Chenghao Xiao @gowitheflow98

There's a wave of omni embedding models (gemini, nemotron, bidirlm). Excited to support this trend with our multimodal mteb versions (mieb, maeb) - video coming soon🎥

Nicolas Boizard @N1colAIs

2 months ago

🚀 New model family release with an OMNIMODAL version ! After Eurobert, I'm excited to introduce BidirLM, a family of 5 frontier bidirectional encoders including an OMNIMODAL encoder at just 2.5B parameters. 🧵👇 huggingface.co/BidirLM

5 12 55 15K 27

Today is the last day of the main conference, and it began with a keynote from Prof Chenghua Lin. We have oral and poster presentations, and the day will end with best paper awards and a closing session. Join us!

🌟 Thrilled to share our NeurIPS 2025 paper: "Scaling Language-Centric Omnimodal Representation Learning"! 🚀 We investigate why MLLM embeddings are so effective and propose a SOTA embedding framework. 💡 Key Findings: 1️⃣ Latent Alignment: We show MLLMs gain powerful cross-modal alignment from generative pretraining, making contrastive learning a lightweight refinement step. 2️⃣ LCO-Emb: We introduce a Language-Centric Omnimodal Embedding framework that achieves SOTA on multimodal benchmarks with minimal text-only contrastive learning data. 3️⃣ Scaling Law (GRSL): We discover a "Generation-Representation Scaling Law," proving that improving an MLLM's generative ability directly enhances its representation quality. 🔓 All code, models, and our novel SeaDoc benchmark will be open-sourced! 👨‍💻 Code: github.com/LCO-Embedding/… 🤗 HF Space: huggingface.co/LCO-Embedding 👍 Please check out and upvote our paper on Hugging Face! 📑 huggingface.co/papers/2510.11… #Multimodal #LLM #RepresentationLearning #AI #MachineLearning

Thanks for sharing! Check out LCO-Embedding - a state-of-the-art omnimodal embedding model family trained using language-centric paradigm!

DailyPapers @HuggingPapers

8 months ago

Alibaba's DAMO Academy introduces LCO-Embedding for omnimodal representation learning It achieves new SOTA on MIEB (Massive Image Embedding Benchmark) & supports audio/video, revealing a new Generation-Representation Scaling Law!

1 8 27 2K 9

✨ Meet our latest multimodal reasoning model — VL-Cogito! Inspired by the Latin word “Cogito” (“I think”), VL-Cogito is built for complex and diverse multimodal reasoning tasks, with a strong focus on autonomous thinking and adaptability 💡 🧠 What makes it special? VL-Cogito adopts Progressive Curriculum Reinforcement Learning (PCuRL) — learning from easy to hard, just like humans. The more it trains, the smarter it gets! 🔥 💡 Two key innovations: 1️⃣ Online difficulty soft weighting: Dynamically adjusts training difficulty based on model capability, enabling personalized and efficient learning. 2️⃣ Dynamic length reward: Encourages the model to respond concisely for simple questions, and thoroughly for complex ones — balancing performance & efficiency. 📊 Strong performance: On major benchmarks spanning math, science, logic, and common sense, VL-Cogito matches or exceeds leading reasoning models. 📎 Check out our paper on HuggingFace👇 huggingface.co/papers/2507.22…

Alibaba just launched VL-Cogito on Hugging Face! This advanced multimodal reasoning model uses a novel Progressive Curriculum RL framework. It tackles complex challenges and improves performance across diverse domains like math, science, and logic.

VL-Cogito Progressive Curriculum Reinforcement Learning for Advanced Multimodal Reasoning

I think this is not only a super dataset but it appears the recipe used to generated it may generalize well & so be useful beyond this task! I need to investigate to be sure. ReasonMed: A 370K Multi-Agent Generated Dataset for Advancing Medical Reasoning huggingface.co/papers/2506.09…

Paper of the day on HF!

Christopher @communicating

a year ago

I think this is not only a super dataset but it appears the recipe used to generated it may generalize well & so be useful beyond this task! I need to investigate to be sure. ReasonMed: A 370K Multi-Agent Generated Dataset for Advancing Medical Reasoning huggingface.co/papers/2506.09…

3 23 85 83K 78

🚀We are thrilled to launch 'Lingshu' – A Generalist Medical Multi-modal Foundation Model! 🩻 🌟 Highlights of Lingshu: ⚕️ Unified knowledge across 12+ imaging modalities (X-Ray, CT, MRI & more!). 🧠 Enhanced reasoning & reduced hallucinations via novel data curation and training pipelines. 🔥 Outperforms GPT-4.1 and Claude Sonnet 4 on most medical QA & report generation tasks! 👍 Please check out and upvote our paper on Hugging Face! 🔗 huggingface.co/papers/2506.07… #MedicalAI #MultiModal #LLM #MLLM #AIinHealthcare

🚀 Discover how LLMs perceive their knowledge boundaries across languages in our #ACL2025 main paper! 🌍 By probing LLMs’ internal representations, we reveal key insights on where knowledge boundaries are encoded & propose a training-free method to combat cross-lingual hallucination. Plus, a new multilingual evaluation suite! 📑 Paper: arxiv.org/pdf/2504.13816 👨‍💻 Code & data: github.com/DAMO-NLP-SG/LL… #LLMs #Hallucination #Multilingual

New method for robust and efficient LLM-based NLG evaluation An inversion learning method that learns effective reverse mappings from model outputs back to their input instructions, enabling the automatic generation of highly effective, model-specific evaluation prompts.

[1/n] Can we generate highly effective, model-specific prompts via inversion learning? Delighted to introduce our new paper: Beyond One-Size-Fits-All: Inversion Learning for Highly Effective NLG Evaluation Prompts. Paper: arxiv.org/abs/2504.21117 HF Daily: huggingface.co/papers/2504.21…

You might've heard of MTEB: Massive Text Embedding Benchmark. Until now, it was famously primarily for text, but now it has been extended to include MIEB: Massive Image Embedding Benchmark! Details in 🧵

Analyzing LLMs Knowledge Boundary Cognition Across Languages Through the Lens of Internal Representations Examines how LLMs internally distinguish known from unknown questions across languages. - signals localize to mid–upper layers - cross-lingual structure is linear—enables training-free alignment - bilingual QA fine-tuning improves recognition, reduces hallucination in low-resource settings - introduces the first multilingual benchmark for knowledge boundary evaluation

Introducing the 3rd edition of BioLaySumm shared task hosted at the BioNLP Workshop @ #ACL2025NLP ! biolaysumm.org BioLaySumm is a shared task that focuses on generating easy-to-understand lay summaries for complex biomedical texts. Building on the success of the first two editions with over 80 participating teams globally, we’re excited to launch a brand new multi-modal track 🚀: generating radiology reports in accessible language for all readers directly from X-ray images! Join us in bridging accessibility, science and healthcare! #MachineLearning #NLP

LB: huggingface.co/spaces/mteb/le… Great work by @KCEnevoldsen @isaacchung1217 Imene Kerboua Márton Kardos @risolomatin @tomaarsen @gowitheflow98 @vaibhav_adlakha @orionweller @sivareddyg & many others ❤️

Check out Yiqi's work -- the first systematic research of the "self-preference" bias of LLMs as evaluators

Yiqi Liu @yiqi_617

2 years ago

[1/n] Do language model-driven evaluation metrics inherently favour texts generated by the same underlying model? Our #ACL2024 Findings paper (arxiv.org/abs/2311.09766) investigates this bias, focusing on metrics such as BARTScore, T5Score, and GPTScore in summarisation tasks. To

3 5 37 4K 4

1/ Excited to announce the release of our new paper "SciMMIR: Benchmarking Scientific Multi-modal Information Retrieval". This benchmark comprises 530K meticulously curated image-text pairs extracted from scientific documents(arXiv Paper). arxiv.org/abs/2401.13478