iiinova — where your digital self keeps evolving

Juraj Vladika @jvladika.bsky.social · 06/09/2025

Delighted to share "Facts Fade Fast: Evaluating Memorization of Outdated Medical Knowledge in LLMs", accepted to Findings of #EMNLP2025 !🐼 Wit a novel dataset of changed medical knowledge, we discover the alarming presence of obsolete advice in eight popular LLMs.⌛ 📝: arxiv.org/abs/2509.04304 #NLP

0100

Juraj Vladika @jvladika.bsky.social · 23/02/2025

Also happy to share that “On the Influence of Context Size and Model Choice in RAG Systems” was accepted to Findings of #NAACL2025! 🇺🇸🏜️ We test how the RAG performance on QA tasks changes (and plateaus) with increasing context size across different LLMs and retrievers. 📝 arxiv.org/abs/2502.14759

line diagram showing the RAG performance of different base LLM models

090

Juraj Vladika @jvladika.bsky.social · 23/02/2025

Thrilled to share that "Step-by-Step Fact Verification for Medical Claims with Explainable Reasoning" was accepted to #NAACL2025! 🇺🇸🏜️ This system iteratively collects new knowledge via generated Q&A pairs, making the verification process more robust and explainable. 📜 arxiv.org/abs/2502.14765 #NLP

architecture of the step-by-step fact verification system

060

Juraj Vladika @jvladika.bsky.social · 16/02/2025

More than 8500 submissions to ACL 2025 (ARR February 2025 cycle)! That is an increase of 3000 submissions compared to ACL 2024. It will be a fun reviewing period. 😅💯 @aclmeeting.bsky.social #ACL2025 #ACL2025nlp #NLP

1205

Juraj Vladika @jvladika.bsky.social · 21/12/2024

Most exciting update to encoder-only models in a long time! Love to use them for classification tasks where LLMs are an overkill #ModernBERT

140

Juraj Vladika @jvladika.bsky.social · 26/11/2024

Organizing hackaTUM 2024 was an incredible experience! Around 1000 participants with 3 days full of intense coding, new experiences, exciting sponsor challenges and workshops, fun side activities, tasty food, creative final solutions, and overall awesome fun! 😊 Join us next year 💙🧑‍💻🔜 hack.tum.de

020

Reposted by Juraj Vladika

Ai2 @ai2.bsky.social · 26/11/2024

Meet OLMo 2, the best fully open language model to date, including a family of 7B and 13B models trained up to 5T tokens. OLMo 2 outperforms other fully open models and competes with open-weight models like Llama 3.1 8B — As always, we released our data, code, recipes and more 🎁

The OLMo 2 models sit at the Pareto frontier of training FLOPs vs model average performance.

515235