Wenyan Li

Happy to share (with a bit of delay tho) our paper on quantifying visual information loss in VLMs --- "Lost in Embeddings: Information Loss in Vision-Language Models" is accepted to EMNLP 2025 findings: arxiv.org/pdf/2509.11986 💃code is also released: github.com/lyan62/vlm-inf…

@AbrahamOwos thanks Abraham. That’s so glad to hear!

@Wenyan62 I really like this paper 👏👏. Read it and shared with some friends.

I will be presenting our Lost in embeddings poster at EMNLP! Hope to see many old and new friends in Suzhou!🤗🤗 📍Time/Date: Fri. Nov 7 at 12:30-13:30 Location: Hall C Also happy to chat anything about VLMs, RAG and recently get in the domain of fintech. #EMNLP2025

@codewithimanshu thanks for the warm words Himanshu!

@Wenyan62 Sounds exciting, Wenyan! Suzhou sounds lovely, and your work on VLMs is always top-notch, I must say.

📢Our new paper critically examines arena-style LLM evaluation, e.g., LMArena, questioning whether draws actually mean equal model ability. TL;DR: simply ignoring draws improves rating systems by 1-3%, and query difficulty/subjectivity relate more strongly to draws than model ratings do.

@_srishtiyadav thanks Srishti ❤️

@Wenyan62 Congrats, Dr. Wenyan! 💚

Happy to share that I’ve successfully defended my PhD today 🎉 A big thank you to my committee members Manex aguirrezabal zabaleta, Anna Korhonen, and Charlie Clark ❤️ Very grateful to all the support and encouragement from my supervisor Anders Søgaard and colleagues at Coastal❤️

@gietema here we could probably phrase it better. A large drop in overlap ratio indicates that the connector is changing neighborhood structure substantially, which may suggest geometric distortion beyond what is required for task alignment.

@Wenyan62 Hi, thanks for this, congrats! Not sure I understand why an optimal connector would maintain the same k-NN sets? Isn't the purpose of the connector to align the embedding for a downstream task, in which case the structure of the original embedding cannot be assumed to be optimal?

Happy to share (with a bit of delay tho) our paper on quantifying visual information loss in VLMs --- "Lost in Embeddings: Information Loss in Vision-Language Models" is accepted to EMNLP 2025 findings: arxiv.org/pdf/2509.11986 💃code is also released: github.com/lyan62/vlm-inf…

@gietema Hi Jochem, thanks for the question. we agree that the purpose of the connector is to align the embeddings spaces. Idea is not that preserving kNN neighborhoods is the end goal, but that overlap ratio provides a way to quantify how much the local structure is perturbed.

@_TobiasLee thank you lei! ☺️@_TobiasLee

@Wenyan62 Thanks for sharing! Very insightful for understanding the ViT embeddings

Kudos to my coauthors @ralph_tang @caiqizh @li_chengzu

Excited to share our multimodal temporal culture benchmark is released 🚀🚀🚀 Dataset is public on 🤗 huggingface Check it out!! arxiv.org/abs/2506.01565 huggingface.co/datasets/lizho…

Check out our new benchmark RAVENEA for VLM culture understanding with retrieval augmentation! Code and data all released!🚀🚀🚀

Current KL estimation practices in RLHF can generate high variance and even negative values! We propose a provably better estimator that only takes a few lines of code to implement.🧵👇 w/ @xtimv and Ryan Cotterell code: arxiv.org/pdf/2504.10637 paper: github.com/rycolab/kl-rb

Forget just thinking in words. 🚀 New Era of Multimodal Reasoning🚨 🔍 Imagine While Reasoning in Space with MVoT Multimodal Visualization-of-Thought (MVoT) revolutionizes reasoning by generating visual "thoughts" that transform how AI thinks, reasons, and explains itself.

🎉🎉

@nicolayr_ Thanks Nicolay!!

@Wenyan62 all the best at #EMNLP2024 , well done!👏

🍗🍗I will present FoodieQA in person at #EMNLP2024😋😋 Looking forward to meeting old and new friends! Feel free to drop by! (and have some snacks) ⏰ Nov, 13th (Wed) 16:00, In-Person Poster Session E (Riverfront Hall) I'm also on the job market and would be happy to chat :)

@roopalgarg Thanks Roopal! Sure, would love to😊

@Wenyan62 @Wenyan62 really cool work. Let's grab some time to sync

1/7 🚨non-LLM paper alert!🚨 Human's perception of the sentence is quite robust against interchanging words with similar meanings, not even mentioning the semantically equivalent words across different languages. How about the language models? In our recent work, we measure the role of subword-level shared semantics in multilingual LMs. tl;dr: semantically similar subwords could largely *share the same word embedding!* arxiv link: arxiv.org/abs/2411.04530