Changsheng Wang（C.S.Wang）

Heading to Vienna for #ICML2024! Excited to explore how current "zeroth-order "optimization techniques are shaping up for memory-efficient fine-tuning of LLMs; Feel free to check our ICML paper, "Revisiting Zeroth-Order Optimization for Memory-Efficient LLM Fine-Tuning: A…

Thank you @INNSociety for this great honor. I am deeply grateful to my nominator, students, and collaborators who made this recognition possible. Excited to keep advancing the frontiers of scalable and trustworthy AI! @OptML_MSU

🤔Come to check Chongyu @ChongyuFan and Jinghan Jia @jia_jinghan ‘s work! 👨‍💻Jinghan has been an incredible mentor to work with—smart, supportive, and inspiring. He’s on the job market now, feel free to reach out and chat with him!

🎯 ICML 2025 Poster! 📄 “Invariance Makes LLM Unlearning Resilient Even to Unanticipated Downstream Fine-Tuning” 🔗 arxiv.org/abs/2506.01339 🗓️ July 15, 4:30–7:00 pm PT 📍 East Exhibition Hall A-B, #E-1108 🧍‍♂️ I won’t be there in person — but feel free to drop by and chat with my…

An LLM generates an article verbatim—did it “train on” the article? It’s complicated: under n-gram definitions of train-set inclusion, LLMs can complete “unseen” texts—both after data deletion and adding “gibberish” data. Our results impact unlearning, MIAs & data transparency🧵

Congratulations to @zyh2022 on being awarded the IBM PhD Fellowship 2024! Yihua is truly one of the most outstanding PhD students and an exceptional mentor I’ve had the pleasure to work with. Eagerly anticipating your next steps and future achievements!

Machine unlearning ("removing" training data from a trained ML model) is a hard, important problem. Datamodel Matching (DMM): a new unlearning paradigm with strong empirical performance! w/ @kris_georgiev1 @RoyRinberg @smsampark @shivamg_13 @aleks_madry @SethInternet (1/4)

BREAKING: Taylor Swift's Eras Tour just did what AI couldn’t—pushed NeurIPS by a whole day! 🤖 🤣🤣🤣 #NeurIPS 2024 Conference Date Change The conference start date has been changed to Tuesday December 10 in order to support delegates arriving on Monday. neurips.cc

Jailbreak images for multimodal fusion models Unlike LLaVA, newer fusion models, like GPT-4o and Gemini, map all modalities into a shared tokenized space. Existing image jailbreak methods fail because tokenization is not differentiable!

Excited to attend #ECCV2024 to present two of our recent works on the adversarial aspects of machine unlearning in computer vision! One explores "challenging forgets", while the other studies unlearning robustness in text-to-image generation. My student @ChongyuFan will lead the…

Are watermarks for generative models strong enough to be useful? Or are they too easy to wipe away? If you think watermarks are no good, then prove it! Our @NeurIPSConf competition on erasing watermarks is off and running. Participating is easy. We give you a batch of…

Wrote about extrinsic hallucinations during the July 4th break. lilianweng.github.io/posts/2024-07-… Here is what ChatGPT suggested as a fun tweet for the blog: 🚀 Dive into the wild world of AI hallucinations! 🤖 Discover how LLMs can conjure up some seriously creative (and sometimes…

📢📢📢 When will machine unlearning methods fail to remove unwanted data/concept influences in models? Our #ECCV2024 papers investigate this by exploring "challenging scenarios". One study identifies the worst-case forget sets from a data selection perspective [1], while the…

Honored to be selected in the Program! I’ll be presenting my study on Machine Unlearning for LLMs and diffusion models at NVIDIA’s HQ in Santa Clara 🎉 MU is a rapidly evolving field in trustworthy ML. Interested in this area or have ideas? DM me for collaboration opportunities!

[1] “Finish the paper, release the code and data, and then tweet about this paper” —Can language model handle this kind of sequential instructions well? Check our paper: arxiv.org/pdf/2403.07794 Data and model: seqit.github.io Code: github.com/hanxuhu/SeqIns

LLMs can memorize training data, causing copyright/privacy risks. Goldfish loss is a nifty trick for training an LLM without memorizing training data. I can train a 7B model on the opening of Harry Potter for 100 gradient steps in a row, and the model still doesn't memorize.