Emma Scharfman

I hear a lot of people talking about how LLMs will revolutionize research. But I don't see many quantitative measures to support this. Until I came across this paper which compares LLM-generated ideas against published research: arxiv.org/abs/2607.01233 Key findings: LLMs produce reasonable ideas, but they cluster around synthesis over novel problem framing. It is worth reading it!

@bofeng @huggingface Thanks @bofenghuang1 for your insights, loved working with your on this!

Just released a new blog post I co-authored with @bofeng: "𝘗𝘶𝘵𝘵𝘪𝘯𝘨 𝘋𝘰𝘤𝘵𝘰𝘉𝘌𝘙𝘛 𝘵𝘰 𝘞𝘰𝘳𝘬: 𝘈 𝘗𝘳𝘢𝘤𝘵𝘪𝘤𝘢𝘭 𝘎𝘶𝘪𝘥𝘦" on @huggingface! 🤗 We walk you through our findings on how to fine-tune DoctoBERT, a medical domain encoder recently released by @doctolib, with step-by-step tutorials. If you're working on domain-specific encoders or making AI more practical for healthcare, 𝗳𝗲𝗲𝗹 𝗳𝗿𝗲𝗲 𝘁𝗼 𝗿𝗲𝗮𝗰𝗵 𝗼𝘂𝘁, 𝗜'𝗱 𝗹𝗼𝘃𝗲 𝘁𝗼 𝗵𝗲𝗮𝗿 𝘆𝗼𝘂𝗿 𝗶𝗻𝗽𝘂𝘁! 𝗖𝗵𝗲𝗰𝗸 𝗼𝘂𝘁 𝘁𝗵𝗲 𝗳𝘂𝗹𝗹 𝗽𝗼𝘀𝘁: huggingface.co/blog/hugging-s… 𝗕𝗼𝗻𝘂𝘀: Bofeng Huang finetuned DoctoModernBERT as a text-diffusion model: …ang-doctobert-diffusion-demo.hf.space

@SwishMoe @huggingface Thanks a lot!

@EmmaScharfmann @huggingface Huge Emma congrats!!!!

I'm excited to share that I'm joining @huggingface as a ML Research Engineer on the science team 🚀🤗 My goal is to bridge the gap between researchers and the Hugging Face tools by collaborating with researchers and making it easier for the scientific community to use open-source data and models! If you're working at the intersection of AI and research or you want to help growing the AI4science community, feel free to reach out!

@Thom_Wolf @huggingface Thanks, really excited to be part of the team!

@EmmaScharfmann @huggingface Welcome!

Really cool paper on AI peer review systems: arxiv.org/abs/2605.03202 Papers are getting harder to review and results are often impossible to replicate. They rely on massive amounts of data and code that would make any SWE choke. And current AI peer review systems don't seem to be the solution. So what would actually work? Any ideas?

Why are there so few domain-specific encoder pretraining pipelines while every search engine, classifier or RAG system depends on good encoders? → Domain-specific encoder pretraining still relies on hand-curated corpora, while decoders LLMs rely on scalable web-curation methods. The gap is specifically important in fields where much of the vocabulary is domain-specific, like medicine. Strong representations of this vocabulary come from seeing terms at scale and in varied contexts, exactly what the web-curation offers and hand-curated corpora lack. 🤗 𝗚𝗿𝗲𝗮𝘁 𝗻𝗲𝘄𝘀: @doctolib Research Lab just released an open-source pipeline for encoder pretraining applied to the medical field. The pipeline explained here: huggingface.co/blog/bofenghua… 💪 𝗪𝗵𝗮𝘁'𝘀 𝗿𝗲𝗮𝗹𝗹𝘆 𝗰𝗼𝗼𝗹: the pipeline is entirely open-source, so this approach can scale to other languages and domains.

How can agents help structure the existing research literature? Join with your agents and contribute to this deep research wiki: huggingface.co/spaces/rl-llm-… Very excited to see how far we can go!

Exciting release from Doctolib Research Lab 👩‍⚕️💉on Hugging Face 🤗! A medical encoder for clinical NLP trained on FineMed, our large-scale French medical corpus. It leads the public benchmark and performs on real clinical tasks. Blog: huggingface.co/blog/bofenghua… Model and data: huggingface.co/doctolib-lab

our book is out and I got the first copy! @andimarafioti @micuelll @orr_zohar

🏆 A template to easily launch your own community challenge 🏆 Build our own challenge using this template where participants can: 🔍Preview the challenge ✍️ Register 📤 Submit results 🏅 Track the leaderboard Demo: huggingface.co/spaces/EmmaSch… Repo: github.com/EmmaScharfmann…

Be a part of finding alien life with... The dataset: huggingface.co/datasets/Astro… The blog: huggingface.co/blog/hugging-s… The code: github.com/astroautomata/… The pre-print: arxiv.org/abs/2606.18338

The AI hunt for alien life has just begun. Welcome to ThousandsWorlds, a wild new dataset from researchers at Oxford/Cambridge++, for detecting faint signatures in the atmospheres of potentially habitable exoplanets. This is the first step towards finding life beyond earth. The plan is basically: 1) scan the galaxy for as many potentially habitable planets as possible 2) detect the gases in their atmospheres with powerful telescopes like JWST 3) infer from these gases whether life is present or not. ThousandWorlds is a benchmark for emulating these exoplanet climates: 1760 simulations across 5 GCMs, 8 planet parameters, and atmospheric variables on a 32 x 64 x 10 latitude-longitude-pressure grid. It includes three nested benchmark subsets, two evaluation protocols, and eight released baseline methods. incredible work from @MilesCranmer and many more 👽👽👽

We let an AI write our release notes, then wrote code to prove it didn't lie. That's how huggingface_hub ships weekly now: the model drafts, code flags any dropped or invented PR, a human approves. New blog post on how we built it 👇