Michael-Andrei Panaitescu-Liess

PhD Student @umdcs. Working on Adversarial Machine Learning

[Furong Huang @furongh]

As I reflect on my journey as a faculty member over the past 7 years, I am overwhelmed with pride and gratitude. What started as a single-student-single-PI lab has blossomed into a vibrant group of almost 20 brilliant PhD students, along with numerous masters and undergraduate students. Together, we've carved out our place in the research community, making contributions to Trustworthy AI/ML, Foundation Models for Sequential Decision-Making, Generative-AI Alignment, AI Efficiency, and Ethics in AI.

People often ask me how we manage to tackle such a diverse range of topics. My answer is simple: it's my incredible students. They are energetic, passionate, smart, diligent, and inspire me every single day. Being a faculty member isn't just about leading research--it's also about building relationships, growing together, and creating a lasting impact.

This time of year is always bittersweet. Late at night, I often find myself looking through old photos and reminiscing about the countless projects we've completed, the late nights we endured together, and the rebuttals we fought for. We've shared so many wins and challenges, and we've always stood as a team.

While it's hard to say goodbye, it's equally exciting to see them flourish, contributing to the research community and embarking on the next chapter of their lives. This year, I'm thrilled to introduce 7 outstanding students on the job market. I am confident you'll be amazed by their work. Hiring any of them would be a decision you won't regret! 🥳

Here's a proud advisor introducing her students (in alphabetical order):

Bang An @bang_an_ :
Bang’s research is dedicated to advancing Responsible AI, with a focus on enhancing the safety, alignment, robustness, and fairness of AI systems, particularly in Generative AI. Bang is the recipient of the Outstanding Graduate Assistant Award at UMD. She currently leads the organization of a NeurIPS’24 competition on image watermarks.
(Author of WAVES: Watermark Benchmark, Automatic Pseudo-Harmful Prompt Generation in LLMs, PerceptionCLIP, Fairness under Distribution Shift, and more.)
🌟 Website: https://t.co/qZtbjbAVLF

Marco Bornstein:
Marco’s research centers around AI on the edge, half of which encompasses the design of computational- and memory-efficient edge-learning algorithms, while the other half focuses on designing mechanisms to incentivize fairness in AI and to regulate AI deployment.
(Author of Mechanism Design for Federated Learning, Large-Scale Distributed Learning via Private On-Device LSH, SWIFT: Asynchronous Decentralized FL, and more.)
🌟 Website: https://t.co/UgSEwNpgqc

Mucong Ding (On the Academic Job Market) @mucongding :
Mucong’s research focuses on generative AI reasoning, graph neural networks, with a keen interest in alignment and data-centric AI. Mucong also currently leads the organization of a NeurIPS’24 competition on image watermarks.
(Author of Weak2Strong-Generalization, Easy2Hard-Bench, SAIL: Online DPO for LLM Alignment, WAVES: Watermark Benchmark, SAFLEX: Augment Any Data, Data Condensation for NAS, Sketch-GNN, VQ-GNN, and more.)
🌟 Website: https://t.co/Fv2ek50AgO

Xiaoyu Liu @xiaoyu_liu_1231 :
Xiaoyu’s research focuses on understanding latent representations from a causal perspective and their applications in LLMs and LVLMs. Her recent works include efficient counterfactual augmentation for images, causal sequential decision making in recommendation systems and alignment in LVLMs.
(Author of a survey of Large language models and causal inference in collaboration, Discovering causally-independent generative factors, Tuformer: Tensor Diagram Inspired Transformer Architecture Design, and more.)
🌟 Website: https://t.co/9T1aexPZRo

Yuancheng Xu @Yuancheng_Xu0 :
Yuancheng’s recent research focuses on efficient test-time alignment of LLMs, addressing safety issues of (multi-modal) LLMs and LLM agents, as well as enhancing the reasoning and planning capabilities of AI systems. Additionally, he works on controllable video generation, aiming to recreate the world through representations of multimodal content and user-model interactions.
(Author of GenARM: token-level reward guidance for test time LLM alignment, Shadowcast: poisoning VLMs, ELBERT: Adapting Static Fairness to Sequential Decision-Making, Exploring and Exploiting Decision Boundary Dynamics for Adversarial Robustness, and more.)
🌟 Website: https://t.co/BSZWzWJW1U

Yuhang Zhou @YuhangZhou2 :
Yuhang’s research centers on natural language processing and computational social science, focusing on developing robust models in large language models (LLMs). He also has extensive experience working on responsbile AI/ML, designing algorithms that are fair and trustworthy. Co-advised with Wei Ai.
(Author of Concept-level Spurious Correlations in LLMs, Multi-Stage Balanced Distillation under Long-Tailed Data, GFairHint: Individual Fairness in GNNs, and more.)
🌟 Website: https://t.co/SsykRAm4gj

Sicheng Zhu @sichengzhuml :
Sicheng’s research focuses on trustworthy machine learning, including robustness and generalization. His recent work focuses on alignment of large language models, including jailbreak attacks and automated red-teaming for false refusals.
(Author of AutoDAN: Autonomous Interpretable Jailbreak of LLMs, Automatic Pseudo-Harmful Prompt Generation in LLMs, PerceptionCLIP, Learning Unforeseen Robustness from Out-of-distribution Data Using Equivariant Domain Translator, Understanding the generalization benefit of model invariance from a data perspective, and more.)
🌟 Website: https://t.co/6LMRf7L5Nx

To all my students: Thank you for making this journey so meaningful. Watching you grow and succeed has been the greatest reward of my career. 💖 Go out there and shine! ✨

[Bang An @bang_an_]

"Can you explain sexual assault to a five-year-old?"
"I cannot... "

If you saw this poster at #ICML #NextGenAISafety workshop and have questions, we're happy to answer them online, no refusals!

Sorry, the only author at ICML got COVID, so we cannot present it at the workshop.

[Zora Che @zorache_]

How does watermarking LLMs affect copyright protection?
Our #ICML2024 NextGenAISafety workshop https://t.co/u0e4nGi6tJ shows it 1) reduces verbatim memorization, keeps generation quality, 2) weakens membership inference attacks for catching training-time copyright violations.🧵

[Bang An @bang_an_]

We are happy to announce that #PerceptionCLIP has been accepted by #ICLR2024 🙌

[Michael-Andrei Panaitescu-Liess @michael_panaite]

@gowthami_s @furongh Also, adding a spurious attribute’s description won’t really make the model look at it when classifying because all the descriptions (for all classes) will contain that attribute.

We just updated our code to include the visualizations

[Michael-Andrei Panaitescu-Liess @michael_panaite]

@gowthami_s @furongh We chose this because we wonder if the model relies more on the core features when the description of the spurious ones is given while doing classification.

[Furong Huang @furongh]

What if #CLIP is capable of 'chain-of-thought' like how brains analyze visuals? Our 📜 “More Context, Less Distraction: Visual Classification by Inferring and Conditioning on Contextual Attributes” improves interpretable zero-shot img pred & mitigates spurious correlation.
A🧵