Mathew Koretsky

NEW Resource: CARDBiomedBench: a benchmark for evaluating the performance of #LLMs in biomedical research. Read it here: buff.ly/FQm6OEu

🚀 We just launched DTAgent-AD, a scientific reasoning agent built to accelerate biomedical research and specialized Alzheimer’s & neurodegenerative disease research. DTAgent-AD is a domain-tuned agent trained on trace data from real biomedical tasks, not a general chatbot.

Stay tuned to see how the biomedical agents we’re building at @DataTecnica stack up against these models!

Despite boasting impressive performance across a range of categories, the latest frontier LLMs (Gemini 3 Pro, Claude Opus 4.5, and GPT-5.2) still struggle to balance accuracy and safety on CARDBiomedBench, our biomedical QA benchmark 👀

Stay tuned to see how our Agents stack up against the latest base LLMs on biomedical question-answering tasks!

👀 Our new Knowledge Agents make BiomedArena.AI more powerful than ever. 🧬 Get insights backed by the journals and databases that biomedical researchers use daily. 📄 Read the blog more more info: datatecnica.com/blog/-biomedar…

We continue to evaluate these new models on our benchmark, CARDBiomedBench. Despite significant progress, there are still no models that balance response accuracy and safety on biomedical questions 👀

Check out the latest models in @BiomedArena for all of your biomedical research questions!

Super proud of all the hard work from our team including @tanaynayak99, @owenbianchi_, Shayan Shahand, @DanielKhashabi, and @FarazFaghri!!!

🚨BiomedArena is live🚨 In a partnership with @lmarena_ai, our team at @DataTecnica has released a feedback-rich platform to evaluate LLM performance on real-world biomedical questions. ⚔️Access the arena: biomedarena.ai 📄Read the blog post: news.lmarena.ai/introducing-bi…

🚨New LLM benchmark🚨 We're releasing BiomedSQL🔬 for tabular reasoning over large-scale biomedical databases. This includes questions based on implicit scientific conventions—like statistical thresholds, effect direction, and drug approval status. 📄 Preprint: arxiv.org/pdf/2505.20321 📊 Dataset: huggingface.co/datasets/NIH-C… Lead by Matt Koretsky at @DataTecnica

📄Read the preprint: arxiv.org/pdf/2505.20321 📊Dataset: huggingface.co/datasets/NIH-C… 💻Code: github.com/NIH-CARD/biome… Thanks to my teammates at NIH/CARD and @DataTecnica including Maya Willey, Adi Asija, @owenbianchi, Chelsea Alvarado, @mike_nalls, @DanielKhashabi, and @FarazFaghri

We believe this benchmark is a critical step towards building trustworthy text-to-SQL systems that can increase efficiency of lookups for PIs and SMEs, democratize access to biomedical knowledge, and accelerate discovery

Can LLMs perform reliably as biomedical data analysts? TL;DR: We created the first benchmark designed to challenge LLMs ability to apply scientific reasoning in text-to-SQL generation over biomedical databases, revealing a 30-40% gap between SOTA models and expert performance

Long-form inputs (e.g., needle-in-haystack setups) are the crucial aspect of high-impact LLM applications. While previous studies have flagged issues like positional bias and distracting documents, they've missed a crucial element: the size of the gold/relevant context. In our latest study, we look into how the size of these gold contexts impacts LLM performance in needle-in-a-haystack scenarios. The verdict? **Smaller gold contexts severely amplify positional bias.** Why should you care? If you're developing LLMs to sift through large number of documents of varying sizes, beware: a smaller gold document among larger distractions can throw your pipeline off course. Basically, practitioners needs to keep an eye not only on the position of the likely gold document but also on its size relative to others. 📄Read the preprint: arxiv.org/abs/2505.18148 Work lead by Owen Bianchi and other collaborators at @DataTecnica

Great work from many of our teammates! Let's accelerate data harmonization!

GenoTools: An Open-Source Python Package for Efficient Genotype Data Quality Control and Analysis biorxiv.org/cgi/content/sh… #biorxiv_genomic

Koretsky et al. use genome-wide data to cluster patients based on genetic status across risk variants for five neurodegenerative disorders. The results suggest that neurodegenerative diseases have more overlapping genetic aetiology than previously assumed. tinyurl.com/yt39brkt