Cindy Zeng

@hanzhao_ml @evangelinejy99 Big congrats 🎉🎉🎉

Glad to share that our paper has been awarded the Outstanding Paper Award at EMNLP' 25!! I am not attending the conference, but please find @evangelinejy99 and talk to her if you want to know more details!

🚀 Our model merging benchmark "MergeBench" is accepted at #NeurIPS2025 DB Track! See you in SD! TL;DR: We introduce a benchmark using modern LLMs across 5 diverse domains to comprehensively evaluate model merging algorithms. 📎 Website: yifei-he.github.io/mergebench 🧵👇

[N/N] We offer: 📐 A theoretical foundation for task arithmetic 💾 A scalable merging strategy 📊 Strong empirical results across domains 🔗 Code & details: github.com/uiuctml/TaskVe… 🔁 RT appreciated if you’re also excited about scalable LLM/VLM merging!

[8/N] 🧪 Experiments show: 1. Save up to 95.71% of memory 2. Strong bases arithmetic across vision & NLP tasks 3. Outperforms full vector merging in some settings 4. Reaches the best accuracy and storage with sparse merging methods

[1/N] #ML #LLM #ModelMerging Paper: arxiv.org/pdf/2502.01015 We build theory to explain why task arithmetic works, and propose Task Vector Bases, a scalable model editing method grounded in it. With @heyifei99, @youweiqiu, Yifan, Hubert, @myamada0, @hanzhao_ml. 🧵 Dive in below:

[7/N] 🚀 [Paper] openreview.net/forum?id=AnL6B… 🚀 [Code] github.com/UIUCTML/OTCPCC 🚀 [Video] iclr.cc/virtual/2025/p… #ICLR2025 #MachineLearning #OptimalTransport [N/N]

[6/N] Come say hi at the poster! Always happy to chat about optimal transport, representation learning, or anything in between 😊 🗓️ Sat Apr 26 🕒 3–5:30PM SGT 📍 Hall 3 + Hall 2B #454

[1/N] 🚨 #ICLR2025 We’ll present “Learning Structured Representations by Embedding Class Hierarchy with Fast Optimal Transport”, a new approach for hierarchy-aware representation learning using Optimal Transport with amazing collaborators: Sixian Du, @myamada0, and @hanzhao_ml

[NeurIPS '24 paper alert 🎉] Excited to present my thesis research on Learning Structured representations with Hyperbolic Embeddings at NeurIPS '24 in Vancouver! Joint work with amazing collaborators Siqi Zeng, Prof. @myamada0 and Prof. @hanzhao_ml 🧵 Let's dive in! [1/N]