I recently analyzed the topics of user instructions from Chatbot Arena and WildChat, with 10K instructions for each language. The patterns observed across different languages were generally similar. However, the English users asked lots of health-related questions, while

If a paper receives a review stating that "the idea is novel and the experiments are comprehensive," what is the minimum score the paper should receive?

support 10M context size ≠ perform well on 10M context size. 🤫

🥳🥳🥳New dataset: huggingface.co/datasets/mingh… We recently received some free compute, so we created a synthetic dataset of 10M realistic personas using meta-llama/Llama-3.3-70B-Instruct & Qwen/Qwen2.5-72B-Instruct. Each persona includes features like name, DOB, personality, and

🌺GPT-4o’s image generation is stunning — but how well does it handle complex scenarios? 🤔 We introduce 🚀CIGEVAL🚀, a novel method to evaluate models' capabilities in Conditional Image Generation 🖼️➕🖼️🟰🖼️. Find out how top models perform when conditions get truly

🌏 New position paper analyzes 2,000+ multilingual LLM benchmarks from 148 countries (2021-2024) and reveals key challenges in multilingual AI evaluation. 🌺

Alibaba presents The Bitter Lesson Learned from 2,000+ Multilingual Benchmarks

I suddenly realize that, given that llama4 doesn't perform well, my reviewers (probably) will not ask me to include additional experiments using llama4 into my resubmissions. yay!

Our upgraded O1-like Anim-Drector is nearly ready! Witness the magic as it transforms a few simple words into a complete animated short film Girl & Cat with zero human intervention! Click the image below to watch the film! Youtube: youtube.com/watch?v=O8XLR1…

🎉 Exciting news! Welcome to CulturalLingo, our new VideoVista series that bridges cultures (China, North America, and Europe), languages, and domains (140+) in video comprehension. Join us on this journey of video understanding! 🌍📽️ Link of paper: researchgate.net/publication/39…

To be honest, this is my new favorite reading of the year.

🚀 Excited to share that our paper "GraphFilter" is accepted at #ICML2025! We introduce a novel bipartite graph approach for data selection in LLM fine-tuning, balancing quality & diversity for optimal performance. GraphFilter outperforms 9 baselines across 6 benchmarks—setting

Thanks for your kind words, mate! The first round of review was somewhat helpful, but after that, most of the reviews were something like "the model backbones are outdated." 🤦🏻‍♂️ I’ve resubmitted this paper for several times but haven't gotten an acceptance. 😮‍💨 I guess the 40

accepted by ACL2025!!! 🎉🎉🎉 many thanks to Weixuan Wang Barry Haddow alexandra birch

📢 New Paper Alert 📢 🍓HBO: A Novel Balancing Method for LLM SFT! 🎯 Bilevel strategy for tackling data imbalance across and within datasets. 🌍 Global Actor: Dynamically adjusts sampling across heterogeneous subsets. 🔍 Local Actors: Optimize subset training by difficulty.

🚨 What if you could hijack any LLM's brain using external expert models? ExpertSteer does exactly that! 🧠⚡ Meet ExpertSteer: a breakthrough that lets you inject expert knowledge into any LLM, guiding its responses without updating model parameters.

Just arrived in Vienna for #ACL2025. I will be presenting a few papers and participating in one of the panels. Friends, feel free to catch up!

Yeah, we won an award.#ACL2025 😉😉😉

Yeah, we won an award! Weixuan Wang Barry Haddow alexandra birch #ACL2025