🔥The 9th Round of Easy Loan, Earn $40 Reward is in progress❗️ ⏰ Promotion Period: January 15th - Feburary 15th, 2025 👉 Register now and check more details at gate.io/campaigns/358

A University of Maryland faculty-student team has cracked a mathematical quandary that stumped scientists for more than 35 years, but it’s more than just a clever solution to a puzzle: The discovery could boost the power and energy efficiency of the chips used to run computers

A team led by Mohammad Hajiaghayi (Mohammad Hajiaghayi (IG@mhajiaghayi,YT:hajiaghayi)) has solved the Steiner Forest problem—a 35-year-old mathematical quandary. The resulting 126-page proof and algorithm could boost chip design for computing and AI, cutting energy use and costs. Read more: go.umd.edu/1zd2

A recent The New York Times article quoted Soheil Feizi (Soheil Feizi) and cited his study that found A.I.-detection services erroneously flag human-written text as A.I.-generated about 6.8% of the time. That same study has also been accepted to #ACL2025. nytimes.com/2025/05/17/sty…

🥳Excited to share our paper RESTOR: Knowledge Recovery in Machine Unlearning is accepted to #TMLR! We introduce a benchmark where unlearning means not just forgetting target data, but recovering the behavior of an ideal model that never saw them. Paper: openreview.net/pdf?id=BbwlJpN…

LLMs choose tools based on how they’re described—not how they work. 🤔 We show that simple edits to tool descriptions can shift usage to 10× between functionally identical tools. This reveals a key flaw in today’s tool/function-calling protocols. 📜arxiv.org/abs/2505.18135

1/ 🌎 Real-world datasets are messy—some domains dominate while others barely show up. This imbalance leads traditional #RLHF (like #GRPO) to favor prevalent domains, hurting fairness and generalization. So, how do we fix this? 🔥 Our new method, 🪩DISCO 🪩, elegantly tackles

Across college campuses, more and more students are using artificial intelligence tools to complete their assignments. Some schools are now using AI detection software to increase academic integrity. However, as Kelly O'Grady found out, the programs are not always accurate.

In a first of its kind study, UMIACS researchers (Furong Huang, Yizheng Chen & Jia-Bin Huang) uncovered crucial factors underlying vulnerabilities of web AI agents, highlighting the need for enhanced security measures. Learn more: go.umd.edu/1zqr

An interview w/Soheil Feizi by Kelly O'Grady on AI text detection was featured CBS Saturday Morning. Feizi explains how current AI detection tools are often unreliable, & should be used with caution in high-stakes settings involving academic integrity. Watch: go.umd.edu/1zt5

🎥 Multi-modal LLMs are great at real-world video understanding… But what happens when they’re shown generated videos from Sora, Veo 2/3, or KLing? Turns out: Text-to-video generation still breaks reality 🤯 A quail morphs into a rooster mid-frame 🐓. A bullet shoots back into

(3/n) We design two types of questions to uncover failures: ❓ Targeted Probes – Ask about specific, counterintuitive events: 🗣️ “The watermelon breaks in the middle. Is it intact at the end?” 🧠 Open-ended Abnormality Checks – Prompt critical thinking: 🗣️ “What is unusual in

(7/n) Follow our work for more! 📄 Paper: arxiv.org/abs/2505.01481 💻 Code: github.com/zli12321/Video… 🌐 Website: wuxiyang1996.github.io/videohallu_pag… 📦 Dataset: huggingface.co/datasets/Intel… 🧠 Co-led by Zongxia Li with amazing collaborators: Guangyao Shi, Yubin Qin, Hongyang Du, Tianyi Zhou,

We’re hiring! 🎓 Feizi Lab has an open Postdoctoral Position in LLM Reasoning, supported by our PECASE award and other projects. If you're excited about foundational research in LLMs, apply here: forms.gle/uupq4wPz2EQnC4…

😀 Excited to share that my PhD student Faisal Hamman has received the ECE UMD Distinguished Dissertation Award! Faisal's research focuses on trustworthy and interpretable machine learning with multiple first-author publications at leading AI conferences #ICML2025 #ICLR2024

Excited to speak at the Workshop on Computer Vision in the Wild #CVPR2025 2025! 🎥🌍 🗓️ June 11 | 📍 Room 101 B, Music City Center, Nashville, TN 🎸 🧠 Talk: From Perception to Action: Building World Models for Generalist Agents Let’s connect if you're around! #CVPR2025 #robotics

Researchers Univ. of Maryland have developed a novel system that teaches complex tasks to robots using simple human drawings. Their work will be presented later this month Robotics: Science and Systems. #RSS2025 Learn more: umiacs.umd.edu/news-events/ne…

📢 New paper at #ICML2025 :Quantifying Prediction Consistency Under Model Multiplicity in Tabular LLMs ICML Conference #TabLLM #robustness #consistency #TrustworthyAI #ResponsibleAI 🔗arxiv.org/html/2407.0417… Joint work with: Faisal Hamman Pasan Dissanayake J.P. Morgan

🚨 New Paper: Adversarial Paraphrasing: A Universal Attack for Humanizing AI-Generated Text AI-text detection has become an essential topic due to the increasing power of large language models. Many detectors rely on subtle stylistic signals to flag AI-generated text but these

#CVPR2025 Our latest work LEMON will appear at CVPR CVinW spotlight talks tomorrow - a unified, scalable, efficient 3D LMM with elegant architecture design. Everything about 🍋 will be open-sourced this month. 🔗: computer-vision-in-the-wild.github.io/cvpr-2025/

Honored to have our research featured! Excited to present our work at #RSS2025, where we show how simple sketches can empower robots to learn complex tasks. Grateful to my amazing collaborators and mentors for making this possible! #robotics #humanrobotinteraction #innovation