Soroosh Tayebi Arasteh retweeted
Academia has a problem:
Every author wants to use LLMs to write papers, but no reviewer wants to read AI (assisted) content.
And every reviewer wants to use LLMs for reviews, but no author wants their paper judged by an AI.
Authors & reviewers are the same people though...
We are hiring a fully funded PhD candidate at RWTH Aachen🇩🇪!
Work on differential #privacy, fairness & medical foundation models, with access to enough GPUs and data.
Apply: arasteh.ai/opening-phd-diffe…
#PhD #machinelearning #dp #ai #position
Soroosh Tayebi Arasteh retweeted
We will soon be hiring postdocs, engineers and researchers (on site in Heidelberg, DE) in:
- Biomedical AI (multimodal foundation models, biomarkers, agents)
- Robotics (lab automation, physical action surfaces for agents)
- AI social sciences (multi-agent systems, agent "civilizations")
- Deploying AI in a large hospital system
More info at imai-heidelberg.de and kather.ai
If you are interested in learning more, reach out to [email protected]
Soroosh Tayebi Arasteh retweeted
Thrilled that #paper2agent is published in @nature today!
Scientific knowledge is traditionally stored in passive papers. Paper2Agent transforms papers into virtual authors that answer questions, apply its methods and collaborate w/ other paper agents to make new discoveries 🧵
🩻 Our recent paper in @CommsMedicine presents a benchmarking study of chest radiograph classification using different publicly available initialization weights, architectures, and input resolutions.
🔗 nature.com/articles/s43856-0…
How reliable are radiology LLMs when users introduce assumptions into a prompt? New research shows that cognitive biases can significantly reduce model accuracy, especially for multimodal tasks, underscoring the need for safeguards in clinical AI applications. Read more: bit.ly/4ggRPwI
@UofTMedIm @starasteh
#ImagingAI #RadSafety
Soroosh Tayebi Arasteh retweeted
NeurIPS discussion portal is a disaster.
Massive waste of time tracking who responded where.
I assume this was designed by a Linux hacker working for free, but these days with coding agents so good at frontend design, there's really no excuse for this bad a UX.
Plus, CVPR's 1-page rebuttal limit and not involving authors in the "discussion" is way more efficient. I strongly suspect there's almost zero marginal utility of such limitless discussion.
Soroosh Tayebi Arasteh retweeted
NeurIPS should not let authors reply to reviewers’ responses to their own rebuttal if they haven’t responded to the rebuttals of papers they reviewed. The authors of every paper I reviewed replied within an hour after I responded to their rebuttal, yet not a single reviewer has responded to mine.
Soroosh Tayebi Arasteh retweeted
NeurIPS reviews are out. At this point it's an arms race where both sides have the same supplier. You write the paper with Opus 4.6, the reviewer writes the review with Opus 4.7, and now we get to rebut with Opus 4.8. Let's just hope Opus 5 doesn't ship before the discussion period closes.
Preprints don’t change much after peer review — and are rarely retracted.
go.nature.com/3R05ZZj
Accepted to #MICCAI 2026: Differential privacy representation geometry for medical image analysis.
DP usually keeps the discriminative signal intact, but private training fails to use it, a "utilization gap." The fix is often better optimization.
arxiv.org/abs/2603.01098
3
101
Clinical LLMs are often scaled assuming better accuracy means safer deployment.
Our new study challenges that.
In radiology, safety and accuracy can follow different scaling laws: a few confident, high-risk errors can matter more than average performance.
arxiv.org/abs/2605.04039
66
Soroosh Tayebi Arasteh retweeted
Clinical LLMs keep getting bigger. The assumption: better benchmarks = safer deployment.
New preprint challenges that directly. In radiology, a few confident errors outweigh average accuracy. The authors call it "different scaling laws" for safety vs. performance.
Worth flagging if you're shipping clinical AI.
arxiv.org/abs/2605.04039
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/s44482-0…
7
388
Soroosh Tayebi Arasteh retweeted
Replying to @starasteh
@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/s41746-0…
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/s41746-0…
Soroosh Tayebi Arasteh retweeted
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/s41746-0…
🚨 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/s41746-0…
















