@zy27962986 Profile picture

Zongyu Lin

@zy27962986

CS Ph.D @UCLADM @UCLANLP | B.E@Tsinghua | Ex. LLM Researcher at https://t.co/neKTTJf2zO | Generalizable and scalable generative models | Research Intern @Apple

Pinned

🚨Looking for a fast and effective contradiction retrieval method? In our new paper, we propose a sparse-aware contrastive learning to beat traditional bi-encoder methods with 37.2% performance gain on average! 🚀Project: sparsecl.github.io 📰Paper: arxiv.org/pdf/2406.10746

Tweet Image 1
Tweet Image 2

Zongyu Lin Reposted

⚡️ SANA is released: github.com/NVlabs/Sana Demo: nv-sana.mit.edu Highlight: - 20 x smaller & 100x faster than FLUX - Generate 4K image. - Deployable on laptop GPU.


Zongyu Lin Reposted

Introducing LongMemEval: a comprehensive, challenging, and scalable benchmark for testing the long-term memory of chat assistants. 📊 LongMemEval features: • 📝 164 topics • 💡 5 core memory abilities • 🔍 500 manually created questions • ⏳ Freely extensible chat history…

Tweet Image 1

Zongyu Lin Reposted

Introducing an approach to directly ground video generation models to policy execution without needing any action labels! video-to-action.github.io Our approach uses a generic goal-conditioned exploration procedure to learn a policy that works across robots / embodiments!


Zongyu Lin Reposted

🤗What’s up #EMNLP2024! 🌟In tomorrow’s poster session 2 (11:00-12:30), I’ll be presenting our work, ‘The Factuality Tax of Diversity-Intervened Text-to-Image Generation: Benchmark and Fact-Augmented Intervention’! 😆Come hang out at Jasmine – Lower Terrace Level and let’s chat!!

Tweet Image 1

🚨WARNING: Instructing T2I models to depict “diverse” people will severely harm historical factuality!🤯In this work, we benchmark the evaluation of this “FACTUALITY TAX” of diversity intervention prompts, and propose Factuality-Augmented Intervention (FAI) to resolve the issue.

Tweet Image 1


Zongyu Lin Reposted

If you are at #EMNLP, check out Ashima's poster! Also talk to her about hate speech detection for LLMs and AI safety! 🦺

I will be presenting my first first-author paper at #EMNLP2024! Come check out our poster on 📅November 12, 2024 at ⏰3 PM in the 🏦Riverfront Hall!!

Tweet Image 1


Zongyu Lin Reposted

😆Check out our poster at #EMNLP2024 ! 🔥I’ll be presenting 𝗩𝗗𝗲𝗯𝘂𝗴𝗴𝗲𝗿, a debugging algorithm for 🖼️ visual reasoning. Check it out at ⏰ 𝗪𝗲𝗱. 𝟭𝟲:𝟬𝟬-𝟭𝟳:𝟯𝟬 at Riverfront Hall! 🤗Happy to chat and connect!

Looking for a debugging algorithm for visual programming? Take a look at 𝗩𝗗𝗲𝗯𝘂𝗴𝗴𝗲𝗿🔥🔥🔥shirley-wu.github.io/vdebugger By tracking execution step by step, VDebugger boosts the accuracy by up to 𝟯.𝟮% on 6 visual reasoning tasks!

Tweet Image 1


Zongyu Lin Reposted

Glad to share that VDebugger is accepted to #EMNLP2024 ! Our 𝐝𝐞𝐦𝐨 is out🔥at huggingface.co/spaces/VDebugg…. Check out our 𝐝𝐞𝐛𝐮𝐠𝐠𝐢𝐧𝐠 𝐚𝐥𝐠𝐨𝐫𝐢𝐭𝐡𝐦 𝐟𝐨𝐫 𝐯𝐢𝐬𝐮𝐚𝐥 𝐩𝐫𝐨𝐠𝐫𝐚𝐦𝐬 in Miami🌴!

Looking for a debugging algorithm for visual programming? Take a look at 𝗩𝗗𝗲𝗯𝘂𝗴𝗴𝗲𝗿🔥🔥🔥shirley-wu.github.io/vdebugger By tracking execution step by step, VDebugger boosts the accuracy by up to 𝟯.𝟮% on 6 visual reasoning tasks!

Tweet Image 1


Zongyu Lin Reposted

🎬Meet SlowFast-VGen: an action-conditioned long video generation system that learns like a human brain! 🧠Slow learning builds the world model, while fast learning captures memories - enabling incredibly long, consistent videos that respond to your actions in real-time.…

Tweet Image 1

Zongyu Lin Reposted

Hello, world of LLM researchers! We are thrilled to introduce the official Twitter account of the 2025 @SemEvalWorkshop Challenge: Unlearning sensitive content from Large Language Models. Stay tuned for challenge announcements! Homepage 🔗: llmunlearningsemeval2025.github.io


Zongyu Lin Reposted

New paper📢 LLM folks have been supervised finetuning their models with data from large and expensive models (e.g., Gemini Pro). However, we achieve better perf. by finetuning on the samples from the smaller and weaker LLMs (e.g., Flash)! w/@kazemi_sm @arianTBD @agarwl_ @vqctran

Tweet Image 1

Zongyu Lin Reposted

Thrilled to receive the KDD Dissertation Award Runner-Up, for my PhD works on Neural-Symbolic Reasoning. Sincerely thanks to my PhD advisors @YizhouSun and @kaiwei_chang, my letter supporters @yisongyue and @jhamrick Thanks to the award committee @kdd_news for such honor.


VideoGen + Physics Commensense🚀We updated the leaderboard with the latest open-source model: CogVideoX~

📜We evaluate the videos for their text-adherence and physical commonsense on our VideoPhy (videophy.github.io) prompts, and update our human and automatic leaderboard! Good to see that it is better than many models, but there is good room for improvements🥳 Examples👇

Tweet Image 1


Zongyu Lin Reposted
Tweet Image 1

Zongyu Lin Reposted

Introducing Diffusion Forcing, which unifies next-token prediction (eg LLMs) and full-seq. diffusion (eg SORA)! It offers improved performance & new sampling strategies in vision and robotics, such as stable, infinite video generation, better diffusion planning, and more! (1/8)


Zongyu Lin Reposted

🚨WARNING: Instructing T2I models to depict “diverse” people will severely harm historical factuality!🤯In this work, we benchmark the evaluation of this “FACTUALITY TAX” of diversity intervention prompts, and propose Factuality-Augmented Intervention (FAI) to resolve the issue.

Tweet Image 1

Zongyu Lin Reposted

Video generation models do not understand basic physics. Let alone the human body.

CW: Body Horror? This AI video attempt to show gymnastics is one of the best examples I have seen that AI doesn’t actually understand the human body and it’s motion but is just regurgitating available data. (Which appears to be minimal for gymnastics) twitter.com/Przemek8739456…



Zongyu Lin Reposted

Probably the perfect time to check out our work, VideoPhy (x.com/hbxnov/status/…), on evaluating a more tractable physical commonsense in the generated videos! This differs from performing precise physics eval which is often very complicated for real-world interactions.

Video generation models do not understand basic physics. Let alone the human body.



United States Trends
Loading...

Something went wrong.


Something went wrong.