Soroosh Tayebi Arasteh

Speech is clinically rich, but deeply identifying. The new paper argues that privacy in speech-based AI should be a system property, not a preprocessing fix: 🎙️ identity diversification 🧠 leakage-resistant models 🌐 governance-aware infrastructure nature.com/articles/s4448…

> Claude drops a new update > Software engineers:

@starasteh shows us differential privacy can protect patient data in medical deep learning, but it isn’t “free.” This review finds DP can preserve clinically useful performance under moderate privacy budgets, while strict privacy often reduces accuracy, especially in smaller/heterogeneous datasets, and may widen subgroup performance gaps. Takeaway: Privacy needs standards and fairness audits. nature.com/articles/s4174…

Differential privacy in medical AI is often seen as too theoretical or too costly. We tried to close that gap. Our new paper in @npjDigitalMed covers fundamentals, practical tradeoffs, and deployment considerations, and is open-access for everyone: nature.com/articles/s4174…

Question to consider: What if radiology AI could reason? @npjDigitalMed study introduces RaR (Retrieval & Reasoning): a multi-step framework that boosts radiology Q&A accuracy, cuts hallucinations, & helps both LLMs and radiologists think more reliably. nature.com/articles/s4174…

🚨 Just published in @npjDigitalMed 🎉 💡 We introduce #RaR, an agentic multi-step RAG framework for radiology QA ✅ Improves diagnostic accuracy vs standard online RAG 📉 Reduces hallucinations to <10% 📈 Biggest gains in small and mid-sized LLMs 🔍nature.com/articles/s4174…

Peer review is one of the core mechanisms that sustains rigor and progress in science, even though much of it happens quietly. In 2025, I did 102 peer reviews across journals in AI and medicine. Sharing a brief year-in-review snapshot of this often unseen but essential work

🚨 Just published in @Nature_NPJ Artificial Intelligence! 💡 Speech data can be as sensitive as images. ✅ We successfully applied DP to DL training on medical speech data ⚖️ Utility tradeoff remains high, but fairness behavior is more complex 🔍 nature.com/articles/s4438…

Are you ready for web-scale pre-training with RL ? 🚀 🔥 New paper: RLP : Reinforcement Learning Pre‑training We flip the usual recipe for reasoning LLMs: instead of saving RL for post‑training, we bring exploration into pretraining. Core idea: treat chain‑of‑thought as an action. Reward it by the information gain it provides for the very next token: This gives a verifier‑free, dense reward on ordinary text with no task checkers, no labels, no filtering. Why this matters ? * 🧠 Models think before predicting during pretraining, not just after alignment. * 📈 Position‑wise credit at every token = stable signal at full web‑scale. * 🔁 No proxy filters or “easy‑token” heuristics. Trains on the entire stream. Results: On the 8‑benchmark math+science suite (AIME’25, MATH‑500, GSM8K, AMC’23, Minerva Math, MMLU, MMLU‑Pro, GPQA): • Qwen3-1.7B-Base: RLP improves the overall average by 24% ! • Nemotron-Nano-12B-v2-Base: RLP improves the overall average by 43% ! 📄Paper: tinyurl.com/rlp-pretraining ✍️Blog: research.nvidia.com/labs/adlr/RLP/ #AI #LLM #ReinforcementLearning #ChainOfThought #Pretraining #RLP

📢 New #preprint: Agentic LLMs improve retrieval-based radiology QA 🧠 Our #agentic #RAG framework lets #LLMs iteratively retrieve & synthesize evidence. ✅ 25 models analyzed (500M→670B) 📈 Accuracy ↑ in mid/small models 🤖 Hallucinations ↓ ~9% 📖 arxiv.org/abs/2508.00743

We're proud to welcome the journal's 2025-2027 Trainee Editorial Board! Learn more about these talented individuals: pubs.rsna.org/page/ai/teb

#RadioRAG introduces real-time data retrieval to support the accuracy & factuality of LLMs in radiologic diagnosis doi.org/10.1148/ryai.2… @starasteh @DanielTruhn @laim_uka #retrieval #Radiopaedia #radiology

„PhD-Level“ Intelligence GPT5 still fails the „first do A, then B(A)“ test. A PhD student in Germany clearly does have such capabilities. GPT unfortunately, not. I don’t feel the AGI today.

🚨 Just published in @Radiology_AI! #RadioRAG: Online Retrieval–Augmented Generation for #Radiology QA 💡Goes beyond static #LLMs by integrating live data from @Radiopaedia 🔓Open-source code & datasets! 📖 pubs.rsna.org/doi/10.1148/ry… #AI #RAG #ChatGPT #LangChain #Mistral #Llama3

RadioRAG retrieves context-specific information from #Radiopaedia in real-time doi.org/10.1148/ryai.2… @starasteh @DanielTruhn @laim_uka #RAG #LLM #RadioRAG

RadioRAG retrieves context-specific information from #Radiopaedia in real-time doi.org/10.1148/ryai.2… @starasteh @DanielTruhn @laim_uka #LLMs #radiology #AI

#LLMs using radiology retrieval-augmented generation (RadioRAG) showed variable performance in answering case-based radiology questions doi.org/10.1148/ryai.2… @starasteh @DanielTruhn @laim_uka #Radiopaedia #radiology #AI

#RadioRAG shows potential to improve LLM accuracy in radiology question answering by integrating real-time domain-specific data doi.org/10.1148/ryai.2… @SNM_MI #SNMMI25 #NucMed #MolRad #radiology

Just crossed 1,000 citations on Google Scholar! Huge thanks to everyone I've collaborated with and learned from along the way, couldn’t have done it without you! scholar.google.com/citations?hl=e…

@peterhan91 @PennMedicine @PennAI2D Congrats Tianyu!

After 11 years in Germany, I'm thrilled to share that I've joined @PennMedicine as an Assistant Professor on the @PennAI2D team! Excited to meet new friends and collaborators. I'm hiring postdocs in radiological image analysis & self-supervised learning—DM if interested! #hiring