David Ifeoluwa Adelani 🇳🇬

Hallelujah! I’m excited to share that I’ve been selected as a 2025 AI2050 Early Career Fellow by @Schmidtsciences This year’s fellows represent 42 institutions across eight countries, working to ensure AI benefits humankind. Learn more at: lnkd.in/eZA5FHci

New Schmidt Sciences RFP on AI Interpretability: We need new tools for detecting and mitigating deceptive behaviors exhibited by LLMs. Funding for $300k-$1M projects Deadline: May 26th, AoE RFP: schmidtsciences.smapply.io/prog/2026_inte… Please share with anyone who may be interested!

Happy to announce that today we released OmniSONAR (tinyurl.com/3j9rn2u8) and OmniMT. In OmniSONAR, we have been able to really push the edge on largely mutlilingual embedding models, where representations across all languages are aligned like never before! 🧵1/n

🚨🎉Excited to announce that our paper “Grok in the Wild: Characterizing the Roles and Uses of Large Language Models on Social Media” is accepted at @icwsm 2026! In this paper, we investigate how, when, and to what effect Grok is used on X. 🧵1/n thread

Four years ago, NLLB set a milestone with MT for 200 languages. Today we present OMT: a family of models that extend support to 1600 languages while delivering competitive results in high/mid-resource language, with our 1B-8B models matching frontier and open 70B LLMs. 🧵(1/n)

Introducing Meta Omnilingual Machine Translation: the first MT system benchmarked on more than 1,600 languages. Beyond MT, our experimentation brings up strong motivations towards embracing omnilinguality by design when building foundational models shorturl.at/Ik9UJ

I (finally) put together a new LLM Architecture Gallery that collects the architecture figures all in one place! sebastianraschka.com/llm-architectu…

I am hiring: candidate.hr-manager.net/ApplicationIni… So if you are interested in training/evaluating LMs and language/linguistics, consider applying!

LLM2Vec-Gen represents a major paradigm shift for embeddings/retrieval. Why encode the query when the LLM already knows what to look for and can directly produce an embedding for it? Best part: it’s self-supervised, and it does all of this while the LLM remains completely frozen. Think about it: "solve x² + 3x − 4 = 0" has zero reasoning in it. But the LLM's response does. By encoding the response, the embedding captures the reasoning --- and the better the LLM reasons, the better the embedding. This is why our results scale with model size. As LLMs get smarter, our embeddings automatically get better. LLM2Vec-Gen is also the first demonstration of the promise of @ylecun's JEPA for text embeddings. The alignment loss is JEPA — predict in representation space, not token space. The reconstruction loss goes beyond --- it keeps embeddings decodable. This paradigm shift opens new frontiers: 🔬 Can we build a full JEPA for language where the teacher and student are the same LLM? ⚡ Can LLMs reason in compressed space without ever generating text? 🤖 Can agents reason in compression tokens and carry that directly into retrieval? 💬 Can agents talk to each other in compression tokens instead of text --- dense, fast, and still human-readable? LLM2Vec-Gen is a first step toward all four.

Fish audio s2 pro released with 80+ supported languages and fine grained inline control. Link on @huggingface huggingface.co/fishaudio/s2-p…

Your deduplication pipeline was built for small models. At scale, it's broken. New preprint: "Scale Dependent Data Duplication" 1/10

Today we release PostTrainBench v1.0 and the accompanying paper! We expect this to be a key indicator for AI R&D automation -- and eventually recursive self-improvement. 🧵

Come participate in the Machine Translation for Vision Challenge! The winners will be announced at our MAPS workshop (sites.google.com/corp/view/maps…) at CVPR 2026!

We release our insight on "Understanding the African Next Voices Datasets" which provides the first holistic, evidence-driven analysis of the African Next Voices datasets — covering geography, speaker demographics, gender balance, dialects, and many more: docs.lanfrica.com/insights/under… --- We may be standing at the beginning of a new era for African speech technology — and many people don’t realize it yet. Just a few years ago, having hundreds of hours of speech data for an African language was rare. Today, the African Next Voices (ANV) project alone has produced close to 18,000 hours of speech data across 24 languages in 7 countries. This scale could power real-world systems — voice assistants, speech interfaces, and AI tools that address African problems. But their potential impact is not yet being fully realized. Most people still don’t know: • where these datasets are • what they contain and how they differ • or how they can be used -- and what they mean for society At Lanfrica Insights, we set out to change that. If you want to understand the breadth and depth of the African Next Voices project — and what it means for the future of AI in Africa — start here: docs.lanfrica.com/insights/under… We also make it trivial to find all the datasets in one place with a single click: lanfrica.com/en/discover?ta…

Metareviews are due on the 4th of March. Should an area chair fail to provide their checklists and metareviews, they may be prevented from committing their own papers. Please read the incentives 2026: aclrollingreview.org/incentives2026 @ReviewAcl #ACL2026NLP #NLProc

🚀 Introducing the Qwen 3.5 Small Model Series Qwen3.5-0.8B · Qwen3.5-2B · Qwen3.5-4B · Qwen3.5-9B ✨ More intelligence, less compute. These small models are built on the same Qwen3.5 foundation — native multimodal, improved architecture, scaled RL: • 0.8B / 2B → tiny, fast, great for edge device • 4B → a surprisingly strong multimodal base for lightweight agents • 9B → compact, but already closing the gap with much larger models And yes — we’re also releasing the Base models as well. We hope this better supports research, experimentation, and real-world industrial innovation. Hugging Face: huggingface.co/collections/Qw… ModelScope: modelscope.cn/collections/Qw…

Technical interview question: Suppose you have 5 TB worth of text data and you want to count the total number of words, how will you do this?

Midtraining is a new part of many training pipelines, but when does it help and can it backfire? 🤔 In our new preprint, we use controlled experiments to pin this down. TL;DR; midtraining helps the most when it “bridges” pretraining and posttraining, and mitigates forgetting after posttraining. Timing is also very important. 🧵

What if your tokenizer could tell you your text's language? We're excited to introduce 𝗨𝗻𝗶𝗟𝗜𝗗: a simple, data-efficient method for language identification (LID) that builds on the UnigramLM tokenization algorithm. 📄 arxiv.org/abs/2602.17655

🚀 Introducing the Qwen 3.5 Medium Model Series Qwen3.5-Flash · Qwen3.5-35B-A3B · Qwen3.5-122B-A10B · Qwen3.5-27B ✨ More intelligence, less compute. • Qwen3.5-35B-A3B now surpasses Qwen3-235B-A22B-2507 and Qwen3-VL-235B-A22B — a reminder that better architecture, data quality, and RL can move intelligence forward, not just bigger parameter counts. • Qwen3.5-122B-A10B and 27B continue narrowing the gap between medium-sized and frontier models — especially in more complex agent scenarios. • Qwen3.5-Flash is the hosted production version aligned with 35B-A3B, featuring: – 1M context length by default – Official built-in tools 🔗 Hugging Face: huggingface.co/collections/Qw… 🔗 ModelScope: modelscope.cn/collections/Qw… 🔗 Qwen3.5-Flash API: modelstudio.console.alibabacloud.com/ap-southeast-1… Try in Qwen Chat 👇 Flash: chat.qwen.ai/?models=qwen3.… 27B: chat.qwen.ai/?models=qwen3.… 35B-A3B: chat.qwen.ai/?models=qwen3.… 122B-A10B: chat.qwen.ai/?models=qwen3.… Would love to hear what you build with it.