🔥The 9th Round of Easy Loan, Earn $40 Reward is in progress❗️ ⏰ Promotion Period: January 15th - Feburary 15th, 2025 👉 Register now and check more details at gate.io/campaigns/358

If you are interested in multimodal RAG, check out our new paper! - We propose a unified multimodal retriever trained with instruction-tuning on 8 tasks (🖼️image/📝text query -> target) arxiv.org/abs/2311.17136 #RAG #multimodal

Upcoming tutorial at ACL 2024 on Automatic & Human-AI Interactive Text Generation tutorial with Yao Dou Philippe Laban Claire Gardent. Stay tuned :) arxiv.org/pdf/2310.03878

Do you know GPT-4o is super good at image geolocate🌎? I've been taking random photos and testing it. The model just keep giving me surprise. What does this mean to user privacy when they post photos on social media? Check out our new paper and benchmark GPTGeoChat!👇

Check out our recent work that aims to protect privacy by helping people make more informed decisions about what they share online. Drop by our poster tomorrow morning at 11am if you are at ACL in Bankok. ACL 2025 arxiv.org/abs/2311.09538

Excited to be in Thailand ☀️😎 🌺🥥 to attend #ACL2024! If you run into my PhD students from Georgia Tech NLP group - Yao Dou Yao Dou, Tarek Naous Tarek Naous, and Jonathan Zheng Jonathan Zheng, please say hi! 👋 Please also say hi to my other collaborators!

1/N) Are current large multimodal models like #GPT4o really good at video understanding? 🚀 We are thrilled to introduce TemporalBench to examine temporal dynamics understanding for LMMs! Our TemporalBench reveals even the SOTA LMM #GPT4o achieves only 38.5, far from

Now TemporalBench is fully public! See how your video understanding model performs on TemporalBench before CVPR! 🤗 Dataset: huggingface.co/datasets/micro… 📎 Integrated to lmms-eval (systematic eval): github.com/EvolvingLMMs-L… (great work by Chunyuan Li Yuanhan (John) Zhang ) 📗 Our

What causes entity-related cultural biases in LMs? Is it just pre-training data? Our latest paper shows how varying linguistic phenomena exhibited by entities (such as word sense in Arabic) impact the cross-cultural performance of LMs. arxiv.org/abs/2501.04662

excited to introduce AceMath, a series of math LLMs and reward models 🤗ckpt and training data are on huggingface: huggingface.co/collections/nv… paper: arxiv.org/abs/2412.15084 accelerating

🚨New Paper: Better search for reasoning (e.g., web tasks) usually requires costly💰demos/rewards What if we only self-improve LLMs on state transitions—capturing a classic RL method in natural language? Spoiler: It works (⬆️39% over base model) & enables efficient search!🚀 🧵

🚨o3-mini vastly outperforms DeepSeek-R1 on an unseen probabilistic reasoning task! Introducing k-anonymity estimation: a novel task to assess privacy risks in sensitive texts Unlike conventional math and logical reasoning, this is difficult for both humans and AI models. 1/7

🚀Big WebDreamer update! We train 💭Dreamer-7B, a small but strong world model for real-world web planning. 💥Beats Qwen2-72B ⚖️Matches #GPT-4o Trained on 3M synthetic examples — and yes, all data + models are open-sourced.

Thrilled to share our new preprint: "Beyond the Reported Cutoff: Where LLMs Fall Short on Financial Knowledge" We evaluated 197,011 revenue questions across 17,621 U.S. companies (1980–2022) using 6 top LLMs. Key insights 🧵

Had a lot of fun to scale up RL to improve math reasoning! Excited to introduce AceMath-RL-Nemotron-7B with a scalable training recipe 📑Full blog: research.nvidia.com/labs/adlr/acem… 🔗Model: huggingface.co/nvidia/AceMath…

I am giving a keynote at PrivateNLP Workshop (sites.google.com/view/privatenl…) at #NAACL2025 (Sunday 9am CT). * GPT4-v is a performant geolocator, predicting exact GPS coordinates of image > any SOTA * LLMs can estimate privacy risk based on probabilistic reasoning > chain-of-thoughts

with just math-RL, AceReason-Nemotron-14B surpass DeepCoder-14B on LiveCodeBench v5. we then did code-RL and found training becomes so much easier

Does RL incentive reasoning capability over the starting SFT model? We show an interesting result with our recent published AceReason-Nemotron-7B model, which was trained with RL pass@K from 1 to 1024 consistently +10% on LiveCodeBench v6 perhaps scaling RL is the key