       BIB     Abstract                     Copy       @misc{zhang2026tilequanttileprecisionallocationunit,       title={TileQuant: Tile as the Precision-Allocation Unit for LLM Mixed-Precision Post-Training Quantization},       author={Hanzhi Zhang and Qiao Zhang and Qinglei Cao and Heng Fan and Yan Huang and Kewei Sha and Yunhe Feng},       year={2026},       eprint={2611.07217},       archivePrefix={arXiv},       primaryClass={cs.LG},       url={https://arxiv.org/abs/2611.07217}, }                           Post-training quantization (PTQ) reduces large language model (LLM) deployment costs, but preserving quality requires accounting for uneven weight sensitivity. Uniform PTQ uses one weight bit-width, importance-aware PTQ adjusts scales or reconstructs weights within that width, and column-wise mixed precision retains selected columns at higher precision. However, sensitivity varies within columns, so whole-column retention also protects less sensitive weights, while runtime outlier detection adds decoding overhead. We therefore introduce TileQuant, a training-free framework that assigns precision to hardware-aligned 2D weight tiles. Offline calibration provides input-activation-aware scoring from forward statistics for lower calibration cost and gradient-informed scoring from loss sensitivity for stronger quality preservation. Within each projection&#39;s retention budget, the highest-scoring tiles remain at 16-bit and the rest use 8-bit. The resulting map guides reconstruction into compact weight buffers, which the kernel reuses for native integer and 16-bit computation with local accumulation to reduce memory traffic. Experiments show that TileQuant reaches 2.66x the original model&#39;s generation throughput. Across language modeling, reasoning, long-context retrieval, and question answering, it preserves quality and surpasses quantized and original models on multiple tasks. These results support selecting precision units that capture local sensitivity while enabling efficient hardware execution. The implementation is: https://github.com/HanzhiZhang-Ulrica/TileQuant     Source                             Oct 2026        H. Zhang, Q. Zhang, Q. Cao, H. Fan, Y. Huang, K. Sha, and Y. Feng.     AlignQuant: Tile-Aligned Mixed-Precision Quantization for Efficient LLM Generation.     arXiv preprint, 2026.               arXiv   Code       BIB     Abstract                     Copy       @misc{zhang2026alignquanttilealignedmixedprecisionquantization,       title={AlignQuant: Tile-Aligned Mixed-Precision Quantization for Efficient LLM Generation},       author={Hanzhi Zhang and Qiao Zhang and Qinglei Cao and Heng Fan and Yan Huang and Kewei Sha and Yunhe Feng},       year={2026},       eprint={2610.07457},       archivePrefix={arXiv},       primaryClass={cs.LG},       url={https://arxiv.org/abs/2610.07457}, }                           Fine-grained mixed-precision quantization promises efficient large language model inference, but local precision choices can conflict with regular GPU storage and computation units. This precision-boundary mismatch limits the translation of compression into practical acceleration. We introduce AlignQuant, a post-training quantization method that uses GPU-compatible two-dimensional weight tiles as the common unit of precision allocation, compact storage, and execution. This shared partition lets precision follow sensitivity within output channels. Joint prefill/decode calibration scores precision reductions using projection-output perturbations weighted by language-model loss gradients under quantized activations. Phase-normalized scores prioritize higher precision for tiles important to either phase under a model-wide weight-storage budget. Each tile stores one selected representation, while phase-specialized kernels reuse the packed model and expand lower-bit weights for INT8 computation with 8-bit activations. Across four LLMs spanning 3B to 14B parameters, AlignQuant achieves up to 2.50× generation speedup over BF16 while preserving model quality. Evaluations further cover three GPUs and contexts up to 64K tokens. These results show that local precision flexibility and regular GPU execution can coexist through a shared tile unit. The implementation is available at https://github.com/HanzhiZhang-Ulrica/AlignQuant.     Source · CC0 1.0 (arXiv metadata)                             Aug 2026        H. Zhang, Q. Zhang, Q. Cao, H. Fan, Y. Huang, K. Sha, and Y. Feng.     TileMix: Tile-Centric Mixed-Precision Attention for LLM Inference Acceleration.     arXiv preprint, 2026.               arXiv   Code       BIB     Abstract                     Copy       @misc{zhang2026tilemixtilecentricmixedprecisionattention,       title={TileMix: Tile-Centric Mixed-Precision Attention for LLM Inference Acceleration},       author={Hanzhi Zhang and Qiao Zhang and Qinglei Cao and Heng Fan and Yan Huang and Kewei Sha and Yunhe Feng},       year={2026},       eprint={2608.17336},       archivePrefix={arXiv},       primaryClass={cs.AI},       url={https://arxiv.org/abs/2608.17336}, }                           Long-context prefill in large language models (LLMs) incurs substantial computation and memory traffic because dense self-attention computes quadratic query-key scores. Existing methods either use a uniform low-precision path or select token interactions, leaving spatial precision routing over hardware-aligned score tiles outside fused dense attention. We introduce TileMix, a tile-centric precision-routing kernel that makes numerical precision an executable spatial decision over score-tile groups within fused dense attention. TileMix partitions the attention matrix into hardware-aligned score tiles, packs routing decisions into compact bitmasks, and dispatches each tile group through FP16 or INT8 score computation while both paths update a shared online-softmax state. Scalable precision grouping lets each routing bit govern multiple adjacent key tiles, preserving hardware-aligned compute tiles and compact metadata at long contexts. By routing all legal tile groups, TileMix preserves dense token connectivity, requires no training, and supports grouped-query attention, variable-length batches, and INT8 key/value caches. Across LongEval, LV-Eval, and A100 prefill benchmarks on LLaMA, Qwen, and Vicuna, TileMix recovers long-context quality lost under uniform INT8 and improves prefill throughput over FP16, yielding a controllable accuracy-efficiency frontier across model families. The implementation is available at https://github.com/HanzhiZhang-Ulrica/TileMix.     Source                                          Refereed Conference Proceedings                     Dec 2026        B. Fan*, M. Li*, H. Zhang*, S. Dong, N. P. Mareedu, X. Liu, W. Shi, Y. Feng, Y. Huang, and H. Fan.     Towards Visual Query Segmentation in the Wild.     Advances in Neural Information Processing Systems (NeurIPS), 2026.               arXiv   Code  Dataset      BIB     Abstract                     Copy       @misc{fan2026visualquerysegmentationwild,       title={Towards Visual Query Segmentation in the Wild},       author={Bing Fan and Minghao Li and Hanzhi Zhang and Shaohua Dong and Naga Prudhvi Mareedu and Xiaoqiong Liu and Weishi Shi and Yunhe Feng and Yan Huang and Heng Fan},       year={2026},       eprint={2603.08898},       archivePrefix={arXiv},       primaryClass={cs.CV},       url={https://arxiv.org/abs/2603.08898}, }                           In this paper, we introduce visual query segmentation (VQS), a new paradigm of visual query localization (VQL) that aims to segment all pixel-level occurrences of an object of interest in an untrimmed video, given a visual query. Compared with existing VQL locating only the last appearance of a target using bounding boxes, VQS enables more comprehensive (i.e., all object occurrences) and precise (i.e., pixel-level masks) localization, making it more practical for real-world scenarios. To foster research on this task, we introduce VQS-4K, a large-scale benchmark dedicated to VQS. Specifically, VQS-4K consists of 4,111 videos with more than 1.3 million frames, and covers a diverse set of 222 object categories. Each video is paired with a visual query defined by a frame outside search video and its target mask, and annotated with spatial-temporal masklets corresponding to the queried target. To ensure high quality, all videos in VQS-4K are manually labeled with meticulous inspection and iterative refinement. To our best knowledge, VQS-4K is the first benchmark specifically designed for VQS. Furthermore, to stimulate future research, we present a simple yet effective method, named VQ-SAM, which extends SAM 2 by leveraging target-specific and background distractor cues from the video to progressively evolve the memory through a novel multi-stage framework with an adaptive memory generation (AMG) module for VQS, significantly improving the performance. In our experiments on VQS-4K, VQ-SAM achieves promising results and surpasses existing methods, evidencing its effectiveness. With VQS-4K and VQ-SAM, we expect to inspire more future research and applications on VQS. Our benchmark and code are released at https://github.com/VAIL-Fan/VQS.     Source · CC0                             Oct 2026        H. Zhang, S. Anjum, H. Fan, W. Zheng, Y. Huang, and Y. Feng.     Poly-FEVER: A Multilingual Hallucination Detection Benchmark.     Findings of the Association for Computational Linguistics (EMNLP Findings), 2026.               Paper    Dataset      BIB     Abstract                     Copy       @inproceedings{zhang-etal-2026-poly,     title = &quot;Poly-{FEVER}: A Multilingual Hallucination Detection Benchmark&quot;,     author = &quot;Zhang, Hanzhi  and       Anjum, Sumera  and       Fan, Heng  and       Zheng, Weijian  and       Huang, Yan  and       Feng, Yunhe&quot;,     editor = &quot;Lee, Hung-Yi  and       Dev, Sunipa  and       Li, Jessy  and       Elliott, Desmond&quot;,     booktitle = &quot;Findings of the {A}ssociation for {C}omputational {L}inguistics: {EMNLP} 2026&quot;,     month = oct,     year = &quot;2026&quot;,     address = &quot;Budapest, Hungary&quot;,     publisher = &quot;Association for Computational Linguistics&quot;,     url = &quot;https://aclanthology.org/2026.findings-emnlp.540/&quot;,     doi = &quot;10.18653/2026.findings-emnlp.540&quot;,     pages = &quot;11353--11369&quot;,     ISBN = &quot;979-8-89176-462-0&quot;,     abstract = &quot;We present Poly-FEVER, an open-source large-scale multilingual benchmark for hallucination detection through controlled factuality judgment in large language models. Poly-FEVER extends FEVER, Climate-FEVER, and SciFact to 77,973 Supported/Refuted claims aligned across 11 languages. We augment each claim with topic metadata from a 22-topic LDA model, enabling topic-aware analysis of multilingual factuality behavior. Translations are produced with Google Cloud Translation and validated by a targeted human audit, GEMBA, and COMETKiwi. We benchmark ChatGPT-3.5, LLaMA-2/3, Phi-3.5, and Mistral across three prompt families and evaluate robustness to model-generated rephrasing. We further analyze resource imbalance using web presence as a proxy for online claim availability, and evaluate retrieval-augmented factuality judgment with DPR over Wikipedia. Results show pronounced cross-lingual disparities under aligned claims, with stronger accuracy in high-resource languages such as English and Chinese and weaker accuracy in lower-resource languages such as Amharic and Tamil. Topic-aware prompting and retrieval provide selective gains, especially in Arabic and Amharic, while their effects vary across languages and can conflict with model priors. Poly-FEVER establishes a rigorous resource for responsible, language-adaptive hallucination detection and language-sensitive factuality evaluation. The proposed Poly-FEVER is available at: https://huggingface.co/datasets/HanzhiZhang/Poly-FEVER.&quot; }                           We present Poly-FEVER, an open-source large-scale multilingual benchmark for hallucination detection through controlled factuality judgment in large language models. Poly-FEVER extends FEVER, Climate-FEVER, and SciFact to 77,973 Supported/Refuted claims aligned across 11 languages. We augment each claim with topic metadata from a 22-topic LDA model, enabling topic-aware analysis of multilingual factuality behavior. Translations are produced with Google Cloud Translation and validated by a targeted human audit, GEMBA, and COMETKiwi. We benchmark ChatGPT-3.5, LLaMA-2/3, Phi-3.5, and Mistral across three prompt families and evaluate robustness to model-generated rephrasing. We further analyze resource imbalance using web presence as a proxy for online claim availability, and evaluate retrieval-augmented factuality judgment with DPR over Wikipedia. Results show pronounced cross-lingual disparities under aligned claims, with stronger accuracy in high-resource languages such as English and Chinese and weaker accuracy in lower-resource languages such as Amharic and Tamil. Topic-aware prompting and retrieval provide selective gains, especially in Arabic and Amharic, while their effects vary across languages and can conflict with model priors. Poly-FEVER establishes a rigorous resource for responsible, language-adaptive hallucination detection and language-sensitive factuality evaluation. The proposed Poly-FEVER is available at: https://huggingface.co/datasets/HanzhiZhang/Poly-FEVER.     Source · CC BY 4.0                             Mar 2026        M. Li, H. Zhang, H. Fan, J. Ding, and Y. Feng.     Harmful Factuality: LLMs Correcting What They Shouldn&#39;t.     Findings of the Conference of the European Chapter of the Association for Computational Linguistics (EACL Findings), 2026.               Paper   Code       BIB     Abstract                     Copy       @inproceedings{li-etal-2026-harmful,     title = &quot;Harmful Factuality: {LLM}s Correcting What They Shouldn{&#39;}t&quot;,     author = &quot;Li, Mingchen  and       Zhang, Hanzhi  and       Fan, Heng  and       Ding, Junhua  and       Feng, Yunhe&quot;,     editor = &quot;Demberg, Vera  and       Inui, Kentaro  and       Marquez, Llu{\\&#39;i}s&quot;,     booktitle = &quot;Findings of the {A}ssociation for {C}omputational {L}inguistics: {EACL} 2026&quot;,     month = mar,     year = &quot;2026&quot;,     address = &quot;Rabat, Morocco&quot;,     publisher = &quot;Association for Computational Linguistics&quot;,     url = &quot;https://aclanthology.org/2026.findings-eacl.46/&quot;,     doi = &quot;10.18653/v1/2026.findings-eacl.46&quot;,     pages = &quot;896--912&quot;,     ISBN = &quot;979-8-89176-386-9&quot;,     abstract = &quot;While Large Language Models (LLMs) are trained for factual accuracy, this objective can directly conflict with the critical demand for source fidelity. This paper isolates and formalizes this conflict as Harmful Factuality Hallucination (HFH): a previously overlooked failure mode where an LLM{&#39;}s attempt to ``correct&#39;&#39; perceived source errors results in an output that is factually true but unfaithful to the input. Unlike traditional hallucination research focused on models generating falsehoods, we investigate the harm of misplaced correctness. We introduce a reproducible framework to elicit and measure HFH using controlled entity-level perturbations (both soft, embedding-based and hard, instruction-based) paired with strategic entity selection. Across summarization, rephrasing, and QA tasks, our evaluation of diverse LLMs reveals that HFH is a prevalent behavior that worsens with model scale. We identify three underlying mechanisms and demonstrate that a simple instructional prompt can reduce HFH rates by approximately 50{\\%}. Our framework turns the abstract factuality{--}faithfulness tension into a measurable, actionable target for building more reliable LLM systems. Our code is publicly available at \\url{https://github.com/ResponsibleAILab/Harmful-Factuality-Hallucination}.&quot; }                           While Large Language Models (LLMs) are trained for factual accuracy, this objective can directly conflict with the critical demand for source fidelity. This paper isolates and formalizes this conflict as Harmful Factuality Hallucination (HFH): a previously overlooked failure mode where an LLM&#39;s attempt to “correct” perceived source errors results in an output that is factually true but unfaithful to the input. Unlike traditional hallucination research focused on models generating falsehoods, we investigate the harm of misplaced correctness. We introduce a reproducible framework to elicit and measure HFH using controlled entity-level perturbations (both soft, embedding-based and hard, instruction-based) paired with strategic entity selection. Across summarization, rephrasing, and QA tasks, our evaluation of diverse LLMs reveals that HFH is a prevalent behavior that worsens with model scale. We identify three underlying mechanisms and demonstrate that a simple instructional prompt can reduce HFH rates by approximately 50%. Our framework turns the abstract factuality–faithfulness tension into a measurable, actionable target for building more reliable LLM systems. Our code is publicly available at https://github.com/ResponsibleAILab/Harmful-Factuality-Hallucination.     Source · CC BY 4.0                             Jul 2025        H. Zhang, H. Fan, K. Sha, Y. Huang, and Y. Feng.     DAM: Dynamic Attention Mask for Long-Context Large Language Model Inference Acceleration.     Findings of the Annual Meeting of the Association for Computational Linguistics (ACL Findings), 2025.               Paper   Code       BIB     Abstract                     Copy       @inproceedings{zhang-etal-2025-dam,     title = &quot;{DAM}: Dynamic Attention Mask for Long-Context Large Language Model Inference Acceleration&quot;,     author = &quot;Zhang, Hanzhi  and       Fan, Heng  and       Sha, Kewei  and       Huang, Yan  and       Feng, Yunhe&quot;,     editor = &quot;Che, Wanxiang  and       Nabende, Joyce  and       Shutova, Ekaterina  and       Pilehvar, Mohammad Taher&quot;,     booktitle = &quot;Findings of the {A}ssociation for {C}omputational {L}inguistics: {ACL} 2025&quot;,     month = jul,     year = &quot;2025&quot;,     address = &quot;Vienna, Austria&quot;,     publisher = &quot;Association for Computational Linguistics&quot;,     url = &quot;https://aclanthology.org/2025.findings-acl.242/&quot;,     doi = &quot;10.18653/v1/2025.findings-acl.242&quot;,     pages = &quot;4663--4676&quot;,     ISBN = &quot;979-8-89176-256-5&quot;,     abstract = &quot;Long-context understanding is crucial for many NLP applications, yet transformers struggle with efficiency due to the quadratic complexity of self-attention. Sparse attention methods alleviate this cost but often impose static, predefined masks, failing to capture heterogeneous attention patterns. This results in suboptimal token interactions, limiting adaptability and retrieval accuracy in long-sequence tasks. This work introduces a dynamic sparse attention mechanism that assigns adaptive masks at the attention-map level, preserving heterogeneous patterns across layers and heads. Unlike existing approaches, our method eliminates the need for fine-tuning and predefined mask structures while maintaining computational efficiency. By learning context-aware attention structures, it achieves high alignment with full-attention models, ensuring minimal performance degradation while reducing memory and compute overhead. This approach provides a scalable alternative to full attention, enabling the practical deployment of large-scale Large Language Models (LLMs) without sacrificing retrieval performance. DAM is available at: \\url{https://github.com/HanzhiZhang-Ulrica/DAM}.&quot; }                           Long-context understanding is crucial for many NLP applications, yet transformers struggle with efficiency due to the quadratic complexity of self-attention. Sparse attention methods alleviate this cost but often impose static, predefined masks, failing to capture heterogeneous attention patterns. This results in suboptimal token interactions, limiting adaptability and retrieval accuracy in long-sequence tasks. This work introduces a dynamic sparse attention mechanism that assigns adaptive masks at the attention-map level, preserving heterogeneous patterns across layers and heads. Unlike existing approaches, our method eliminates the need for fine-tuning and predefined mask structures while maintaining computational efficiency. By learning context-aware attention structures, it achieves high alignment with full-attention models, ensuring minimal performance degradation while reducing memory and compute overhead. This approach provides a scalable alternative to full attention, enabling the practical deployment of large-scale Large Language Models (LLMs) without sacrificing retrieval performance. DAM is available at: https://github.com/HanzhiZhang-Ulrica/DAM.     Source · CC BY 4.0                             Jun 2025        S. Anjum, H. Zhang, W. Zhou, E. J. Paek, X. Zhao, and Y. Feng.     HALO: Hallucination Analysis and Learning Optimization to Empower LLMs with Retrieval-Augmented Context for Guided Clinical Decision Making.     IEEE/ACM Conference on Connected Health: Applications, Systems and Engineering Technologies (IEEE/ACM CHASE), 2025.               Paper   Code       BIB     Abstract                     Copy       @inproceedings{10.1145/3721201.3721385, author = {Anjum, Sumera and Zhang, Hanzhi and Zhou, Wenjun and Paek, Eun Jin and Zhao, Xiaopeng and Feng, Yunhe}, title = {HALO: Hallucination Analysis and Learning Optimization to Empower LLMs with Retrieval-Augmented Context for Guided Clinical Decision Making}, year = {2025}, isbn = {9798400715396}, publisher = {Association for Computing Machinery}, address = {New York, NY, USA}, url = {https://doi.org/10.1145/3721201.3721385}, doi = {10.1145/3721201.3721385}, abstract = {Large language models (LLMs) have significantly advanced natural language processing tasks, yet they are susceptible to generating inaccurate or unreliable responses, a phenomenon known as hallucination. In critical domains such as health and medicine, these hallucinations can pose serious risks. This paper introduces HALO, a novel framework designed to enhance the accuracy and reliability of medical question-answering (QA) systems by focusing on the detection and mitigation of hallucinations. Our approach generates multiple variations of a given query using LLMs and retrieves relevant information from external open knowledge bases to enrich the context. We utilize maximum marginal relevance scoring to prioritize the retrieved context, which is then provided to LLMs for answer generation, thereby reducing the risk of hallucinations. The integration of LangChain further streamlines this process, resulting in a notable and robust increase in the accuracy of both open-source and commercial LLMs, such as Llama-3.1 (from 44\\% to 66\\%) and ChatGPT (from 56\\% to 70\\%). This framework underscores the critical importance of addressing hallucinations in medical QA systems, ultimately improving clinical decision-making and patient care. The open-source HALO is available at: https://github.com/ResponsibleAILab/HALO.}, booktitle = {Proceedings of the ACM/IEEE International Conference on Connected Health: Applications, Systems and Engineering Technologies}, pages = {187–198}, numpages = {12}, keywords = {large language models (LLMs), AI hallucination, retrieval augmented generation (RAG), chain of thought (CoT), few-shot question answering, clinical decision making}, location = {New York, NY, USA}, series = {CHASE &#39;25} }                           Large language models (LLMs) have significantly advanced natural language processing tasks, yet they are susceptible to generating inaccurate or unreliable responses, a phenomenon known as hallucination. In critical domains such as health and medicine, these hallucinations can pose serious risks. This paper introduces HALO, a novel framework designed to enhance the accuracy and reliability of medical question-answering (QA) systems by focusing on the detection and mitigation of hallucinations. Our approach generates multiple variations of a given query using LLMs and retrieves relevant information from external open knowledge bases to enrich the context. We utilize maximum marginal relevance scoring to prioritize the retrieved context, which is then provided to LLMs for answer generation, thereby reducing the risk of hallucinations. The integration of LangChain further streamlines this process, resulting in a notable and robust increase in the accuracy of both open-source and commercial LLMs, such as Llama-3.1 (from 44% to 66%) and ChatGPT (from 56% to 70%). This framework underscores the critical importance of addressing hallucinations in medical QA systems, ultimately improving clinical decision-making and patient care. The open-source HALO is available at: https://github.com/ResponsibleAILab/HALO.     Source                             May 2025        H. Zhang, H. Lucero, K. Sha, H. Fan, S. Fu, and Y. Feng.     D2SO: Detecting Distant and Small Objects for Vision-Based Vehicle Autonomous Systems.     IEEE International Conference on Mobility: Operations, Services, and Technologies (IEEE MOST), 2025.               Paper         BIB     Abstract                     Copy       @INPROCEEDINGS{11071446,   author={Zhang, Hanzhi and Lucero, Harold and Sha, Kewei and Fan, Heng and Fu, Song and Feng, Yunhe},   booktitle={2025 IEEE 3rd International Conference on Mobility, Operations, Services and Technologies (MOST)},   title={D2SO: Detecting Distant and Small Objects for Vision-Based Vehicle Autonomous Systems},   year={2025},   volume={},   number={},   pages={35-42},   keywords={YOLO;Image segmentation;Computer vision;Visualization;Autonomous systems;Image color analysis;Urban areas;Transformers;Hazards;Reliability;Small Object Detection;Distant Object Detection;Image Segmentation;Vehicle Autonomous Systems;Segment Anything Model (SAM);Vision Transformer (ViT)},   doi={10.1109/MOST65065.2025.00013}}                           Detecting distant and small objects is a critical capability for vision-based vehicle autonomous systems, particularly in safety-critical scenarios such as assisted driving, where constant alertness, early reaction, and safe operation are very important. However, accurately recognizing tiny objects at a distance presents significant challenges due to limited pixel information, preprocessing-induced downscaling, and restricted detector resolution. To address these issues, this paper introduces D2SO, a Vision Transformer (ViT)-based framework specifically designed to enhance the detection of distant and small objects for autonomous systems. D2SO integrates multiple fine-tuned AI models, derived from the open-source Segment Anything Model (SAM), to detect distant objects such as static structures, humans, and vehicles, thereby improving situational awareness. The system employs visual cues through color mask overlays to efficiently convey essential information, ensuring users remain well-informed about detection outcomes. D2SO is explicitly optimized to detect distant entities that occupy fewer than 24 × 24 pixels on display, a scale often imperceptible to humans. Experimental results demonstrate that D2SO significantly outperforms baseline models, including SegFormer, YOLO v11 Segmentation, and U-Net, on a real-world street scene dataset spanning 50 cities, establishing its effectiveness in enhancing autonomous system performance.     Source                             Dec 2024        T. Xie, H. Zhang, S. Wang, Y. Wan, I. Razzak, C. Kit, W. Zhang, and B. Hoex.     ByteScience: Bridging Unstructured Scientific Literature and Structured Data with Auto Fine-Tuned Large Language Model in Token Granularity.     IEEE International Conference on Data Mining Workshops (IEEE ICDMW), 2024.               Paper         BIB     Abstract                     Copy       @INPROCEEDINGS{10917307,   author={Xie, Tong and Zhang, Hanzhi and Wang, Shaozhou and Wan, Yuwei and Razzak, Imran and Kit, Chunyu and Zhang, Wenjie and Hoex, Bram},   booktitle={2024 IEEE International Conference on Data Mining Workshops (ICDMW)},   title={ByteScience: Bridging Unstructured Scientific Literature and Structured Data with Auto Fine-tuned Large Language Model in Token Granularity},   year={2024},   volume={},   number={},   pages={907-911},   keywords={Cloud computing;Accuracy;Web services;Large language models;Computational modeling;Propulsion;Natural language processing;Data models;Data mining;Informatics;component;formatting;style;styling;insert},   doi={10.1109/ICDMW65004.2024.00126}}                           Natural Language Processing (NLP) is widely used to supply summarization ability from long context to structured information. However, extracting structured knowledge from scientific text by NLP models remains a challenge because of its domain-specific nature to complex data preprocessing and the granularity of multi-layered device-level information. To address this, we introduce ByteScience, a non-profit cloud-based auto fine-tuned Large Language Model (LLM) platform, which is designed to extract structured scientific data and synthesize new scientific knowledge from vast scientific corpora. The platform capitalizes on DARWIN, an open-source, fine-tuned LLM dedicated to natural science. The platform was built on Amazon Web Services (AWS) and provides an automated, user-friendly workflow for custom model development and data extraction. The platform achieves remarkable accuracy with only a small amount of well-annotated articles. This innovative tool streamlines the transition from the science literature to structured knowledge and data and benefits the advancements in natural informatics.     Source · CC BY 4.0                                  No publications match your search.   "
}







,{
  "title": "Professional Services",
  "url": "/service/",
  "content": " Professional Services Conference Reviewer: WWW 2027, ICLR 2027, NeurIPS 2026 (Top Reviewer), ICML 2026, ACL Rolling Review 2026 (October cycle), ACL Rolling Review 2026 (August cycle), ACL Rolling Review 2026 (May cycle), CIKM 2025, ICDM 2024, IEEE ICPADS 2023. Journal Reviewer: IEEE IoT-J 2026, IEEE JBHI 2026, Smart Health 2026, Multimedia Tools and Applications 2026. External Reviewer: ACM SEC 2026, ICDM 2025, WWW 2024, AAAI 2024, WACV 2024, ICNC 2024, EISA 2023, Inscrypt 2023. Student Volunteer, NSF CISE MSI Aspiring PI Workshop, University of North Texas, Dec 2024. Student Volunteer, IEEE MOST 2024, , May 2024.  "
}





,{
  "title": "Sitemap",
  "url": "/sitemap/",
  "content": " {% include base_path %}  A list of all the posts and pages found on the site. For you robots out there is an XML version available for digesting as well.  Pages {% for post in site.pages %}   {% include archive-single.html %} {% endfor %}  Posts {% for post in site.posts %}   {% include archive-single.html %} {% endfor %}  {% capture written_label %}’None’{% endcapture %}  {% for collection in site.collections %} {% unless collection.output == false or collection.label == “posts” %}   {% capture label %}{{ collection.label }}{% endcapture %}   {% if label != written_label %} {{ label }} {% capture written_label %}{{ label }}{% endcapture %}   {% endif %} {% endunless %} {% for post in collection.docs %}   {% unless collection.output == false or collection.label == “posts” %}   {% include archive-single.html %}   {% endunless %} {% endfor %} {% endfor %}  "
}





,{
  "title": "Posts by Tags",
  "url": "/tags/",
  "content": " {% include base_path %} {% include group-by-array collection=site.posts field=”tags” %}  {% for tag in group_names %}   {% assign posts = group_items[forloop.index0] %} {{ tag }} {% for post in posts %}     {% include archive-single.html %}   {% endfor %} {% endfor %}  "
}





,{
  "title": "Talk map",
  "url": "/talkmap.html",
  "content": " This map is generated from a Jupyter Notebook file in /_talks/talkmap.ipynb, which mines the location fields in the .md files in _talks/.   "
}





,{
  "title": "Talks & Presentations",
  "url": "/talks/",
  "content": " Talks &amp; Presentations MixLLM: LLM Quantization with Global Mixed-precision between Output-features and Highly-efficient System Design. Research presentation, LLaVi Lab AI Seminar, UNT, Sep 2026.  Responsible AI. Invited talk, Presidential AI Challenge (Denton ISD and UNT), May 2026. Audience: educators and industry professionals. Gated Attention for Large Language Models: Non-linearity, Sparsity, and Attention-Sink-Free. Research presentation, LLaVi Lab AI Seminar, UNT, Apr 2026.  SageAttention: Accurate 8-Bit Attention for Plug-and-play Inference Acceleration. Research presentation, LLaVi Lab AI Seminar, UNT, Nov 2025.  Responsible AI: The Hidden Curriculum. Invited talk, AIxHEART, Oct 2025. Professional development session for middle-school teachers. Efficient Long-Context Inference through Dynamic Sparse Attention. Poster presentation, UNT University Research Day, Oct 2025.  DAM: Dynamic Attention Mask for Long-Context Large Language Model Inference Acceleration. Conference presentation, ACL 2025 (Findings), Jul 2025.  D2SO: Detecting Distant and Small Objects for Vision-Based Vehicle Autonomous Systems. Conference presentation, IEEE MOST 2025, May 2025.  Adaptive Attention Masks for Large Language Model Acceleration. Poster presentation, North Texas Symposium on Generative AI, UT Arlington, Mar 2025.  Large Language Diffusion Models. Research presentation, LLaVi Lab AI Seminar, UNT, Feb 2025.  Efficient Small Language Models for Edge Devices via Task-Aware Sparse Attention. Proposal presentation, NSF IUCRC eCAT meeting, Nov 2024.  Poly-FEVER: A Multilingual Hallucination Detection Benchmark. Poster presentation, UNT University Research Day, Oct 2024.  MoA: Mixture of Sparse Attention for Automatic Large Language Model Compression. Research presentation, LLaVi Lab AI Seminar, UNT, Oct 2024.  Small Language Model. Research presentation, LLaVi Lab AI Seminar, UNT, Apr 2024.  Investigating Biased Cross-Lingual Hallucination Detection Capabilities of LLMs: POLY-FEVER Benchmark &amp; Mitigating Approaches. Poster presentation, UNT Engineering Showcase, Mar 2024. First Place.  "
}





,{
  "title": "Teaching & Mentoring",
  "url": "/teaching/",
  "content": "    {{ page.title }}         Teaching Experience     {% for entry in site.data.teaching.teaching %}       {% include editorial/teaching-entry.html entry=entry %}     {% endfor %}            Mentoring &amp; Advising     {% for entry in site.data.teaching.mentoring %}       {% include editorial/teaching-entry.html entry=entry %}     {% endfor %}            What My Students Say     {% for group in site.data.student_feedback %}            {{ group.title | escape }}       {% for quote in group.quotes %}                “{{ quote.paragraphs | join: ' ' | escape }}”              {% endfor %}          {% endfor %}      "
}





,{
  "title": "Terms and Privacy Policy",
  "url": "/terms/",
  "content": " {% include base_path %} {% include toc %}  Privacy Policy  The privacy of my visitors is extremely important. This Privacy Policy outlines the types of personal information that is received and collected and how it is used.  First and foremost, I will never share your email address or any other personal information to anyone without your direct consent.  Log Files  Like many other websites, this site uses log files to help learn about when, from where, and how often traffic flows to this site. The information in these log files include:     Internet Protocol addresses (IP)   Types of browser   Internet Service Provider (ISP)   Date and time stamp   Referring and exit pages   Number of clicks   All of this information is not linked to anything that is personally identifiable.  Cookies and Web Beacons  When you visit this site “convenience” cookies are stored on your computer when you submit a comment to help you log in faster to Disqus the next time you leave a comment.  Third-party advertisers may also place and read cookies on your browser and/or use web beacons to collect information. This site has no access or control over these cookies. You should review the respective privacy policies on any and all third-party ad servers for more information regarding their practices and how to opt-out.  If you wish to disable cookies, you may do so through your web browser options. Instructions for doing so can be found on the specific web browsers’ websites.  Google Analytics  Google Analytics is a web analytics tool I use to help understand how visitors engage with this website. It reports website trends using cookies and web beacons without identifying individual visitors. You can read Google Analytics Privacy Policy.  "
}





,{
  "title": "Blog posts",
  "url": "/year-archive/",
  "content": " {% include base_path %} {% capture written_year %}’None’{% endcapture %} {% for post in site.posts %}   {% capture year %}{{ post.date | date: ‘%Y’ }}{% endcapture %}   {% if year != written_year %}     &lt;h2 id=\"{{ year | slugify }}\" class=\"archive__subtitle\"&gt;{{ year }}&lt;/h2&gt;     {% capture written_year %}{{ year }}{% endcapture %}   {% endif %}   {% include archive-single.html %} {% endfor %}  "
}





























,{
  "title": "Editorial academic template assets",
  "url": "/assets/vendor/sbryngelson/",
  "content": " Editorial academic template assets  Adapted from sbryngelson/academic-website-template at commit 1fe64d1c1e8a6465ac267534b095119101dfa550.  Sources and output     _sass/editorial/ contains the template’s complete _sass/ tree with local color and readability adaptations.   _sass/editorial/theme.scss is the template’s assets/main.scss with only Jekyll front matter removed. This is the compilation entry point.   assets/editorial.css is the complete, compressed output, compiled with Dart Sass 1.108.1. It includes the template’s selected Bootstrap 5.3.8 modules and editorial styles. It can be served by the existing Jekyll 3 / GitHub Pages build without adding Ruby plugins.   assets/fonts/ contains the template’s exact Source Serif 4, DM Sans, and JetBrains Mono WOFF2 files, including Latin and Latin Extended subsets and their original license files.   Font URLs such as fonts/dm-sans-latin-normal-wght.woff2 are relative to /assets/editorial.css, so they resolve to /assets/fonts/ without modification. All fonts and styles are served locally.  The upstream tree contains a legacy _sass/bootstrap.scss file based on Bootswatch/Bootstrap 4.5.2. It is retained as source but is not imported by theme.scss. The compiled editorial CSS does not contain Lato or the previous template’s Bootstrap 4.6 styles.  Rebuild  From the repository root, using the official Sass npm package:  npx --yes --package sass@1.108.1 sass --load-path=_sass/editorial --style=compressed --no-source-map --quiet-deps --silence-deprecation=import _sass/editorial/theme.scss assets/editorial.css   Alternatively, install Dart Sass 1.108.1 outside the website repository and run its sass command with the same arguments. The migration used a workspace-only pnpm install and node &lt;compiler-directory&gt;/node_modules/sass/sass.js; no package dependencies were added to this repository.  The unmodified Bootstrap source emits Sass deprecation warnings about legacy functions with this compiler; compilation succeeds. Commit the updated assets/editorial.css after changing Sass sources.  The warm/cool palette is defined in _sass/editorial/base/_variables.scss. The light-mode cool base is also set by accent_color in _config.yml. Warm accents mark internal links, in-page disclosures, research keywords, and heading rules; cool accents mark external links, resource buttons, and year badges. Navigation branding and inactive labels stay neutral; only active and hovered navigation items use the warm accent. All pill labels and buttons are borderless, with a warm tint for disclosures and keywords. Link destinations are classified in assets/js/editorial.js, using both the current and canonical site origins so previews behave like the live site. Dark-mode values are explicit, print colors remain neutral, and text selection uses the theme’s contrast foreground.  Licenses and provenance     Template MIT license: LICENSE.   Bootstrap 5.3.8 MIT license: Bootstrap-LICENSE.   Legacy Bootswatch MIT license: Bootswatch-LICENSE.   Fonts: ../../fonts/LICENSE-source-serif-4.md, ../../fonts/LICENSE-dm-sans.txt, and ../../fonts/LICENSE-jetbrains-mono.txt (SIL Open Font License 1.1).   manifest.json records source paths, the pinned template revision, and SHA-256 hashes of copied sources and the initial compiled stylesheet.   The website’s existing root license remains in place.  Trailing spaces in the template MIT license and the DM Sans / JetBrains Mono license notices were removed for repository whitespace checks; their notice text is unchanged. The manifest records these whitespace-only adaptations.  "
}





,{
  "title": "Jupyter notebook markdown generator",
  "url": "/markdown_generator/",
  "content": " Jupyter notebook markdown generator  These .ipynb files are Jupyter notebook files that convert a TSV containing structured data about talks (talks.tsv) or presentations (presentations.tsv) into individual markdown files that will be properly formatted for the academicpages template. The notebooks contain a lot of documentation about the process. The .py files are pure python that do the same things if they are executed in a terminal, they just don’t have pretty documentation.  "
}











,{
  "title": "Docker Setup Instruction",
  "url": "/posts/2023/10/16/docker_setup_instruction/",
  "content": " This guide provides instructions on installing the Docker engine on Ubuntu, alongside essential commands for frequent use. It also covers setting up environments for PyTorch and TensorFlow, optimizing GPU usage, and managing the export and import of containers.  1. Install Docker Engine on Ubuntu  1.1 Prerequisites  1.1.1 OS requirements  To install Docker Engine, you need the 64-bit version of one of these Ubuntu versions:     Ubuntu Lunar 23.04   Ubuntu Kinetic 22.10   Ubuntu Jammy 22.04 (LTS)   Ubuntu Focal 20.04 (LTS)   Docker Engine for Ubuntu is compatible with x86_64 (or amd64), armhf, arm64, s390x, and ppc64le (ppc64el) architectures.  # check ubuntu version lsb_release -a   No LSB modules are available. Distributor ID:\tUbuntu Description:\tUbuntu 22.04.2 LTS Release:\t    22.04 Codename:\t    jammy   1.1.2 Uninstall conflicting packages  for pkg in docker.io docker-doc docker-compose docker-compose-v2 podman-docker containerd runc; do sudo apt-get remove $pkg; done   apt-get might report that you have none of these packages installed.  If the following results appear:  You might want to run 'apt --fix-broken install' to correct these. The following packages have unmet dependencies:  nvidia-dkms-535 : Depends: nvidia-kernel-common-535 (&gt;= 535.104.05) but 535.86.10-0ubuntu1 is to be installed  nvidia-driver-535 : Depends: libnvidia-gl-535 (= 535.104.05-0ubuntu0.22.04.4) but 535.86.10-0ubuntu1 is to be installed                      Depends: nvidia-kernel-common-535 (&gt;= 535.104.05) but 535.86.10-0ubuntu1 is to be installed                      Recommends: libnvidia-compute-535:i386 (= 535.104.05-0ubuntu0.22.04.4)                      Recommends: libnvidia-decode-535:i386 (= 535.104.05-0ubuntu0.22.04.4)                      Recommends: libnvidia-encode-535:i386 (= 535.104.05-0ubuntu0.22.04.4)                      Recommends: libnvidia-fbc1-535:i386 (= 535.104.05-0ubuntu0.22.04.4)                      Recommends: libnvidia-gl-535:i386 (= 535.104.05-0ubuntu0.22.04.4) E: Unmet dependencies. Try 'apt --fix-broken install' with no packages (or specify a solution).   Try:  # First sudo apt update &amp;&amp; sudo apt upgrade -y # Second sudo apt --fix-broken install # Third sudo apt autoremove   1.2 Installation     Docker Engine comes bundled with Docker Desktop for Linux. This is the easiest and quickest way to get started.   Set up and install Docker Engine from Docker’s Apt repository.   1.2.1 Docker’s Apt repository  Before you install Docker Engine for the first time on a new host machine, you need to set up the Docker repository. Afterward, you can install and update Docker from the repository.     Set up Docker’s Apt repository.   # Add Docker's official GPG key: sudo apt-get update sudo apt-get install ca-certificates curl gnupg sudo install -m 0755 -d /etc/apt/keyrings curl -fsSL https://download.docker.com/linux/ubuntu/gpg | sudo gpg --dearmor -o /etc/apt/keyrings/docker.gpg sudo chmod a+r /etc/apt/keyrings/docker.gpg  # Add the repository to Apt sources: echo \\   \"deb [arch=\"$(dpkg --print-architecture)\" signed-by=/etc/apt/keyrings/docker.gpg] https://download.docker.com/linux/ubuntu \\   \"$(. /etc/os-release &amp;&amp; echo \"$VERSION_CODENAME\")\" stable\" | \\ sudo tee /etc/apt/sources.list.d/docker.list &gt; /dev/null sudo apt-get update      Install the Docker packages.   # Latest Version sudo apt-get install docker-ce docker-ce-cli containerd.io docker-buildx-plugin docker-compose-plugin # Specific Version # List the available versions: apt-cache madison docker-ce | awk '{ print $3 }' # Select the desired version and install: VERSION_STRING=5:24.0.0-1~ubuntu.22.04~jammy sudo apt-get install docker-ce=$VERSION_STRING docker-ce-cli=$VERSION_STRING containerd.io docker-buildx-plugin docker-compose-plugin      Verify that the Docker Engine installation is successful by running the hello-world image.   sudo docker run hello-world   1.3 Upgrade and Uninstall  1.3.1 Upgrade Docker Engine  To upgrade Docker Engine, follow step 2 of the installation instructions, choosing the new version you want to install.  1.3.2 Uninstall Docker Engine     Uninstall the Docker Engine, CLI, containerd, and Docker Compose packages:   sudo apt-get purge docker-ce docker-ce-cli containerd.io docker-buildx-plugin docker-compose-plugin docker-ce-rootless-extras      Images, containers, volumes, or custom configuration files on your host aren’t automatically removed. To delete all images, containers, and volumes:   sudo rm -rf /var/lib/docker sudo rm -rf /var/lib/containerd   You have to delete any edited configuration files manually.  2 Fundamental Commands     display the containers currently running:   docker ps      display all the containers (even those not running anymore):   docker ps -a       display the images locally saved:   docker images      remove a docker container:   docker stop container_name # if container is running docker rm container_name      remove all docker containers (not running anymore):   docker container prune      remove an image:   docker rmi image_name      remove all docker images (be very careful with this one!):   docker image prune -a   3 PyTorch on Docker  3.1 PyTorch Image  In Github, there are some repositories created by good developers. PyTorch Docker image is repository has many PyTorch images for most of the previous versions with compatible CUDA versions.  # docker pull image_name:image_tag docker pull pytorch/pytorch:latest docker pull pytorch/pytorch:2.1.0-cuda12.1-cudnn8-runtime docker pull pytorch/pytorch:2.1.0-cuda12.1-cudnn8-devel   The latest image comes with the latest stable versions of PyTorch, CUDA and cuDNN. There are also other tags of the form X-cuda-Y-cudnn-Z-runtime/devel, where X is the pytorch version, Y is the CUDA version and Z is the cuDNN version. The images tagged with devel come preinstalled with various compiler configurations.  4 TensorFlow on Docker  TensorFlow provides a number of images depending on your use case such as latest, nightly and devel. Assuming you have Docker installed on your computer we can download these images using commands such as  docker pull tensorflow/tensorflow  docker pull tensorflow/tensorflow:latest-gpu   These commands will install the latest stable release and the latest GPU compatible release respectively. For more detailed instructions please refer to the official documentation.  5 Using GPU on PyTorch Image  5.1 Prerequisites  Create an environment where the host can use the GPU. To test the installation, run:  nvidia-smi   5.2 Installing Nvidia Container Toolkit  Docker containers share your host’s kernel but bring their own operating system and software packages. This means they lack the NVIDIA drivers used to interface with your GPU. Docker doesn’t even add GPUs to containers by default, so a plain Docker run won’t see your hardware at all.  To get your GPU to work, you’ll have to install the drivers within your image and then instruct Docker to add GPU devices to your containers at runtime.  You’ll add NVIDIA Container Toolkit to your machine to enable your docker container to use your GPUs. This integrates into Docker Engine to automatically configure your containers for GPU support.  Setup the package repository and the GPG key:  distribution=$(. /etc/os-release;echo  $ID$VERSION_ID) curl -s -L https://nvidia.github.io/nvidia-docker/gpgkey | sudo apt-key add -   curl -s -L https://nvidia.github.io/nvidia-docker/$distribution/nvidia-docker.list | sudo tee /etc/apt/sources.list.d/nvidia-docker.list   Install the nvidia-container-toolkit package (and dependencies) after updating the package listing:  sudo apt-get update sudo apt-get install -y nvidia-container-toolkit   Now, configure the Docker daemon to recognize the NVIDIA Container Runtime:  sudo nvidia-ctk runtime configure --runtime=docker   Restart the Docker daemon to complete the installation after setting the default runtime:  sudo systemctl restart docker   At this point, the NVIDIA Container Toolkit is up and running, you’re ready to test its operation.  5.3 Docker with GPU  Docker doesn’t provide your system’s GPUs by default, you need to create containers with the --gpus flag for your hardware to show up. The nvidia/cuda images are preconfigured with the CUDA binaries and GPU tools.  sudo docker run -t -i -d --gpus all pytorch/pytorch   docker run = docker pull + docker create + docker start  You’ll get the CONTAINER ID like:  4056fed7496b7b88872f43f6087c7b76415eceff06f86a72f61f994b0357da63   Enter the container:  sudo docker exec -it 4056 /bin/bash # 4056 is the CONTAINER ID   To check if Docker can access your GPU, run the nvidia-smi command in your container.  nvidia-gpu   6 Docker Export and Import  If you want to save the entire image, including all its history and tags, then you should use the docker save and docker load commands. And if you just want to save the current state of a container, then you should use the docker export and docker import commands.  6.1 Docker Container Export and Import  This operation will package all the image layers and metadata into a tar file. You can then use the docker load command to import this tar file into any Docker environment.  sudo docker export &lt;container ID&gt; &gt; mycontainer.tar # sudo docker export e2ca7008de40 &gt; hallu_container.tar   This approach is primarily used to share or migrate an entire image, including all versions, tags, and history.  6.1.2 Import and Use  Import as an image  sudo docker import mycontainer.tar &lt;image name&gt; # sudo docker import hallu_container.tar hallu_backup   Create a container with the same command as original one  sudo docker run &lt;image name&gt; /bin/bash # sudo docker run -t -i -d --gpus all hallu_backup:latest /bin/bash   6.2 Docker Image Export and Import  The docker export command exports the file system of a running container to a tar file. You can then use the docker import command to import this tar file as a new image.  sudo docker save -o &lt;save path&gt;/myimage.tar myimage:latest   sudo docker load -i &lt;path&gt;/myimage.tar   This approach is primarily used to share or migrate the current state of a container. This does not include the container’s history or metadata, such as environment variables, so it is often used to take a snapshot of the container.  References     Install Docker Engine on Ubuntu   Why use Docker containers for machine learning development?   Using Docker for Deep Learning projects   PyTorch With Docker   pytorch/pytorch   Setting Up TensorFlow And PyTorch Using GPU On Docker   TensorFlow Docker   How to Install PyTorch on the GPU with Docker   "
}





,{
  "title": "TACC Lonestar6 Instruction",
  "url": "/posts/2025/03/04/TACC_Longstar6_instruction/",
  "content": " This is the quick start instruction for beginners on Texas Advanced Computing Center (TACC) Lonestar6 High Performance Computing (HPC) Systems. For more details information, please refer to the official website.  At first, you should have an account for TACC. Log In or Create Account.  Remember your username and password. Only when you have/are added to certain project, you have access to corresponding system. For this document, we use Lonestar6 system.  0. Not the First Time Cheat Sheet     Login: ssh &lt;username&gt;@ls6.tacc.utexas.edu   Go to work directory: cd $WORK   File Transfer: Use scp or rsync   Environment setup: source ~/.bashrc   Interactive Job Session: srun --partition=gpu-a100-dev --nodes=1 --time=00:30:00 --ntasks=1 --pty bash   1. Access the System  The ssh command (SSH protocol) is the standard way to connect to Lonestar6 (ls6.tacc.utexas.edu). SSH also includes support for the file transfer utilities scp and sftp.  The Linux command line:  localhost$ ssh &lt;username&gt;@ls6.tacc.utexas.edu   The above command will rotate connections across all available login nodes, login1-login3, and route your connection to one of them. To connect to a specific login node, use its full domain name:  localhost$ ssh &lt;username&gt;@login2.ls6.tacc.utexas.edu   To connect with X11 support on Lonestar6 (usually required for applications with graphical user interfaces), use the -X or -Y switch:  localhost$ ssh -X &lt;username&gt;@ls6.tacc.utexas.edu   To report a connection problem, execute the ssh command with the -vvv option and include the verbose output when submitting a help ticket. Do not run the ssh-keygen command on Lonestar6.  SSH Config Example:  Host TACC     HostName ls6.tacc.utexas.edu     User &lt;username&gt;   When connecting, you’ll be asked to enter your password and TACC Token Code.  To access the system:  1) If not using ssh-keys, please enter your TACC password at the password prompt 2) At the TACC Token prompt, enter your 6-digit code followed by &lt;return&gt;.  If you are facing issues logging in, please use our login wizard at https://accounts.tacc.utexas.edu/login_support to troubleshoot.  (&lt;username&gt;@ls6.tacc.utexas.edu) Password: &lt;your password&gt; (&lt;username&gt;@ls6.tacc.utexas.edu) TACC Token Code: &lt;Duo Mobile Token&gt;   2. Working Directory  Lonestar6’s startup mechanisms define corresponding account-level environment variables $HOME, $SCRATCH and $WORK that store the paths to directories that you own on each of these file systems.  Your home directory $HOME do not have enough space, go to your account-specific working directory $WORK once you login.  $ cd $WORK $ pwd /work/&lt;number&gt;/&lt;username&gt;/ls6                  File System       Quota       Key Features                       $HOME       10 GB       200,000 files Not intended for parallel or high-intensity file operations. NFS file system Backed up regularly. Overall capacity 7 TB Not purged.                 $WORK       1 TB 3,000,000 files Across all TACC systems       Not intended for high-intensity file operations or jobs involving very large files. Lustre file system On the Global Shared File System that is mounted on most TACC systems. See Stockyard system description for more information. Defaults: 1 stripe, 1MB stripe size Not backed up. Not purged.                 $SCRATCH       none       Overall capacity 8 PB Defaults: 4 targets, 512 KB chunk size Not backed up Files are subject to purge if access time* is more than 10 days old.*                 /tmp on nodes       288 GB       Data purged at the end of each job. Access is local to the node. Data in /tmp is not shared across nodes.           3. File Transfer  You can transfer files between Lonestar6 and Linux-based systems using either scp or rsync. Both scp and rsync are available in the Mac Terminal app. Windows SSH clients typically include scp-based file transfer capabilities.  Using scp  scp &lt;local_file&gt; &lt;username&gt;@ls6.tacc.utexas.edu:$WORK   Using rsync (for large or multiple files)  rsync -av &lt;local_dir&gt; &lt;username&gt;@ls6.tacc.utexas.edu:$WORK   For a more user-friendly experience, consider using an application like Termius or cloud storage platforms such as GitHub or Dropbox. For example  $ git clone &lt;github repository link&gt; $ # Or $ wget -O &lt;file name&gt; &lt;dropbox link&gt;   4. Environment Setup  Put all your customizations in ~/.bashrc. Take mine as an example, anything that default in $HOME, change it to $WORK  export PYTHONPATH=\"$WORK/python-packages:$PYTHONPATH\" export NLTK_DATA=$WORK/python-packages/nltk_data export HF_HOME=$WORK/huggingface_cache/ export PATH=$WORK/python3.11/bin:$PATH export PIP_CACHE_DIR=$WORK/.cache/pip   After making these changes, run source ~/.bashrc to apply them.  Take your time with this step to ensure your environment is set up correctly. If needed, consider using AI tools for assistance.  5. Running Jobs  Lonestar6 uses the Simple Linux Utility for Resource Management (Slurm) batch environment                 Queue Name       Min/Max Nodes per Job (assoc’d cores)*       Max Job Duration       Max Nodes per User       Max Jobs per User       Charge Rate (per node-hour)                       development       4 nodes (512 cores)       2 hours       6       1       1 SU                 gpu-a100       8 nodes (1024 cores)       48 hours       12       8       4 SUs                 gpu-a100-dev       2 nodes (256 cores)       2 hours       2       1       4 SUs                 gpu-a100-small       1 node       48 hours       2       2       1.5 SUs                 gpu-h100       1 node       48 hours       1       1       6 SUs                 large       65/256 nodes (65536 cores)       48 hours       256       1       1 SU                 normal       1/64 nodes (8192 cores)       48 hours       75       20       1 SU                 vm-small       1/1 node (16 cores)       48 hours       4       4       0.143 SU           * Access to the large queue is restricted. To request more nodes than are available in the normal queue, submit a consulting (help desk) ticket through the TACC User Portal. Include in your request reasonable evidence of your readiness to run under the conditions you’re requesting. In most cases this should include your own strong or weak scaling results from Lonestar6.  ** The gpu-a100-small and vm-small queues contain virtual nodes with fewer resources (cores) than the nodes in the other queues.  Copy and customize the following scripts to specify and refine your job’s requirements.     specify the maximum run time with the -t option.   specify number of nodes needed with the -N option   specify total number of MPI tasks with the -n option   specify the project to be charged with the -A option.   An example for the Interactive Job Session:  $ srun --partition=gpu-a100-dev --nodes=1 --time=00:30:00 --ntasks=1 --pty bash $ # Or simply $ srun -p gpu-a100-dev -N 1 -t 00:30:00 -n 1 --pty bash   Slurm Cheat Sheet  Basic Commands                 Command       Description                       sinfo       Show partitions and node status.                 squeue       List queued and running jobs.                 squeue -u $USER       Show your jobs.                 sbatch script.sh       Submit a batch job.                 srun       Run a command or script interactively.                 scancel JOBID       Cancel a job.                 sacct       Show completed jobs.                 scontrol show job JOBID       Detailed job info.                 scontrol show node NODENAME       Node details.                 seff JOBID       Job efficiency details (if available).           Job Monitoring                 Command       Description                       squeue       List all jobs.                 squeue -u $USER       Your jobs only.                 scontrol show job JOBID       Detailed job info.                 sacct -u $USER       Completed job history.                 sacct -j JOBID --format=JobID,JobName,State,Elapsed       Summary for a job.           Resource Specifications                 Option       Description                       --nodes=2       Request 2 nodes.                 --ntasks=4       Request 4 tasks (MPI).                 --cpus-per-task=2       Request 2 CPU cores per task (OpenMP).                 --gres=gpu:2       Request 2 GPUs per node.                 --mem=16G       Request 16 GB RAM per node.                 --time=02:00:00       Max runtime of 2 hours.                 --partition=gpu-a100       Specify partition/queue.                 --account=your_account       Charge to a specific account.           References     Lonestar6 User Guide   Slurm Workload Manager   "
}





,{
  "title": "CSCE 5300: Introduction to Big Data and Data Science",
  "url": "/teaching/2024-2025-csce-5300/",
  "content": " Graduate course, 79 students in Fall 2024 and 64 students in Spring 2025, University of North Texas.     Received an overall teaching evaluation of Excellent in both semesters.   Managed grading and office hours, coordinated randomized-question exams, and analyzed score distributions.   "
}





,{
  "title": "Long-Context Evaluation of Large Language Models",
  "url": "/teaching/2026-summer-nsf-reu/",
  "content": " Delivered a guest lecture on Long-Context Evaluation of Large Language Models in UNT’s NSF Research Experiences for Undergraduates (REU) program, alongside her work as Academic Coordinator and Research Mentor.  "
}





,{
  "title": "CSCE 4201: Introduction to Artificial Intelligence",
  "url": "/teaching/2026-fall-csce-4201/",
  "content": " Undergraduate course, 74 students, University of North Texas.     Independently designed and delivered three 3-hour lectures on CNNs, NLP, and Transformers.   Student feedback highlighted interactive whiteboard examples and explanations that helped students better understand the material.   "
}





,{
  "title": "ByteScience: Bridging Unstructured Scientific Literature and Structured Data with Auto Fine-Tuned Large Language Model in Token Granularity",
  "url": "/publications/#pub-2024-bytescience-bridging-literature",
  "content": " "
}





,{
  "title": "D2SO: Detecting Distant and Small Objects for Vision-Based Vehicle Autonomous Systems",
  "url": "/publications/#pub-2025-d2so-detecting-distant-small-objects",
  "content": " "
}





,{
  "title": "HALO: Hallucination Analysis and Learning Optimization to Empower LLMs with Retrieval-Augmented Context for Guided Clinical Decision Making",
  "url": "/publications/#pub-2025-halo-hallucination-analysis",
  "content": " "
}





,{
  "title": "DAM: Dynamic Attention Mask for Long-Context Large Language Model Inference Acceleration",
  "url": "/publications/#pub-2025-dam-dynamic-attention-mask",
  "content": " "
}





,{
  "title": "Harmful Factuality: LLMs Correcting What They Shouldn't",
  "url": "/publications/#pub-2026-harmful-factuality",
  "content": " "
}





,{
  "title": "TileMix: Tile-Centric Mixed-Precision Attention for LLM Inference Acceleration",
  "url": "/publications/#pub-2026-tilemix",
  "content": " "
}





,{
  "title": "Poly-FEVER: A Multilingual Hallucination Detection Benchmark",
  "url": "/publications/#pub-2025-poly-fever-multilingual-fact-verification",
  "content": " "
}





,{
  "title": "AlignQuant: Tile-Aligned Mixed-Precision Quantization for Efficient LLM Generation",
  "url": "/publications/#pub-2026-alignquant",
  "content": " "
}





,{
  "title": "TileQuant: Tile as the Precision-Allocation Unit for LLM Mixed-Precision Post-Training Quantization",
  "url": "/publications/#pub-2026-tilequant",
  "content": " "
}





,{
  "title": "TriAttention: Unified Sparse Mixed-Precision Attention for Long-Context LLM Inference",
  "url": "/publications/#pub-2026-triattention",
  "content": " "
}





,{
  "title": "Towards Visual Query Segmentation in the Wild",
  "url": "/publications/#pub-2026-visual-query-segmentation",
  "content": " "
}





,{
  "title": "IEEEE MOST 2024",
  "url": "/portfolio/MOST2024/",
  "content": " IEEE MOST 2024  So lucky to meet Dr. Weisong Shi    Q&amp;A with Mr. Jeff White      New friends!    Many memories!        "
}





,{
  "title": "Generative AI Prompt A Thon",
  "url": "/portfolio/prompt_a_thon/",
  "content": " Read more about the event on the Generative AI Prompt-A-Thon blog post.  "
}





,{
  "title": "UNT College of Engineering Research Showcase",
  "url": "/talks/2024-03-27-poster",
  "content": " This poster has been selected to receive an outstanding poster award on the 1st place! As a winner of the Engineering Showcase Poster Presentation, Hanzhi Zhang have been selected to represent College of Engineering to present at the University Research Day annual poster session.    See poster details below:    "
}




,{"title": "News · Nov 2026", "url": "/news/#news-1", "content": "Our research paper, “TriAttention: Unified Sparse Mixed-Precision Attention for Long-Context LLM Inference”, is now available as a preprint on arXiv."}


,{"title": "News · Nov 2026", "url": "/news/#news-2", "content": "Our research paper, “TileQuant: Tile as the Precision-Allocation Unit for LLM Mixed-Precision Post-Training Quantization”, is now available as a preprint on arXiv."}


,{"title": "News · Oct 2026", "url": "/news/#news-3", "content": "Selected as a “Top Reviewer” for NeurIPS 2026"}


,{"title": "News · Oct 2026", "url": "/news/#news-4", "content": "Served as a reviewer for ICLR 2027"}


,{"title": "News · Oct 2026", "url": "/news/#news-5", "content": "Served as a reviewer for ACL Rolling Review 2026 (October cycle)"}


,{"title": "News · Oct 2026", "url": "/news/#news-6", "content": "Gave lectures on CNNs, NLP, and Transformers for CSCE 4201"}


,{"title": "News · Oct 2026", "url": "/news/#news-7", "content": "Started research collaboration with Amazon scientist Qian Qu"}


,{"title": "News · Oct 2026", "url": "/news/#news-8", "content": "Our research paper, “AlignQuant: Tile-Aligned Mixed-Precision Quantization for Efficient LLM Generation”, is now available as a preprint on arXiv."}


,{"title": "News · Sep 2026", "url": "/news/#news-9", "content": "Gave a presentation on “MixLLM: LLM Quantization with Global Mixed-precision between Output-features and Highly-efficient System Design” at UNT's LLaVi Lab AI Seminar"}


,{"title": "News · Sep 2026", "url": "/news/#news-10", "content": "Our research paper, “Towards Visual Query Segmentation in the Wild”, has been accepted for publication in Advances in Neural Information Processing Systems (NeurIPS 2026). I am a co-first author of this work."}


,{"title": "News · Sep 2026", "url": "/news/#news-11", "content": "I contributed to the Amazon Research Awards (ARA) project “TileRoute-Trainium: Tile-Level Mixed-Precision Kernels for Long-Context LLM Inference” as lead student researcher and proposal co-writer."}


,{"title": "News · Aug 2026", "url": "/news/#news-12", "content": "Served as a reviewer for ACL Rolling Review 2026 (August cycle)"}


,{"title": "News · Aug 2026", "url": "/news/#news-13", "content": "Our research paper, “Poly-FEVER: A Multilingual Hallucination Detection Benchmark”, has been accepted for publication in Findings of the Association for Computational Linguistics (EMNLP 2026)."}


,{"title": "News · Aug 2026", "url": "/news/#news-14", "content": "Our research paper, “TileMix: Tile-Centric Mixed-Precision Attention for LLM Inference Acceleration”, is now available as a preprint on arXiv."}


,{"title": "News · Summer 2026", "url": "/news/#news-15", "content": "Served as Academic Coordinator and Research Mentor for nine undergraduates in UNT's NSF REU program"}


,{"title": "News · Summer 2026", "url": "/news/#news-16", "content": "Gave a guest lecture on “Long-Context Evaluation of Large Language Models” at UNT's NSF REU program"}


,{"title": "News · Jun 2026", "url": "/news/#news-17", "content": "Started research collaboration with ORNL's AAIMS Group"}


,{"title": "News · May 2026", "url": "/news/#news-18", "content": "Served as an external reviewer for ACM SEC 2026"}


,{"title": "News · May 2026", "url": "/news/#news-19", "content": "Served as a reviewer for NeurIPS 2026"}


,{"title": "News · May 2026", "url": "/news/#news-20", "content": "Gave an invited talk on “Responsible AI” at the Presidential AI Challenge"}


,{"title": "News · May 2026", "url": "/news/#news-21", "content": "Served as a reviewer for ACL Rolling Review 2026 (May cycle)"}


,{"title": "News · Apr 2026", "url": "/news/#news-22", "content": "Gave a presentation on “Gated Attention for Large Language Models: Non-linearity, Sparsity, and Attention-Sink-Free” at UNT's LLaVi Lab AI Seminar"}


,{"title": "News · Mar 2026", "url": "/news/#news-23", "content": "Our research paper, “Towards Visual Query Segmentation in the Wild”, is now available as a preprint on arXiv."}


,{"title": "News · Mar 2026", "url": "/news/#news-24", "content": "Our research paper, “Harmful Factuality: LLMs Correcting What They Shouldn’t”, has been published in Findings of the Association for Computational Linguistics (EACL 2026)."}


,{"title": "News · Feb 2026", "url": "/news/#news-25", "content": "Served as a reviewer for ICML 2026"}


,{"title": "News · Nov 2025", "url": "/news/#news-26", "content": "Gave a presentation on “SageAttention: Accurate 8-Bit Attention for Plug-and-play Inference Acceleration” at UNT's LLaVi Lab AI Seminar"}


,{"title": "News · Oct 2025", "url": "/news/#news-27", "content": "Gave an invited talk on “Responsible AI: The Hidden Curriculum” at AIxHEART"}


,{"title": "News · Oct 2025", "url": "/news/#news-28", "content": "Presented a poster on “Efficient Long-Context Inference through Dynamic Sparse Attention” at UNT University Research Day"}


,{"title": "News · Jul 2025", "url": "/news/#news-29", "content": "Gave a presentation on “DAM: Dynamic Attention Mask for Long-Context Large Language Model Inference Acceleration” at ACL 2025 (Findings)"}


,{"title": "News · Jun 2025", "url": "/news/#news-30", "content": "Served as an external reviewer for ICDM 2025"}


,{"title": "News · May 2025", "url": "/news/#news-31", "content": "Served as a reviewer for CIKM 2025"}


,{"title": "News · May 2025", "url": "/news/#news-32", "content": "Our research paper, “DAM: Dynamic Attention Mask for Long-Context Large Language Model Inference Acceleration”, has been accepted for publication in Findings of the Association for Computational Linguistics (ACL 2025)."}


,{"title": "News · May 2025", "url": "/news/#news-33", "content": "Gave a presentation on “D2SO: Detecting Distant and Small Objects for Vision-Based Vehicle Autonomous Systems” at IEEE MOST 2025"}


,{"title": "News · Apr 2025", "url": "/news/#news-34", "content": "Received a Student Travel Grant Award from IEEE MOST 2025"}


,{"title": "News · Apr 2025", "url": "/news/#news-35", "content": "Participated in the Women-in-Computing session at IEEE MOST 2025"}


,{"title": "News · Mar 2025", "url": "/news/#news-36", "content": "Our research paper, “Poly-FEVER: A Multilingual Fact Verification Benchmark for Hallucination Detection in Large Language Models”, is now available as a preprint on arXiv."}


,{"title": "News · Mar 2025", "url": "/news/#news-37", "content": "Our research paper, “D2SO: Detecting Distant and Small Objects for Vision-Based Vehicle Autonomous Systems”, has been accepted for publication in the Proceedings of the IEEE International Conference on Mobility, Operations, Services and Technologies (IEEE MOST 2025)."}


,{"title": "News · Mar 2025", "url": "/news/#news-38", "content": "Presented a poster on “Adaptive Attention Masks for Large Language Model Acceleration” at the North Texas Symposium on Generative AI at UT Arlington"}


,{"title": "News · Feb 2025", "url": "/news/#news-39", "content": "Gave a presentation on “Large Language Diffusion Models” at UNT's LLaVi Lab AI Seminar"}


,{"title": "News · Feb 2025", "url": "/news/#news-40", "content": "Passed PhD Qualifying Exam"}


,{"title": "News · Feb 2025", "url": "/news/#news-41", "content": "Our research paper, “HALO: Hallucination Analysis and Learning Optimization to Empower LLMs with Retrieval-Augmented Context for Guided Clinical Decision Making”, has been accepted for publication in the Proceedings of the IEEE/ACM Conference on Connected Health: Applications, Systems and Engineering Technologies (CHASE 2025)."}


,{"title": "News · Spring 2025", "url": "/news/#news-42", "content": "Served as a Teaching Assistant for CSCE 5300: Introduction to Big Data and Data Science (64 graduate students)"}


,{"title": "News · Dec 2024", "url": "/news/#news-43", "content": "Served as a student volunteer at the NSF CISE MSI Aspiring PI Workshop at UNT"}


,{"title": "News · Dec 2024", "url": "/news/#news-44", "content": "Our research paper, “ByteScience: Bridging Unstructured Scientific Literature and Structured Data with Auto Fine-Tuned Large Language Model in Token Granularity”, has been published in the Proceedings of the IEEE International Conference on Data Mining Workshops (ICDMW 2024)."}


,{"title": "News · Nov 2024", "url": "/news/#news-45", "content": "Presented a proposal on “Efficient Small Language Models for Edge Devices via Task-Aware Sparse Attention” at the NSF IUCRC eCAT meeting"}


,{"title": "News · Oct 2024", "url": "/news/#news-46", "content": "Gave a presentation on “MoA: Mixture of Sparse Attention for Automatic Large Language Model Compression” at UNT's LLaVi Lab AI Seminar"}


,{"title": "News · Oct 2024", "url": "/news/#news-47", "content": "Presented a poster on “Poly-FEVER: A Multilingual Hallucination Detection Benchmark” at UNT University Research Day"}


,{"title": "News · Fall 2024", "url": "/news/#news-48", "content": "Served as a Teaching Assistant for CSCE 5300: Introduction to Big Data and Data Science (79 graduate students)"}


,{"title": "News · Jul 2024", "url": "/news/#news-49", "content": "Participated in the Argonne Training Program on Extreme-Scale Computing (ATPESC) 2024"}


,{"title": "News · Jul 2024", "url": "/news/#news-50", "content": "Started serving as PI for ALCF Polaris computing allocations for “Large Language Models’ Hallucination Detection and Mitigation Ability across Multi Languages” (5,000 node-hours per allocation)"}


,{"title": "News · Jun 2024", "url": "/news/#news-51", "content": "Served as a reviewer for ICDM 2024"}


,{"title": "News · May 2024", "url": "/news/#news-52", "content": "Served as a student volunteer at IEEE MOST 2024"}


,{"title": "News · May 2024", "url": "/news/#news-53", "content": "Participated in the AI Prompt-a-Thon at UT Arlington with Microsoft"}


,{"title": "News · Apr 2024", "url": "/news/#news-54", "content": "Gave a presentation on “Small Language Model” at UNT's LLaVi Lab AI Seminar"}


,{"title": "News · Mar 2024", "url": "/news/#news-55", "content": "I won First Place for the poster “Investigating Biased Cross-Lingual Hallucination Detection Capabilities of LLMs: POLY-FEVER Benchmark & Mitigating Approaches” at UNT Engineering Showcase."}


,{"title": "News · Oct 2023", "url": "/news/#news-56", "content": "Served as an external reviewer for WWW 2024"}


,{"title": "News · Oct 2023", "url": "/news/#news-57", "content": "Served as an external reviewer for ICNC 2024"}


,{"title": "News · Sep 2023", "url": "/news/#news-58", "content": "Served as a reviewer for IEEE ICPADS 2023"}


,{"title": "News · Aug 2023", "url": "/news/#news-59", "content": "Served as an external reviewer for AAAI 2024"}


,{"title": "News · Aug 2023", "url": "/news/#news-60", "content": "Started PhD in Computer Science at UNT and joined Responsible AI Lab"}


,{"title": "News · Jul 2023", "url": "/news/#news-61", "content": "Served as an external reviewer for EISA 2023"}


,{"title": "News · Jul 2023", "url": "/news/#news-62", "content": "Served as an external reviewer for Inscrypt 2023"}


,{"title": "News · Jul 2023", "url": "/news/#news-63", "content": "Served as an external reviewer for WACV 2024"}


]
